Modeling categorization dynamics through conversation by constructive approach
多模态生物特征识别模型研究
多模态生物特征识别模型研究随着科技的不断发展，人类的生活变得越来越便利，同时也出现了越来越多的安全问题。
为了保障人们的生命安全和财产安全，生物特征识别技术应运而生。
生物特征识别技术是指利用人体生理特征或行为特征来识别身份的一种技术，其中多模态生物特征识别模型是当前研究的热点之一。
一、多模态生物特征识别模型的基本原理多模态生物特征识别模型是指通过多种生物特征数据进行判定。
在进行多模态生物特征识别模型的研究时，我们通常会选取多个特征数据集，如常见的指纹、虹膜、面部、声音、手掌等等。
将多个特征数据集进行组合，可以大幅提高识别准确率，实现“一图胜千言”的效果。
多模态生物特征识别模型的基本原理是将参考样本的数据放入存储器中，并提供对应多个特征的样本图片。
当开启多模态生物特征识别模型后，对输入数据的特征进行提取、归一化处理后，在生物特征数据集的存储库中进行比对，通过比对算法寻找最佳匹配。
最终输出匹配的结果并给出匹配的相似度。
二、多模态生物特征识别模型的发展趋势目前多模态生物特征识别模型在研究方面还存在一些问题，如不同模态特征的选择、管理和整合等方面。
随着大数据与人工智能的应用以及多模态生物特征识别技术的不断发展，未来的多模态生物特征识别模型将会朝着以下几个方向发展：1. 融合大数据与人工智能技术进行优化，提高多模态生物特征识别模型的准确率和实际应用。
2. 建立更加完善的生物特征数据集，多元化的生物特征数据集与可靠性较高的生物特征数据集的融合可以提高多模态生物特征识别模型的准确性，丰富生物特征数据集的种类。
3. 加强对多模态生物特征识别模型的安全性、稳定性、隐私保护等方面的研究，进一步完善多模态生物特征识别技术。
三、多模态生物特征识别模型在实际应用中的优势多模态生物特征识别模型在现实生活中有着广泛的应用场景。
它可以广泛应用于警务、金融、电子商务、智能家居等领域，其主要优势包括：1. 准确性高：多模态生物特征识别模型基于多种生物特征数据进行判定，识别准确性更高，假冒、欺诈等行为更难进行。
虚拟动点运动捕捉技术赋能阿兹海默症研究,OptiTrack还有这样的作用!
虚拟动点运动捕捉技术赋能阿兹海默症研究，OptiTrack还有这样的作用！2018年9月，罗切斯特大学医疗中心的神经科学家们开始使用一种世界前沿研究平台——MoBI系统（全称Mobile Brain/Body Imaging system）——来帮助研究治疗自闭症、阿兹海默症及外伤性脑损伤。
MoBI系统结合了虚拟现实、大脑监测技术以及受好莱坞启发的运动捕捉技术。
它能有效帮助神经科学家研究人类脑神经失调带来的身体运动协调困难，以及为什么我们的大脑在同时进行多任务时会产生挣扎。
原理解析目前MoBI平台被安置在德尔蒙特学院认知神经生理学实验室，集合了几种高科技系统，其中之一便是电影工作室制作CGI特效（Computer Graphics Interface）时常用的运动捕捉技术。
实验参与者需身穿附着反射式标记点的黑色运动捕捉服，并在一间安装有16台亚毫米精度的OptiTrack高速运动捕捉摄像机的房间中完成“在跑步机上走路”和“物体操作”的动作。
运动捕捉服装上的反射式标记点随参与者的运动而不断变换空间位置，并可实时将实验室内16台OptiTrack摄像机发出的红外光反射回至所有摄像机，摄像机接收到每个标记点的空间坐标数据后，将数据实时传输到软件中进行解算，形成刚体，并可将6DoF信息实时传输，驱动研究人员电脑端的3D虚拟模型。
实验参与者在走动过程中，他们面前的屏幕上会投射出沉浸式可互动虚拟场景（如城市街景），需要参与者带路前行。
他们也可以被要求执行任务，做出决策并回应屏幕上的内容。
实验进行时，通过使用高密度EEG（Electroencephalogram：脑动电流图）——参与者头戴金属电极帽，电极帽上的小金属片紧贴头皮——科研人员可实时监测参与者脑电活动。
当运动捕捉数据与EEG监测实时同步，科研人员就可以在实验参与者走路和执行任务时观察其大脑哪些区域被激活，并研究其在移动、执行任务以及同时进行以上两种动作时，大脑如何给予反馈。
基于注意力机制和动态卷积的滚珠螺杆表面缺陷识别
Science and Technology & Innovation｜科技与创新2024年第04期DOI：10.15913/ki.kjycx.2024.04.009基于注意力机制和动态卷积的滚珠螺杆表面缺陷识别符诗语1，高齐2（1.江苏理工学院汽车与交通工程学院，江苏常州213001；2.江苏理工学院机械工程学院，江苏常州213001）摘要：针对滚珠螺杆表面缺陷识别特征信息提取困难和识别精度较低等问题，提出了一种基于注意力机制和动态卷积的滚珠螺杆表面缺陷识别方法。
首先，利用AlexNet网络进行缺陷样本图像的特征提取；然后，引入卷积注意力机制模块，在通道维度和空间维度增强缺陷区域位置权重，有效抑制相似背景干扰；最后，采用动态卷积模块对不同尺度的缺陷特征图像进行融合，提升模型特征提取能力，捕获丰富上下文信息。
实验表明，该方法在滚珠螺杆表面缺陷测试集上精确率为96.6%，召回率为96.5%，具有良好的实际工业应用价值。
关键词：缺陷识别；AlexNet；注意力机制；动态卷积中图分类号：TP183 文献标志码：A 文章编号：2095-6835（2024）04-0039-03滚珠螺杆作为转换机械运动的精密传动组件[1]，其质量将会直接影响仪器性能。
但在加工制造或运行时，受不利条件影响，滚动螺杆易出现“点蚀”现象[2]，最终导致机械设备损坏。
因此，实现高效的滚珠螺杆表面缺陷识别，对于提高工业产能和生产制造水平具有重要意义[3]。
近年来，随着深度学习模型的发展，诸多计算机视觉领域应用卷积神经网络解决了各类挑战，例如人脸识别、自然语言处理和目标检测等，因此，许多学者尝试将深度学习方法应用于表面缺陷检测。
文献[4]设计了一种新颖的级联式自动编码器结构，实现了金属表面缺陷有效识别，相比于传统视觉方法，实时性和精确率都取得了较大的提升，但当缺陷目标较小时，模型易受干扰，导致表现一般。
文献[5]利用迁移学习和特征重用技术，提出了一种基于DenseNet121网络的齿轮表面缺陷识别方法，该算法综合性能优异，但针对划痕和无缺陷2类相似图片的识别效果有待提升。
基于深度学习和PnP模型的激光跟踪仪自动姿态测量
第30卷第9期2022年5月Vol.30No.9May2022光学精密工程Optics and Precision Engineering基于深度学习和PnP模型的激光跟踪仪自动姿态测量周道德1，2，高豆豆1，董登峰1，2*，周维虎1，2，崔成君1（1.中国科学院微电子研究所，北京100029；2.中国科学院大学，北京100049）摘要：针对航空航天、汽车装配等高端制造领域对姿态测量的迫切需求，提出一种面向激光跟踪仪的快速高精度姿态测量方法，利用深度学习结合视觉PnP模型实现了激光跟踪过程中被测件姿态的自动测量。
针对PnP姿态求解模型所需的3D特征点和2D特征点之间的对应关系难以直接确定的问题，设计了一个特征提取网络用于提取特征点对应的高维特征，采用最优传输理论确定特征向量之间的联合概率分布，从而完成3D-2D特征点的自动匹配；使用Ransac-P3P结合EPnP算法对匹配好的3D特征点和2D像素点进行姿态求解，获得高精度的姿态信息；在此基础上，利用隐式微分理论计算PnP求解过程的雅克比矩阵，从而将PnP姿态求解模型集成到网络中并指导网络训练，实现了深度网络匹配能力与PnP模型姿态求解能力的优势互补，提高了解算精度。
最后，制作了一个含有丰富标注信息的数据集，用于训练面向激光跟踪仪的姿态测量网络。
基于高精度二维转台进行了姿态测量实验，结果表明，该方法在3m处对俯仰角的测量精度优于0.31°，横滚角精度优于0.03°，单次测量耗时约40ms，能够实现激光跟踪仪的高精度姿态测量。
关键词：激光跟踪仪；姿态测量；单目视觉；深度学习中图分类号：TP391.4；TH744文献标识码：A doi：10.37188/OPE.20223009.1047Automatic attitude measurement of laser tracker based ondeep learning and PnP modelZHOU Daode1，2，GAO Doudou1，DONG Dengfeng1，2*，ZHOU Weihu1，2，CUI Chengjun1（1.Institute of Microelectronics of the Chinese Academy of Sciences，Beijing100029，China；2.University of Chinese Academy of Sciences，Beijing100049，China）*Corresponding author，E-mail：Dongdengfeng@Abstract：In view of the urgent demand for attitude measurement in high-end manufacturing applications，such as aerospace and automobile assembly，a fast and high-precision attitude measurement method for a laser tracker was proposed.The method employed deep learning in conjunction with the visual PnP model to realize automatic attitude measurement of the laser tracker.The correspondence between3D feature points and2D feature points required by the traditional PnP model were directly determined through a fea⁃ture extraction network designed to extract high-dimensional features.The joint probability distribution be⁃tween feature vectors was determined using optimal transmission theory to complete the matching of3D-文章编号1004-924X（2022）09-1047-11收稿日期：2022-03-04；修订日期：2022-03-16.基金项目：国家重点研发计划资助项目（No.2019YFB1310100）第30卷光学精密工程2D feature points.Subsequently，Ransac-P3P combined with EPnP algorithm was used to obtain high-pre⁃cision attitude information；Based on this，the Jacobian matrix of PnP solution process was calculated us⁃ing implicit differential theory，and the PnP attitude solution model was integrated into the network to guide the training of the network.The complementary advantages of strong depth network matching abili⁃ty and high attitude solution accuracy of the PnP model improved the solution accuracy of the network.In addition，a dataset with rich annotation information was used to train the attitude measurement network for the laser tracker.Finally，an attitude measurement test was conducted using a high-precision two-dimen⁃sional turntable.The experimental results show that the calculation error of pitch angle is less than0.31°，the rolling angle error is less than0.03°，and the single measurement takes approximately40ms.The pro⁃posed method can potentially be applied to attitude measurement scene of the laser tracker.Key words：laser tracker；attitude measurement；monocular vision；deep learning1引言随着制造业的快速发展，在航空航天、汽车装配等领域，大尺寸高精度姿态测量技术越来越重要。
基于自适应多尺度形态梯度变换的滚动轴承故障特征提取
动
与
冲
击
Ｊ ＯＵＲＮＡＬ Ｂ ＯＦ ＶＩ ＲＡＴＩ ＯＮ ＡＮＤ Ｈ０ＣＫ Ｓ
基 于 自适应 多尺 度 形态 梯 度变 换 的滚 动轴 承 故 障特 征提 取
李 兵 ， 张培林 刘东升 米双 山 任 国全 ， ， ，
（． １ 石家庄军械工程学院 自行火炮教研室， 石家庄 ０００ ； ． ５０３ ２ 石家庄军械工程学院 导弹机电工程教研室， 石家庄 ０００ ） ５０３
来提出的另一种基于数学形态学 的形态 闭变换方法 相比较 ， 自适应多尺度形态梯度变换具有更 强的噪声抑制和脉冲提取
能力 ， 并且计算 简单 、 快速 ， 为滚 动轴承故障特征提取提供 了一种有效 的方法 。
关键词 ：数学形态学 ； 自适应多尺度形态梯度 ； 滚动轴承 ； 故障诊断 ； 特征提取
［ ８—１ ］ 用 形 态 闭算 子 对 轴 承 、 轮 进 行 了特 征 提 ０采 齿 取 ， 献 ［ １ 提 出 了一 种 形 态 非 抽样 小 波对 转 子 振 动 文 １］ 冲击 特征信 号 提取 ， 得 了一 定 的效果 。 取 但上 述方 法均 采用 单 一 尺度 的结 构 元 素对 信 号 进 行处 理 ， 小尺 度下 能够 保 留更 多 的信 号 细节 ， 同 时会 但
Ａｂ ｔ ａ ｔ Ｉ ｕｌｉ ｅ ｔｐｅ ｓｇ ａ ｓｔ ｅ ｃ ｒ ｃ ｅｉｔｃ ｒ ｓ ｎ ｅ ｏ ｅｅ ｔ ｄ ｒ ｌｅ ｅ ｒｎ ．Ｈｏ ｔ ｘ ｒ ｃ ｍｐ ｌｉ ｅ ｓ ｒ ｃ ： ｍｐ ｓｖ ｙ ｉｎ ｌｉ ｈ ｈａ ａ ｔ ｒｓｉ ｅ ｐｏ ｓ ｆａ ｄ ｆ ｃｅ ｏｌｒｂ ａ ｇ ｉ ｗ ｏ ｅ ｔａ ｔｉ ｕ ｓｖ ｓｇ ｌｆｏ ａ ｎｏｓ ｄ ｖｂｒｔｏ ｉ ｎ ｌｂｅ ｏ ｓ ｔ ｅ ｔ ｐ ｆ ｒ ｂ ａ ｉ ｇ ｆ ｕ ｔｄ ａ ｎｏ ｉ． Ａ ｏ ｅ ｔ ｏ ｍｅ ｄａ ｔｖ ｉｎａ ｒ ｍ ｉｅ ｉ ａｉ ｎ ｓｇ ａ ｃ ｍｅ ｈｅ ｋ ｙ ｓｅ ｏ ｅ ｒｎ ａ ｌ ｉ ｇ ｓｓ ｎ ｖ ｌ ｍｅ ｈ ｄ ｎａ ｄ 梯 度 变 换
基于肢体和微表情的多模态特征的抑郁倾向识别方法[发明专利]
![基于肢体和微表情的多模态特征的抑郁倾向识别方法[发明专利]](https://img.taocdn.com/s3/m/fe464be9a48da0116c175f0e7cd184254b351bcc.png)
(19)中华人民共和国国家知识产权局(12)发明专利申请(10)申请公布号 (43)申请公布日 (21)申请号 202010763656.4(22)申请日 2020.07.31(71)申请人 华南理工大学地址 510640 广东省广州市天河区五山路381号(72)发明人 杜广龙 (74)专利代理机构 广州粤高专利商标代理有限公司 44102代理人 何淑珍 江裕强(51)Int.Cl.G06K 9/00(2006.01)G06K 9/32(2006.01)G06K 9/62(2006.01)G06N 3/04(2006.01)G06N 3/08(2006.01)(54)发明名称基于肢体和微表情的多模态特征的抑郁倾向识别方法(57)摘要本发明公开了一种基于肢体和微表情的多模态特征的抑郁倾向识别方法。
所述方法包括以下步骤：借助非接触式测量传感器Kinect检测人体运动，生成运动文本描述；采用非接触式测量传感器Kinect捕捉人脸图像帧，对人脸感兴趣区域进行Gabor小波和线性判别分析，进行特征提取和降维，然后采用三层神经网络实现人脸表情分类，生成表情文本描述；通过一个具有自组织映射层的融合神经网络提取的文本描述进行融合并生成带有情感特征的信息；使用Softmax分类器将S3中生成的特征信息在情感类别中进行分类，分类结果用于评估该患者是否具有抑郁倾向。
本发明考虑到静态身体运动和动态身体运动，达到了更高的效率。
身体运动有助于识别抑郁症患者的情绪。
权利要求书3页 说明书6页 附图1页CN 111967354 A 2020.11.20C N 111967354A1.基于肢体和微表情的多模态特征的抑郁倾向识别方法，其特征在于，包括以下步骤：S1、借助非接触式测量传感器Kinect检测人体运动，分别采用卷积神经网络(CNN)和双向长短时记忆条件随机场(Bi-LSTM-CRF)对人体静态运动和动态运动进行分析，生成运动文本描述；S2、采用非接触式测量传感器Kinect捕捉人脸图像帧，对人脸感兴趣区域(ROI)进行Gabor小波和线性判别分析(LDA)，进行特征提取和降维，然后采用三层神经网络实现人脸表情分类，生成表情文本描述；S3、通过一个具有自组织映射层的融合神经网络对步骤S1和步骤S2中提取的文本描述进行融合并生成带有情感特征的信息；S4、使用Softmax分类器将S3中生成的特征信息在情感类别中进行分类，分类结果用于评估该患者是否具有抑郁倾向。
基于特征模型的永磁同步直线电机自适应控制
第28卷㊀第3期2024年3月㊀电㊀机㊀与㊀控㊀制㊀学㊀报Electri c ㊀Machines ㊀and ㊀Control㊀Vol.28No.3Mar.2024㊀㊀㊀㊀㊀㊀基于特征模型的永磁同步直线电机自适应控制曹阳,㊀郭健(南京理工大学自动化学院,江苏南京210094)摘㊀要:为了解决永磁同步直线电机系统的参数不确定性㊁建模不确定性及饱和非线性等问题,提出一种基于特征模型的自适应控制器㊂依据特征模型理论描述永磁同步直线电机系统,采用自适应和鲁棒控制方法设计控制器㊂建立永磁同步直线电机的特征模型,并给出具体建立步骤,使得控制器设计变得简单,易于工程实现㊂通过设计参数自适应律对系统未知特征参数进行估计,可实现对系统模型的精确补偿,同时在控制器中添加带有误差积分的鲁棒控制项,提高系统对不确定参数及未知干扰的鲁棒性㊂此外,由于饱和特性的存在,导致控制器产生windup 问题,给系统的控制性能和稳定性造成不利影响㊂因此,该控制器中还带有抗饱和控制项,能够提升系统的抗饱和能力㊂最后,通过对比实验验证了所提控制器的有效性㊂关键词:永磁同步直线电机;参数不确定性;建模不确定性;饱和非线性;特征模型;自适应控制;抗饱和DOI :10.15938/j.emc.2024.03.013中图分类号:TM351文献标志码:A文章编号:1007-449X(2024)03-0131-10㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀收稿日期:2022-07-04基金项目:国家自然科学基金(61673219)作者简介:曹㊀阳(1993 ),男,博士研究生,研究方向为电机系统分析与控制;郭㊀健(1974 ),男,博士,教授,博士生导师,研究方向为智能系统与智能控制㊁机器人系统㊁高精度电机控制等㊂通信作者:郭㊀健Adaptive control of permanent magnet synchronous linear motorbased on characteristic modelCAO Yang,㊀GUO Jian(School of Automation,Nanjing University of Science and Technology,Nanjing 210094,China)Abstract :To address the problems of parameter uncertainty,modeling uncertainty and saturation nonlin-earity in the permanent magnet synchronous linear motor system,an adaptive controller based on charac-teristic model was proposed.A characteristic model was used to describe the permanent magnet synchro-nous linear motor system,and the controller was designed using adaptive and robust control methods.The characteristic model was established based on the system dynamics and parameters,and the specific steps were presented.This simplifies the controller design and facilitates the engineering implementation.An online parameter adaptation law was employed to estimate the unknown characteristic parameters of the system and achieve accurate compensation for the system model.Furthermore,an integral-type robust control term was incorporated into the controller,which improves the robustness of the system against un-certain parameters and unknown disturbances.In addition,the saturation nonlinearity leads to the windup problem in the controller,which has adverse effects on the control performance and stability of the sys-tem.Therefore,an anti-windup control scheme was devised for the controller,which can enhance the an-ti-saturation ability of the system.Finally,comparative experiments with other control methods were con-ducted to verify effectiveness of the proposed controller.Keywords:permanent magnet synchronous linear motor;friction nonlinearity;saturation nonlinearity;ar-mature mass variation;characteristic model;adaptive control;anti-windup0㊀引㊀言相比于旋转同步电机,永磁同步直线电机(per-manent magnet synchronous linear motor,PMSLM)具有更高的推力密度和更快的动态响应,特别适用于对速度和精度要求较高的场合,已被广泛应用在高精密加工㊁轨道交通传输等现代工业领域[1-2]㊂但是由于采用直接驱动方式,PMSLM控制系统对参数摄动及扰动等因素变得更加敏感[3],这会严重影响系统的控制性能㊂因此,保证PMSLM系统的高精度跟踪性能与抗扰动能力十分重要,对提高机床加工精度㊁提升交通传输效率具有重要的意义㊂针对PMSLM系统的高精度跟踪问题,国内外已有众多学者对其进行了研究㊂文献[4]设计了一种带模型参考自适应观测器的预测电流控制策略,经过实验验证该控制策略可以实现对速度进行在线准确辨识,进而提高电流的跟踪性能㊂文献[5]利用扩张状态观测器和非线性状态误差反馈对PMSLM的自抗扰控制器进行优化,提高了系统的动态响应性能和抗干扰能力㊂文献[6]提出一种基于周期性扰动学习的自适应滑模控制方法,采用滑模控制确保PMSLM系统对不确定性因素具有较强的鲁棒性㊂文献[7]在系统模型反馈线性化的基础上,将Hɕ鲁棒控制方法与D-K迭代法相结合,提高了系统对不确定性因素影响的抑制能力㊂姚斌等[8]提出一种自适应鲁棒控制方法,所开发的控制器成功应用在多种控制系统中[9-11]㊂为了解决非光滑饱和非线性的影响,文献[12]构造了一种新的近似饱和模型,该模型能够以任意规定的精度平滑地逼近实际饱和㊂此外,通过添加积分器技术,使得控制器可以消除与表面误差和边界层误差有关的耦合项㊂但是该方法在控制器的设计中需要对虚拟控制量重复微分,如果系统模型阶数高,会增加设计的复杂性㊂文献[13]提出一种考虑LuGre 摩擦的自适应鲁棒控制方法,针对陀螺框架伺服系统未知惯量和阻尼系数㊁LuGre摩擦参数不确定性及未知外部干扰上界,设计参数更新律对其进行估计,该控制律提高了系统的跟踪精度并通过仿真结果验证了所提方法的有效性㊂但该方法需要被控对象的精确数学模型,另外估计的未知参数过多,多个自适应参数需要反复调试,增加了实际应用时的难度㊂自适应鲁棒控制可以估计系统未知参数,但如果系统模型复杂㊁未知参数多㊁某些状态不可测时,控制器的设计将面临巨大挑战㊂针对这些问题,吴宏鑫院士等[14-15]提出特征建模的思想,特征模型一般用一阶或二阶差分方程/微分方程来描述,有关信息都压缩到几个特征参数中,并不丢失原有的信息㊂特征模型建立的形式比原对象动力学方程简单,为实际复杂系统的建模问题提供了一条途径㊂文献[16]基于永磁同步电机的特征模型,设计一个以非线性黄金分割自适应控制为主的控制方案㊂通过安排过渡过程和特征模型参数的在线辨识,该控制方案实现了控制器参数的在线自适应调节㊂文献[17]将特征建模方法推广到具有惯性变化的齿轮传动伺服系统中,设计了一个自适应二阶离散终端滑模控制器,并实现了有限时间有界性㊂然而上述基于特征模型所设计的控制器没有进行抗饱和(anti-windup)研究㊂windup现象是指由于被控对象的输入限制,使得被控对象的实际输入与控制器的输出不等,引起系统闭环响应变差(如超调变大,调节时间变长,甚至使系统失去稳定)的现象㊂实际的PMSLM是个物理限制系统,转速控制器的输出必须限定在一定的范围内,使得实际电机的控制输入量不能大于一个预先设定值㊂当控制器输出受到饱和限制时,特别是含有积分项的控制信号仍然增加时,就会出现windup现象,使实际闭环系统的性能下降,因此对PMSLM系统设计抗饱和控制是有必要的[18-19]㊂基于上述分析,针对PMSLM系统存在的参数不确定性㊁建模不确定性及饱和非线性等问题,提出一种基于特征模型的抗饱和自适应鲁棒控制器(an-ti-windup adaptive robust control based on characteris-tic model,AARC)㊂利用特征模型简化PMSLM系统的描述,并对其进行验证㊂然后,设计一种基于参数投影的自适应律,实现对系统模型的在线补偿㊂同时,将系统的不确定参数和未知干扰视为集总的干231电㊀机㊀与㊀控㊀制㊀学㊀报㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀第28卷㊀扰项,引入误差积分的鲁棒控制项进行抑制㊂此外,为了解决积分环节可能引起的windup 现象,加入抗饱和控制项,提高系统的抗饱和能力㊂最后,基于Lyapunov 函数证明闭环系统的稳定性,并通过实验验证所提控制器的有效性和鲁棒性㊂1㊀PMSLM 的特征建模与验证1.1㊀PMSLM 模型PMSLM 的运动方程为m d y d t =3π2τn p i q [ψf+(L d -L q )i d ]-F fric (y )㊂(1)式中:m 为等效质量;ψf 为磁链;y 为动子速度;i d ㊁i q 分别为d㊁q 轴电流;τ为极距;n p 为极对数;L d ㊁L q 分别为d㊁q 轴电感;F fric (y )为摩擦力㊂由式(1)可得y ㊃㊃=1.5πn p mτ[ψf i ㊃q +(L d -L q )(i ㊃d i q +i ㊃q i d )]- F fric y㊃m y㊂(2)设PMSLM 的采样周期为T ,将式(2)离散化可得㊀y (k +1)-2y (k )+y (k -1)T 2=[1.5πmTτn p ψf +1.5n p (L d -L q )i d (k )mTτ]i q (k )-[1.5πmTτn p ψf +1.5n p (L d -L q )i d (k )mTτ]i q (k -1)+1.5πn p (L d -L q )i q (k )mTτ[i d (k )-i d (k -1)]-1mT F firc (y (k )-y (k -1))y ㊂(3)在式(3)两边同乘T 2,可以重新写为y (k +1)=[1.5πmτn p ψfT +1.5n p (L d -L q )i d (k )Tmτ]i q (k )+[2-1m F firc T v ]y (k )+[1m F firc T v-1]y (k -1)+[1.5n p (L d -L q )i d (k )T mτ-1.5πmτn p ψfT ]i q (k -1)+1.5πn p (L d -L q )i q (k )Tmτˑ[i d (k )-i d (k -1)]=β1(k )i q (k )+α1(k )y (k )+α2(k )y (k -1)+Δ(k )㊂(4)式中:y (k )为系统输出;i q (k )为系统输入;α1㊁α2㊁β1为系统的特征参数,定义为:α1(k )=[2-1m F firc Tv];α2(k )=[1m F firc Tv -1];β1(k )=[1.5πmτn p ψf T +1.5n p (L d -L q )i d (k )T mτ]㊂üþýïïïïïïï(5)Δ(k )表示集总未知非线性函数,包括建模误差和未知扰动,定义为Δ(k )=[1.5n p (L d -L q )i d (k )Tmτ-1.5πmτn p ψfT ]i q (k -1)+1.5πn p (L d -L q )i q (k )Tmτˑ[i d (k )-i d (k -1)]㊂(6)通过式(4)可以看出,特征模型是将模型结构的模型不确定性和参数摄动等不确定信息压缩成几个未知的特征参数,使其与实际模型等价而不是近似㊂使用特征建模不仅能简化控制器设计,而且更利于工程应用㊂1.2㊀特征模型验证特征模型验证过程如图1所示㊂首先,分别给予PMSLM 系统和特征模型相同的输入信号u ㊂然后,采样PMSLM 的输入输出信号,采用传统投影梯算法[16]在线辨识特征参数,并计算特征模型输出㊂最后,通过比较特征模型输出y ^与PMSLM 系统输出y ,得到误差e 0㊂将输入设为1sin(2.09t )A 的正弦信号,并且设PMSLM 的采样频率为80μs㊂特征模型验证结果如图2所示㊂实验结果表明,在相同的控制输入作用下,特性模型输出与实际系统输出的误差很小,说明特征模型可以很好地描述PMSLM 系统的输入输出特征,可以利用该特征模型来设计控制器㊂331第3期曹㊀阳等:基于特征模型的永磁同步直线电机自适应控制图1㊀特征模型验证Fig.1㊀Verification block diagram of characteristicmodel图2㊀特征模型验证结果Fig.2㊀Verification results of characteristic model2㊀非线性自适应控制器设计2.1㊀自适应控制设计针对PMSLM 系统中存在的参数不确定㊁饱和非线性以及外界干扰,设计基于特征模型的自适应鲁棒控制律,对系统的不确定性和干扰进行估计和补偿,实现PMSLM 的速度跟踪控制㊂设计的自适应控制结构如图3所示,控制器包括模型补偿项u a ㊁线性反馈项u s1㊁积分鲁棒控制律u s2和抗饱和控制律k cw η,i qmax =0.03㊁i qmin =-0.03为饱和限制上下界㊂图3㊀自适应抗饱和控制结构框图Fig.3㊀Structure diagram of adaptive anti-windupcontroller将特征模型写成如下二阶时变辨识模型:y (k +1)=φ(k )T θ(k )㊂(7)式中:φ(k )=[y (k )y (k -1)u (k )]T ;θ(k )=[α1(k )α2(k )β1(k )]T ㊂在下面的部分中,㊃j 表示向量㊃的第j 个分量,并且针对2个向量的运算 < 是根据向量的相应元素来执行的㊂用θ^表示θ的估计值,θ~表示估计误差(θ~=θ^-θ)㊂结合式(7),一种不连续投影可以定义为proj θ^j {㊃j }=0,if θ^j =θj max and㊃j >0;0,if θ^j =θj min and㊃j <0;㊃j ,otherwise㊂ìîíïïïï(8)式中:j =1,2,3;proj θ^j{㊃j }可以保证估计参数在有界凸闭集D s 内㊂为保证参数估计值的有界性,设计未知参数估计自适应律为:θn (k )=θ^(k -1)+Γτλ+φT(k -1)φ(k -1);θ^(k )=proj θ^(θn(k ))㊂}(9)式中:Γ>0,λ>0为待设计的可调参数;τ为待合成的自适应函数;θ^(k )为系统参数θ(k )的估计值,利用基于不连续投影的参数自适应律可以估计出未知的特征参数α1㊁α2㊁β1㊂特征模型式(4)可被重写为y (k +1)=[α^1(k )-α~1(k )]y (k )+[α^2(k )-α~2(k )]y (k -1)+[β^1(k )-β~1(k )]u (k )+β1η(k )+Δ(k )㊂(10)式中α~1(k )=α^1(k )-α1(k ),α~2(k )=α^2(k )-α2(k ),β~1(k )=β^1(k )-β1(k )为辨识误差㊂所以式(10)可以改写为431电㊀机㊀与㊀控㊀制㊀学㊀报㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀第28卷㊀y(k+1)=α^1(k)y(k)+α^2(k)y(k-1)+β^1(k)u(k)+β1η+Δ(k)-θ~(k)φ(k)㊂(11)其中θ~(k)φ(k)=α~1(k)y(k)+α~2(k)y(k)+β~1(k)u(k)表示模型估计误差㊂假设1:从工程实践中可知,对于稳定对象,参数不确定性和不确定非线性的程度已知,即θɪΩθ {θ:θminɤθɤθmax};ΔɪΩd {Δ:|Δ(k)-Δ(k-1)|ɤδd(k)}㊂}(12)式中:θmin=[θ1min, ,θ3min]T;θmax=[θ1max, ,θ3max]T;δd是已知的㊂控制目标是设计自适应控制器使得系统的输出y(k)跟踪期望输出y d(k),定义跟踪误差函数为e(k)=y(k)-y d(k)㊂(13)定义s(k)为s(k)=e(k)-k1e(k-1)㊂(14)其中0<k1<1为待设计的可调参数㊂所以有s(k+1)=e(k+1)-k1e(k)㊂(15)自适应抗饱和控制律可以设计为:u(k)=1β^1(k)[u a(k)+u s1(k)+u s2(k)];u a(k)=-α^1(k)y(k)-α^2(k)y(k-1)+ y d(k+1)+k1e(k)-k cwη;u s1(k)=k s s(k);u s2(k)=-E1(k)㊂üþýïïïïïïïï(16)式中:k cwȡ β1 max为抗饱和反馈增益;|k s|<1是待设计的可调参数;E1(k)表达式为E1(k)=E1(k-1)+k s k2s(k-1)+βsat(s(k-1))㊂(17)式中:k2>0为可调系数;sat(㊃)为饱和函数㊂设计参数自适应律τ=s(k)φ(k-1),将式(9)改写为:θn(k)=θ^(k-1)+Γs(k)φ(k-1)λ+φT(k-1)φ(k-1);θ^(k)=projθ^(θn(k))㊂üþýïïï(18) 2.2㊀稳定性分析定理1:对于特征模型式(10)所描述的PMSLM,所有信号都是有界的㊂采用自适应控制律式(16)和参数更新规律式(18),能使闭环系统的跟踪误差渐近收敛至0㊂证明:将式(16)代入式(10)中,并结合式(18)可得s(k+1)=[y(k+1)-y d(k+1)]-k1e(k)=α^1(k)y(k)+α^2(k)y(k-1)+β^1(k)u(k)-α~1(k)y(k)-α~2(k)y(k-1)-β~1(k)u(k)+Δ(k)=-θ~T(k)φ(k)+β1η(k)-k cwη(k)+k s s(k)-E1(k)+Δ(k)㊂(19)取k cwȡ β1 max,然后对式(19)进行差分可得s(k+1)-s(k)=-(θ~T(k)φ(k)-θ~T(k-1)φ(k-1))+k s(s(k)-s(k-1))-(E1(k)-E1(k-1))+Δ(k)-Δ(k-1)㊂(20)考虑到采样周期很小,通过线性外推法预测可知s(k+1)=2s(k)-s(k-1)㊂(21)构建Lyapunov函数为V(k)=s(k)λ+φT(k-1)φ(k-1)+θ~(k) 2Γ㊂(22)首先考虑式(22)的第2项,根据投影参数自适应律式(18)可得θ~(k) 2ɤ θn(k)-θ(k) 2= θ~(k-1) 2+2Γs(k)φT(k-1)θ~(k-1)λ+ φ(k-1)Tφ(k-1) +(Γs(k))2 φ(k-1) 2(λ+ φ(k-1) 2)2ɤ2Γs(k)φT(k-1)θ~(k-1)λ+ φ(k-1) 2+Γ2s2(k)λ+ φ(k-1) 2+ θ~(k-1) 2㊂(23)将式(16)㊁式(20)和式(21)代入式(23)可得 θ~(k) 2- θ~(k-1) 2ɤ2Γs(k)[-(s(k)-s(k-1))+k s(s(k-1)-s(k-2))]λ+ φ(k-1) 2+ 2Γs(k)[-θ~T(k-2)φ(k-2)+k s k2s(k-1)-βsign(s(k-1))]λ+ φ(k-1) 2+531第3期曹㊀阳等:基于特征模型的永磁同步直线电机自适应控制2Γs (k )[(Δ(k -1)-Δ(k -2)]λ+ φ(k -1) 2+Γ2s 2(k )λ+ φ(k -1) 2㊂(24)选取βȡ| θM φmax +δd |,进一步可得 θ~(k ) 2- θ~(k -1) 2ɤ2Γs (k )(k s -1)(s (k )-s (k -1))+2Γk s k 2s (k )s (k -1)λ+ φ(k -1) 2+Γ2s 2(k )λ+ φ(k -1) 2㊂(25)引理1[20]:(Young 不等式)假设a ㊁b 为非负实数,P >1,1p +1q =1,那么ab ɤa p p +b pq ,当且仅当a p=b q时,等号成立㊂根据引理1可得:2s (k )s (k -1)ɤ s (k ) 2+ s (k -1) 2; θ~(k ) 2- θ~(k -1) 2ɤ-Γ(3-3k s -k s k 2)s 2(k )λ+ φ(k -1) 2+Γ(k s +k s k 2-1)s 2(k -1)λ+ φT (k -1) 2㊂üþýïïïïïï(26)对Lyapunov 函数式(22)进行差分,并联立式(26)可得ΔV (k )=V (k )-V (k -1)ɤs 2(k )λ+ φT (k -1) 2-s 2(k -1)λ+ φT (k -2) 2+-(3-3k s -k s k 2)s 2(k )λ+ φ(k -1) 2+(k s +k s k 2-1)s 2(k -1)λ+ φT (k -1) 2+Γs 2(k )λ+ φT (k -1) 2ɤ-(2-3k s -k s k 2-Γ)s 2(k )λ+ φT (k -1) 2+(k s +k s k 2-1)s 2(k -1)λ+ φT (k -1) 2-s 2(k -1)λ+ φT (k -2) 2ɤ-As 2(k )-Bs 2(k -1)㊂(27)式中:A =2-3k s -k s k 2-Γλ+ φT (k -1) 2;B =1λ+ φT (k -2) 2-1-k s -k s k 2λ+ φT (k -1) 2㊂通过选取合适的参数k s ㊁k 2㊁Γ㊁λ使得A >0,B >0㊂根据式(27),对Δ(k )从1到k 求和可得ðki =1[As 2(k )+Bs 2(k -1)]ɤV (1)-V (k )ɤV (1)㊂(28)当k ңɕ时,As 2(k )+Bs 2(k -1)ȡ0,由于φ(k ) 有界,可知lim k ңɕ|s (k )|=0㊂(29)根据式(29)可知,∃N ,当k >N 时,有|s (k )|ɤ0㊂(30)由式(15)可得|e (k )|ɤ|k 1||e (k -1)|+|s (k )|ɤ|k 1|k -N|e (N )|+|k 1|k -N -1|s (N +1)|+ +s (k )ɤ|k 1|k -N|e (N )|+0㊂(31)因为|k s |<1,所以有lim k ңɕsup |e (k )|=0㊂(32)3㊀实验结果比较为了说明上述方法的可行性和有效性,在实验室建立一个验证平台如图4所示,PMSLM 的基本参数列于表1㊂该平台由MOSFET 三相逆变桥㊁磁栅尺㊁相电流采样电路㊁TMS320F28062(DSP)及外围电路㊁IR2181S 驱动电路㊁系统电源电路组成㊂此外,为了模拟不同的工作条件,对直线电机的动子进行了调整㊂通过直接在动子上安装标准化铁块,准确地改变其质量m ,以模拟不同的惯性效应㊂图4㊀PMSLM 实验平台Fig.4㊀PMSLM experimental platform631电㊀机㊀与㊀控㊀制㊀学㊀报㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀第28卷㊀表1㊀PMSLM 的基本参数Table 1㊀Parameters of PMSLM㊀㊀参数数值极对数n p7极距τ/mm(180ʎ)12d 轴电感L d /mH 8q 轴电感L q /mH 8永磁体磁链ψf /Wb0.61PMSLM 矢量控制系统框架如图5所示㊂它由PMSLM㊁空间矢量脉宽调制(space vector pulse widthmodulation,SVPWM)模块㊁Park 和Clark 坐标变换㊁电压源逆变器㊁电流调节器和速度控制器组成㊂本文设计一种速度控制器,电流控制器采用PI 控制㊂图5㊀矢量控制总体结构框图Fig.5㊀Overall structure diagram of vector control为了验证所提控制器的可行性和有效性,本文对以下3种控制器进行比较㊂1)AARC㊂本文设计的抗饱和自适应鲁棒控制器参数设置如下:k 1=0.15,k 2=0.0006,k s =0.1,β=0.04,k cw =0.1,Γ=0.05,λ=0.995,θ^(0)=[1.9,-0.9,0.00001]T ㊂2)抗饱和自适应控制器(anti-windup adaptivecontrol based on characteristic model,AAC)㊂未添加鲁棒项u s2的抗饱和自适应控制器,其他参数与AARC 一致㊂3)抗饱和PID 控制器(anti-windup proportional-integral-differential,APID)㊂控制器的增益设置为k p =150,k i =1,k d =0,k cw =0.1㊂此外,将使用跟踪误差的最大值㊁平均值和标准差来衡量每个控制算法的质量,定义如下:1)最大跟踪误差的绝对值为M e =max i =1, ,N{|e (i )|}㊂(33)2)平均跟踪误差定义为μ=1N ðNi =1|e (i )|㊂(34)3)跟踪误差的标准差为δ=1N ðNi =1[|e (i )|-μ]2㊂(35)其中N 是所记录的数字信号的个数㊂首先将给定速度设置为y d =0.56sin(3.14t)m/s㊂系统跟踪结果如图6所示,性能指标如表2所示㊂从这些实验结果可以看出,所提出的AARC 控制器在瞬态和最终跟踪误差方面优于其他两种控制器,因为AARC 采用了基于参数自适应的补偿和鲁棒控制项,可以同时处理参数和未建模不确定性㊂虽然AAC 中也包含参数自适应,但对于建模的不确定性和未知扰动的抑制效果不佳㊂通过表2可以看出,AARC 添加鲁棒项后各种误差指标会比AAC 小,验证了鲁棒控制项u s2的有效性㊂在3种控制器中,线性抗饱和PID 的误差指标最差,达到了AARC 的2倍以上,这说明基于非线性模型的控制器设计方法具有更大的优势㊂图6㊀无铁块情况下PMSLM 的跟踪结果Fig.6㊀Tracking results of PMSLM without iron表2㊀最后两个周期的性能指标Table 2㊀Performance indexes during the last two cycles控制方法M e /(m /s)μ/(m /s)δ/(m /s)APID 0.055420.013360.00971AAC0.026890.008100.00572AARC 0.025220.006000.00490731第3期曹㊀阳等:基于特征模型的永磁同步直线电机自适应控制为了进一步验证控制器对参数变化的自适应能力,设定了不同的动子质量来进行实验㊂给PMSLM 的动子上添加1.33kg 的铁块㊂系统跟踪结果如图7所示,表3列出了最后两个周期的性能指标㊂从图7可以看出,使用AARC 控制方法的控制系统,在面对动子质量变化时,其反应速度快,并且波动较小㊂从表3可知,APID 的最大跟踪误差没有增大,意味着APID 中存在大的积分增益对该扰动也有一定的抑制效果㊂但与上一个实验情况相比,APID 的μ和δ指标增大明显,仍然比其他2个控制器差㊂适当的参数自适应在一定程度上也可以削弱动子质量变化给系统带来的参数不确定性影响,就像AAC 那样㊂AARC 的各项误差指标是3个控制器中最好的,再次证明了该控制器的有效性㊂图7㊀铁块质量为1.33kg 时PMSLM 的跟踪结果Fig.7㊀Tracking results of PMSLM when iron massis 1.33kg表3㊀最后两个周期的性能指标Table 3㊀Performance indexes during the last two cycles控制方法M e /(m /s)μ/(m /s)δ/(m /s)APID 0.043890.015370.01061AAC0.029620.008440.00605AARC 0.025320.005980.00496最后将动子上的铁块增加到2.64kg,此时PMSLM 受到的摩擦非线性和扰动进一步增大,3个控制器的跟踪性能都有所变差㊂实验结果如图8所示,误差指标见表4㊂在这个测试用例中,APID 中的跟踪误差抖动变大,而AARC 的跟踪误差则相当平滑㊂APID 控制器表现出最差的跟踪性能,最大跟踪误差为0.094,表明APID 在该跟踪任务中已经达到了其局限性㊂另外,即使在增大动子质量情况下,所提出的AARC 控制器仍然可以对模型进行补偿并衰减未建模的扰动,从而在所有比较的控制器中达到最好的跟踪性能㊂图8㊀铁块质量增加到2.64kg 情况下PMSLM 的跟踪结果Fig.8㊀Tracking results of PMSLM when the mass ofiron is increased to 2.64kg 表4㊀最后两个周期的性能指标Table 4㊀Performance indexes during the last two cycles控制方法M e /(m /s)μ/(m /s)δ/(m /s)APID 0.093700.027090.01934AAC0.034620.008410.00643AARC 0.028870.005860.005054㊀结㊀论本文针对PMSLM 系统提出一种基于特征模型的自适应控制方法,该方法能够有效地解决PMSLM 系统的参数不确定性㊁建模误差和外部干扰等问题㊂首先利用二阶变差分方程对PMSLM 系统进行简化831电㊀机㊀与㊀控㊀制㊀学㊀报㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀第28卷㊀建模,然后设计了一种基于特征模型的自适应控制器,仅利用系统的输入和输出信号,实现了对PMSLM系统的精确速度跟踪控制㊂为了提高系统的鲁棒性和抗饱和能力,还引入了鲁棒补偿项和抗饱和控制项,并严格证明了闭环系统的稳定性㊂最后,通过实验结果验证了所提控制方法的有效性㊂本文控制器的参数是固定的,需要通过反复调试来确认㊂当实验条件和环境发生改变时,可能导致参数不一定是最优的㊂因此,在未来工作中将考虑进一步研究控制器参数的自动调整技术[21],采用自学习的方法来替代控制器中参数的人工调整部分㊂参考文献:[1]㊀龚夕霞,李焱鑫,卢琴芬.模块化永磁直线同步电机考虑制造公差的推力鲁棒性优化[J].电工技术学报,2024,39(2):465.GONG Xixia,LI Yanxin,LU Qinfen.Thrust robustness optimiza-tion of modular permanent magnet linear synchronous motor ac-counting for manufacture tolerance[J].Transactions of China Electrotechnical Society,2024,39(2):465.[2]㊀张春雷,张辉,叶佩青.高霍尔位置检测精度的圆筒型永磁同步直线电机设计[J].电工技术学报,2022,37(10):2481.ZHANG Chunlei,ZHANG Hui,YE Peiqing.Design of tubular permanent magnet synchronous linear motor by reliability-based ro-bust design optimization[J].Transactions of China Electrotechni-cal Society,2022,37(10):2481.[3]㊀缪仲翠,苏乙,张磊,等.梯形Halbach交替极无铁心永磁同步直线电机特性分析与优化设计[J].电机与控制学报, 2024,28(1):164.MIAO Zhongcui,SU Yi,ZHANG Lei,et al.Characteristic analy-sis and optimization design of air-core permanent magnet linear synchronous motor with trapezoidal Halbach array consequent-pole [J].Electric Machines and Control,2024,28(1):164. [4]㊀李争,安金峰,肖宇,等.基于自适应观测器的永磁同步直线电机模型预测控制系统设计[J].电工技术学报,2021,36(6):1190.LI Zheng,AN Jinfeng,XIAO Yu,et al.Design of model predic-tive control system for permanent magnet synchronous linear motor based on adaptive observer[J].Transactions of China Electrotech-nical Society,2021,36(6):1190.[5]㊀李争,张梓豪,王康涛,等.基于无模型的PMLSM改进自适应滑模自抗扰控制[J].电机与控制学报,2024,28(1):142.LI Zheng,ZHANG Zihao,WANG Kangtao,et al.Improved adap-tive sliding mode active disturbance rejection control for PMLSM based on model-free theory[J].Electric Machines and Control, 2024,28(1):142.[6]㊀张康,王丽梅.基于周期性扰动学习的永磁直线电机自适应滑模位置控制[J].电机与控制学报,2021,25(8):132.ZHANG Kang,WANG Limei.Adaptive sliding mode position con-trol for permanent magnet linear motor based on periodic disturb-ance learning[J].Electric Machines and Control,2021,25(8): 132.[7]㊀孙宜标,毛爽,夏加宽.直线电机悬浮平台的μ-Hɕ鲁棒控制[J].沈阳工业大学学报,2014,36(1):7.SUN Yibiao,MAO Shuang,XIA Jiakuan.μ-Hɕrobust control for linear motor levitation platform[J].Journal of Shenyang Uni-versity of Technology,2014,36(1):7.[8]㊀YAO B,BU F,REEDY J,et al.Adaptive robust control of sin-glerod hydraulic actuators:theory and experiments[J].IEEE/ ASME Transactions on Mechatronics,2000,5(1):79. [9]㊀CHEN S,CHEN Z,YAO B,et al.Adaptive robust cascade forcecontrol of1-DOF hydraulic exoskeleton for human performance augmentation[J].IEEE/ASME Transactions on Mechatronics, 2017,22(2):589.[10]㊀HELIAN B,CHEN Z,YAO B.Adaptive robust motion control ofa pump direct drive electro-hydraulic system with meter-out pres-sure regulation[J].IFAC-Papers OnLine,2020,53(2):9005.[11]㊀HAN J,YANG S,XIA L,et al.Deterministic adaptive robustcontrol with a novel optimal gain design approach for a fuzzy2doflower limb exoskeleton robot system[J].IEEE Transactions onFuzzy Systems,2020,29(8):2373.[12]㊀LIU Y.Saturated robust adaptive control for uncertain nonlinearsystems using a new approximate model[J].IET Control Theoryand Applications,2017,11(6):870.[13]㊀王璐,郭毓,钟晨星,等.控制力矩陀螺框架伺服系统期望补偿自适应鲁棒控制[J].控制理论与应用,2017,34(9):1143.WANG Lu,GUO Yu,ZHONG Chenxing,et al.Desired com-pensation adaptive robust control for control moment gyroscopesgimbal servo system[J].Control Theory and Applications,2017,34(9):1143.[14]㊀CHANG Y,JIANG T,PU Z.Adaptive control of hypersonic ve-hicles based on characteristic models with fuzzy neural networkestimators[J].Aerospace Science and Technology,2017,68:475.[15]㊀常亚菲.一类不确定非线性系统基于特征模型的复合自适应控制[J].控制理论与应用,2019,36(7):1137.CHANG Yafei.Characteristic model-based composite adaptivecontrol for a class of uncertain nonlinear systems[J].ControlTheory and Applications,2019,36(7):1137. [16]㊀王永,窦晓华,方浩,等.永磁同步电机非线性黄金分割自适应转速控制[J].电机与控制学报,2017,21(10):23.WANG Yong,DOU Xiaohua,FANG Hao,et al.Nonlineargolden-section adaptive speed control of permanent magnet syn-chronous motor[J].Electric Machines and Control,2017,21(10):23.931第3期曹㊀阳等:基于特征模型的永磁同步直线电机自适应控制[17]㊀WANG X,WU Y,ZHANG E,et al.Adaptive terminal sliding-mode controller based on characteristic model for gear transmis-sion servo systems[J].Transactions of the Institute of Measure-ment and Control,2019,41(1):219.[18]㊀QI L,BAO S,SHI H.Permanent-magnet synchronous motor ve-locity control based on second-order integral sliding mode controlalgorithm[J].Transactions of the Institute of Measurement andControl,2015,37(7):875.[19]㊀张兴华,姚丹.感应电机直接转矩控制系统的 抗饱和 控制器设计[J].电工技术学报,2014,29(5):181.ZHANG Xinghua,YAO Dan.Anti-windup speed controller de-sign for direct torque controlled induction motor drives[J].Trans-actions of China Electrotechnical Society,2014,29(5):181.[20]㊀IGHACHANE M A,AKKOUCHI M.Further refinements ofYoung's type inequality for positive linear maps[J].Revista de laReal Academia de Ciencias Exactas,Físicas y Naturales.SerieA.Matemáticas,2021,115(2):10.[21]㊀BENOSMAN M.Learning-based adaptive control for nonlinearsystems[C]//European Control Conference(ECC),June24-27,2014,Strasbourg,France.2014:920-925.(编辑:邱赫男)041电㊀机㊀与㊀控㊀制㊀学㊀报㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀第28卷㊀。
基于边缘检测的抗遮挡相关滤波跟踪算法
基于边缘检测的抗遮挡相关滤波跟踪算法唐艺北方工业大学 北京 100144摘要：无人机跟踪目标因其便利性得到越来越多的关注。
基于相关滤波算法利用边缘检测优化样本质量，并在边缘检测打分环节加入平滑约束项，增加了候选框包含目标的准确度，达到降低计算复杂度、提高跟踪鲁棒性的效果。
利用自适应多特征融合增强特征表达能力，提高目标跟踪精准度。
引入遮挡判断机制和自适应更新学习率，减少遮挡对滤波模板的影响，提高目标跟踪成功率。
通过在OTB-2015和UAV123数据集上的实验进行定性定量的评估，论证了所研究算法相较于其他跟踪算法具有一定的优越性。
关键词：无人机 目标追踪 相关滤波 多特征融合 边缘检测中图分类号：TN713;TP391.41;TG441.7文献标识码：A 文章编号：1672-3791(2024)05-0057-04 The Anti-Occlusion Correlation Filtering Tracking AlgorithmBased on Edge DetectionTANG YiNorth China University of Technology, Beijing, 100144 ChinaAbstract: For its convenience, tracking targets with unmanned aerial vehicles is getting more and more attention. Based on the correlation filtering algorithm, the quality of samples is optimized by edge detection, and smoothing constraints are added to the edge detection scoring link, which increases the accuracy of targets included in candi⁃date boxes, and achieves the effects of reducing computational complexity and improving tracking robustness. Adap⁃tive multi-feature fusion is used to enhance the feature expression capability, which improves the accuracy of target tracking. The occlusion detection mechanism and the adaptive updating learning rate are introduced to reduce the impact of occlusion on filtering templates, which improves the success rate of target tracking. Qualitative evaluation and quantitative evaluation are conducted through experiments on OTB-2015 and UAV123 datasets, which dem⁃onstrates the superiority of the studied algorithm over other tracking algorithms.Key Words: Unmanned aerial vehicle; Target tracking; Correlation filtering; Multi-feature fusion; Edge detection近年来，无人机成为热点话题，具有不同用途的无人机频繁出现在大众视野。
运动性肌肉疲劳
运动生n适理学用研范究发围现：，动主态要和静用态于运动手负部荷诱肌发肉肌肉和疲部劳过分程上中主下要肢运动肌肌肉的平力均量功率的（评Mea价n po。wer frequency, MPF）和中位频率（ M主e要dia强nn优调fre发q点u生en：在cy,中M最枢F）大神均经电呈系单刺统调（激递Ce减肌nt变ra力l化ne，能rv变ou够化s s率y在st大em一小,C取N定S决）的于中运程的动神度强经度上生并理减与、肌小生肉化M疲过V劳程C度及或测其肌在试肉运耐过动力性程明肌显中肉相疲的关劳。主发生观发努展过力程程中的作用（ 中运枢动机 性度制肌效）肉。疲应劳，（E但xer是cise由-ind于uce个d m人uscl最e fat大igue电）：刺运激动引主起观的肌耐肉受收缩性产的生最差大异主动，收仍缩力然量无（M法ax消imal除vol该unta因ry c素ont对racti疲on,MVC）或 者运最动大 性劳输肌检出肉功疲测率劳结的(M检a果xi测m的a通l p常影ow分e响r为o直u。tp接ut检)暂测时法性和下间降接的检生测理法现，象前。者直接检测运动肌最大运动抗阻能力，后者则主要依据肌肉疲劳过程中 其适他用相 范3、关围的：表生主面理要生适肌化用电指于标四检变肢测化肌来肉评，价而疲对劳于的其程他度一，些此力外学，关人系体比主较观复疲杂劳部感位也（是如常肩用部的、评腰价部指）标的。肌肉疲劳的评价则效度较低。 新的中n表枢疲面劳肌模型电：（由南S非u开rf普a敦ce大学el的eＮctａroｏmｋyｅoｓgTr.aph,sEMG）信号：是神经肌肉系统活动时的 早神期经的 递生中质物枢对疲于电劳有变理关论神化：经经中细枢胞表疲的面劳活是动电由而极运言动是引时具导与有“C、N毒S放兴性奋”大性的、有副关产显的物示中，枢它和神们记经可递以录质降所，低如C获N5-S得H募T集、的运氨一动和单细维位胞电的因能子压力等时和含影量间响增序大加脑所列的引信运起动，号控这，制些功中能枢 ，②从长而 时其在间振大耐脑力幅发运为动动随中0-意血5运氨0动主00时要μ无来V法自，动BC员频AA相在率应运数3动0量肌-的3中运5的0动降H单解z位。。参加活动和控制相关肌肉的协调活动，进而造成肌肉疲劳。 骨儿骼茶肌 酚n研、胺腱能究鞘够表周抑围制明的色：组氨织酸s和羟E化CMN酶SG是的信运活动性号中，源血因液而于I儿L大-茶6增酚脑多胺的运具来有动源减。皮少5层-HT控生制物合之成下的作的用脊。 髓α运动神经元的生物电 在亚极活量动最大，负形荷的成耐于力运众动多过程外中周，肌运肉动和肝单脏位糖原在储时备的间耗和竭被空认间为与上疲的劳的总发和生有。关信。 号的振幅和频率特征
Geometric Modeling
Geometric ModelingGeometric modeling is a crucial aspect of computer graphics, engineering, and design. It involves creating digital representations of objects and environments using mathematical and computational techniques. Geometric modeling plays a significant role in various industries, including architecture, automotive design, video game development, and virtual reality. The process of geometric modeling allows designers and engineers to visualize and analyze complex structures, simulate real-world scenarios, and create realistic visualizations of their ideas. One of the key challenges in geometric modeling is achieving a balance between accuracy and efficiency. Designers and engineers often need to create highly detailed models with complex geometries, which can be computationally expensiveand time-consuming. At the same time, they also need to ensure that the models can be manipulated and rendered in real-time for interactive applications. This trade-off between accuracy and efficiency requires careful consideration of the modeling techniques and algorithms used to represent and manipulate geometric data.Another important consideration in geometric modeling is the representation of curved surfaces and freeform shapes. While simple geometric primitives such as cubes, spheres, and cylinders can be easily defined using mathematical equations, representing more complex shapes like human bodies, organic forms, and natural landscapes requires more advanced techniques. B-spline and NURBS (Non-Uniform Rational B-Splines) are commonly used to represent and manipulate curved surfacesin geometric modeling, allowing for smooth and flexible deformation of shapes. Geometric modeling also involves the creation of 3D models from 2D sketches or images. This process, known as 3D reconstruction, requires the use of computer vision and image processing techniques to extract depth and spatial information from 2D data. 3D reconstruction has applications in fields such as medical imaging, remote sensing, and augmented reality, where 3D models are generated from 2D images to facilitate analysis and visualization. In addition to creating static3D models, geometric modeling also encompasses the simulation and animation of dynamic objects and environments. Physics-based modeling techniques are used to simulate the behavior of physical systems, such as the motion of rigid bodies, the deformation of elastic materials, and the interaction of fluids and gases. Thesesimulations are essential for applications like virtual prototyping, computer-aided engineering, and special effects in movies and video games. Moreover, geometric modeling is closely related to the field of computational geometry, which focuses on the development of algorithms and data structures for solving geometric problems. Computational geometry has applications in areas such as computer-aided design, robotics, geographic information systems, and computer graphics. It addresses fundamental problems like geometric intersection, proximity queries, convex hull computation, and mesh generation, which are essential for many geometric modeling tasks. In conclusion, geometric modeling is a multifaceted discipline that encompasses a wide range of techniques and applications. It plays a critical role in various industries and research fields, enabling the creation, analysis, and visualization of complex geometric data. The challenges in geometric modeling, such as balancing accuracy and efficiency, representing curved surfaces, reconstructing 3D models from 2D data, simulating dynamic systems, and solving fundamental geometric problems, require innovative solutions and advancements in computational and mathematical techniques. As technology continues to evolve, geometric modeling will continue to be anessential tool for shaping the virtual and physical world around us.。
Finite element model updating in structural dynamics by using the response surface method
Engineering Structures32(2010)2455–2465Contents lists available at ScienceDirectEngineering Structures journal homepage:/locate/engstructFinite element model updating in structural dynamics by using the response surface methodWei-Xin Ren∗,Hua-Bing ChenDepartment of Civil Engineering,Central South University,Changsha,410075,ChinaNational Engineering Laboratory for High Speed Railway Construction,Changsha,410075,Chinaa r t i c l e i n f o Article history:Received3February2009 Received in revised form6April2010Accepted7April2010 Available online4May2010 Keywords:Response surfaceFinite elementModel updatingOptimizationDesign of experiment Regression analysis a b s t r a c tFast-running response surface models that approximate multivariate input/output relationships of time-consuming physical-based computer models enable effective finite element(FE)model updating analyses. In this paper,a response surface-based FE model updating procedure for civil engineering structures in structural dynamics is presented.The key issues to implement such a model updating are discussed such as sampling with design of experiments,selecting the significant updating parameters and constructing a quadratic polynomial response surface.The objective function is formed by the residuals between analytical and measured natural frequencies.Single-objective optimization with equal weights of natural frequency residual of each mode is used for optimization computation.The proposed procedure is illustrated by a simulated simply supported beam and a full-size precast continuous box girder bridge tested under operational vibration conditions.The results have been compared with those obtained from the traditional sensitivity-based FE model updating method.The real application to a full-size bridge has demonstrated that the FE model updating process is efficient and converges fast with the response surface to replace the original FE model.With the response surface at hand,an optimization problem is formulated explicitly.Hence,no FE calculation is required in each optimization iteration.The response surface-based FE model updating can be easily implemented in practice with available commercial FE analysis packages.©2010Elsevier Ltd.All rights reserved.1.IntroductionNowadays the finite element(FE)method has become an important and practical numerical analysis tool.It is commonly used in almost all areas of engineering analysis.However,the FE model of a structure is normally constructed on the basis of highly idealized engineering blueprints and designs that may not truly represent all the aspects of an actual structure.As a result, the analytical predictions from a FE model often differ from the results of a real structure.These discrepancies originate from the uncertainties in simplifying assumptions of structural geometry, materials as well as inaccurate boundary conditions.It is often required to update or calibrate the uncertain parameters of a FE model that leads to the better predictions of the responses of an actual structural.Finite element model updating is such a procedure that mod-ifies or updates the uncertainty parameters in the initial finite element model based on the experimental results so that a more realistic or refined model can be achieved[1].In other words,FE ∗Corresponding author at:Department of Civil Engineering,Central South University,Changsha,410075,China.Tel.:+8673182654349;fax:+86731 85571736.E-mail addresses:renwx@,renwx@(W.-X.Ren).model updating is the process of using experimental results to re-fine a mathematical model of a physical structure.Basically,FE model updating is an inverse problem to identify or correct the uncertain parameters of FE models.It is usually posed as an op-timization problem.Setting-up of an objective function,selecting updating parameters and using robust optimization algorithm are the three crucial steps in FE model updating.In a model updating process,not only the satisfactory correlation is required between analytical and experimental results,but also the updated param-eters should preserve the physical significance.The updated FE models are used in many applications for civil engineering struc-tures such as damage detection,health monitoring,structural con-trol,structural evaluation and assessment.Finite element model updating is a topic of significant interest in the field of structural dynamics.A number of FE model updating methods in structural dynamics have been proposed.The direct updating methods compute a closed-form solution for the global stiffness and/or mass matrices using the structural equations of motion and the orthogonality equations.These non-iterative methods that directly update the elements of stiffness and mass matrices are one-step procedures[2,3].The resulting updated matrices reproduce the measured structural modal properties well but do not generally maintain structural connectivity and the corrections suggested are not always physically meaningful.0141-0296/$–see front matter©2010Elsevier Ltd.All rights reserved. doi:10.1016/j.engstruct.2010.04.0192456W.-X.Ren,H.-B.Chen/Engineering Structures32(2010)2455–2465The iterative parameter updating methods involve using the sensitivity of the parameters to find their changes.The sensitivity-based FE model updating methods are often posed as optimization problems.These methods set the errors of the structural response features between analytical and experimental results as an objective function and try to minimize the objective function by making changes to the pre-selected set of physical parameters of the FE model.Link[4]gave a clear overview of the sensitivity-based updating methods in structural dynamics.It is noted that the mathematical model used in the model updating is usually ill posed and the special attention is required for an accurate solution[5].Jaishi and Ren[6–8]used either single-objective or multi-objective optimization technique to update the FE models of civil engineering structures in structural dynamics.However,the sensitivity-based method involving in the determination of local gradients at points may cause not only computational intensive, but also convergence difficulty.If the structure of interest is represented by,e.g.a large finite element model,the large number of computations involved can rule out many approaches due to the expense of carrying out many runs.For such a large FE model where so many elements are involved,there are huge of both geometric and physical possible parameters to be updated.In addition,there are now many commercial finite element analysis packages available such as ANSYS,ABAQUS and SAP2000et al..The structural FE models are often constructed by using these packages.In all the iterative parameter updating methods,each iteration needs to go back to run the finite element analysis package with any parameter updated,which limits the popular applications of structural FE model updating in practice.One way of circumnavigating the time-consuming and FE analysis package-related problems during the sensitivity-based model updating is to replace the FE model by an approximate surrogate/replacement meta-model that is fast-running and less parameters involved[9].Such a meta-model is easier to compute with,because it is controlled only by a few explanatory variables. The FE model updating is carried out on the meta-model instead of the analytical FE model.Response surface is one of the commonly used meta-models.Response surface methodology is originally an experimental design approach for selecting design parameters for experiments with the objective of optimizing some function of a response[10–12].It provides a mechanism for guiding experimentation in search of optimal settings for design parameters or optimal values of unknown response.Many additional applications(largely a consequence of the increased use of computational analyses)have broadened the range of application of response surface methods in the statistical and engineering literature.Recent literature has addressed more flexible functional forms for modeling the response,new methods of response surface construction[13],alternate approaches to updating the surface estimate[14],new sampling methods[15], etc.In many fields of engineering,the term‘‘response surface’’is used synonymously with‘‘meta-model’’or‘‘surrogate model’’, which refer to any relatively simple mathematical relationship between parameters and a response,often based on limited data[16].In the case of structural finite element model updating, once the response surface of a structure has been constructed, updating the model is reduced to the task of finding the smallest value on the response surface.The parameter values that correspond to this smallest value are those that are used to update the model.Recently,the response surface that is represented by a simple least-squares multinomial model has been adopted in structural FE model updating,verification and validation[17–20].The response surface method for damage detection and reliability analysis is not quite new[21,22].However,the response surface method for structural finite element model updating is somewhat new,especially with the civil engineering communities. This paper is intended to present a response surface-based finite element model updating procedure in structural dynamics and to take advantages of response surfaces for the FE model updating of civil engineering structures in practice.Its purpose is to estimate the values of structural parameters(moment of inertia,cross-sectional area and modulus of elasticity)based on the measured response quantities(natural frequencies).The proposed procedure is based on constructing quadratic response surfaces.Those surfaces represent an estimate for the relation between the unknown parameters of the finite element model and response quantities of interest.With the response surface at hand,an optimization problem,whose solution is the estimate for the values of the structural parameters,is formulated explicitly. Hence,no further finite element simulations are required.The objective function is formed by the residuals between analytical and measured natural frequencies.Single-objective optimization with equal weights of each natural frequency is implemented for optimization computation.The presented procedure is illustrated by a simulated simply supported beam and a full-size precast continuous box girder bridge tested under operational vibration conditions.The results have been compared with those obtained from the traditional sensitivity-based FE model updating method. The real application to a full-size bridge has demonstrated that the response surface-based FE model updating procedure is simple and fast so that it can be easily implemented in practice.2.Response surface-based finite element model updatingResponse surface-based finite element model updating is an approach to achieve the global approximations of the structural response feature objectives and constrains based on functional evaluations at various points in the design space.It often involves experimental strategies,mathematical methods,probability and statistical inference that enable an experimenter to make efficient empirical exploration of the structure of interest.The flowchart of response surface-based finite element model updating is shown in Fig.1.The main steps include the following.•The selection of the structural parameters and the definition ofa number of‘‘level’’for each selected parameters by using thedesign of experiments(DOE)techniques.•In design space,the response features are calculated by carrying out finite element analysis(FEA).•Performing the final regression followed by a regression error analysis to create the response surface model of the structure.•The response features of the structure are measured and corresponding objective functions(feature residuals)to be minimized are constructed.•The finite element model updating(iteration)is carried out within the established response surface model.Updated parameters are obtained and transferred to the original finite element model.2.1.Sampling and parameter selectionTo create a response surface that will serve as a surrogate for the FE simulation model,the basic process is one of calculating predicted values of the response features at various sample points in the parameter space by performing a experiment at each of those points.A number of feature values from the experiment ran across the parameter domain are fit with a response surface. The key is to select the parameters carefully,to minimize the number of dimensions in the parameter space,and then to selectW.-X.Ren,H.-B.Chen /Engineering Structures 32(2010)2455–24652457Fig.1.Flowchart of response surface-based FE model updating.the combinations of parameter values where the experiment is performed.The term experiment herein refers to either physical experi-ments or computer experiments.The planning of experimentation is further referred to design of experiments (DOE).The selection of sample points is related to the accuracy and cost of a response sur-face to be constructed.Less sample points may reduce the surface accuracy,while more sample points may improve the surface accu-racy but increase the work load.In the real application,the sample points mainly depend on the problem to be solved,the response feature values of interest and the selected method of DOE.DOE plays an important role in constructing a response surface.Two efficient DOE methods are commonly used.They are central composite design (CCD)method and D-optimal design method.There are several other DOE methods that can be used in constructing a response surface such as box-behnken design,Latin Square design,fractional factor design [12].For the purpose of structural finite element model updating in structural dynamics,Guo and Zhang [18]found that both CCD and D-optimal design methods can achieve almost the same accuracy in the creation of polynomial surfaces.In the current study,the CCD method in DOE is used in the paper as it is simple in constructing the response surfaces of a polynomial type.Central composite design uses the orthogonal table to perform the experimentation to determine the sample points of selected parameters.It contains a fractional factorial design 2k (levels are ±1and k is quantity of factors)with central points that are augmented with a group of 2k star points that allow for the estimation of curvature.2k star points are (±α,0,...,0),(0,±α,...,0),...,(0,0,...,±α).For the purpose of constructing high-order surfaces,another 2k star points (±α1,0,...,0),(0,±α1,...,0),...,(0,0,...,±α1)are appended.The precise values of αand α1rely on certain properties such as orthogonality and rotatability desired for the design and on the number of factors involved [12].The selection of updating parameters to calculate the response features of a structure is an important issue in FE model updating.The structural parameters selected for updating should be able to clarify the ambiguity of the model,and in that case it is necessary for the model response to be sensitive to these parameters.The problem always arises:How many parameters should be selected and which parameters from many possible parameters are used in the FE model updating?If too many parameters are included in the FE model updating in structural dynamics,the optimized problem may appear ill-conditioned because only limited vibration modes can be effectively identified from the field vibration measurements.The parameter selection requires a considerable physical insight into the target structure,and trial-and-error approaches are often used with different set of selected parameters for complicated structures.In the frame of response surface-based FE model updating,the selected updating parameters can be evaluated from the sampled data by performing a parameter effect analysis (hypothesis testing)based on statistical variance (square of the standard deviation)analysis.Compared with tests on means for a hypothesis testing,tests on variances are rather sensitive to the normality assumption.Variance analysis formulates all individual square of deviations of sample data.The basic idea of variance analysis is to divide the total square deviation of sampling features into two parts:S A —a square of deviation caused by design parameter A (system deviation)and S e —a square of deviation caused by the experiment.The F-test method [12]can be used to carry out the hypothesis testing to check the significance of parameter A:F A =S A /f A S e /f e∼F (f A ,f e )(1)where f A and f e are the degrees of freedom of S A and S e respectively.For a given significance level p ,if F A ≥F 1−p (f a ,f e ),the effect of parameter A is significant.2.2.Response surface regressionThe family of response surface forms selected to represent a response can have a substantial impact on the results of an analysis.The selected response surface form should be capable of attaining surfaces that meet specific smoothness requirements of an application.There is often a balance between assumptions and data requirements.Different surface families may be preferred for different applications.In the case of structural finite element model updating in structural dynamics,polynomials are popular forms representing a response surface because the calculations are simple and the resulting function is closed-form algebraic expression.For example,a quadratic polynomial response surface has the form:y =β0+k i =1βi x i +k i =1k j =1βij x i x j(2)where β0,βi ,βij are the regression coefficients to be estimatedfrom the experimental data.In such a way,a response surface y is represented as a function of the parameters or variables x of the design space (generally some subset of Euclidian k -space,R k ).In the case of k =2,the vector of surface parameters is six dimensional.The number of sampling points n must be greater than or equal to the number of terms in the polynomial.When there are more sampling points,the equation is over-determined and the regression techniques are required to fit the response surface to the sampled data.In this case,the surface does not,in general,exactly match the response values at every sample points.The method2458W.-X.Ren,H.-B.Chen/Engineering Structures32(2010)2455–2465 Table1of least-square fitting is usually used in the coefficient estimationprocess to create a response surface.Before the regressed response surface is put forward to be usedin structural FE model updating,it should be verified to checkwhether the regressed surface has enough accuracy.If not,theparameters are adjusted(based on the response data)to achievea balance between variability and bias of the regressed surface.R2(ranged from0.0to1.0)criterion and empirical integrated squaredError(EISE)criterion can be used in response surface verification:R2=1−Ni=1[y RS(j)−y(j)]2Nj=1[y(j)−¯y]2(3)EISE=1N∗Nj=1[y RS(j)−y(j)]2(4)where y RS is the response value of the confirmation samples;y is the true value of the confirmation samples;¯y is the mean of all true values.The larger the value of R2,the more accurate the regressed response surface.On the contrary,the smaller the value of EISE,the closer the fit is to the data.Another criterion is the root mean squared error(RMSE).It is the square root of the mean square error(MSE)where the distance vertically from the point is taken to the corresponding point on the curve fit(the error).The squaring is done so negative values do not cancel positive values.The RMSE is thus the distance,on average,of a data point from the fitted line,measured along a vertical line.The smaller the value of root mean squared error,with 0corresponding to the ideal,the closer the fit is to the data.In the real application,the R2criterion and EISE or the RMSE criterion can be used in a complementary way to check the accuracy of the regressed response surface.2.3.Model updatingIn the finite element model updating in structural dynamics,the structural response features of interest are often eigen solutions related to such as natural frequencies and mode shapes.In this study,the structural natural frequency is employed as a response feature.Therefore,the optimized objective function is formulated in terms of the residuals between analytical and measured natural frequenciesΠ(x)=mi=1w i(λai−λei)0≤w i≤1(5)whereλai andλei are the analytical(finite element calculated) and measured natural frequencies of the i th mode respectively. w i is the weight factor to impose to the different order of naturalfrequencies.m is the number of modes involved and x is the design set.The FE model updating can then be posed as a constrained minimization problem to find the satisfied design set such that Min Π(x) 22(6)subjected tox lk≤x k≤x uk,k=1,2,3,...where x lk and x uk are the upper and lower bounds on the k th design variable to be set.N is the total number of design variables (updated parameters).Single-objective optimization with equal weight of each natural frequency is used in this paper for optimization computation.3.Numerical verificationA simulated simply supported beam without damage and with an assumed damage element is considered to demonstrate the procedure of response surface-based FE model updating.The beam of6m length is equally divided into15two-dimensional beam elements as shown in Fig.2.The density and elastic modulus of the beam are2500kg/m3and3.2×1010Pa,respectively,while the area and moment of inertia of cross-section are0.05m2and 1.66×10−4m4,respectively.FE modal analysis is first carried out on the beam without damage to get the analytical natural frequencies.To simulate the actual(target)beam,one damage location is assumed at beam element10where the element bending stiffness is reduced by50%. The FE modal analysis is again carried out on this damaged beam to get the simulated measured modal parameters.The initial values of the first10natural frequencies selected and corresponding differences are shown in Table1.The maximum and minimum errors that appeared in the natural frequency are15.63%and1.89% respectively.Updating of the FE model of the undamaged beam is to achieve a goal to correlate the natural frequencies with the damaged beam. In this study,moment of inertia I10and elastic modulus E10of individual element10are chosen as the updating parameters that affect the independent variable,natural frequencies.The range selected for each parameter should reflect the change that one expects to observe for the domain of the prediction of interest with 1.920≤E10≤3.465(×1010Pa)and0.863≤I10≤1.797 (×10−4m4).The central composite design(CCD)method in design of experiments is then used to get the sample points of the parameter values selected.A total of9runs of experiments are carried out.The sampled parameter values and corresponding natural frequencies calculated from FE model are listed in Table2.By using least-square fitting with these sample values,a quadratic polynomial response surface(Eq.(2))can be regressed where parameter x1refers to E10while parameter x2refers to I10.Fig.3illustrates the regressed response surfaces of the first and second natural frequencies.The horizontal axes are parameters selected while the vertical axis gives the response(natural frequency)at any point or location.To evaluate the above-selected parameters of the FE model,a statistics-based parameter effects analysis is performed.Parts of the general regression significance of selected parameters(x1and x2)calculated from Eq.(1)on each mode of natural frequencies comparing with the significance level of p=0.05are shown in Fig.4.The results show that both elastic modulus E10and moment of inertia I10are high significance on the structural natural frequencies.It is observed that the cross-quadratic term x1x2inW.-X.Ren,H.-B.Chen /Engineering Structures 32(2010)2455–24652459Fig.2.Simulated simply supported beam.Table 2Central composite design and sample values.9.598.58F r e q u e n c y7.576.53.532.5x 1×1010x 2×10–4221.51.510.5F r e q u e n c yx 1×1011.522.533.5x 2×10–40.511.523836343230(a)Surface of first natural frequency.(b)Surface of second natural frequency.Fig.3.Typical regressed response surfaces.Table 3the surface expression is less significant on the structural natural frequencies.For the simplification of a quadratic polynomial response surface,the cross-quadratic term can be omitted in practice to increase the efficiency of structural FE model updating.To check the accuracy of the regressed surface,the RMSE is calculated for each mode as shown in Table 3.It is demonstrated that all RMSE values are close to zero,which indicates that the created response surface has a high regression accuracy.Now it is time to carry out the FE model updating where the FE model is replaced by the regressed quadratic polynomial.The residuals between analytical (undamaged beam)and measured (damaged beam)natural frequencies are used as the optimized objective function.Single-objective optimization algorithm with equal weight of each natural frequency is implemented to achieve the best minimization of natural frequency residuals.The optimization algorithm used to minimize the objective function is a standard Trust Region Newton method in MATLAB.The tuning minimization process is over when the tolerances are achieved or pre-defined number of iterations is reached.The updated natural frequencies and differences of the simulated simply supported beam are also summarized in Table 1.It can be observed that a good agreement of natural frequencies has been achieved after carrying out the response surface-based FE model updating.The2460W.-X.Ren,H.-B.Chen/Engineering Structures32(2010)2455–2465(a)Parameter significance on the first natural frequency.(b)Parameter significance on the second natural frequency.Fig.4.The regression significance of selected parameters on the naturalfrequencies.Fig.5.Convergence of objective function of response surface-based updating.final updated results on the parameters are E10=1.57×1010Pa(true value E10=1.60×1010Pa)and I10=0.85×10−4m4(truevalue I10=0.83×10−4m4).Fig.5illustrates the convergence of the objective functionduring each optimization iteration.The horizontal axis is thenumber of iteration while the vertical axis is the square sum ofnatural frequency residuals of each mode.To demonstrate theefficiency of current response surface-based FE model updating,the objective function convergence of the traditional sensitivity-based FE model updating is shown in Fig.6.For such a numericalsimple beam,it is shown that the convergence of the objectivefunction is fast and iteration number is dramatically reduced toreach the same residual level of the objective function by using theresponse surface method.4.A precast concrete bridge tested in the fieldA real case study is carried out on the Hongtang Bridge,locatedin Fuzhou city,the capital of Fujian province,China.The bridge is amulti-span continues-deck precast concrete highway bridge.Theconstruction was completed in December1990.The total lengthof the bridge is1843m,with a span layout of(16+27+4∗30+60+120+60+31∗40+8∗25)m.It includesthreeFig.6.Convergence of objective function of sensitivity-basedupdating.Fig.7.Potion of Hongtang bridge.types of spans:simply supported spans,precast truss supportedspans and precast continuous girder spans.The photograph ofthe part of bridge at present condition is shown in Fig.7.Forthe purpose of dynamics-based condition assessment,one portionof six continuous girder spans with a span length of40m weretested on the site under operational vibration conditions.On-sitedynamic testing of a structure provides an accurate and reliabledescription of its current dynamic characteristics.。
特征更新的动态图卷积表面损伤点云分割方法
第41卷 第4期吉林大学学报(信息科学版)Vol.41 No.42023年7月Journal of Jilin University (Information Science Edition)July 2023文章编号:1671⁃5896(2023)04⁃0621⁃10特征更新的动态图卷积表面损伤点云分割方法收稿日期:2022⁃09⁃21基金项目:国家自然科学基金资助项目(61573185)作者简介:张闻锐(1998 ),男,江苏扬州人,南京航空航天大学硕士研究生,主要从事点云分割研究,(Tel)86⁃188****8397(E⁃mail)839357306@;王从庆(1960 ),男,南京人,南京航空航天大学教授,博士生导师,主要从事模式识别与智能系统研究,(Tel)86⁃130****6390(E⁃mail)cqwang@㊂张闻锐,王从庆(南京航空航天大学自动化学院,南京210016)摘要:针对金属部件表面损伤点云数据对分割网络局部特征分析能力要求高,局部特征分析能力较弱的传统算法对某些数据集无法达到理想的分割效果问题,选择采用相对损伤体积等特征进行损伤分类,将金属表面损伤分为6类,提出一种包含空间尺度区域信息的三维图注意力特征提取方法㊂将得到的空间尺度区域特征用于特征更新网络模块的设计,基于特征更新模块构建出了一种特征更新的动态图卷积网络(Feature Adaptive Shifting⁃Dynamic Graph Convolutional Neural Networks)用于点云语义分割㊂实验结果表明,该方法有助于更有效地进行点云分割,并提取点云局部特征㊂在金属表面损伤分割上,该方法的精度优于PointNet ++㊁DGCNN(Dynamic Graph Convolutional Neural Networks)等方法,提高了分割结果的精度与有效性㊂关键词:点云分割;动态图卷积;特征更新;损伤分类中图分类号:TP391.41文献标志码:A Cloud Segmentation Method of Surface Damage Point Based on Feature Adaptive Shifting⁃DGCNNZHANG Wenrui,WANG Congqing(School of Automation,Nanjing University of Aeronautics and Astronautics,Nanjing 210016,China)Abstract :The cloud data of metal part surface damage point requires high local feature analysis ability of the segmentation network,and the traditional algorithm with weak local feature analysis ability can not achieve the ideal segmentation effect for the data set.The relative damage volume and other features are selected to classify the metal surface damage,and the damage is divided into six categories.This paper proposes a method to extract the attention feature of 3D map containing spatial scale area information.The obtained spatial scale area feature is used in the design of feature update network module.Based on the feature update module,a feature updated dynamic graph convolution network is constructed for point cloud semantic segmentation.The experimental results show that the proposed method is helpful for more effective point cloud segmentation to extract the local features of point cloud.In metal surface damage segmentation,the accuracy of this method is better than pointnet++,DGCNN(Dynamic Graph Convolutional Neural Networks)and other methods,which improves the accuracy and effectiveness of segmentation results.Key words :point cloud segmentation;dynamic graph convolution;feature adaptive shifting;damage classification 0 引 言基于深度学习的图像分割技术在人脸㊁车牌识别和卫星图像分析领域已经趋近成熟,为获取物体更226吉林大学学报(信息科学版)第41卷完整的三维信息,就需要利用三维点云数据进一步完善语义分割㊂三维点云数据具有稀疏性和无序性,其独特的几何特征分布和三维属性使点云语义分割在许多领域的应用都遇到困难㊂如在机器人与计算机视觉领域使用三维点云进行目标检测与跟踪以及重建;在建筑学上使用点云提取与识别建筑物和土地三维几何信息;在自动驾驶方面提供路面交通对象㊁道路㊁地图的采集㊁检测和分割功能㊂2017年,Lawin等[1]将点云投影到多个视图上分割再返回点云,在原始点云上对投影分割结果进行分析,实现对点云的分割㊂最早的体素深度学习网络产生于2015年,由Maturana等[2]创建的VOXNET (Voxel Partition Network)网络结构,建立在三维点云的体素表示(Volumetric Representation)上,从三维体素形状中学习点的分布㊂结合Le等[3]提出的点云网格化表示,出现了类似PointGrid的新型深度网络,集成了点与网格的混合高效化网络,但体素化的点云面对大量点数的点云文件时表现不佳㊂在不规则的点云向规则的投影和体素等过渡态转换过程中,会出现很多空间信息损失㊂为将点云自身的数据特征发挥完善,直接输入点云的基础网络模型被逐渐提出㊂2017年,Qi等[4]利用点云文件的特性,开发了直接针对原始点云进行特征学习的PointNet网络㊂随后Qi等[5]又提出了PointNet++,针对PointNet在表示点与点直接的关联性上做出改进㊂Hu等[6]提出SENET(Squeeze⁃and⁃Excitation Networks)通过校准通道响应,为三维点云深度学习引入通道注意力网络㊂2018年,Li等[7]提出了PointCNN,设计了一种X⁃Conv模块,在不显著增加参数数量的情况下耦合较远距离信息㊂图卷积网络[8](Graph Convolutional Network)是依靠图之间的节点进行信息传递,获得图之间的信息关联的深度神经网络㊂图可以视为顶点和边的集合,使每个点都成为顶点,消耗的运算量是无法估量的,需要采用K临近点计算方式[9]产生的边缘卷积层(EdgeConv)㊂利用中心点与其邻域点作为边特征,提取边特征㊂图卷积网络作为一种点云深度学习的新框架弥补了Pointnet等网络的部分缺陷[10]㊂针对非规律的表面损伤这种特征缺失类点云分割,人们已经利用各种二维图像采集数据与卷积神经网络对风扇叶片㊁建筑和交通工具等进行损伤检测[11],损伤主要类别是裂痕㊁表面漆脱落等㊂但二维图像分割涉及的损伤种类不够充分,可能受物体表面污染㊁光线等因素影响,将凹陷㊁凸起等损伤忽视,或因光照不均匀判断为脱漆㊂笔者提出一种基于特征更新的动态图卷积网络,主要针对三维点云分割,设计了一种新型的特征更新模块㊂利用三维点云独特的空间结构特征,对传统K邻域内权重相近的邻域点采用空间尺度进行区分,并应用于对金属部件表面损伤分割的有用与无用信息混杂的问题研究㊂对邻域点进行空间尺度划分,将注意力权重分组,组内进行特征更新㊂在有效鉴别外邻域干扰特征造成的误差前提下,增大特征提取面以提高局部区域特征有用性㊂1 深度卷积网络计算方法1.1 包含空间尺度区域信息的三维图注意力特征提取方法由迭代最远点采集算法将整片点云分割为n个点集:{M1,M2,M3, ,M n},每个点集包含k个点:{P1, P2,P3, ,P k},根据点集内的空间尺度关系,将局部区域划分为不同的空间区域㊂在每个区域内,结合局部特征与空间尺度特征,进一步获得更有区分度的特征信息㊂根据注意力机制,为K邻域内的点分配不同的权重信息,特征信息包括空间区域内点的分布和区域特性㊂将这些特征信息加权计算,得到点集的卷积结果㊂使用空间尺度区域信息的三维图注意力特征提取方式,需要设定合适的K邻域参数K和空间划分层数R㊂如果K太小,则会导致弱分割,因不能完全利用局部特征而影响结果准确性;如果K太大,会增加计算时间与数据量㊂图1为缺损损伤在不同参数K下的分割结果图㊂由图1可知,在K=30或50时,分割结果效果较好,K=30时计算量较小㊂笔者选择K=30作为实验参数㊂在分析确定空间划分层数R之前,简要分析空间层数划分所应对的问题㊂三维点云所具有的稀疏性㊁无序性以及损伤点云自身噪声和边角点多的特性,导致了点云处理中可能出现的共同缺点,即将离群值点云选为邻域内采样点㊂由于损伤表面多为一个面,被分割出的损伤点云应在该面上分布,而噪声点则被分布在整个面的两侧,甚至有部分位于损伤内部㊂由于点云噪声这种立体分布的特征,导致了离群值被选入邻域内作为采样点存在㊂根据采用DGCNN(Dynamic Graph Convolutional Neural Networks)分割网络抽样实验结果,位于切面附近以及损伤内部的离群值点对点云分割结果造成的影响最大,被错误分割为特征点的几率最大,在后续预处理过程中需要对这种噪声点进行优先处理㊂图1 缺损损伤在不同参数K 下的分割结果图Fig.1 Segmentation results of defect damage under different parameters K 基于上述实验结果,在参数K =30情况下,选择空间划分层数R ㊂缺损损伤在不同参数R 下的分割结果如图2所示㊂图2b 的结果与测试集标签分割结果更为相似,更能体现损伤的特征,同时屏蔽了大部分噪声㊂因此,选择R =4作为实验参数㊂图2 缺损损伤在不同参数R 下的分割结果图Fig.2 Segmentation results of defect damage under different parameters R 在一个K 邻域内,邻域点与中心点的空间关系和特征差异最能表现邻域点的权重㊂空间特征系数表示邻域点对中心点所在点集的重要性㊂同时,为更好区分图内邻域点的权重,需要将整个邻域细分㊂以空间尺度进行细分是较为合适的分类方式㊂中心点的K 邻域可视为一个局部空间,将其划分为r 个不同的尺度区域㊂再运算空间注意力机制,为这r 个不同区域的权重系数赋值㊂按照空间尺度多层次划分,不仅没有损失核心的邻域点特征,还能有效抑制无意义的㊁有干扰性的特征㊂从而提高了深度学习网络对点云的局部空间特征的学习能力,降低相邻邻域之间的互相影响㊂空间注意力机制如图3所示,计算步骤如下㊂第1步,计算特征系数e mk ㊂该值表示每个中心点m 的第k 个邻域点对其中心点的权重㊂分别用Δp mk 和Δf mk 表示三维空间关系和局部特征差异,M 表示MLP(Multi⁃Layer Perceptrons)操作,C 表示concat 函数,其中Δp mk =p mk -p m ,Δf mk =M (f mk )-M (f m )㊂将两者合并后输入多层感知机进行计算,得到计算特征系数326第4期张闻锐,等:特征更新的动态图卷积表面损伤点云分割方法图3 空间尺度区域信息注意力特征提取方法示意图Fig.3 Schematic diagram of attention feature extraction method for spatial scale regional information e mk =M [C (Δp mk ‖Δf mk )]㊂(1) 第2步,计算图权重系数a mk ㊂该值表示每个中心点m 的第k 个邻域点对其中心点的权重包含比㊂其中k ∈{1,2,3, ,K },K 表示每个邻域所包含点数㊂需要对特征系数e mk 进行归一化,使用归一化指数函数S (Softmax)得到权重多分类的结果,即计算图权重系数a mk =S (e mk )=exp(e mk )/∑K g =1exp(e mg )㊂(2) 第3步,用空间尺度区域特征s mr 表示中心点m 的第r 个空间尺度区域的特征㊂其中k r ∈{1,2,3, ,K r },K r 表示第r 个空间尺度区域所包含的邻域点数,并在其中加入特征偏置项b r ,避免权重化计算的特征在动态图中累计单面误差指向,空间尺度区域特征s mr =∑K r k r =1[a mk r M (f mk r )]+b r ㊂(3) 在r 个空间尺度区域上进行计算,就可得到点m 在整个局部区域的全部空间尺度区域特征s m ={s m 1,s m 2,s m 3, ,s mr },其中r ∈{1,2,3, ,R }㊂1.2 基于特征更新的动态图卷积网络动态图卷积网络是一种能直接处理原始三维点云数据输入的深度学习网络㊂其特点是将PointNet 网络中的复合特征转换模块(Feature Transform),改进为由K 邻近点计算(K ⁃Near Neighbor)和多层感知机构成的边缘卷积层[12]㊂边缘卷积层功能强大,其提取的特征不仅包含全局特征,还拥有由中心点与邻域点的空间位置关系构成的局部特征㊂在动态图卷积网络中,每个邻域都视为一个点集㊂增强对其中心点的特征学习能力,就会增强网络整体的效果[13]㊂对一个邻域点集,对中心点贡献最小的有效局部特征的边缘点,可以视为异常噪声点或低权重点,可能会给整体分割带来边缘溢出㊂点云相比二维图像是一种信息稀疏并且噪声含量更大的载体㊂处理一个局域内的噪声点,将其直接剔除或简单采纳会降低特征提取效果,笔者对其进行低权重划分,并进行区域内特征更新,增强抗噪性能,也避免点云信息丢失㊂在空间尺度区域中,在区域T 内有s 个点x 被归为低权重系数组,该点集的空间信息集为P ∈R N s ×3㊂点集的局部特征集为F ∈R N s ×D f [14],其中D f 表示特征的维度空间,N s 表示s 个域内点的集合㊂设p i 以及f i 为点x i 的空间信息和特征信息㊂在点集内,对点x i 进行小范围内的N 邻域搜索,搜索其邻域点㊂则点x i 的邻域点{x i ,1,x i ,2, ,x i ,N }∈N (x i ),其特征集合为{f i ,1,f i ,2, ,f i ,N }∈F ㊂在利用空间尺度进行区域划分后,对空间尺度区域特征s mt 较低的区域进行区域内特征更新,通过聚合函数对权重最低的邻域点在图中的局部特征进行改写㊂已知中心点m ,点x i 的特征f mx i 和空间尺度区域特征s mt ,目的是求出f ′mx i ,即中心点m 的低权重邻域点x i 在进行邻域特征更新后得到的新特征㊂对区域T 内的点x i ,∀x i ,j ∈H (x i ),x i 与其邻域H 内的邻域点的特征相似性域为R (x i ,x i ,j )=S [C (f i ,j )T C (f i ,j )/D o ],(4)其中C 表示由输入至输出维度的一维卷积,D o 表示输出维度值,T 表示转置㊂从而获得更新后的x i 的426吉林大学学报(信息科学版)第41卷特征㊂对R (x i ,x i ,j )进行聚合,并将特征f mx i 维度变换为输出维度f ′mx i =∑[R (x i ,x i ,j )S (s mt f mx i )]㊂(5) 图4为特征更新网络模块示意图,展示了上述特征更新的计算过程㊂图5为特征更新的动态图卷积网络示意图㊂图4 特征更新网络模块示意图Fig.4 Schematic diagram of feature update network module 图5 特征更新的动态图卷积网络示意图Fig.5 Flow chart of dynamic graph convolution network with feature update 动态图卷积网络(DGCNN)利用自创的边缘卷积层模块,逐层进行边卷积[15]㊂其前一层的输出都会动态地产生新的特征空间和局部区域,新一层从前一层学习特征(见图5)㊂在每层的边卷积模块中,笔者在边卷积和池化后加入了空间尺度区域注意力特征,捕捉特定空间区域T 内的邻域点,用于特征更新㊂特征更新会降低局域异常值点对局部特征的污染㊂网络相比传统图卷积神经网络能获得更多的特征信息,并且在面对拥有较多噪声值的点云数据时,具有更好的抗干扰性[16],在对性质不稳定㊁不平滑并含有需采集分割的突出中心的点云数据时,会有更好的抗干扰效果㊂相比于传统预处理方式,其稳定性更强,不会发生将突出部分误分割或漏分割的现象[17]㊂2 实验结果与分析点云分割的精度评估指标主要由两组数据构成[18],即平均交并比和总体准确率㊂平均交并比U (MIoU:Mean Intersection over Union)代表真实值和预测值合集的交并化率的平均值,其计算式为526第4期张闻锐,等:特征更新的动态图卷积表面损伤点云分割方法U =1T +1∑Ta =0p aa ∑Tb =0p ab +∑T b =0p ba -p aa ,(6)其中T 表示类别,a 表示真实值,b 表示预测值,p ab 表示将a 预测为b ㊂总体准确率A (OA:Overall Accuracy)表示所有正确预测点p c 占点云模型总体数量p all 的比,其计算式为A =P c /P all ,(7)其中U 与A 数值越大,表明点云分割网络越精准,且有U ≤A ㊂2.1 实验准备与数据预处理实验使用Kinect V2,采用Depth Basics⁃WPF 模块拍摄金属部件损伤表面获得深度图,将获得的深度图进行SDK(Software Development Kit)转化,得到pcd 格式的点云数据㊂Kinect V2采集的深度图像分辨率固定为512×424像素,为获得更清晰的数据图像,需尽可能近地采集数据㊂选择0.6~1.2m 作为采集距离范围,从0.6m 开始每次增加0.2m,获得多组采量数据㊂点云中分布着噪声,如果不对点云数据进行过滤会对后续处理产生不利影响㊂根据统计原理对点云中每个点的邻域进行分析,再建立一个特别设立的标准差㊂然后将实际点云的分布与假设的高斯分布进行对比,实际点云中误差超出了标准差的点即被认为是噪声点[19]㊂由于点云数据量庞大,为提高效率,选择采用如下改进方法㊂计算点云中每个点与其首个邻域点的空间距离L 1和与其第k 个邻域点的空间距离L k ㊂比较每个点之间L 1与L k 的差,将其中差值最大的1/K 视为可能噪声点[20]㊂计算可能噪声点到其K 个邻域点的平均值,平均值高出标准差的被视为噪声点,将离群噪声点剔除后完成对点云的滤波㊂2.2 金属表面损伤点云关键信息提取分割方法对点云损伤分割,在制作点云数据训练集时,如果只是单一地将所有损伤进行统一标记,不仅不方便进行结果分析和应用,而且也会降低特征分割的效果㊂为方便分析和控制分割效果,需要使用ArcGIS 将点云模型转化为不规则三角网TIN(Triangulated Irregular Network)㊂为精确地分类损伤,利用图6 不规则三角网模型示意图Fig.6 Schematic diagram of triangulated irregular networkTIN 的表面轮廓性质,获得训练数据损伤点云的损伤内(外)体积,损伤表面轮廓面积等㊂如图6所示㊂选择损伤体积指标分为相对损伤体积V (RDV:Relative Damege Volume)和邻域内相对损伤体积比N (NRDVR:Neighborhood Relative Damege Volume Ratio)㊂计算相对平均深度平面与点云深度网格化平面之间的部分,得出相对损伤体积㊂利用TIN 邻域网格可获取某损伤在邻域内的相对深度占比,有效解决制作测试集时,将因弧度或是形状造成的相对深度判断为损伤的问题㊂两种指标如下:V =∑P d k =1h k /P d -∑P k =1h k /()P S d ,(8)N =P n ∑P d k =1h k S d /P d ∑P n k =1h k S ()n -()1×100%,(9)其中P 表示所有点云数,P d 表示所有被标记为损伤的点云数,P n 表示所有被认定为损伤邻域内的点云数;h k 表示点k 的深度值;S d 表示损伤平面面积,S n 表示损伤邻域平面面积㊂在获取TIN 标准包络网视图后,可以更加清晰地描绘损伤情况,同时有助于量化损伤严重程度㊂笔者将损伤分为6种类型,并利用计算得出的TIN 指标进行损伤分类㊂同时,根据损伤部分体积与非损伤部分体积的关系,制定指标损伤体积(SDV:Standard Damege Volume)区分损伤类别㊂随机抽选5个测试组共50张图作为样本㊂统计非穿透损伤的RDV 绝对值,其中最大的30%标记为凹陷或凸起,其余626吉林大学学报(信息科学版)第41卷标记为表面损伤,并将样本分类的标准分界值设为SDV㊂在设立以上标准后,对凹陷㊁凸起㊁穿孔㊁表面损伤㊁破损和缺损6种金属表面损伤进行分类,金属表面损伤示意图如图7所示㊂首先,根据损伤是否产生洞穿,将损伤分为两大类㊂非贯通伤包括凹陷㊁凸起和表面损伤,贯通伤包括穿孔㊁破损和缺损㊂在非贯通伤中,凹陷和凸起分别采用相反数的SDV 作为标准,在这之间的被分类为表面损伤㊂贯通伤中,以损伤部分平面面积作为参照,较小的分类为穿孔,较大的分类为破损,而在边缘处因腐蚀㊁碰撞等原因缺角㊁内损的分类为缺损㊂分类参照如表1所示㊂图7 金属表面损伤示意图Fig.7 Schematic diagram of metal surface damage表1 损伤类别分类Tab.1 Damage classification 损伤类别凹陷凸起穿孔表面损伤破损缺损是否形成洞穿××√×√√RDV 绝对值是否达到SDV √√\×\\S d 是否达到标准\\×\√\2.3 实验结果分析为验证改进的图卷积深度神经网络在点云语义分割上的有效性,笔者采用TensorFlow 神经网络框架进行模型测试㊂为验证深度网络对损伤分割的识别准确率,采集了带有损伤特征的金属部件损伤表面点云,对点云进行预处理㊂对若干金属部件上的多个样本金属面的点云数据进行筛选,删除损伤占比低于5%或高于60%的数据后,划分并装包制作为点云数据集㊂采用CloudCompare 软件对样本金属上的损伤部分进行分类标记,共分为6种如上所述损伤㊂部件损伤的数据集制作参考点云深度学习领域广泛应用的公开数据集ModelNet40part㊂分割数据集包含了多种类型的金属部件损伤数据,这些损伤数据显示在510张总点云图像数据中㊂点云图像种类丰富,由各种包含损伤的金属表面构成,例如金属门,金属蒙皮,机械构件外表面等㊂用ArcGIS 内相关工具将总图进行随机点拆分,根据数据集ModelNet40part 的规格,每个独立的点云数据组含有1024个点,将所有总图拆分为510×128个单元点云㊂将样本分为400个训练集与110个测试集,采用交叉验证方法以保证测试的充分性[20],对多种方法进行评估测试,实验结果由单元点云按原点位置重新组合而成,并带有拆分后对单元点云进行的分割标记㊂分割结果比较如图8所示㊂726第4期张闻锐,等:特征更新的动态图卷积表面损伤点云分割方法图8 分割结果比较图Fig.8 Comparison of segmentation results在部件损伤分割的实验中,将不同网络与笔者网络(FAS⁃DGCNN:Feature Adaptive Shifting⁃Dynamic Graph Convolutional Neural Networks)进行对比㊂除了采用不同的分割网络外,其余实验均采用与改进的图卷积深度神经网络方法相同的实验设置㊂实验结果由单一损伤交并比(IoU:Intersection over Union),平均损伤交并比(MIoU),单一损伤准确率(Accuracy)和总体损伤准确率(OA)进行评价,结果如表2~表4所示㊂将6种不同损伤类别的Accuracy 与IoU 进行对比分析,可得出结论:相比于基准实验网络Pointet++,笔者在OA 和MioU 方面分别在贯通伤和非贯通伤上有10%和20%左右的提升,在整体分割指标上,OA 能达到90.8%㊂对拥有更多点数支撑,含有较多点云特征的非贯通伤,几种点云分割网络整体性能均能达到90%左右的效果㊂而不具有局部特征识别能力的PointNet 在贯通伤上的表现较差,不具备有效的分辨能力,导致分割效果相对于其他损伤较差㊂表2 损伤部件分割准确率性能对比 Tab.2 Performance comparison of segmentation accuracy of damaged parts %实验方法准确率凹陷⁃1凸起⁃2穿孔⁃3表面损伤⁃4破损⁃5缺损⁃6Ponitnet 82.785.073.880.971.670.1Pointnet++88.786.982.783.486.382.9DGCNN 90.488.891.788.788.687.1FAS⁃DGCNN 92.588.892.191.490.188.6826吉林大学学报(信息科学版)第41卷表3 损伤部件分割交并比性能对比 Tab.3 Performance comparison of segmentation intersection ratio of damaged parts %IoU 准确率凹陷⁃1凸起⁃2穿孔⁃3表面损伤⁃4破损⁃5缺损⁃6PonitNet80.582.770.876.667.366.9PointNet++86.384.580.481.184.280.9DGCNN 88.786.589.986.486.284.7FAS⁃DGCNN89.986.590.388.187.385.7表4 损伤分割的整体性能对比分析 出,动态卷积图特征以及有效的邻域特征更新与多尺度注意力给分割网络带来了更优秀的局部邻域分割能力,更加适应表面损伤分割的任务要求㊂3 结 语笔者利用三维点云独特的空间结构特征,将传统K 邻域内权重相近的邻域点采用空间尺度进行区分,并将空间尺度划分运用于邻域内权重分配上,提出了一种能将邻域内噪声点降权筛除的特征更新模块㊂采用此模块的动态图卷积网络在分割上表现出色㊂利用特征更新的动态图卷积网络(FAS⁃DGCNN)能有效实现金属表面损伤的分割㊂与其他网络相比,笔者方法在点云语义分割方面表现出更高的可靠性,可见在包含空间尺度区域信息的注意力和局域点云特征更新下,笔者提出的基于特征更新的动态图卷积网络能发挥更优秀的作用,而且相比缺乏局部特征提取能力的分割网络,其对于点云稀疏㊁特征不明显的非贯通伤有更优的效果㊂参考文献:[1]LAWIN F J,DANELLJAN M,TOSTEBERG P,et al.Deep Projective 3D Semantic Segmentation [C]∥InternationalConference on Computer Analysis of Images and Patterns.Ystad,Sweden:Springer,2017:95⁃107.[2]MATURANA D,SCHERER S.VoxNet:A 3D Convolutional Neural Network for Real⁃Time Object Recognition [C]∥Proceedings of IEEE /RSJ International Conference on Intelligent Robots and Systems.Hamburg,Germany:IEEE,2015:922⁃928.[3]LE T,DUAN Y.PointGrid:A Deep Network for 3D Shape Understanding [C]∥2018IEEE /CVF Conference on ComputerVision and Pattern Recognition (CVPR).Salt Lake City,USA:IEEE,2018:9204⁃9214.[4]QI C R,SU H,MO K,et al.PointNet:Deep Learning on Point Sets for 3D Classification and Segmentation [C]∥IEEEConference on Computer Vision and Pattern Recognition (CVPR).Hawaii,USA:IEEE,2017:652⁃660.[5]QI C R,SU H,MO K,et al,PointNet ++:Deep Hierarchical Feature Learning on Point Sets in a Metric Space [C]∥Advances in Neural Information Processing Systems.California,USA:SpringerLink,2017:5099⁃5108.[6]HU J,SHEN L,SUN G,Squeeze⁃and⁃Excitation Networks [C ]∥IEEE Conference on Computer Vision and PatternRecognition.Vancouver,Canada:IEEE,2018:7132⁃7141.[7]LI Y,BU R,SUN M,et al.PointCNN:Convolution on X⁃Transformed Points [C]∥Advances in Neural InformationProcessing Systems.Montreal,Canada:NeurIPS,2018:820⁃830.[8]ANH VIET PHAN,MINH LE NGUYEN,YEN LAM HOANG NGUYEN,et al.DGCNN:A Convolutional Neural Networkover Large⁃Scale Labeled Graphs [J].Neural Networks,2018,108(10):533⁃543.[9]任伟建,高梦宇,高铭泽,等.基于混合算法的点云配准方法研究[J].吉林大学学报(信息科学版),2019,37(4):408⁃416.926第4期张闻锐,等:特征更新的动态图卷积表面损伤点云分割方法036吉林大学学报(信息科学版)第41卷REN W J,GAO M Y,GAO M Z,et al.Research on Point Cloud Registration Method Based on Hybrid Algorithm[J]. Journal of Jilin University(Information Science Edition),2019,37(4):408⁃416.[10]ZHANG K,HAO M,WANG J,et al.Linked Dynamic Graph CNN:Learning on Point Cloud via Linking Hierarchical Features[EB/OL].[2022⁃03⁃15].https:∥/stamp/stamp.jsp?tp=&arnumber=9665104. [11]林少丹,冯晨,陈志德,等.一种高效的车体表面损伤检测分割算法[J].数据采集与处理,2021,36(2):260⁃269. LIN S D,FENG C,CHEN Z D,et al.An Efficient Segmentation Algorithm for Vehicle Body Surface Damage Detection[J]. Journal of Data Acquisition and Processing,2021,36(2):260⁃269.[12]ZHANG L P,ZHANG Y,CHEN Z Z,et al.Splitting and Merging Based Multi⁃Model Fitting for Point Cloud Segmentation [J].Journal of Geodesy and Geoinformation Science,2019,2(2):78⁃79.[13]XING Z Z,ZHAO S F,GUO W,et al.Processing Laser Point Cloud in Fully Mechanized Mining Face Based on DGCNN[J]. ISPRS International Journal of Geo⁃Information,2021,10(7):482⁃482.[14]杨军,党吉圣.基于上下文注意力CNN的三维点云语义分割[J].通信学报,2020,41(7):195⁃203. YANG J,DANG J S.Semantic Segmentation of3D Point Cloud Based on Contextual Attention CNN[J].Journal on Communications,2020,41(7):195⁃203.[15]陈玲,王浩云,肖海鸿,等.利用FL⁃DGCNN模型估测绿萝叶片外部表型参数[J].农业工程学报,2021,37(13): 172⁃179.CHEN L,WANG H Y,XIAO H H,et al.Estimation of External Phenotypic Parameters of Bunting Leaves Using FL⁃DGCNN Model[J].Transactions of the Chinese Society of Agricultural Engineering,2021,37(13):172⁃179.[16]柴玉晶,马杰,刘红.用于点云语义分割的深度图注意力卷积网络[J].激光与光电子学进展,2021,58(12):35⁃60. CHAI Y J,MA J,LIU H.Deep Graph Attention Convolution Network for Point Cloud Semantic Segmentation[J].Laser and Optoelectronics Progress,2021,58(12):35⁃60.[17]张学典,方慧.BTDGCNN:面向三维点云拓扑结构的BallTree动态图卷积神经网络[J].小型微型计算机系统,2021, 42(11):32⁃40.ZHANG X D,FANG H.BTDGCNN:BallTree Dynamic Graph Convolution Neural Network for3D Point Cloud Topology[J]. Journal of Chinese Computer Systems,2021,42(11):32⁃40.[18]张佳颖,赵晓丽,陈正.基于深度学习的点云语义分割综述[J].激光与光电子学,2020,57(4):28⁃46. ZHANG J Y,ZHAO X L,CHEN Z.A Survey of Point Cloud Semantic Segmentation Based on Deep Learning[J].Lasers and Photonics,2020,57(4):28⁃46.[19]SUN Y,ZHANG S H,WANG T Q,et al.An Improved Spatial Point Cloud Simplification Algorithm[J].Neural Computing and Applications,2021,34(15):12345⁃12359.[20]高福顺,张鼎林,梁学章.由点云数据生成三角网络曲面的区域增长算法[J].吉林大学学报(理学版),2008,46 (3):413⁃417.GAO F S,ZHANG D L,LIANG X Z.A Region Growing Algorithm for Triangular Network Surface Generation from Point Cloud Data[J].Journal of Jilin University(Science Edition),2008,46(3):413⁃417.(责任编辑:刘俏亮)。
高精度定位系统的摩擦力自适应前馈补偿
电气传动2021年第51卷第4期ELECTRIC DRIVE 2021Vol.51No.4摘要：为了有效抑制机电系统摩擦力等外部扰动对系统动态性能的影响，针对直驱伺服系统中往复定位存在的摩擦力，提出了一种基于自适应前馈控制器的摩擦力补偿策略，此方法能够有效利用参考模型与被控对象的位置跟踪误差等信息，在线实时确定自适应控制率，在保证系统稳定的条件下，能够有效克服系统摩擦力及模型慢时变等引起的系统动态性能异常。
针对直驱伺服系统建立其数学模型，根据数学模型确定自适应补偿环节的数学形式，并利用Lyapunov 函数证明了自适应控制率的稳定性。
最后通过试验表明，在跟踪正弦位置指令时，基于自适应前馈补偿的方法动态跟踪误差的均方根值为4.8μm ，与PID 无摩擦补偿控制方法相比，提高了47.3%，与传统模型参考自适应控制方法相比，提高了17.9%。
综上所述，所提方法可以有效抑制系统摩擦力干扰，提高系统动态跟踪精度。
关键词：自适应；前馈补偿；定位系统；摩擦力中图分类号：TP273+.5文献标识码：ADOI ：10.19457/j.1001-2095.dqcd20741Friction Adaptive Feedforward Compensation for High Precision Positioning SystemYANG Hong ，LI Shengming（College of Electromechanic and Automotive Engineering ，Qingyuan Polytechnic ，Qingyuan 511510，Guangdong ，China ）Abstract:In order to suppress the influence of external disturbances such as friction on the dynamic performance of electromechanical systems ，considering friction force for reciprocating positioning in direct drive servo system ，a friction compensation strategy based on adaptive feedforward controller was proposed.The method can effectively utilize the information of reference model and position tracking error of controlled object and determine the control rate on line.Under the condition of guaranteeing the stability of the system ，it can effectively overcome the abnormal dynamic performance of the system caused by system friction and slow time-varying model.The mathematical model of direct drive servo system and adaptive compensation component were established.Lyapunov function was adopted to guarantee the stability of the adaptive control rate.Finally ，experiments show that the root mean square value of dynamic tracking error based on adaptive feedforward compensation is 4.8μm when tracking sinusoidal position commands ，which is 47.3%higher than that of PID friction-free compensation control method and 17.9%higher than that of traditional model reference adaptive control method.In conclusion ，the proposed method can effectively suppress the friction interference and improve the dynamic tracking accuracy of the system.Key words:adaptive ；feedforward compensation ；positioning system ；friction高精度定位系统的摩擦力自适应前馈补偿杨红，李生明（清远职业技术学院机电与汽车工程学院，广东清远511510）基金项目：广东省高等职业教育教学改革研究与实践项目（GDJG2019380）作者简介：杨红（1969—），女，硕士，副教授，Email ：高精度定位系统广泛应用于航空航天、军工打印机、医疗器械及IC 装备等领域，定位系统的精度和响应速度等指标直接影响军用设备的加工精度及医疗器械的治疗效果等，因此研究有效提高定位系统的精度对国防军工、医疗卫生和生产生活的各领域有着重要意义。
dynamical model 动力学模型 英文说法
dynamical model 动力学模型英文说法1. 引言1.1 概述在科学研究和工程实践中，动力学模型是描述系统行为和演化的重要工具。
它们被广泛应用于生物学、物理学、社会科学以及许多其他领域。
动力学模型可以帮助我们理解自然现象背后的基本原理，揭示系统内部的相互作用和变化规律。
本文将介绍动力学模型的定义与原理，并探讨其在科学研究中的应用。
文章还将覆盖构建动力学模型的方法和技巧，并对未来动力学模型发展趋势进行展望。
1.2 文章结构本文分为五个主要部分：第一部分是引言，概述了动力学模型的重要性和应用领域，并介绍了文章的组织结构。
第二部分将阐述动力学模型的定义与原理。
我们将讨论动力学模型概念的含义，以及如何通过动力学方程和变量定义来描述系统演化过程。
此外，我们还将探讨动力学模型所依赖的基本假设和限制条件。
第三部分将详细介绍动力学模型在科学研究中的应用。
我们将以生物学、物理学和社会科学领域为例，说明动力学模型如何被用来解释和预测自然界和社会现象的行为。
第四部分将探讨构建动力学模型的方法和技巧。
我们将讨论实验数据收集与分析方法，参数估计与拟合技术，以及模拟和预测验证方法。
这些技术将有助于研究人员从实际观测中提取出系统的关键动力学特征，并验证模型的准确性。
最后一部分是结论和未来展望。
我们将总结主要研究结果，并对动力学模型发展趋势进行展望，探讨可能的研究方向和新兴应用领域。
1.3 目的本文的目的是提供一个全面且清晰的介绍动力学模型及其在科学研究中应用的文章。
通过阐述动力学模型背后的原理和基本假设，读者可以更好地理解系统演化过程中内在机制和相互作用规律。
同时，本文还旨在帮助读者掌握构建动力学模型所需的方法与技巧，并对未来该领域的发展趋势进行展望。
2. 动力学模型的定义与原理2.1 动力学模型的概念动力学模型是科学研究中常用的工具，用于描述和预测系统随时间演化的行为。
它是基于物理、生物或社会系统内部因素之间相互作用关系的数学表达式。
基于粒子群优化和极限学习机的底盘测功机加载扭矩模型
10.16638/ki.1671-7988.2020.11.039基于粒子群优化和极限学习机的底盘测功机加载扭矩模型覃健（广西壮族自治区计量检测研究院，广西南宁530000）摘要：汽车底盘测功机通过电涡流测功器产生加载扭矩模拟汽车行驶在道路上时所受到的阻力。
针对电涡流测功器输出扭矩的非线性特性，提出采用极限学习机神经网络建立电涡流测功器输出扭矩模型的方法，并利用粒子群算法优化网络结构提高模型的准确性。
仿真结果显示，模型的平均相对误差为1.3%。
通过与BP神经网络模型相比较，极限学习机模型显示出了更高的准确性。
关键词：底盘测功机；加载力矩；超限学习机；粒子群优化中图分类号：U467.3 文献标识码：A 文章编号：1671-7988(2020)11-121-03Modeling Chassis Dynamometer Load Torque Using Particle SwarmOptimization and Extreme Learning MachineQin Jian( Guangxi Zhuang Autonomous Region Institute of Metrology and Testing, Guangxi Nanning 530000 )Abstract: The chassis dynamometer generates the load torque by the eddy current dynamometer to simulate the on road resistance of the vehicles. For the nonlinear characteristics of the load torque, a method that using extreme learning machine (ELM) neural network to model the load torque of the chassis dynamometer is proposed, and the Particle Swarm Optimization algorithm is applied to optimize the model structure to improve the model accuracy. The simulation results show that the mean absolute relatively error of the proposed model is 1.3%. Compared with the BP model，the ELM model shows better accuracy.Keywords: Chassis dynamometer; Load torque; Extreme learning machine; Particle Swarm OptimizationCLC NO.: U467.3 Document Code: A Article ID: 1671-7988(2020)11-121-031 前言汽车底盘测功机通过加载装置产生加载扭矩实现对机动车在实际路面上行驶阻力的模拟。
自动化控制工程外文翻译外文文献英文文献
Team-Centered Perspective for Adaptive Automation DesignLawrence J.PrinzelLangley Research Center, Hampton, VirginiaAbstractAutomation represents a very active area of human factors research. Thejournal, Human Factors, published a special issue on automation in 1985.Since then, hundreds of scientific studies have been published examiningthe nature of automation and its interaction with human performance.However, despite a dramatic increase in research investigating humanfactors issues in aviation automation, there remain areas that need furtherexploration. This NASA Technical Memorandum describes a new area ofIt discussesautomation design and research, called “adaptive automation.” the concepts and outlines the human factors issues associated with the newmethod of adaptive function allocation. The primary focus is onhuman-centered design, and specifically on ensuring that adaptiveautomation is from a team-centered perspective. The document showsthat adaptive automation has many human factors issues common totraditional automation design. Much like the introduction of other new technologies and paradigm shifts, adaptive automation presents an opportunity to remediate current problems but poses new ones forhuman-automation interaction in aerospace operations. The review here isintended to communicate the philosophical perspective and direction ofadaptive automation research conducted under the Aerospace OperationsSystems (AOS), Physiological and Psychological Stressors and Factors (PPSF)project.Key words:Adaptive Automation; Human-Centered Design; Automation;Human FactorsIntroduction"During the 1970s and early 1980s...the concept of automating as much as possible was considered appropriate. The expected benefit was a reduction inpilot workload and increased safety...Although many of these benefits have beenrealized, serious questions have arisen and incidents/accidents that have occurredwhich question the underlying assumptions that a maximum availableautomation is ALWAYS appropriate or that we understand how to designautomated systems so that they are fully compatible with the capabilities andlimitations of the humans in the system."---- ATA, 1989The Air Transport Association of America (ATA) Flight Systems Integration Committee(1989) made the above statement in response to the proliferation of automation in aviation. They noted that technology improvements, such as the ground proximity warning system, have had dramatic benefits; others, such as the electronic library system, offer marginal benefits at best. Such observations have led many in the human factors community, most notably Charles Billings (1991; 1997) of NASA, to assert that automation should be approached from a "human-centered design" perspective.The period from 1970 to the present was marked by an increase in the use of electronic display units (EDUs); a period that Billings (1997) calls "information" and “management automation." The increased use of altitude, heading, power, and navigation displays; alerting and warning systems, such as the traffic alert and collision avoidance system (TCAS) and ground proximity warning system (GPWS; E-GPWS; TAWS); flight management systems (FMS) and flight guidance (e.g., autopilots; autothrottles) have "been accompanied by certain costs, including an increased cognitive burden on pilots, new information requirements that have required additional training, and more complex, tightly coupled, less observable systems" (Billings, 1997). As a result, human factors research in aviation has focused on the effects of information and management automation. The issues of interest include over-reliance on automation, "clumsy" automation (e.g., Wiener, 1989), digital versus analog control, skill degradation, crew coordination, and data overload (e.g., Billings, 1997). Furthermore, research has also been directed toward situational awareness (mode & state awareness; Endsley, 1994; Woods & Sarter, 1991) associated with complexity, coupling, autonomy, and inadequate feedback. Finally, human factors research has introduced new automation concepts that will need to be integrated into the existing suite of aviationautomation.Clearly, the human factors issues of automation have significant implications for safetyin aviation. However, what exactly do we mean by automation? The way we choose to define automation has considerable meaning for how we see the human role in modern aerospace s ystems. The next section considers the concept of automation, followed by an examination of human factors issues of human-automation interaction in aviation. Next, a potential remedy to the problems raised is described, called adaptive automation. Finally, the human-centered design philosophy is discussed and proposals are made for how the philosophy can be applied to this advanced form of automation. The perspective is considered in terms of the Physiological /Psychological Stressors & Factors project and directions for research on adaptive automation.Automation in Modern AviationDefinition.Automation refers to "...systems or methods in which many of the processes of production are automatically performed or controlled by autonomous machines or electronic devices" (Parsons, 1985). Automation is a tool, or resource, that the human operator can use to perform some task that would be difficult or impossible without machine aiding (Billings, 1997). Therefore, automation can be thought of as a process of substituting the activity of some device or machine for some human activity; or it can be thought of as a state of technological development (Parsons, 1985). However, some people (e.g., Woods, 1996) have questioned whether automation should be viewed as a substitution of one agent for another (see "apparent simplicity, real complexity" below). Nevertheless, the presence of automation has pervaded almost every aspect of modern lives. From the wheel to the modern jet aircraft, humans have sought to improve the quality of life. We have built machines and systems that not only make work easier, more efficient, and safe, but also give us more leisure time. The advent of automation has further enabled us to achieve this end. With automation, machines can now perform many of the activities that we once had to do. Our automobile transmission will shift gears for us. Our airplanes will fly themselves for us. All we have to dois turn the machine on and off. It has even been suggested that one day there may not be aaccidents resulting from need for us to do even that. However, the increase in “cognitive” faulty human-automation interaction have led many in the human factors community to conclude that such a statement may be premature.Automation Accidents. A number of aviation accidents and incidents have been directly attributed to automation. Examples of such in aviation mishaps include (from Billings, 1997):DC-10 landing in control wheel steering A330 accident at ToulouseB-747 upset over Pacific DC-10 overrun at JFK, New YorkB-747 uncommandedroll,Nakina,Ont. A320 accident at Mulhouse-HabsheimA320 accident at Strasbourg A300 accident at NagoyaB-757 accident at Cali, Columbia A320 accident at BangaloreA320 landing at Hong Kong B-737 wet runway overrunsA320 overrun at Warsaw B-757 climbout at ManchesterA310 approach at Orly DC-9 wind shear at CharlotteBillings (1997) notes that each of these accidents has a different etiology, and that human factors investigation of causes show the matter to be complex. However, what is clear is that the percentage of accident causes has fundamentally shifted from machine-caused to human-caused (estimations of 60-80% due to human error) etiologies, and the shift is attributable to the change in types of automation that have evolved in aviation.Types of AutomationThere are a number of different types of automation and the descriptions of them vary considerably. Billings (1997) offers the following types of automation:?Open-Loop Mechanical or Electronic Control.Automation is controlled by gravity or spring motors driving gears and cams that allow continous and repetitive motion. Positioning, forcing, and timing were dictated by the mechanism and environmental factors (e.g., wind). The automation of factories during the Industrial Revolution would represent this type of automation.?Classic Linear Feedback Control.Automation is controlled as a function of differences between a reference setting of desired output and the actual output. Changes a re made to system parameters to re-set the automation to conformance. An example of this type of automation would be flyball governor on the steam engine. What engineers call conventional proportional-integral-derivative (PID) control would also fit in this category of automation.?Optimal Control. A computer-based model of controlled processes i s driven by the same control inputs as that used to control the automated process. T he model output is used to project future states and is thus used to determine the next control input. A "Kalman filtering" approach is used to estimate the system state to determine what the best control input should be.?Adaptive Control. This type of automation actually represents a number of approaches to controlling automation, but usually stands for automation that changes dynamically in response to a change in state. Examples include the use of "crisp" and "fuzzy" controllers, neural networks, dynamic control, and many other nonlinear methods.Levels of AutomationIn addition to “types ” of automation, we can also conceptualize different “levels ” of automation control that the operator can have. A number of taxonomies have been put forth, but perhaps the best known is the one proposed by Tom Sheridan of Massachusetts Institute of Technology (MIT). Sheridan (1987) listed 10 levels of automation control:1. The computer offers no assistance, the human must do it all2. The computer offers a complete set of action alternatives3. The computer narrows the selection down to a few4. The computer suggests a selection, and5. Executes that suggestion if the human approves, or6. Allows the human a restricted time to veto before automatic execution, or7. Executes automatically, then necessarily informs the human, or8. Informs the human after execution only if he asks, or9. Informs the human after execution if it, the computer, decides to10. The computer decides everything and acts autonomously, ignoring the humanThe list covers the automation gamut from fully manual to fully automatic. Although different researchers define adaptive automation differently across these levels, the consensus is that adaptive automation can represent anything from Level 3 to Level 9. However, what makes adaptive automation different is the philosophy of the approach taken to initiate adaptive function allocation and how such an approach may address t he impact of current automation technology.Impact of Automation TechnologyAdvantages of Automation . Wiener (1980; 1989) noted a number of advantages to automating human-machine systems. These include increased capacity and productivity, reduction of small errors, reduction of manual workload and mental fatigue, relief from routine operations, more precise handling of routine operations, economical use of machines, and decrease of performance variation due to individual differences. Wiener and Curry (1980) listed eight reasons for the increase in flight-deck automation: (a) Increase in available technology, such as FMS, Ground Proximity Warning System (GPWS), Traffic Alert andCollision Avoidance System (TCAS), etc.; (b) concern for safety; (c) economy, maintenance, and reliability; (d) workload reduction and two-pilot transport aircraft certification; (e) flight maneuvers and navigation precision; (f) display flexibility; (g) economy of cockpit space; and (h) special requirements for military missions.Disadvantages o f Automation. Automation also has a number of disadvantages that have been noted. Automation increases the burdens and complexities for those responsible for operating, troubleshooting, and managing systems. Woods (1996) stated that automation is "...a wrapped package -- a package that consists of many different dimensions bundled together as a hardware/software system. When new automated systems are introduced into a field of practice, change is precipitated along multiple dimensions." As Woods (1996) noted, some of these changes include: ( a) adds to or changes the task, such as device setup and initialization, configuration control, and operating sequences; (b) changes cognitive demands, such as requirements for increased situational awareness; (c) changes the roles of people in the system, often relegating people to supervisory controllers; (d) automation increases coupling and integration among parts of a system often resulting in data overload and "transparency"; and (e) the adverse impacts of automation is often not appreciated by those who advocate the technology. These changes can result in lower job satisfaction (automation seen as dehumanizing human roles), lowered vigilance, fault-intolerant systems, silent failures, an increase in cognitive workload, automation-induced failures, over-reliance, complacency, decreased trust, manual skill erosion, false alarms, and a decrease in mode awareness (Wiener, 1989).Adaptive AutomationDisadvantages of automation have resulted in increased interest in advanced automation concepts. One of these concepts is automation that is dynamic or adaptive in nature (Hancock & Chignell, 1987; Morrison, Gluckman, & Deaton, 1991; Rouse, 1977; 1988). In an aviation context, adaptive automation control of tasks can be passed back and forth between the pilot and automated systems in response to the changing task demands of modern aircraft. Consequently, this allows for the restructuring of the task environment based upon (a) what is automated, (b) when it should be automated, and (c) how it is automated (Rouse, 1988; Scerbo, 1996). Rouse(1988) described criteria for adaptive aiding systems:The level of aiding, as well as the ways in which human and aidinteract, should change as task demands vary. More specifically,the level of aiding should increase as task demands become suchthat human performance will unacceptably degrade withoutaiding. Further, the ways in which human and aid interact shouldbecome increasingly streamlined as task demands increase.Finally, it is quite likely that variations in level of aiding andmodes of interaction will have to be initiated by the aid rather thanby the human whose excess task demands have created a situationrequiring aiding. The term adaptive aiding is used to denote aidingconcepts that meet [these] requirements.Adaptive aiding attempts to optimize the allocation of tasks by creating a mechanism for determining when tasks need to be automated (Morrison, Cohen, & Gluckman, 1993). In adaptive automation, the level or mode of automation can be modified in real time. Further, unlike traditional forms of automation, both the system and the pilot share control over changes in the state of automation (Scerbo, 1994; 1996). Parasuraman, Bahri, Deaton, Morrison, and Barnes (1992) have argued that adaptive automation represents the optimal coupling of the level of pilot workload to the level of automation in the tasks. Thus, adaptive automation invokes automation only when task demands exceed the pilot's capabilities. Otherwise, the pilot retains manual control of the system functions. Although concerns have been raised about the dangers of adaptive automation (Billings & Woods, 1994; Wiener, 1989), it promises to regulate workload, bolster situational awareness, enhance vigilance, maintain manual skill levels, increase task involvement, and generally improve pilot performance.Strategies for Invoking AutomationPerhaps the most critical challenge facing system designers seeking to implement automation concerns how changes among modes or levels of automation will be accomplished (Parasuraman e t al., 1992; Scerbo, 1996). Traditional forms of automation usually start with some task or functional analysis and attempt to fit the operational tasks necessary to the abilities of the human or the system. The approach often takes the form of a functional allocation analysis (e.g., Fitt's List) in which an attempt is made to determine whether the human or the system is better suited to do each task. However, many in the field have pointed out the problem with trying to equate the two in automated systems, as each have special characteristics that impede simple classification taxonomies. Such ideas as these have led some to suggest other ways of determining human-automation mixes. Although certainly not exhaustive, some of these ideas are presented below.Dynamic Workload Assessment.One approach involves the dynamic assessment o fmeasures t hat index the operators' state of mental engagement. (Parasuraman e t al., 1992; Rouse,1988). The question, however, is what the "trigger" should be for the allocation of functions between the pilot and the automation system. Numerous researchers have suggested that adaptive systems respond to variations in operator workload (Hancock & Chignell, 1987; 1988; Hancock, Chignell & Lowenthal, 1985; Humphrey & Kramer, 1994; Reising, 1985; Riley, 1985; Rouse, 1977), and that measures o f workload be used to initiate changes in automation modes. Such measures include primary and secondary-task measures, subjective workload measures, a nd physiological measures. T he question, however, is what adaptive mechanism should be used to determine operator mental workload (Scerbo, 1996).Performance Measures. One criterion would be to monitor the performance of the operator (Hancock & Chignel, 1987). Some criteria for performance would be specified in the system parameters, and the degree to which the operator deviates from the criteria (i.e., errors), the system would invoke levels of adaptive automation. For example, Kaber, Prinzel, Clammann, & Wright (2002) used secondary task measures to invoke adaptive automation to help with information processing of air traffic controllers. As Scerbo (1996) noted, however,"...such an approach would be of limited utility because the system would be entirely reactive."Psychophysiological M easures.Another criterion would be the cognitive and attentional state of the operator as measured by psychophysiological measures (Byrne & Parasuraman, 1996). An example of such an approach is that by Pope, Bogart, and Bartolome (1996) and Prinzel, Freeman, Scerbo, Mikulka, and Pope (2000) who used a closed-loop system to dynamically regulate the level of "engagement" that the subject had with a tracking task. The system indexes engagement on the basis of EEG brainwave patterns.Human Performance Modeling.Another approach would be to model the performance of the operator. The approach would allow the system to develop a number of standards for operator performance that are derived from models of the operator. An example is Card, Moran, and Newell (1987) discussion of a "model human processor." They discussed aspects of the human processor that could be used to model various levels of human performance. Another example is Geddes (1985) and his colleagues (Rouse, Geddes, & Curry, 1987-1988) who provided a model to invoke automation based upon system information, the environment, and expected operator behaviors (Scerbo, 1996).Mission Analysis. A final strategy would be to monitor the activities of the mission or task (Morrison & Gluckman, 1994). Although this method of adaptive automation may be themost accessible at the current state of technology, Bahri et al. (1992) stated that such monitoring systems lack sophistication and are not well integrated and coupled to monitor operator workload or performance (Scerbo, 1996). An example of a mission analysis approach to adaptive automation is Barnes and Grossman (1985) who developed a system that uses critical events to allocate among automation modes. In this system, the detection of critical events, such as emergency situations or high workload periods, invoked automation.Adaptive Automation Human Factors IssuesA number of issues, however, have been raised by the use of adaptive automation, and many of these issues are the same as those raised almost 20 years ago by Curry and Wiener (1980). Therefore, these issues are applicable not only to advanced automation concepts, such as adaptive automation, but to traditional forms of automation already in place in complex systems (e.g., airplanes, trains, process control).Although certainly one can make the case that adaptive automation is "dressed up" automation and therefore has many of the same problems, it is also important to note that the trend towards such forms of automation does have unique issues that accompany it. As Billings & Woods (1994) stated, "[i]n high-risk, dynamic environments...technology-centered automation has tended to decrease human involvement in system tasks, and has thus impaired human situation awareness; both are unwanted consequences of today's system designs, but both are dangerous in high-risk systems. [At its present state of development,] adaptive ("self-adapting") automation represents a potentially serious threat ... to the authority that the human pilot must have to fulfill his or her responsibility for flight safety."The Need for Human Factors Research.Nevertheless, such concerns should not preclude us from researching the impact that such forms of advanced automation are sure to have on human performance. Consider Hancock’s (1996; 1997) examination of the "teleology for technology." He suggests that automation shall continue to impact our lives requiring humans to co-evolve with the technology; Hancock called this "techneology."What Peter Hancock attempts to communicate to the human factors community is that automation will continue to evolve whether or not human factors chooses to be part of it. As Wiener and Curry (1980) conclude: "The rapid pace of automation is outstripping one's ability to comprehend all the implications for crew performance. It is unrealistic to call for a halt to cockpit automation until the manifestations are completely understood. We do, however, call for those designing, analyzing, and installing automatic systems in the cockpit to do so carefully; to recognize the behavioral effects of automation; to avail themselves of present andfuture guidelines; and to be watchful for symptoms that might appear in training andoperational settings." The concerns they raised are as valid today as they were 23 years ago.However, this should not be taken to mean that we should capitulate. Instead, becauseobservation suggests that it may be impossible to fully research any new Wiener and Curry’ｓtechnology before implementation, we need to form a taxonomy and research plan tomaximize human factors input for concurrent engineering of adaptive automation.Classification of Human Factors Issues. Kantowitz and Campbell (1996)identified some of the key human factors issues to be considered in the design of advancedautomated systems. These include allocation of function, stimulus-response compatibility, andmental models. Scerbo (1996) further suggested the need for research on teams,communication, and training and practice in adaptive automated systems design. The impactof adaptive automation systems on monitoring behavior, situational awareness, skilldegradation, and social dynamics also needs to be investigated. Generally however, Billings(1997) stated that the problems of automation share one or more of the followingcharacteristics: Brittleness, opacity, literalism, clumsiness, monitoring requirement, and dataoverload. These characteristics should inform design guidelines for the development, analysis,and implementation of adaptive automation technologies. The characteristics are defined as: ?Brittleness refers to "...an attribute of a system that works well under normal or usual conditions but that does not have desired behavior at or close to some margin of its operating envelope."?Opacity reflects the degree of understanding of how and why automation functions as it does. The term is closely associated with "mode awareness" (Sarter & Woods, 1994), "transparency"; or "virtuality" (Schneiderman, 1992).?Literalism concern the "narrow-mindedness" of the automated system; that is, theflexibility of the system to respond to novel events.?Clumsiness was coined by Wiener (1989) to refer to automation that reduced workload demands when the demands are already low (e.g., transit flight phase), but increases them when attention and resources are needed elsewhere (e.g., descent phase of flight). An example is when the co-pilot needs to re-program the FMS, to change the plane's descent path, at a time when the co-pilot should be scanning for other planes.?Monitoring requirement refers to the behavioral and cognitive costs associated withincreased "supervisory control" (Sheridan, 1987; 1991).?Data overload points to the increase in information in modern automated contexts (Billings, 1997).These characteristics of automation have relevance for defining the scope of humanfactors issues likely to plague adaptive automation design if significant attention is notdirected toward ensuring human-centered design. The human factors research communityhas noted that these characteristics can lead to human factors issues of allocation of function(i.e., when and how should functions be allocated adaptively); stimulus-response compatibility and new error modes; how adaptive automation will affect mental models,situation models, and representational models; concerns about mode unawareness and-of-the-loop” performance problem; situation awareness decay; manual skill decay and the “outclumsy automation and task/workload management; and issues related to the design of automation. This last issue points to the significant concern in the human factors communityof how to design adaptive automation so that it reflects what has been called “team-centered”;that is, successful adaptive automation will l ikely embody the concept of the “electronic team member”. However, past research (e.g., Pilots Associate Program) has shown that designing automation to reflect such a role has significantly different requirements than those arising in traditional automation design. The field is currently focused on answering the questions,does that definition translate into“what is it that defines one as a team member?” and “howUnfortunately, the literature also shows that the designing automation to reflect that role?” answer is not transparent and, therefore, adaptive automation must first tackle its own uniqueand difficult problems before it may be considered a viable prescription to currenthuman-automation interaction problems. The next section describes the concept of the electronic team member and then discusses t he literature with regard to team dynamics, coordination, communication, shared mental models, and the implications of these foradaptive automation design.Adaptive Automation as Electronic Team MemberLayton, Smith, and McCoy (1994) stated that the design of automated systems should befrom a team-centered approach; the design should allow for the coordination betweenmachine agents and human practitioners. However, many researchers have noted that automated systems tend to fail as team players (Billings, 1991; Malin & Schreckenghost,1992; Malin et al., 1991;Sarter & Woods, 1994; Scerbo, 1994; 1996; Woods, 1996). Thereason is what Woods (1996) calls “apparent simplicity, real complexity.”Apparent Simplicity, Real Complexity.Woods (1996) stated that conventional wisdomabout automation makes technology change seem simple. Automation can be seen as simply changing the human agent for a machine agent. Automation further provides for more optionsand methods, frees up operator time to do other things, provides new computer graphics and interfaces, and reduces human error. However, the reality is that technology change has often。
Modeling methods for modeling the three-dimensiona
专利名称：Modeling methods for modeling the three-dimensional topography and radioactivity ofthe environment using mobile terminals,computer programs, digital data media,mobile terminals发明人：デュバール フィリップ,モリシ マッシモ申请号：JP2019084335申请日：20190425公开号：JP6766218B2公开日：20201007专利内容由知识产权出版社提供摘要：To find out a novel method that models an environment accompanied by a risk of nuclear contamination.SOLUTION: A method for modeling an environment accompanied by a risk of nuclear contamination includes the steps of: (a) acquiring information relevant to topography of an environment and radioactivity measurement data on the environment, using detection means 10 and through a three-dimensional displacement of the detection means in the environment; (b) in association of the radioactivity measurement data with the location data in the environment via computer processing means 20, the location data being estimated from route data on the detection means; and (c) gradually preparing at least one matrix in which radioactivity data relevant to topography data on the environment and the location data are edited using information and via the computer processing means, and three-dimensional mapping, representing the environment, in which the topography data and the radioactivity data are combinedly displayed.SELECTED DRAWING: Figure 10申请人：エステーエムイー ソシエテ デ テクニーク アン ミリウ イオニゾン地址：フランス国，エフ－９１１９６，ジフ シュル イヴェット，１ ルート ド ラ ヌ，ザックド クルセル国籍：FR代理人：特許業務法人 信栄特許事務所更多信息请下载全文后查看。
翻译与现代汉语中的欧化语法
摘要｛翻译是语言接触的一种特殊形式，在翻译过程中，目标语和源语通过译者个人的言语行为相互作用、相互影响，给各自的语言规范带来一定的冲击，导致语言借用和语言的变异。
通过翻译过程中的语言接触争外域文化的渗透可以推动目的语的发展与完善，这在中外翻译史上是一种较普遍的现象。
ｆ五·四运动以来近百年时间是现代汉语形成和发展的关键期，包括社会生活的急剧变化，语言翻译引起的语言接触在内的诸多外部因素在特定的历史时期对汉语言的发展与变化带来了巨大的影响。
伴随着大量的西方文学作品的汉译，汉语不可避免的受到印欧语的冲击，大量的欧化结构和表达方式涌入汉语，形成了现代汉语中的欧化语法现象。
这种目标语语言结构在翻译影响下发生的变异，是一种积极的历史现象。
汉语“白话文”从五一四时代的雏形状态发展为稳定而成熟的现代汉语，在某种程度上便是一个欧化的过程。
对于汉语欧化的讨论和研究，可以追溯到上世纪二三十年代。
但长期以来关于欧化语法的研究大多是规定－｝生的，学者们大都局限于讨论汉语欧化的必要性或合理·陛，而在很大程度上忽视了对欧化语法这一客观存在的语言现象进行细致客观的描述。
另外，描述也并不是最终目的，最重要的是在观察的基础上给出合理的阐释，以帮助我们能了解一个欧化结构是如何从译者个人的言语表达方式最终融入目标语语言体系的，从而推动语言变化研究的深八发展。
ｙ－ｙ－ＯＩ本文试图对现代汉语中的欧化语法进行系统的描述和分类，并以此为基础对汉语的欧化进行多角度的阐释，以期能揭示欧化的形成机制以及欧化和汉语发展演变的内在联系。
本文首先回顾了汉语欧化的历史和现状，以及前人对汉语中的欧化语法进行的相关研究。
ｆ汉语史上曾经有过三次翻译高潮，每次都给汉语带来了深刻的影响。
汉语的欧化最早发端于五四时期，并一直持续到当代，对于欧化的研究和讨论至今未有结果。
欧化语法的规定性研究忽视了对语言现象的客观描述和阐释，最终不可能得到令人信服的结论。
机器人动力学研究常用方法
机器人动力学研究常用方法
机器人动力学研究的常用方法包括动力学建模、解算、仿真、优化及控制等。
一、动力学建模
依据物理原理，利用诸如Lagrange、Newton-Euler等方法，根据一个机器人的结构参数（如转动副、固定副和驱动元件）建立其动力学模型。
该模型是机器人轨迹规划中的基础，用于计算机器人各部件的运动学及动力学特性。
二、解算
解算是指利用动力学模型对机器人内部参数进行解算，以得出机器人关节角度或二维位置坐标。
解算方法中，基本的逆运动学可以利用正解算或近似解算的方式来求解机器人关节角度，而非正解法则常用来求解机器人的二维坐标位置。
三、仿真
仿真是用计算机模拟机器人实物运动，以此来优化机器人控制程序，验证机器人控制方案，评估系统性能，并研究机器人运动学和动力学问题。
仿真的模型可以是动力学模型，也可以是视觉系统、所在环境的模型。
四、优化
优化是在低层控制问题中采用的常用技术，它旨在使机器人的运动更加高效有效，并且能够较好地应对内外部环境的变化。
优化技术常用于为机器人规划路径，即将机器人从起点移动到终点，在最短时间内完成任务。
五、控制
控制是机器人动力学系统最为重要的部分，它可以有效地调节机器人的位置，使其高效地控制机器人的运动。
在机器人控制过程中，
与传感器结合使用的有感控制技术是一种常用的技术，其能够更精确地定位机器人的状态，并使用相应的控制方法来控制机器人的运动方式。
多模态运动分析与动作识别
多模态运动分析与动作识别第一章：引言多模态运动分析与动作识别是计算机视觉和机器学习领域的一个重要研究方向。
随着计算机视觉和机器学习技术的不断发展，多模态运动分析与动作识别在人工智能、智能交互、虚拟现实等领域具有广泛的应用前景。
本章将介绍多模态运动分析与动作识别的研究背景和意义，以及本文的研究目标和内容安排。
第二章：多模态运动数据获取与处理在进行多模态运动分析与动作识别之前，首先需要获取和处理多种不同类型的运动数据。
本章将介绍常用的数据获取方法，包括传感器、摄像头等设备，并介绍如何对获取到的数据进行预处理和特征提取。
第三章：传感器融合技术在多模态运动分析中的应用传感器融合是实现多模态运动分析与动作识别的关键技术之一。
本章将介绍传感器融合技术在人体姿势估计、行为识别等方面的应用，并详细讨论传感器融合的方法和算法。
第四章：图像与视频处理在多模态运动分析中的应用图像与视频处理在多模态运动分析中扮演着重要的角色。
本章将介绍图像与视频处理技术在人体运动跟踪、行为识别等方面的应用，并详细讨论常用的图像与视频处理方法和算法。
第五章：声音信号处理在多模态运动分析中的应用声音信号是一种重要的多模态数据，对于多模态运动分析和动作识别具有重要意义。
本章将介绍声音信号处理技术在人体行为识别、情感识别等方面的应用，并详细讨论常用的声音信号处理方法和算法。
第六章：机器学习算法在多模态运动分析中的应用机器学习是实现多模态运动分析与动作识别的核心技术之一。
本章将介绍常见机器学习算法及其在人体姿势估计、行为识别等方面的应用，并详细讨论机器学习算法如何结合传感器数据、图像数据和声音数据进行特征提取和分类。
第七章：实验设计与结果分析本文将设计一系列实验来验证多模态运动分析与动作识别的效果。
本章将介绍实验设计的方法和步骤，并对实验结果进行详细的分析和讨论。
第八章：多模态运动分析与动作识别的应用前景本章将对多模态运动分析与动作识别在人工智能、智能交互、虚拟现实等领域的应用前景进行展望，并讨论未来研究方向和挑战。
- 1、下载文档前请自行甄别文档内容的完整性，平台不提供额外的编辑、内容补充、找答案等附加服务。
- 2、"仅部分预览"的文档,不可在线预览部分如存在完整性等问题,可反馈申请退款(可完整预览的文档不适用该条件!)。
- 3、如文档侵犯您的权益，请联系客服反馈,我们会尽快为您处理(人工客服工作时间：9:00-18:30)。
Modeling Categorization Dynamics throughConversation by Constructive ApproachTakashi HashimotoKnowledge Science,Japan Advanced Institute of Science and Technology(JAIST) 1-1,Tatsunokuchi,Ishikawa923-1292,JAPANhash@jaist.ac.jpAbstract.Categorization dynamics as the clustering of words in wordrelation is studied by a constructive approach which is suited to inquireevolutionary linguistics with dynamical view on language.Word meaningis represented by relationship among words.Tthe relationship shouldbe derived from usage of language.Being founded on this usage-basedview,we define an algorithm to evaluate word ing thealgorithm,cluster structure and its dynamics of words are shown in amodel with communicating artificial agents.The relevance of clusteringwith linguistic categorization is discussed.1Dynamical View on LanguageThere are two ways of viewing language:structurally and dynamically.The structural view is a static one in which language structure,for example,syn-tax,dictionaries,or pragmatic rules,offers idealized approaches to language. The alternative view is dynamic.It concentrates on the actual use of language rather than abstract notions of how language ought to be.It is possible to better understand the value of the second approach by thinking of metaphor.Metaphor-icalexpressions are creative and dynamic precisely because they can“bend”or “break”the rules of conventionally structured language.By producing or un-derstanding metaphorical expressions,especially creative or unique metaphors, our internal models should change.We can not say valid or not valid for such creative expressions,since the expressions are so novel that it is not valid for a conventional language structure.We should consider whether the expressions are to be accepted or not.If we accept them,our internal structure changes and language structure might also come to be modified.In the dynamical view,the whole system of such dynamical processes is considered as’language.’Constructive approaches are highly advantageous to understanding complex systems[4].These approaches are also useful for studying evolutionary linguis-tics[7].In contrast to conventional linguistics which attempts to describe various language phenomena,in the constructive approach the emergence of global or-der as language-like behavior is modeled through interaction among individuals. However,not only emergence but also the dynamics of global order should be ob-served in constructive models,since language is indeed an ever-changing system.2Takashi HashimotoPerhaps the internal dynamics of individuals should be taken into considera-tion to study evolutionary language system so as for individuals to change their internal states and relationship to others and circumstances.2Modeling–Word Relationship and ConversationWe have proposed usage-based viewpoint on meaning[2]which have claimed that meanings of words should be discussed in terms of how language is used[9]. Interrelationship among words can be employed as a representation of meanings of words to some extent.This point of view implies that relationship of one word to other words should be derived from analyzing the usage of the word in the language,not entirely from its indication or reference.Moreover,a word in a sentence is understood from not only relation with only entities mentioned by the sentence but position in the whole system of language.Based on this viewpoint,we discuss dynamics of categorization by observing how the relationship among words changes through conversations.Building rela-tionship in use of language is a dynamical process performed by language users. We call this process sense-making process[1]to emphasize its subjective nature. The sense-making process is modeled by positioning a word in the relationship among all words.The algorithm to evaluate relationship between words is basically attributed to Karov and Edelman’s work[5]with two revisions.The one is to calculate relationship dynamically in the course of conversation,since what interests us is not in thefinal form of category but in the dynamics of categorization.The other is to consider’texts’on higher level than sentences1.A text is a stream of sentences uttered and accepted.The relationship between words is defined by the linear combination of the terms of usage-similarity and appearance-similarity using a coefficientαw as2R(w i,w j)=αw(usage-similarity)+(1−αw)(appearance-similarity).Thefirst term is designed to calculate usage similarity of words in sentences by considering the syntagmatic relationships between words,i.e.words used in a sentence are in strong relationship.Since this algorithm is applied iteratively for each sentence in texts,words used not in a sentence but at the same position in different sentences grow their relationship.In other words,this algorithm is able to capture the paradigmatic relationship from the syntagmatic one.The second term seizes the similarity among patterns of appearance of words in texts.Words with resemblant patterns of appearance among texts,e.g.,words used much often in particular texts but not so in other texts,raise their rela-tionship.Conversely,words with different patterns of appearance weaken their relationship.This is realized by calculating the correlation of appearance in texts. 1A text is a set of sentences.In our paper,this is applied to a conversation.2As space is limited,for the detail of the algorithm,see[3].Modeling Categorization Dynamics3 We model a simple conversation process between agents having word relation matrices as their internal structures.Here,we focus on dynamical changes of internal structures of agents through exchanging sentences,the simplest act of using language.A conversation between agents starts with uttering a sentence about a topic displayed to the agents.After the beginning of the conversation, each sentence is not restricted to the topic but there is some relevance with the previous sentence.In this model,to express this relevance,at least one word in the accepted sentence should be used in a reply sentence.The procedure of conversation in a text is as follows:1.A speaker agent produces a sentence about a topic.2.The sentence is modified according to the creativity rate,c,and then uttered to a hearer agent.3.The speaker’s word relation matrix is updated in terms of the uttered sentence.4.The hearer accepts the uttered sentence if there are less than two unknown words in the sentence3.If the sentence is not accepted,the speaker turns to another topic.(go to1.)5.The hearer’s word relation matrix is updated in terms of the accepted sentence. If there is an unknown word,the matrix is expanded to incorporate the new word.6.To reply to the utterance,the role of speaker and hearer are exchanged between them.(go to1.)When the number of accepted sentences or that of rejected sentences in a text reach some values,the text ends up.Then another pair of agents and a topic are selected for a new text.3Summary of Simulation ResultsIn one conversation,two agents fromfive are randomly selected as a speaker and a hearer.Sentences are produced artificially by agents by arranging words in which5different characters are combined4.The number of topics to be displayed to agents is10.The maximum of accepted sentences in a text is100and that of rejected sentences is5.The parameters areαw=0.4,c=0.1.The followings are the major results:1.Agents develop cluster structure in their own word relation matrices.We observe two characteristic types of clusters.One isflat type in which words have strong relationship with each other.The other is gradual type in which word relationships change gradually.As a result of development,these two types of clusters exist in combination.2.Relationship among words drastically changes when a new word is used or a word is used in an unusual way.For example,at the21st text in Fig.1(a) most words with strong relation with a word in new usage weaken their relation value and vice versa.3Note that the criterion for acceptance of uttered sentence by the hearer lays down the limitation of ability to make sense for new words.4The number of words and that of sentences are in principle infinity.4Takashi Hashimoto3.The structure of clusters has stability and adaptability.The change of position of words in cluster structure is examplified in Fig.1(b).The words in a new usage,linked with a dashed arrow,moves its belonging cluster.The other words move so coherently in each cluster that the whole structure of clusters is not modified very much.4.Structure common to agents develops in the course of conversations.5.Agents also develop structure peculiar to individuals,because they go through diversified experiences of conversations.(a)01020304050text 0.200.400.600.801.00w o r d r e l a t i o n s h i p (b)first principal compornent s e c o n d p r i n c i p a l c o m p o r n e n tFig.1.(a)Transition of word relationship.The x and y axes are the number of texts and the relationship of all words with a word in an agent’s word relation matrix,respectively.(b)Dynamics of cluster structure caused by the rapid change shown in (a).This is a scattered diagram from principal component analysis of matrices.4DiscussionsThe clustering can be regarded as categorization through conversations,since words in a cluster have stronger relation with each other and weaker relation with words outside the cluster.Typical clusters are a combination of two types of clusters,flat and gradual,that is,a flat center with gradual expansion into the peripheral.The cluster structure shares some characteristics with the prototype category [6,8].In the traditional notion of category,the membership of a category is thought to be defined rigidly like the set notion.In the prototype category theory,in contrast,the membership is matter of gradient and the boundary of a category is fuzzy.The category of liquid containers provides an example.Bottles and glasses are the typical members of the category.Glasses are similar to bowls,bowls are to soup plates,and soup plates are to flat dishes.Although neighboring members are fairly similar,the last one may not be the member of the category,but the boundary which defines the membership of the category is unclear.Another important feature of prototype categories is stability and adaptability with which languages should equip themselves to establish communication andModeling Categorization Dynamics5 to beflexible about changes.Prototype category and our cluster structure are akin in these traits5.Agents develop both commonality and individuality.The structure common to agents implies the emergence of a social system,in which some words are used in the same way by most agents.The words acquire,in the speculative view, virtual references in the society6.For a developmental enquiry,we should study how word relationship which reflects relation among prepared entities changes or expands with communication.The present algorithm shows not a simple convergence but drastic turnovers, which are usually brought by new combinations of words.The turnover behav-ior locally restructures words in clusters.Such new combinations of words is like metaphorical expressions which often tie different semantic domains by using words from the separated domains in one sentence.And such metaphorical ex-pressions,if they are totally impressive,might modify our internal models,or world view,dramatically.Therefore,it is important for dynamics of linguistic categorization not only to develop clusters but to modify the clusters by a small impact.This is also important for maintaining the dynamics at the global level.The coefficient parameterαw controls nonlinearity of the present system. Although the results reported here are seen in the broad area ofαw,the system is likely to fall intofixed and uniform structure at the too large value ofαw.If the creativity rate c is too large,the system has too strong randomness for us tofind any significant structure in word relation matrices and their dynamics. References1.M.Fukaya and S.Tanaka.Kotoba no Imiduke-ron (A Theory of”Sence Making”of Language)[in Japanese].Kinokuniya Shoten,Tokyo,1996.age-based structuralization of relationships between words.InP.Husbands and I.Harvey,editors,Fourth European Conference on Artificial Life, pages483–492.MIT Press,Cambridge,MA,1997.3.T.Hashimoto.Dynamics of internal and global structure through linguistic interac-tions.In Conte Sichman and Gilbert,editors,Multi-Agent Systems and Agent-Based Simulation,pages124–139.Springer-Verlag,Berlin,1998.4.K.Kaneko and I.Tsuda.Constructive complexity and artificial reality:an intro-duction.Physica,D75:1–10,1994.5.Y.Karov and S.Edelman.Similarity-based word sense puta-tional Linguistics,24:41–59,1998.koff.Women,Fire,and Dangerous Things.The University of Chicago Press,Chicago,1987.7.L.Steels.The synthetic modeling of language origin.Evolution of Communication,1(1):1–34,1997.8.J.R.Taylor.Linguistic Categorization–Prototypes in Linguistic Theory.OxfordUniversity Press,Oxford,1995.9.L.Wittgenstein.Philosophische Untersuchungen.Basil Blackwell,1953.5Prototype category has some other attributes[8].6Some abstract notions might be created in our society and internalized in using language.。