The new fault diagnosis method of wavelet packet neural network on
一种改进一维卷积神经网络的轴承故障诊断方法
第 22卷第 4期2023年 4月Vol.22 No.4Apr.2023软件导刊Software Guide一种改进一维卷积神经网络的轴承故障诊断方法潘琳鑫，巩永旺，晏生莲（盐城工学院信息工程学院，江苏盐城 224051）摘要：经典一维卷积神经网络模型诊断准确率不高且模型存在过拟合问题，难以满足轴承故障诊断时效性和准确率要求。
鉴于此，提出一种基于改进一维卷积神经网络的轴承故障诊断方法，在卷积层之后加入批量归一化层的方法以增加模型泛化能力，并采用Dropout的方法解决模型过拟合问题。
基于凯斯西储大学轴承故障数据集的实验结果显示，相比经典一维卷积神经网络，该方法可显著提升故障诊断准确率，故障诊断准确率可达99.79%，并且整个诊断过程无需手动特征提取，从而减少了特征提取过程中的损失，实现端到端的检测，具有较好的通用性。
关键词：故障诊断；卷积神经网络；深度学习；Dropout方法DOI：10.11907/rjdk.221570开放科学（资源服务）标识码（OSID）：中图分类号：TP183 文献标识码：A文章编号：1672-7800（2023）004-0038-05A Bearing Fault Diagnosis Method Based on Improved One-dimensionalConvolutional Neural NetworkPAN Lin-xin， GONG Yong-wang， YAN Sheng-lian（School of Information Engineering，Yancheng institute of Technology， Yancheng 224051， China）Abstract：Aiming at solving the problems of low diagnostic accuracy and over fitting of the classical one-dimensional convolutional neural net‐work model， a bearing fault diagnosis method based on improved one-dimensional convolutional neural network is proposed. The normaliza‐tion method is used to increase the generalization ability of the model. For the problem of model over fitting， the dropout method is adopted. Fi‐nally， the proposed method is applied to the experimental data of rolling bearing failure of Case Western Reserve University. The experimental results show that compared with the classical convolutional neural network， the proposed method can improve greatly the diagnostic accuracy （as high as 99.79%）. Moreover， the entire diagnosis process does not require any manual feature extraction， which can reduce the loss in the feature extraction process， and the method can realize end-to-end detection and has better versatility.Key Words：fault diagnosis； convolutional neural network； deep learning； Dropout method0 引言滚动轴承是旋转机械设备中最重要的零部件之一，在维持运动精度和提高机械效率上发挥重要作用，在复杂机械装备中有着广泛应用［1］。
外文文献
Fault diagnosis for temperature, flow rate and pressuresensors in V A V systems using wavelet neural network Zhimin Du, a, , Xinqiao Jina and Y unyu YangaaSchool of MechanicalEngineering, Shanghai Jiao TongUniversity,800,Dongchuan Road,Shanghai,ChinaReceived 30 April 2008; revised 14 January 2009;accepted14January2009. Available online 5 March 2009.Abstract Wavelet neural network, the integration of wavelet analysis and neural network, is presented to diagnose the faults of sensors including temperature, flow rate and pressure in variable air volume (V A V) systems to ensure well capacity of energy conservation. Wavelet analysis is used to process the original data collected from the building automation first. With three-level wavelet decomposition, the series of characteristic information representing various operation conditions of the system are obtained. In addition, neural network is developed to diagnose the source of the fault. To improve the diagnosis efficiency, three data groups based on several physical models or balances are classified and constructed. Using the data decomposed by three-level wavelet, the neural network can be well trained and series of convergent networks are obtained. Finally, the new measurements to diagnose are similarly processed by wavelet. And the well-trained convergent neural networks are used to identify the operation condition and isolate the source of the fault.Keywords Wavelet analysis; Neural network; Fault diagnosis; Sensor ; Variable air volumeIntroductionVariable air volume (V A V) systems are widely used in actual buildings to save energy through employing some optimal control strategies. Obviously, energy conservation capacity of a real V A V system deeply depends on the executing efficiency of various control loops including outdoor air flow rate, supply air temperature, supply air static pressure and zone temperature controllers. These controllers modulate related components after comparing the measurements of the control variables with the optimal setpoints. With the effective control, energy conservation and better indoor air quality can be achieved. During the control process, however, one premise can never be ignored: the measurements are accurate. If the sensors are biased, the controller may be misled and give incorrect commands. The related components may be incorrectly modulated. Finally the energy consumption of the system may be unreasonably increased greatly. In summer conditions, for example, the positive biases (the measurements are larger than the true values) of supply air temperature mislead the controller to open the water valve at a larger position. The quantity of chilled water is increased incorrectly, which wastes more energy of the pumps. Similarly, the biases of outdoor air flow rate sensors may increase the chilled water flow rate or decrease the chilled water temperature that may increase energy consumption of the pumps or chillers. The faults of supply air static pressure sensor may require higher rotational speed of the supply fan,which means more energy consumption of the fan. Consequently, the waste of energy is always inevitable under those various faulty conditions although optimal control strategies are applied in the system. Finding a suitable method to detect and diagnose the faults occurred in the V A V system, to avoid the waste of energy, is a significant target.Recently, the study of fault detection and diagnosis (FDD) for sensors in heating, ventilation and air conditioning field are more active than ever after the popularity of research on faults of facilities including chillers [1], [2], [3] and [4] and air-handling units [5], [6], [7], [8], [9] and [10]. Two typical diagnosis methods for sensor faults have been developed. One is the model-based, and the other is the data-driven.The model-based method [11], [12], [13] and [14] is to obtain predicted values of the parameters calculated by the mathematical models first. Then the differences between the outputs of real process and those of predicted ones, so-called residuals, are calculated and used as the fault indexes to diagnose. Stylianou and Nikanour [1] used a first-order model to detect faults of temperature sensors by comparing the actual temperature decay with the model output using the hypothesis testing. Wang and Wang [15] developed model-based strategies to diagnose the faults of commonly used temperature and flow rate sensors in chilling plant. The premise of model-based method is that accurate mathematical models must be built. And this premise is also the difficult point for the application. The model-based method is efficient to discover the abrupt faults of sensors such as the complete failure through analyzing the great change of operation conditions caused by the abrupt faults. Limited by the precision of the prediction models, however, it is insensitive to detect the small fixed or drifting biases since not abrupt change but slow degradation of the operation or control efficiency happens.The data-driven approaches [16], [17] and [18], on the other hand, never construct physical models but just learn the intrinsic relations among variables or parameters through employing the process data including normal and faulty conditions. Recently, principal component analysis [19] and [20] and Fisher discriminant analysis [21] were presented to diagnose the sensor faults in heating, ventilation and air conditioning systems. Besides the statistic method, neural network and wavelet analysis also began to apply in this field. Lee [22] presented general regression neural network models to diagnose the abrupt and performance degradation faults in an air-handling unit. Wang and Chen [23] developed a neural network trained by lots of running data to diagnose the faults of outdoor, supply and return air flow rate sensors. Later, Chen et al. [18] employed wavelet analysis to diagnose the faults of flow rate sensors in central chilling systems. Obviously, the data-driven method highly relies on the quantity and quality of the data obtained. Fortunately, with the popularity of building automation and energy management and control systems, the various historical operation data including normality and fault can be collected and obtained easily.A data-driven diagnosis method combining wavelet analysis with neural network is presented in this paper that can be used to diagnose the faults in the V A V systems. Wavelet decomposition is used to process the original data and then the characteristic data representing the main operation information of the system are obtained. Employing these data decomposed, the neural networks are well trained and then they can identify various faults of commonly used sensors including temperature, flow rate and pressure in the V A V systems2. Wavelet neural networkNeural network technique is a valuable pattern recognition method in theory and application. It is widely used in engineering application [24], [25] and [26] especially to deal those issues concerned in non-linear or complicated systems. It is efficient to learn the certain status or operation condition of the objective systems. And then the well-trained network can recognize these various conditions. Actually, the process of fault diagnosis is essentially a kind of recognition classification or recognition. Therefore, the neural network can be used as a diagnosis method. In fact, it has been well applied to detect and diagnose faults in many fields [27], [28], [29] and [30].2.1. Application opportunity in VAV systemsAs a complicated non-linear system, the VAV system includes many control components and measuring sensors.According to the different control strategies, the variables have changeable control relations. Also, the variables have implicit physical relations because of the physical principles. For this complex system, it is difficult to construct not only general but also precise models for so many variables. As a data-driven method, however, neural network never construct detailed models but continually learn the operation data. Through lots of training, the neural network can capture the important physical and control relations among the different variables in the VAV system. Once the networks obtain the main information of different operation conditions, they can be used for fault diagnosis.Though neural network is capable of learning and judging various operation conditions, its capacity of data processing or analyzing is not satisfied. Especially for the VAV system, there are large quantities of samples for many measuring and control points. Since the noises are always included in the measurements and some uncertainty factors usually disturb the control actions, the pure neural network method may be affected or disturbed. As a result, its diagnosis efficiency is limited. Themistaken-warnings or missing-warnings may happen inevitably. To solve this problem, the data selected for training must be preprocessed to remove those disturbing information.Wavelet analysis, originally developed from the Fourier transform at the end of 1980s [31] and [32], is widely used in various engineering [33], [34] and [35] systems. The wavelet analysis, also called wavelet transform, employs two opposite time intervals: shorter and longer. The shorter time interval can be used by wavelet analysis to analyze the high frequency characteristics of the signals. While the longer one is used to analyze the low frequency characteristics of the signals. Since it is capable of data processing, it can be used as the complement for neural network.Through analyzing the time-varying signals of variables in the VAV system, the wavelet can capture the local time-frequency domain information. The main important information of the system can be seized. Simultaneously, the disturbing factors can be removed. Indeed, the data after processing are much better than those initial signals for related trainings of the networks. Wavelet analysis can improve the diagnosis process of neural network. Consequently, wavelet neural network, the integration of wavelet analysis and neural network that the former is to process data and the latter to diagnose, is a valuable approach.2.2. Wavelet decompositionAccording to wavelet analysis, the measurements signals from sensors in the VAV systems can be decomposed using three-level wavelet packets shown in Fig. 1. S is the original signal from various sensors.S ij means the j th node or decomposition coefficient in the i th level. Consequently, the original signal S can be reconstructed using all of the nodes in the 3rd level and expressed asSince the characteristic of frequency domain representing operation condition is captured using the wavelet packets, the original signal can be replaced by thesedecomposed wavelets. And the eigenvector matrix concluded by the wavelet decomposition can be used as the training data for neural network to diagnose.Description of the VAV systemA typical VAV system, including air-handling unit, supply fan, VAV terminals, return fan, controllers and various sensors,is shown in Fig. 5. The supply air, a mixture of outdoor air and recycle air, is cooled down (in summer condition) in the air-handling unit using the chilled water coming from the chillers. The supply air is circulated to the VAV terminals to meet the demand of zones. The return air from the zones, on the other hand, is divided into two branches: one is circulated as the recycle air to join the next circulation, and the other is exhausted out of the building.In this VAV system, five kinds of controllers are included to improve the operation efficiency of the system so as to save more energy. Outdoor air flow rate (M OA) controller adjusts the air dampers to ensure enough outdoor air for the users. The supply air temperature (T SA) controller modulates the chilled water valve to maintain suitable supply air temperature through comparing the feedback information of T SA sensor with its optimal setpoint. The P SA controller adjusts the variable-speed supply fan to ensure the proper supply air static pressure. The variable-speed return fan is modulated to ensure the positive indoor air pressure. Moreover, many VAV terminal controllers located in multi-zones adjust the terminal dampers to ensure the heat comfort through monitoring the temperature and air flow rate in each zone.The diagnosis capacity of wavelet neural network may be limited if only separated measuring data of the variables are selected and used for training. After all the pure data-driven method does not directly describe the physical relations among variables. Actually, some of the variables are correlated and they influence each other because of the physical models and control relations. The relevant variables can be defined as one data group. Through classifying various groups, the diagnosis process can beimproved greatly. Therefore, some physical balances representing strong relevant relations among variables are employed to improve the diagnosis efficiency.ConclusionThe faults in the V A V system may not only worsen the operation efficiency but also waste the energy of system. To ensure the capacity of energy conservation, wavelet neural network integrating wavelet analysis with neural network is presented to diagnose the faults of sensors in the V A V system.Wavelet analysis is used to process the original data so as to seize the essential operation information of the V A V system. With three-level wavelet decomposition, the characteristic information of the system is obtained. Using these processed data, the neural networks are easily trained and recognize the system well. Once the convergent networks are obtained, the new data to diagnose are decomposed similarly using three-level wavelet. Finally, the new operation conditions, including normality and faults, can be diagnosed one by one using the well-trained networks.Two main contributions have been made in this paper. Firstly, wavelet neural network is used to diagnose the faults in the V A V systems, which is the combination of wavelet analysis and neural network. The actual V A V and its control systems are very complex that include many measuring points and control components. The efficiency of pure neural network may be not satisfied because its prediction or recognition may be disturbed by some uncertain or subordinate factors such as the measuring noises. As a result, more mistaken-warnings or missing-warnings may happen during the diagnosis process. With the wavelet analysis, however, the main information can be seized and those disturbing factors can be well removed. After the wavelet-based data processing, the diagnosis efficiency of neural network can be improved.As the universal mathematic methods, in addition, neither wavelet analysis nor neural network can well describe the physical relations between the variables in the V A V systems. Ifonly separated data are used for training, its diagnosis efficiency may be inevitably limited. The second contribution in this paper is the construction of various data groups using the essential conservation relations and models in the V A V system. Once the data are classified according to those physical models, the recognition for the relevant relations among variables is strengthened. Consequently, the diagnosis capacity of wavelet neural network is improved.。
状态监测术语(中英对照)
诊断名词术语和释义基本术语(1)状态监测(condition monitoring)－对机械设备的工作状态（静的和动的）进行监视和测量（实时的或非实时的），以了解其正常与不正常。
(2)故障诊断(fault diagnosis)又称为技术诊断(technical diagnosis)－采用一定的诊断方法和手段，确定机械设备功能失常的原因、部位、性质、程度和类别，明确故障的存在和发展。
(3)简易诊断(simple diagnosis)－使用简易仪器和方法进行诊断。
(4)精密诊断(meticulous diagnosis)－使用精密仪器进行的诊断（优于精确诊断或精度诊断术语）。
(5)故障征兆(symptom of fault)（或称故障症状）－能反映机械设备功能失常，存在故障的各种状态量。
(6)征兆参数(symptom of parameter)－能有效识别机械设备故障源故障的各种特征量，包括：原始量和处理量。
(7)状态识别(condition recognition/identification)－为判断机械设备工作状态的正常与不正常和通过故障状态量的区别，诊断其故障的方法。
(8)特征提取(feature extraction)－为了正确识别和诊断机械设备故障的存在与否，对征兆参数进行特别的处理。
(9)故障类别(fault classification)－反映机械设备功能失常、结构受损、工作实效的专用分类、名称。
(10)故障性质(nature of fault)－描述故障发生速度、危险程度、发生规律、发生原因等问题。
(11)突发故障(sudden fault)－突然发生的故障。
在故障发生瞬间，必须采用实时监控、保安装置、紧急停机等措施。
(12)渐发故障(slow fault)－故障的形成和发展比较缓慢，能够提供监测与诊断的条件。
(13)破坏性故障(damaging fault)或称灾难性故障(catastrophic fault)－故障的发生影响机械设备功能的全部失去，并造成局部或整体的毁坏，难以修复重新使用。
pathological infant cry
Pathological infant cry analysis using wavelet packet transform and probabilistic neural networkM.Hariharan ⇑,Sazali Yaacob,Saidatul Ardeenaawatie AwangSchool of Mechatronic Engineering,Universiti Malaysia Perlis (UniMAP),02600Perlis,Malaysiaa r t i c l e i n f o Keywords:Acoustic analysis Infant cryWavelet packet transform Probabilistic neural networka b s t r a c tA new approach has been presented based on the wavelet packet transform and probabilistic neural net-work (PNN)for the analysis of infant cry signals.Feature extraction and development of classification algorithms play important role in the area of automatic analysis of infant cry signals.Infant cry signals are decomposed into five levels using wavelet packet transform.Energy and entropy measures are extracted at every level of decomposition and they are used as features to quantify the infant cry signals.A PNN is developed to classify the infant cry signals into normal and pathological and trained with dif-ferent spread factor or smoothing parameter to obtain better classification accuracy.The experimental results demonstrate that the proposed features and classification algorithms give very promising classi-fication accuracy of 99%and it proves that the proposed method can be used to help medical profession-als for diagnosing pathological status of an infant from cry signals.Ó2011Elsevier Ltd.All rights reserved.1.IntroductionCrying is the only way of communication for an infant.From the cry,a trained professional can understand the physical or psycho-logical status of the baby.Infants cry due to some possible reasons such as,hunger,pain,sleepiness,discomfort,feeling too hot or too cold,and too much noise or light.Acoustic analysis of infant cry signal is a non-invasive and has been proven tool for the detection of certain pathological conditions (García &García,2003;Orozco &García,2003;Reyes-Galaviz &Reyes-Garcia,2004;Reyes-Galaviz,Verduzco,Arch-Tirado,&Reyes-García,2005).In the recent years,simple techniques have been proposed for analyzing the infant cry through linear prediction coding,Mel frequency cepstral coef-ficients and pitch information (García &García,2003;Orozco &García,2003;Reyes-Galaviz &Reyes-Garcia,2004;Reyes-Galaviz et al.,2005).Little attention has been paid by the researchers based on wavelet and wavelet packet transform.This paper presents the development of an intelligent classification system for classifying normal and pathological cry using wavelet packet transform (WPT)and probabilistic neural network (PNN).The application of wavelet and wavelet packet transform analysis is diversified and has been used in many signal and image processing applications.Avci and Avci have proposed a novel approach for radio signal classification based on wavelet packet energy and multi-class support vector machine (Avci &Avci,2008).Xian and Zeng have proposed an intelligent fault diagnosing method of rotation machinery based on the wavelet packet analysis and hybrid support vector machines (Xian &Zeng,2009).Wu and Liu have proposed a fault diagnosis system for internal combustion engines using wavelet packet transform (WPT)and artificial neural network (ANN)techniques (Wu &Liu,2009).Wu and Lin have con-ducted an investigation on speaker identification based on discrete wavelet packet transform with irregular decomposition (Wu &Lin,2009).Hanbay et al.have proposed a method for the prediction of wastewater treatment plant performance based on wavelet packet decomposition and neural networks (Hanbay,Turkoglu,&Demir,2008).1.1.Previous worksThis section deals with some of the significant works on infant cry signal analysis.Reyes-Galaviz et al.have presented the development of an automatic infant cry recognizer for the early identification of pathologies with the objective of classifying three classes,normal,hypo acoustics and asphyxia (Reyes-Galaviz et al.,2005).They used Mel frequency cepstral coefficients (MFCCs)for feature extraction and a Feed Forward Input Delay neural network with training based on Gradient Descent with Adaptive Back-propagation for classification.The accuracy of their proposed system varies from 96.08%to 97.39%.Orozco and García have developed a method based on linear prediction technique and scaled conjugate gradient neural networks for the detection of pathologies from infant cry.The classification accuracy of their proposed method was 91.08%for 314samples and 86.20%for0957-4174/$-see front matter Ó2011Elsevier Ltd.All rights reserved.doi:10.1016/j.eswa.2011.06.025⇑Corresponding author.Tel.:+6049798419;fax:+6049885167.E-mail addresses:hari@.my (M.Hariharan),s.yaacob@.my (S.Yaacob),saidatul@.my (S.A.Awang).1036samples(Orozco&García,2003).In another study,the same authors have used MFCC and linear prediction coding techniques for characterizing the infant cry signal and used a feed-forward neural network for classification with several learning methods (García&García,2003).The accuracy of their proposed system was up to97.43%.Várallyay et al.have proposed fundamental fre-quency detection by the smoothed spectrum method(SSM)for the analysis of infant cry signal(Várallyay,Benyó,Illényi,Farkas,& Kovács,2004).From the previous works,it has been observed that the feature extraction plays an important role in the area of auto-matic detection of pathological cries.In this paper,the method based on wavelet packet transform is proposed for analyzing the infant cry signals.The infant cry is signal decomposed intofive lev-els.Energy and entropy features are computed from the wavelet packet coefficients and used to characterize the infant cry signals. In order to test the effectiveness of wavelet packet energy and en-tropy features,a probabilistic neural network is employed.The experimental results elucidate that the wavelet packet features and PNN classifier can be used to detect certain pathological con-ditions of infant cry medically.The synopsis of the paper is as follows:Section2deals with a brief explanation of the infant cry database used in this work.Section3 deals with introduction to the design of wavelet packet subbandfil-ters and feature extraction of energy and Shannon entropy.The brief introduction of PNN classifier is described in Section4.In Section5, inferences from the results of the PNN classifier are presented and it shows the usefulness of the wavelet packet features.Finally,Section 6deals with the conclusion and suggestion for the future work. 2.Data selectionThe database of infant cry is downloaded from the website http://ingenieria.uatx.mx/orionfrg/cry/called Baby Chillanto data-base and is a property of the Instituto Nacional de Astrofisica Opti-ca y Electronica(INAOE)–CONACYT,Mexico.The database is described in Reyes-Galaviz,Cano-Ortiz,Reyes-García,y Electró-nica,and Puebla(2009).From that,340normal cry and340path-ological cry with asphyxia are used for our analysis.The length of cry signal is1s.Asphyxia is defined as the failure to breathe well within1min after delivery of the baby.This disease can cause damage to the brain,organs and tissues or even death if subjected to delayed or improper treatment.The sampling frequency of in-fant cry signals is set to16,000Hz for our analysis.3.Design of wavelet packetfiltersThis section briefly explains the design of wavelet packetfilters and the feature extraction using them.3.1.Wavelet packetsIn discrete wavelet transform(DWT)decomposition procedure, a signal is decomposed into two frequency bands such as lower frequency band(approximation coefficients)and higher frequency band(detail coefficients).Low frequency band is used for further decomposition.Hence DWT gives a left recursive binary tree struc-ture.In wavelet packet(WP)decomposition procedure,both lower and higher frequency bands are decomposed into two sub-bands. Thereby wavelet packet gives a balanced binary tree structure.In the tree,each subspace is indexed by its depth i and the number of subspaces p.The two wavelet packet orthogonal bases at a par-ent node(i,p)are given by the following forms(Burrus&Haitao Guo,1998;Raghuveer&Rao,1998)w2piþ1ðkÞ¼X1n¼À1l½n w piðkÀ2i nÞð1Þwhere l[n]is a low pass(scaling)filterw2pþ1iþ1ðkÞ¼X1n¼À1h½n w piðkÀ2i nÞð2Þwhere h[n]is the high pass(wavelet)filter.Wavelet packet decom-position helps to partition the high frequency side into smaller bands which cannot be achieved by using general discrete wavelet transform.The decomposition coefficient of i th depth can be ob-tained by the(iÀ1)th level,finally we can get the coefficients of all levels through sequential analogy.After it is decomposed by i th depths,the frequency ranges of all subspaces at the i th depth are given as[0,f s/2i+1],[f s/2i+1,2f s/2i+1],...,[(2iÀ1)f s/2i+1,f s/2], where f s is the sampling frequency.In this paper,the infant cry sig-nal is decomposed in tofive levels.Table1summarizes the fre-quency bands of each level which has been decomposed.In our analysis,the sampling frequency f s is16,000Hz.The cry signals are sampled at16kHz giving an8kHz band-width signal.The cry signals arefiltered with the wavelet packet filters and four different orders of Daubechies wavelets db1,db4, db10and db20are used.In this work,Daubechies wavelet has been chosen due to the following properties(Cohen,Daubechies,& Feauveau,2006):time invariance–if the time series is time shifted then its wavelet packet coefficients are only time shifted.Fast com-putation–Daubechies wavelets have fractal-like self-similarity properties that lead to fast wavelet transform techniques.Sharpfil-ter transition bands–Daubechies wavelets have very sharp transi-tion bands which minimizes edge effects between frequency bands.Energy and Shannon entropy are computed using the ex-tracted wavelet packet coefficients in this study.The subband en-ergy can be computed using the extracted wavelet packet coefficients by the following Eq.(3)Energyn¼X ni¼1j C Pn;kj2n¼1;2;...;N;k¼0;1;...;2NÀ1ð3Þwhere P is the scale index,n represents the number of decomposi-tion level.Entropy can be used as a common measure to quantify irregular pattern of the pathological infant cry signals.The measure ofTable1Frequency band of each level.Wavelet packetdecomposition levelFrequency band(Hz)10–4000,4000–800020–2000,2000–4000,4000–6000,6000–800030–1000,1000–2000,2000–3000,3000–4000,4000–5000,5000–6000,6000–7000,7000–8000,40–500,500–1000,1000–1500,1500–2000,2000–2500,2500–3000,3000–3500,3500–4000,4000–4500,4500–5000,5000–5500,5500–6000,6000–6500,6500–7000,7000–7500,7500–800050–250,250–500,500–750,750–1000,1000–1250,1250–1500,1500–1750,1750–2000,2000–2250,2250–2500,2500–2750,2750–3000,3000–3250,3250–3500,3500–3750,3750–4000,4000–4250,4250–4500,4500–4750,4750–5000,5000–5250,5250–5500,5500–5750,5750–6000,6000–6250,6250–6500,6500–6750,6750–7000,7000–7250,7250–7500,7500–7750,7750–800015378M.Hariharan et al./Expert Systems with Applications38(2011)15377–15382Fig.1b.Two level wavelet packet decomposition of the pathological cry signal with‘db4’wavelet.to quantify wavelet packet coefficients such as energy and entropy.A feature database is created,after the computation of entropy measures from each subband wavelet packet coefficients and they are used as input features for the classifiers to distinguish the cry signals as normal or pathological.The block diagram of the feature extraction and classification phase is shown in Fig.2.4.Probabilistic neural networkArtificial neural networks are widely used in pattern recogni-tion and classification problems by learning from examples.Differ-ent neural network paradigms are available for classifying patterns.In this work,PNN structure used for classifying normal and pathological cries.Specht has proposed the probabilistic neu-ral net based on Bayesian classification and classical estimators for probability density function(Specht,1990).PNN comprises of four units,such as input units,pattern units,summation units and output units.All the units are fully interconnected and the pat-tern units are activated by exponential function,instead of sigmoi-dal activation function.The pattern unit computes distances from the input vector to the training input vectors,when an input is pre-sented,and produces a vector whose elements indicate how close the input is to a training input.The summation unit sums these contributions for each class of inputs and produces a net output which is a vector of probabilities.From the maximum of these probabilities,output units produce a1for that class and a0for the other classes using compete transfer function.The net can be used for classification as soon as an example of a pattern from each of the two classes has been presented to it.How-ever,PNN generalizes well as it is trained with more examples. Varying smoothing parameter(r)gives control over the degree of nonlinearity of the decision boundaries for the net.A decision boundary approaches a hyperplane for large values of r and approximates the highly nonlinear decision surface of the nearest neighbour classifier for small values of r that are close to zero.In this paper,PNN architecture is constructed using newpnn()in MATLAB function.The detailed information about the PNN archi-tecture and mathematical equations can be found in the Specht’s paper(Specht,1990).5.InferencesIn this study,conventional validation was performed to evalu-ate the efficacy of the proposed features.680segments(340nor-mal+340pathological)of cry signals were used.Among them, 60%of data(408samples)were used for training and remaining 40%of data(272samples)were used for testing.The PNN was trained with different spread factor or smoothing parameter such as from0.001to0.009and from0.01to0.09.Though,the network was trained with different spread factors,only the best accuracy results were presented.In order to test the classifier performance, several measures such as,sensitivity,specificity,and the overall accuracy were considered.The sensitivity,specificity and overall accuracy were calculated from the measures true positive(TP), true negative(TN),false positive(FP),and false negative(FN)as presented in Table2.TP=true positive,the classifier classified as pathology(Asphy-sia)when pathological samples were present.TN=true negative,the classifier classified as normal when nor-mal samples were present.FN=false negative,the classifier classified as normal when pathological samples were present.FP=false positive,the classifier classified as pathological when normal samples were present.Sensitivity=TP/(TP+FN).Table2Confusion matrix.Actual classification Predicted classificationPathological NormalPathological TP FNNormal FP TNTable3PNN classification results at each level of wavelet packet decomposition using‘db1’.Decomposition level Spread factor Sensitivity Specificity Overall accuracy Spread factor Sensitivity Specificity Overall accuracy Energy features Entropy features10.00982.9178.6580.480.00980.4374.2176.880.0183.6379.3981.320.0180.3373.8776.580.0287.0480.3783.310.0383.3774.8778.460.0388.2079.3082.900.0584.8974.5978.7920.00586.2381.8883.820.00682.6780.5081.430.0185.8882.2683.860.00982.9179.1680.850.0389.1083.0585.770.0184.1479.1081.320.0490.0181.5285.180.0286.1977.0080.8530.00594.2789.4391.650.00588.9586.5787.680.00793.2490.6791.880.00790.4985.4087.720.0194.3790.1892.130.0189.7985.0087.170.0293.4589.3391.250.0292.2983.4387.3240.00494.0294.5494.260.00594.3893.3593.820.00995.3193.7694.490.0195.0192.5393.680.0196.4493.0094.630.0295.1591.8693.420.0597.4592.0994.560.0394.5490.2892.2850.00796.1794.0795.070.00996.4495.2195.770.0197.3892.5594.820.0197.6795.1496.360.0796.8792.7594.670.0298.3294.1896.140.198.8091.2394.670.0396.8494.9195.8515380M.Hariharan et al./Expert Systems with Applications38(2011)15377–15382Specificity=TN/(TN+FP).Overall accuracy=(TP+TN)/(TP+TN+FP+FN).Tables3–6shows the results of PNN for‘db1’,‘db4’,‘db10’,and ‘db20’with energy and entropy features.From tables,it was found that increasing the wavelet packet decomposition level leads to the extraction of a more predominant group of feature vectors,thereby increase in the classification accuracy.The highest classification accuracy was achieved at thefifth level of wavelet packet decom-position.From Table3,the percentage of increase in classification accuracy was12%for energy features and17%for entropy features at thefifth level of decomposition.The best overall accuracy was 95.07%for energy features and96.36%for entropy features.From Table4,the percentage of increase in classification accuracy15% for energy features and23%for entropy features.For Daubechies wavelet‘db4’,entropy features perform better than energy features.From Table5,the percentage of increase in classification accuracy was17%for energy features and23%for entropy features.For‘db10,both the features equally perform well at thefifth le-vel of decomposition.From Table6,the percentage of increase in classification accuracy was12%for energy features and15%for en-tropy features.The number of features at thefifth level of decom-position was32.All the32features were used to provide better representation of cry signal.From the above discussion,it can be observed that the maximum of classification accuracy can be ob-tained regardless of the different order of Daubechies wavelets. In this paper,two-class pattern recognition was carried out such as normal or pathological cry signal.In Section2,some of significant works were reported and the maximum classification accuracy of97.43%was obtained.At thefifth level of wavelet decomposition,the maximum classification accuracy of99.49% for‘db20’was obtained and it shows that the proposed featuresTable4PNN classification results at each level of wavelet packet decomposition using‘db4’.Decomposition level Spread factor Sensitivity Specificity Overall accuracy Spread factor Sensitivity Specificity Overall accuracy Energy features Entropy features10.00977.5282.0379.560.00977.8972.6574.930.0177.1083.7679.960.0180.9273.7476.730.0277.0489.4582.020.0383.9075.3578.900.0375.4692.1381.690.0584.3373.4677.7220.00589.1687.2188.130.00686.5387.1586.800.0189.5287.2388.270.00988.5385.3286.760.0291.3785.8488.240.0187.2685.4186.210.0493.2283.5087.720.0390.9581.5785.6330.00694.9894.0094.450.00792.8090.2691.430.00896.3793.0994.630.00892.7290.4191.470.0497.4292.0894.560.0193.4990.9692.170.0598.4291.8694.890.0395.4989.2892.1340.00595.1195.4295.260.00795.6095.3295.440.00796.9594.1695.440.00996.1895.4995.810.0698.4893.4095.740.0395.8895.4895.660.0798.3293.6295.810.0495.8895.0695.4450.00879.9295.5986.030.00995.0897.4496.210.0188.4495.5791.650.0195.3696.5995.960.0892.3297.6694.820.0298.5096.4297.430.0992.2097.5294.670.0398.4995.8797.13Table5PNN classification results at each level of wavelet packet decomposition using‘db10’.Decomposition level Spread factor Sensitivity Specificity Overall accuracy Spread factor Sensitivity Specificity Overall accuracy Energy features Entropy features10.00982.3779.8981.030.00978.3973.9875.810.0182.5380.4181.360.0179.7773.8276.360.0286.5280.7983.350.0384.9074.6478.820.0487.9081.7384.260.0584.1674.9378.7520.00895.8895.1295.480.00295.2894.0394.600.00996.3094.8295.510.00595.6393.9794.740.0295.6295.4895.510.0295.9492.8494.300.0397.4793.6095.400.0395.2190.0692.4330.00698.0394.5196.180.00697.0696.6596.840.00898.7794.0396.250.00997.9896.1997.060.0198.1094.2596.070.0297.3396.1096.690.0698.6093.6995.990.0396.6196.1396.3640.00898.3595.8797.060.00696.4198.4497.390.00998.2896.4997.350.00897.3897.7297.540.0498.5196.6997.570.0297.6396.6697.130.0799.1796.3697.720.0497.8897.0497.4350.00995.9199.3997.570.00996.6499.1097.830.0198.3797.1197.720.0197.8897.8797.870.0499.7797.2298.460.0799.5597.6398.570.0899.8597.3698.570.0898.9798.1198.53M.Hariharan et al./Expert Systems with Applications38(2011)15377–1538215381and classification algorithm provides better classification com-pared to earlier works.Finally,the experimental results indicates that the propose features has the potential in detecting pathologi-cal problem of an infant from cry signals.6.ConclusionThis paper presents the analysis of infant cry signals based on the wavelet packet transform and PNN.The infant cry signals were decomposed intofive levels.At each level,energy and entropy features were extracted using wavelet packet coefficients.The effect of the proposed features was tested at every level of decomposition of cry signals.The sensitivity,specificity,and overall accuracy were used as performance measures,in order to test the reliability of the PNN classifier.The maximum classification accuracy of99%was ob-tained at thefifth level of decomposition.The classification results indicate that the proposed method could be used as a valuable tool for clinical diagnosis of the infant cry signals.AcknowledgementsThe Baby Chillanto Data Base is a property of the Instituto Nac-ional de Astrofisica Optica y Electronica–CONACYT,Mexico.We like to thank Dr.Carlos A.Reyes-Garcia,Dr.Emilio Arch-Tirado and his INR-Mexico group,and Dr.Edgar M.Garcia-Tamayo for their dedication of the collection of the Infant Cry data base.The authors would like to thank Dr.Carlos Alberto Reyes-Garcia,Re-searcher,CCC-Inaoep,Mexico for providing infant cry database. ReferencesAvci,E.,&Avci,D.(2008).A novel approach for digital radio signal classification: Wavelet packet energy-multiclass support vector machine(WPE-MSVM).Expert Systems with Applications,34(3),2140–2147.Burrus,C.S.,&Haitao Guo,R.A.G.(1998).Introduction to wavelets and wavelet transforms:A primer.Prentice Hall.Cohen,A.,Daubechies,I.,&Feauveau,J.(2006).Biorthogonal bases of compactly supported munications on Pure and Applied Mathematics,45(5), 485–560.García,J.,&García,C.(2003).Acoustic features analysis for recognition of normal and hypoacustic infant cry based on neural networks.Artificial Neural Nets Problem Solving Methods,615–622.Hanbay,D.,Turkoglu,I.,&Demir,Y.(2008).Prediction of wastewater treatment plant performance based on wavelet packet decomposition and neural networks.Expert Systems with Applications,34(2),1038–1043.Orozco,J.,&García,C.(2003).Detecting pathologies from infant cry applying scaled conjugate gradient neural networks.Paper Presented at the European Symposium on Artificial Neural Networks,Bruges(Belgium),23–25April2003.Raghuveer,M.,&Rao,A.S.B.(1998).Wavelet transforms:Introduction to theory and applications.Addison-Wesley.Reyes-Galaviz,O.,Cano-Ortiz,S.,Reyes-García,C.,y Electrónica,O.,&Puebla,M.(2009).Evolutionary-neural system to classify infant cry units for pathologies identification in recently born babies.paper presented at the8th mexican international conference on artificial intelligence,MICAI2009,Guanajuato, Mexico.Reyes-Galaviz,O.,&Reyes-Garcia,C.(2004).A system for the processing of infant cry to recognize pathologies in recently born babies with neural networks.Paper presented at the9th conference speech and computer(SPECOM’2004),St.Petersburg,Russia,September20–22,2004.Reyes-Galaviz,O.,Verduzco,A.,Arch-Tirado,E.,&Reyes-García,C.(2005).Analysis of an infant cry recognizer for the early identification of pathologies.Nonlinear Speech Modeling and Applications,3445,404–409.Specht,D.(1990).Probabilistic neural networks.Neural networks,3(1),109–118. Várallyay,G.,Jr.,Benyó,Z.,Illényi, A.,Farkas,Z.,&Kovács,L.(2004).Acoustic analysis of the infant cry:Classical and new methods.Paper presented at the26th annual international conference of the IEEE EMBS,San Francisco,CA, USA.Wu,J.,&Lin, B.(2009).Speaker identification using discrete wavelet packet transform technique with irregular decomposition.Expert Systems with Applications,36(2),3136–3143.Wu,J.,&Liu,C.(2009).An expert system for fault diagnosis in internal combustion engines using wavelet packet transform and neural network.Expert Systems with Applications,36(3),4278–4286.Xian,G.,&Zeng,B.(2009).An intelligent fault diagnosis method based on wavelet packer analysis and hybrid support vector machines.Expert Systems with Applications,36(10),12131–12136.Table6PNN classification results at each level of wavelet packet decomposition using‘db20’.Decomposition level Spread factor Sensitivity Specificity Overall accuracy Spread factor Sensitivity Specificity Overall accuracy Energy features Entropy features10.00888.8285.5587.020.00884.0884.1284.040.0188.5987.3987.900.0185.6482.4883.900.0291.1187.6689.260.0283.9482.9583.270.0388.7188.6888.600.0484.8780.2182.3220.00698.3097.3297.790.00897.6098.4598.010.0198.7497.0597.870.0197.8998.2498.050.0398.3398.1098.200.0297.9097.8997.870.0597.9898.1998.050.0496.3997.5596.9530.00798.9098.0598.460.00698.3998.0498.200.00998.9898.5598.750.00798.5397.6898.090.0199.1998.4898.820.0298.2598.1898.200.0598.9098.1198.490.0498.3998.3198.3540.00899.6399.3499.490.00898.9098.4798.680.00999.7098.2698.970.00999.0598.4098.710.0499.5698.7099.120.0198.9199.1299.010.0799.6398.4899.040.0598.9798.5498.7550.00991.5099.4495.110.00998.3499.1998.750.0197.1499.3298.200.0199.2799.4999.380.0399.8598.4899.150.0299.6399.2099.410.0999.8598.4299.120.0399.5699.0699.3015382M.Hariharan et al./Expert Systems with Applications38(2011)15377–15382。
汽车毕业论文
汽车毕业论文篇一：汽车专业毕业论文北京工业职业技术学院汽车检测与维修专业毕业设计系别机电工程系专业班级汽车检测与维修0831姓名学号指导教师联系方式北京工业职业技术学院汽车波形分析与检验目录摘要---------------------------------------------------------- 3 第 1 章绪论---------------------------------------------------- 51.1 波形分析法概念--------------------------------------------- 5 1.2 故障诊断机理----------------------------------------------- 5 1.3 信号的类型。
----------------------------------------------- 5 1.4 信号特征--------------------------------------------------- 6 1.5 波形诊断分析机理------------------------------------------- 6 第 2 章波形分析------------------------------------------------ 72.1 次级高压电火波形分析-------------------------------------- 7 2.11 分电器点火次级阵列波形------------------------------------ 7 2.12 分电器点火次级阵列波形（急加速）-------------------------- 8 2.13 分电器点火次级单缸波形------------------------------------ 92.14 电子点火(EI)次级单缸波形----------------------------------112.15 电子点火次级单缸急加速波形--------------------------------132.16 无分电器/电子点火线圈压力试验-----------------------------142.17 电子点火作功及排气点火测试--------------------------------162.2 初级低压点火波形分析--------------------------------------172.21点火初级闭合角波形-----------------------------------------172.22点火初级线圈电流波形---------------------------------------192.23分电器点火初级阵列波形-------------------------------------212.24分电器初级阵列波形(调整时基和触发)-------------------------222.25分电器初级单缸波形-----------------------------------------232.26电子点火初级单缸波形---------------------------------------242.27点火正时及参考信号-----------------------------------------24 第 3 章故障案例-------------------------------------------------283.1 喷油器电阻过大---------------------------------------------283.2 喷油器供电线路虚接-----------------------------------------283.3 喷油器线圈短路---------------------------------------------29 致谢-----------------------------------------------------------30 参考文献---------------------------------------------------------31 摘要随着汽车电子信息技术的迅速发展，汽车上装用的电子设备越来越多，这就对今天的汽车故障诊断提出了新的挑战。
基于动态检测法的提升机钢丝绳在线监测系统白笠言
0引言矿井提升机是联系井下与地面的唯一通道，担负着提升煤炭、矸石，下放材料及升降人员、设备的任务。
从而，提升机运行状况不仅直接影响整个矿井的生产能力，而且还涉及到人员的生命安全，一旦发生故障，将造成巨大的经济损失及恶劣的社会影响。
目前，对提升机系统故障诊断已经有了很多检测原理和方法，如振波法、压轮-力电转换传感器法等，但这些方法都是属于静态检测方法，对实际应用带来很大的不便。
为此，本文讨论基于动态监测法原理，对提升机系统故障诊断进行实时监测。
1主井提升机振动模型研究矿井提升钢丝绳实际上是一个黏、弹性体，而不是刚体。
在提升机装载、卸载、加速、减速以及紧急制动时，钢丝绳会储存或释放能量，引起提升容器剧烈振荡，而且振动过程在箕斗或罐笼的提升运行整个循环中持续存在。
为此，本文将装载、卸载振动过程分别划分为3个阶段进行分析。
（1）装载3个阶段分别为：①箕斗下放到装载点装煤之前；②从装载点处装煤开始，直到装够额定重量的煤为止；③箕斗装煤完成但还未上提的时间段。
（2）卸载3个阶段分别为：①装满煤的箕斗上提到卸载点卸煤之前；②从卸煤点处卸煤开始，直到所装的煤卸完为止；③箕斗卸煤结束但还未下放阶段。
本文将结合九龙矿主井提升机的参数及运行情况进行分析，以装载第1阶段为例，详细地计算出在这一阶段钢丝绳的自由振动角频率、自然频率和单个传感器所受到的力，并将这一阶段定义为单自由度无阻尼自由振动，其他阶段分析过程类似。
如图1所示。
九龙矿提升机参数：绞车型号JKMD－3.25/4（Ⅱ）-7.35提升钢丝绳型号6Δ（34）－32－170－2绳尾型号8×4×9-143×24-140-甲镀图1单自由度无阻尼自由振动简化图m l=m le+m lt1（1）式中m l——装载点在装载之前单根提升钢丝绳的质量，kg；m le———装载之前箕斗和配重的质量，kg;m lt———装载点处尾绳的质量，kg;n1———钢丝绳条数。
Fault diagnosis of rotating machinery based on auto-associative neural networks and wavelet transfor
ARTICLE IN PRESS
982
J. Sanz et al. / Journal of Sound and Vibration 302 (2007) 981–999
that are not exactly synchronous with the gear of interest. The use of the residual signal obtained by removing the regular gear meshing harmonics from the TSA is also being frequently used as a fault diagnosis technique [4,5]. The resulting residual signal contains essentially the portion that is caused by the gear fault and geometrical irregularity.
A time–frequency analysis offers an alternative method to signal analysis by presenting information in the time–frequency domain simultaneously. The method known as short-time Fourier transform (STFT) and proposed by Gabor [6] is probably the most widely used time–frequency representation. The characteristic feature of the STFT is the application of the Fourier transform to a time varying signal when the signal is viewed through a narrow window centred at a time t. In this way, the frequency content is obtained at time t and in any other time if the process is repeated. The resolution depends on the size of the window, and as it is constant, a high resolution in time and frequency cannot be obtained simultaneously. So, the window must be chosen for locating sharp peaks or low frequency features and, therefore, its resolution is often unsatisfactory.
基于贝叶斯网络的TSW2500_型短波发射机故障诊断方法
第4期2024年2月无线互联科技Wireless Internet Science and TechnologyNo.4February,2024作者简介:石珺磊(1985 ),男,工程师,本科;研究方向:短波发射机及天馈线系统㊂基于贝叶斯网络的TSW2500型短波发射机故障诊断方法石珺磊1,张军磊2(1.国家广播电视总局九五一台,河北石家庄050400;2.石家庄水投集团,河北石家庄050051)摘要:文章提出了一种基于贝叶斯网络的短波发射机故障诊断方法,通过采集TSW2500型短波发射机的故障数据,并结合已知的故障信息和系统知识构建故障诊断模型,以快速㊁准确地识别可能出现的故障㊂实验结果表明,与传统方法相比,基于贝叶斯网络的方法能够更快速地识别故障并提供有用的诊断信息,具有较高的可行性和准确性㊂这种方法在实际应用中有望帮助工程师迅速定位和解决短波发射机故障,提高设备可靠性和维护效率,具有重要的应用价值㊂关键词:贝叶斯网络;短波发射机;故障诊断;数据采集中图分类号:D26.4㊀㊀文献标志码:A0㊀引言㊀㊀在故障诊断领域,贝叶斯网络可用于建立故障诊断模型,通过分析历史故障数据,实现对未来故障的预测和识别㊂TSW2500型短波发射机是广播通信领域的常用设备,其性能和可靠性直接关系到通信质量㊂然而,由于发射机复杂的结构和运行环境,故障诊断和维修一直是一个难题㊂传统的故障诊断方法依赖于人工经验和故障模拟试验,具有不确定性和局限性㊂针对这一问题,本文提出一种基于贝叶斯网络的TSW2500型短波发射机故障诊断方法,利用贝叶斯网络的学习和推理能力,分析和学习历史故障数据,建立准确的故障诊断模型㊂通过对模型的分析和处理,可以快速地识别和预测发射机故障,为维修人员及时提供故障解决方案[1]㊂1㊀基于贝叶斯网络故障诊断方法的设计1.1㊀基于贝叶斯网络采集的故障数据㊀㊀本文使用基于贝叶斯网络的方法解决了TSW2500型短波发射机系统的复杂特点㊂通过建立故障诊断模型,可以准确地确定故障原因和位置,并提高通信支持水平㊂利用贝叶斯网络,可以科学有效地推理不完整或不准确的信息,并以可视化的网络展示结果㊂贝叶斯网络是一种有效的不确定信息表示方法,在故障诊断㊁预测和数据挖掘等领域有着重要应用[2]㊂本文采用数据采集卡来收集TSW2500型短波发射机的故障数据,其中传感器将设备工作参数传送给上位机,进行预处理和显示㊂通过传感器采集的故障数据进行分析㊁存储和显示,具体流程如图1所示㊂依据图1所示的流程进行数据采集,TSW2500型短波发射机的传感器硬件部分采用多路信号采集模式,但每次只能测量2个信号㊂为了解决这个问题,在软件处理过程中使用了多个数据处理器,确保每次图1㊀故障数据采集流程只有2个信号被同时输入发射机的示波器㊂可以根据信号的特性设置采样参数[3]㊂1.2㊀构建TSW2500型短波发射机的故障诊断模型㊀㊀在设计TSW2500型短波发射机的故障数据库过程中,数据处理是关键㊂数据知识处理是将各种专家知识按规则转换为计算机可理解的形式,以满足专家推理和引擎要求㊂故障诊断数据库的设计如图2所示[4]㊂从图2可以看出,按照以上所述知识的表示方式,可以根据配置和部件模块之间的连接工作性能㊁工作特殊性和工作时间等实际条件,来构建TSW2500型短波发射机故障诊断的知识库㊂目前,大多数的发射机故障诊断方法都是采用固态功放的输出方式,采用高电压供电的方法对其进行分类[5-6]㊂当TSW2500型短波发射机发生整流桥故障时,会产生1个2.50s 的高电平脉冲信号㊂这个故障会㊀㊀34图2㊀故障诊断数据库导致发射机无法发射信号,造成关机㊂如果在超过2.50s后重新开机,故障将被清除,发射机将恢复正常工作㊂而如果故障未被清除,发射机将无法继续正常工作㊂除此之外,当TSW2500型短波发射机发生其他故障时,经过一系列步骤进行信号处理,最终生成短波信号,可以实现通信功能㊂该单元的脉冲信号特征的表达公式如下所示:S=ðɕn=-ɕα(t-nQ)(1)公式中,S表示短波发射机的单元脉冲信号的特征,t表示故障发生的时刻,α表示周期函数,nQ表示第n个短波信号产生的脉冲序列㊂提取短波发射机的故障特征后,需要对发射机发生故障的位置进行电压检测[7],从而对短波发射机的故障进行诊断㊂设定数据集合{x1i|x1i=1,2, (30)X-1}中的平稳信号为x iε,故障信号为x iϕ,当x iε㊁x iϕ处于同一节点时,x iε为x iϕ最左侧的平稳节点时,则x iε=x iϕ㊂在x1㊁x iϕ-1㊁ϕ-13个节点处,可以形成3点平衡的关系,其表达公式如下:x iε=[x iϕ-1,x iϕ]max(2)公式中,x iε和x iϕ-1㊁x iϕ2节点的最大值形成最大等式关系㊂则x iε任意节点的短值计算公式如下:x iεᶄ={x iϕ-1,x iϕ,x iϕ+1}(3)公式中,x iεᶄ代表短值信号的特征,x iϕ-1㊁x iϕ㊁x iϕ+1为不同形式的故障信号㊂当x iεᶄ通过短值滤波后消失,形成一个 零点 ,进而不会对平稳信号产生影响,则x iεᶄ可以作为短值滤波对短波发射机的故障诊断的约束条件,其故障信号的变化会逐渐平稳㊂基于上述分析,故障检测模型的表达公式如下:P=ðɕn=-ɕʏɕ-ɕX(η-δ)㊃νd xL(4)公式中,η代表冲激脉冲序列的强度系数,ν表示理想谐波,δ代表频率系数,L表示故障位置的特征㊂针对TSW2500型短波发射机,其音频输入端的信号频率应与工作频率保持一致㊂在故障的诊断时,需要将输入端信号频率与工作频率区分开来㊂检测结束后,通过理想低通滤波装置恢复原始音频信号,以确保音频信号的完整性[8]㊂1.3㊀实现短波发射机故障类型的诊断㊀㊀本文使用虚拟仪器技术实现仿真,故障诊断中心处理用户指令并通过驱动程序向硬件发送指令㊂利用状态监控平台对短波发射机进行实时监控,并提供反馈以辅助观察和判断[9]㊂通过软件监测与诊断模块,在LabVIEW的基础上对发射机进行自检㊂该监控平台可实现对TSW2500型短波发射机的实时监控和诊断㊂发射机状态监控和故障诊断中心主要监控参数,并将结果存入数据库㊂若无法侦测某一特殊参数,需在指定范围内实时监测,并对异常情况发出警报并记录于数据库㊂如表1所示为TSW2500型短波发射机的故障类型分析数据[10]㊂表1㊀短波发射机故障类型故障位置故障模式故障原因电源单元无直流电源输出整流器件坏数字信号处理单元射频无输出上变频器件坏前置放大单元模块射频无输出功率放大器坏谐波滤波器某一波段无功率输出继电器故障配电模块发射机无显示配电模块无电源输出检测单元部分频率点不能正常调谐器件损坏在故障诊断过程中,需要尽可能减少检测时间,以确保实时获取数据㊂由于短波通信发射机以电流和电压为主要标志,在使用换能器时通常会采用变压器和电流互感器㊂一般情况下,可以通过直流电流传感器和交流电流传感器来防止或避免对短波发射机传输电缆产生影响,并收集单相交流电流信息进行处理,以判别发射机故障类型㊂2 实验测试与分析2.1㊀实验准备㊀㊀为验证本文提出的基于贝叶斯网络的TSW2500型短波发射机故障诊断方法的使用效果,将其与传统发射机故障诊断方法进行实验对比测试㊂此次测试的供电环境参数设置如表2所示㊂表2㊀实验参数设置发射机单元输出电压DC/V输出电平p-p/VZBFSJ110.020.0ZBFSJ220.021.0ZBFSJ330.021.0ZBFSJ440.022.0ZBFSJ550.022.5ZBFSJ660.023.0ZBFSJ770.024.0442.2㊀实验结果与分析㊀㊀根据上述实验准备,将本文提出的贝叶斯网络算法与传统方法进行对比,在7个短波发射机单元中进行故障诊断,将2种方法正确识别故障的时间作为实验结果并进行测试,结果如表3所示㊂表3㊀实验结果发射机故障单元本文方法的故障诊断所需时间/s传统发射机故障诊断方法的故障诊断所需时间/sZBFSJ1 3.64 5.92 ZBFSJ2 3.07 6.74 ZBFSJ3 2.94 5.39 ZBFSJ4 2.68 6.84 ZBFSJ5 1.527.51 ZBFSJ6 3.59 6.19 ZBFSJ7 2.177.06根据上述测试结果可以看出,基于贝叶斯网络的故障诊断方法能够更快地识别出TSW2500型短波发射机中出现的故障,与传统的故障诊断方法相比,本文方法具有更高的准确性和可靠性㊂3 结语㊀㊀基于贝叶斯网络的故障诊断方法,可以准确地识别和预测TSW2500型短波发射机的故障,提高其可靠性和稳定性㊂该方法利用贝叶斯网络的学习和推理能力,通过分析和学习历史故障数据,建立准确的故障诊断模型㊂通过对模型进行分析和处理,能够快速识别和预测发射机故障,并为维修人员及时提供解决方案㊂实施该方法需要大量的历史故障数据和专业知识支持㊂未来的研究可以进一步探索如何利用深度学习等先进技术提高故障诊断的准确性和效率㊂同时,还可以对发射机的其他方面进行深入研究,如可靠性和维护性等,以提供更多的支持于发射机的设计和改进㊂参考文献[1]段安民,徐皓,孙卫华,等.基于贝叶斯网络的短波发射机故障诊断研究[J].舰船科学技术,2022 (9):142-145.[2]石星.小功率短波发射机开机保护故障判断和维修[J].电子元器件与信息技术,2023(2):157-159,167.[3]阿米娜木㊃于苏普喀迪尔,建良.关于LF2001型短波发射机几种罕见故障案的案例分析及探讨[J].电子制作,2022(24):86-88,38.[4]潘会兰,梁高峰,汪洋.DF100A&418E/F系列短波发射机谐波滤波器调测方法探究[J].现代信息科技, 2022(20):46-49.[5]安建慧.TSW2500型短波发射机远程自动抄表与回放系统的设计与实现[J].广播电视信息,2022 (8):73-75.[6]阿尔孜古丽㊃拜合提.中短波广播发射机间的电磁干扰解决措施探析[J].现代工业经济和信息化, 2023(3):259-261.[7]魏家军.基于PLC控制的DF100A型短波发射机电机运行维护及故障报警控制系统的设计[J].广播电视信息,2021(3):83-87.[8]姚雨杉.解决中短波发射机之间的电磁干扰问题对策探讨[J].电子元器件与信息技术,2021(2):66-67.[9]李明嘉,钱鸿,周爱民.基于演化序搜索的混合贝叶斯网络结构学习方法[J].计算机科学,2023(10): 230-238.[10]佟维妍,林昊,刘宇征,等.基于模糊贝叶斯网络的复杂系统可靠性分析研究[J].制造业自动化,2023 (6):148-153,174.(编辑㊀沈㊀强)Fault diagnosis method of TSW2500shortwave transmitter based on Bayesian networkShi Junlei1Zhang Junlei21.951of the National Radio and Television Administration PRC Shijiazhuang050400 China2.Shijiazhuang Water Investment Group Shijiazhuang050051 ChinaAbstract This article proposes a Bayesian network-based method for diagnosing faults in shortwave transmitters.The method collects fault data from the TSW2500model and combines it with known fault information and system knowledge to quickly and accurately identify potential faults.Experimental results show that compared to traditional methods the Bayesian network-based method can identify faults faster and provide useful diagnostic information with higher feasibility and accuracy.This method is expected to help engineers quickly locate and resolve faults in shortwave transmitters improving equipment reliability and maintenance efficiency and has important practical applications.Key words Bayesian network short-wave transmitter fault diagnosis data acquisition54。
成都电子科技大学优秀工学硕士学位论文
Major: Author: Advisor: School:
Instrument Science and Technology Li Wei Prof. Tian Shulin School of Automation Engineering
独创性声明
本人声明所呈交的学位论文是本人在导师指导下进行的研究工作 及取得的研究成果。据我所知，除了文中特别加以标注和致谢的地方 外，论文中不包含其他人已经发表或撰写过的研究成果，也不包含为 获得电子科技大学或其它教育机构的学位或证书而使用过的材料。与 我一同工作的同志对本研究所做的任何贡献均已在论文中作了明确的 说明并表示谢意。
摘要
摘 要
相控阵技术起源于 20 世纪 30 年代。与传统机械扫描雷达相比，作为新体制 电扫描雷达的代表，相控阵雷达由于其卓越的技战术性能，从诞生之日起就受到 世界各国的高度重视，广泛应用于几乎所有类型的军用雷达中。随着雷达工作电 磁环境的日益恶化、辐射单元数量的与日俱增，作为无线通信系统的核心收发器 件，相控阵天线在日常工作过程中不可避免地会产生通道故障或单元失效现象。 众所周知，现代相控阵天线阵面通常含有大量的辐射单元，如何通过测量相控阵 天线的辐射特性，利用一定的故障诊断算法，定位故障单元的具体位置并判断故 障类型，进而为后续的决策维修提供有效的指导，仍然是一项较为复杂的技术难 题，对相控阵天线的测试与诊断技术提出了迫切的需求。 为了适应这一需求，本文在查阅大量国内外文献的基础上，提出了一种针对 相控阵天线单元进行实时故障诊断的方法。该方法以前苏联专家提出的换相测量 法作为蓝本，通过引入沃尔什函数，形成特殊的实验步骤矩阵，从而有效简化了 测量方程的求解。在中场测试距离下，位于相控阵天线内各通道的所有移相器相 位切换时，探头同步获取不同配相状态下的合成电磁波信号的幅度和相位信息。 通过联立求解由测量方程与由移相器先验信息所构成的补充方程，可复原任意配 相状态下各通道的激励幅相值。继而利用故障诊断算法对施加激励与复原激励进 行差异性分析，即可实现单元移相器的故障定位与类型识别，同时可以重构包括 天线方向图在内的辐射特性。本文主要工作如下： 1. 研究了相控阵天线阵面的故障模式及其对天线辐射特性的影响。研究故障 模式对天线辐射特性的影响是进行测试诊断的前提和基础。当天线阵面中的少量 单元组件失效时，对天线整体性能指标没有较大影响。如果多个单元失效尤其当 这些失效单元具有某种相关性时，其影响较为显著。通过建立相控阵天线的模型， 仿真了三种典型故障模式下天线的辐射特性，并对其变化情况进行了统计分析。 2. 研究了单元在阵中方向图的快速确定方法。单元在阵中方向图会随着单元 在阵中的位置不同有所差异。在辐射器数量众多的背景下，探索了如何快速近似 确定指定规模子阵时单元在阵中的方向图，为故障诊断技术的实施提供了必要的 先验信息。单元在阵中方向图，一方面在各通道激励的复原过程中需要使用，另 一方面在辐射特性的重构阶段也必不可少。通过理论仿真，探讨了在阵面规模一 定的情况下，如何快速确定单元在阵中方向图的方法，并给出了仿真结果。
格力变频空调故障代码（GREEinverterairconditionerfaultcode）
格力变频空调故障代码（GREE inverter air conditioner faultcode）Fault code cabinetE1: compressor high voltage protection, when 3 seconds continuous detection of high voltage protection (greater than 27KG/CM2), turn off the light box, other loads, shielding all buttons and remote control signal, the indicator light flashing and display E1.E2: indoor anti freeze protection, in refrigeration, dehumidification mode, the compressor starts 6 minutes, 3 minutes of continuous detection of T vapor is less than or equal to -5 DEG C, flashing lights and display E2, stop the compressor and outdoor fan: when the T evaporation of more than 6 DEG C, the compressor has been stopped for three minutes, the light is off, liquid crystal display recovery operation, according to the original state. Masking key.E3: low pressure compressor protection, compressor start three minutes after the start of low pressure switch signal detection, if detected 3 minutes of continuous low voltage switch disconnect, the stop lights flashing, E3 display, to remind the user of system leak.E4: high temperature protection of the exhaust pipe. After the compressor is started, the exhaust temperature is higher than 120 DEG C or the exhaust pipe temperature head is shorted (open circuit) for 30 seconds, the indicator flashes and shows E4.E5: low voltage protection (over current protection). After the compressor is running, if the current is detected more than 3 seconds for more than 25A, the indicator flashes and displays E5.E6: electrostatic precipitator protection.Characteristics and application characteristics of frequency converterVariable frequency air conditioner is a kind of air conditioner whose frequency conversion rate can change with the change of load. After starting up, if the room temperature is different from the set temperature, the high frequency and high power operation will make the room temperature reach the set temperature rapidly, which is twice as fast as that of the ordinary air conditioner. Then the body will be in accordance with the requirements of environmental temperature and humidity, automatically using low frequency and low power operation, to maintain the set temperature, avoid frequent starting of ordinary air-conditioning, temperature fluctuation phenomenon, people in a comfortable environment.GREE inverter air conditioning uses a new concept design fan air duct system, greatly reducing the noise of the whole machine. The technology of frequency conversion and fuzzy control adopted by the utility model can automatically adjust the running state according to the change of the environmental temperature, and carry out the energy saving operation with the best output powerDigital temperature sensor.Built in microcomputer temperature sensing device, can accurately perceive the temperature difference of 0.5 degrees celsius. The temperature exceeds a predetermined value, and the digital control system responds immediately to a constant optimum room temperature.Low temperature heating is strongGREE air conditioner outdoor machine adopts two gear motor and electric auxiliary heat device, unit heating high wind open, effectively improve the unit heating capacity, to avoid the disadvantages of insufficient heat in winter, average heat 600 900W higher than the cooling capacity.Digital DC variable frequency compressorGREE digital DC frequency compressor will, according to the change of temperature to adjust the frequency of the compressor is always in the best state power output; adopting digital control, greatly improve the efficiency of refrigeration (heating), reached the set temperature than ordinaryair-conditioning more than twice as fast, saving 30 more than%.Advantages of variable frequency air conditionerFirst of all, energy saving, because of the use of variable frequency control technology, to avoid unnecessary waste of electricity. Because the traditional air conditioner controls the operation of the motor on / off mode, it needs to consumea great deal of energy in the process of starting the compressor motor. But the frequency conversion type compressor depends on the room temperature and sets the temperature and so on parameter to control the rotational speed, when the room temperature achieves the request comfortable temperature, has maintained the low frequency movement, will not stop immediately. Thus, it ensures that the air conditioner does not cause additional energy loss when the compressor is started frequently.Unit is different from ordinary air conditionersDifferent from the traditional air conditioning inverter air conditioner, mainly through the processing of power frequency inverter, the power supply frequency of the compressor can be changed, so that the compressor motor speed change, to control the compressor exhaust control, refrigeration, air conditioning to truly achieve energy-saving effect.In addition to the function of the refrigeration and heating system and the working principle of the inverter type air conditioner, the control system and compressor are different from the general air conditioner, and the frequency converter is added. And the frequency conversion compressor adopted,In the process of operation, the frequency converter is controlled by the frequency converter of the control system, and the refrigerating or heating capacity of the air conditioner will change with the change of the compressor speed. Variable frequency compressor and general compressor is different, it can be arbitrarily high and low speed operation,so that the displacement of the compressor has been effectively changed and controlled. Such compressors are mostly scroll type, double rotor, rotary and other efficient compressor.Installation precautions and maintenance technology of inverter air conditionerInstallation notesThe outdoor unit must be installed in a well ventilated position, otherwise it will cause the air conditioner to work at low frequency. The compressor and the frequency module are easy to be protected frequently, which will lead to the shutdown (the current is too large).A connecting line with wire clamp, pull off, avoid breaking wires, connectors must be reliable, otherwise, easy to cause the fire caused by fire accidents; the signal line will lead to poor contact can not afford to open the machine or compressor frequent stop.All inverter air conditioners must be reliably grounded, and the user's home has no ground wire. When the air conditioner does not have ground wire, the common fault is frequent opening and stopping, and the work is unstable.When all inverter air conditioners are installed, both internal and external machines must be connected by frequency converter.Maintenance Note (part of the control element fault analysis)Interior partEnvironmental temperature package open circuit: when the whole machine does not start or stop when it is cooled, it works normally when heating, and has been operating at high frequency.Open tube temperature package: split machine and light box Guiji, prone to work 6 minutes to 10 minutes to stop the machine, LCD will display E2 and stop outdoor machine.Tube temperature package short circuit: refrigeration is no freeze protection, the external machine does not start; heating is no protection against high temperature; the whole machine stops working.When the resistance value of all temperature control elements is deviation, the frequency will always appear high frequency, no down frequency or low frequency. (some temperature control elements are normal in the case of no power. It is better to check through electricity.)Outdoor partCompressor overheating protector, when it appears to protect: stop the outdoor unit, the external machine board indicator light flashing and long time can not open the machine.Outdoor defrosting pipe temperature head open circuit: Refrigeration normal: heating time will be 45 minutes defrost, 10 minutes defrost, cycle.Outdoor defrosting pipe temperature short circuit: refrigeration, heating will not work.When the outdoor environment is warm and the temperature is open, the operation of the air conditioner is not affected.Outdoor environment temperature sensitive package short circuit: when the refrigeration is not affected, the heating time and space transfer has been low frequency, the frequency does not rise.The air conditioner runs at a high frequency (without decreasing frequency) when the exhaust air temperature of the compressor is opened.When the compressor exhaust air temperature is short, the refrigeration and heating can not be started.Fault diagnosis and replacement of frequency converter (module)After measuring whether a DC voltage of about 300V between p+, nCheck whether the +5V and +12V outputs are normal. You can find some measurements on the outside machineIf normal (a) and (b) are normal, check whether the U, V, and W phases have balanced AC outputs. (to check the voltage between the U, V and W, it is better to remove the compressorconnections)If normal (a), (b) and (c) are normal, check that the compressor coil resistance is normal. (compressor three terminal resistance is equal, resistance value should be 1~3 ohms)When the module is changed and installed, it must be coated with cooling cream. The screws should be tightened evenly and pressed against the heat sink. Otherwise, the temperature is too high, the modules will be protected frequently, and the compressor will stop frequently.The 10 communication lines between the outdoor master board and the module must be carefully inserted, and the +5V and +12V on the controller are all output by the module. 10 communication lines, three of which are: ground, +5V, +12V, and the other 7 are data lines, which need to be tested by oscilloscope.Outdoor unit two rectifier bridge, one for the 220V input, output 300V DC to module p+ and N at both ends: the other is a half wave rectifier filter function.The reactor is a conducting coil, and it is only necessary to check that both ends are connected.Four, frequency converter common faultsThe outdoor unit does not workAfter the boot, check whether the outdoor machine 220V voltage, if not, please check whether the indoor and outdoor machineconnection is correct, the indoor machine board wiring is correct, otherwise change the indoor machine board.Please check the transformer as the electric buzzer doesn't turn off.Such as outside the machine with 220V voltage, check the motherboard red light is on, otherwise check the machine connecting line is loose, the power module p+, N whether there is a DC voltage 300V, if not, then check the reactor, rectifier bridge and wiring. If there is, but the board indicator does not light, check the power module to the motherboard signal cable (total: 10) whether loose or poor contact, and then not, replace the power supply module, module replacement, between the radiator and the module must be coated with thermal grease.If the outdoor machine has the power, the red light machine does not start, you can check is internal and external communication, (inspection methods: press "TEST" button, the indoor machine observation indicator), any kind of flashing lights as normal, otherwise the communication problem; check inside and outside machine connecting line whether the flat line special, or replacement. If the communication is normal, please check the inside and outside machine, whether the temperature package is open or short circuit or the resistance is not normal, whether the overload protector terminal is connected or not. The above two methods can not be solved, then replace the outdoor controller.Such as the boot 11 minutes or so, and can not start, please check the temperature of the indoor tube temperature packageis open; if after the boot and then start, the fan does not start, check the indoor, exogenous temperature head is short-circuit.The air-conditioner has been running at low frequency since it is turned onPlease check the indoor pipe temperature, outdoor environment, compressor and defrosting, whether there is an open circuit or short circuit, the abnormal resistance.P board fault code frequency Guiji and SolutionsE1: compressor current is too large, compressor overheating, exhaust temperature is too high, module protection, overload protector, whether short-circuit, compressor temperature sensitive package is short circuit.E2: indoor machine evaporator antifreeze protection, check whether the temperature sensor package open, can be excluded.E3: room temperature, temperature sensitive package, short circuit or open circuit.E4: interior tube temperature sensitive package open or short circuit.E5: indoor and outdoor communication failure, check whether the indoor and outdoor connection line is wrong (zero, FireWire can not be connected): the signal line and the control connection, whether the socket is loose or not, and whether the controller is damaged or not.During heating, the chamber does not workPlease check whether the power line (inside and outside machine) is correct, whether the power line is grounded or not. If the above is normal, replace the main board of the indoor unit.Troubleshooting guide for variable frequency air conditioner seriesGREE, a 2000 frequency converter and frequency conversion cabinet seriesThe reasons why the outdoor unit does not start are: the power module is bad, the communication between the indoor and outdoor is not normal, the indoor temperature fault is outside, the compressor overload protector is open, the pCB board is broken, and so on, and the concrete judgment and treatment method are also given:The bad: the power supply module after power supply module p+, first check whether there is a N 310VDC voltage, if no, please check the outdoor machine main circuit rectifier, reactor and capacitor is faulty, the wiring is loose, and check pTC resistance is bad, pTC resistance is normal, its resistance is 30 ohms to 60 ohms, open circuit or short circuit is not normal. If there is a 310VDC voltage, and the outdoor machine pCB in the red indicator light does not shine, please check the pCB signal between the ten core board and the power supply module connecting line is good contact, the power module on the needle base pin is bent, such as above normal, while the pCB officeon the red light is not bright, that the power module is bad. If the red indicator light is on, the compressor does not start, the compressor U, V, W three lines unplug, boot, the external air function normal operation (more than three minutes), also shows that the power supply module is bad.Note: when changing the power module, be sure to heat the radiator evenly on the power module and radiator.Indoor, exogenous temperature package failure (failure phenomenon does not start or after a period of time to stop): frequency converter control and indoor, exogenous temperature package are related.Cooling and pumping mode: if there is a stop phenomenon for a few minutesPlease check if the temperature in the pipe is faulty or the temperature is too low to prevent cold protection.Please check whether the outdoor temperature sensing package is faulty or the temperature is too high. When the T is 65 degrees centigrade, the compression opportunity will be stopped and the operation will be resumed when T is less than 58 degrees centigrade.Is the exhaust temperature sensor too high or faulty?. T row > 115 degrees, stop compressor, T row less than 92 degrees to resume operation.Whether the overload protector is open or the compressor isoverloaded or out of contact or out of order.(2) heating model:If the compressor does not start, the fan always does not turn, is a communication failure, please check the indoor connecting line is correct, whether completely reliable grounding.Indoor tube temperature protection, when the T tube is 65 degrees, stop the compressor; T < 52 DEG C to resume operation, the processing method can first Guan Wen bag pull out or change the new temperature sensitive package.Exhaust temperature protection and overload protection with refrigeration.3, indoor and outdoor other fault displayLED1 the light is green and the compressor is stopped and faulty;LED3 yellow light, outdoor environment, temperature package failure;LED2 the red light is on, and the outside pipe feels warm and the temperature is faultyLED1 the green light flashes and the module has a protection signal;LED2, LED3 red light, yellow light flashing, compressoroverload protection signal;The green light, the red light and the yellow light are all bright at the same time, and the exhaust temperature sensor is faulty;Indoor D1 light, compressor operation;Indoor D2 light, communication work properly, otherwise it is not normal;Indoor D3 light, interior temperature package failure.Introduction of power moduleSignal line function introductionTake the red line of the signal line as the baseline. It is line one. The lines are as follows:Line 1: negative phase of W phase control signal; line 2: positive control signal of W phase;Line 3: negative phase of V phase control signal; line 4: positive control signal of V phase;Line 5: U phase negative end control signal; 6 good line: U phase positive end control signal;Line 7: ground wire; line 8: +5V line; line 9: +12V line;Line 10: module protection signal line;Two 、 module protection type: when the power module has over / over current (short circuit) and under voltage protection, the power module has microsecond output signal.GREE accessories model identification1. Identification of temperature sensitive package and tube temperature packageGREE began using a temperature range of 15K/25 degrees 20K/25 degrees celsius. The temperature sensing head with a white heat shrinkable sleeve about 20mm on a black word, such as the temperature of 15K/25 DEG C (15K), such as on the temperature of 5K/25 DEG C hit (5K); if you do not see, look at the terminal can also identify the white color, the terminal is 5K/ DEG C, red the terminal is 10K/25 Deg. C (such as the four core temperature, environment temperature of 5K/25 DEG C, temperature package for 10K/25 C), the black terminal is 15K/25 Deg. C, C 20K/25 (15K for the environmental temperature, 20K tube tube temperature package), yellow terminal is 50K/ DEG C; at the same time the temperature of 10K/25 DEG C sense lead to blue, the temperature of 50K/25 DEG C sense leads to black, 5K/25 C. 15K/25 DEG C. 20K/25, the lead of temperature sensitive package is dark black. (two core temperature sensitive package terminal is XH, usually environment temperature sensitive package, terminal is EH, generally pipe temperature sensitive package) Just a little attention can be identified.。
结合FTU实现配电网故障诊断的行波定位方法
随着配电网电力系统的应用和规模的不断扩大，电网中输电线路的负荷量呈剧烈增加的趋势[1-3]。
其中电网配电网系统中线路故障测距技术对于配电网的安全稳定运行发挥着极其重要的结合FTU 实现配电网故障诊断的行波定位方法顾健1†，郭元萍1，李波2（1.贵州电网有限责任公司毕节供电局，贵州毕节551700；2.贵州电网有限责任公司毕节市郊供电局，贵州毕节551700）摘要：针对电力配电网系统的故障情况，提出了FTU 采集单元与行波定位法相结合的配电网接地故障定位诊断方法。
通过FTU 采集单元对配电网线路中的故障信号进行采集，实时获取配电网系统中的不同监测节点的暂态电压和暂态电流数据，并通过A/D 转换单元将采集到的原始故障波电压、电流模拟信号转换成数字信号，计算机处理系统利用行波定位方法对接收到的数字信号进行分析计算，利用EMD 算法分析出信号中的模态混叠现象和端点效应，采用VMD 算法对获取的故障信号分解，通过该方法大大减少配电网故障信号中的伪分量，有效地去除信号噪音，再利用行波定位公式计算配电网故障位置，得出故障信息。
实验数据表示，设计的配电网故障诊断方法误差较小。
关键词：电力配电网系统；FTU 采集单元；A/D 转换单元；EMD 算法；VMD 算法中图分类号：TM63文献标识码：ATraveling Wave Positioning Method for Fault Diagnosisof Distribution Network Based on FTUGU Jian 1†，GUO Yan-ping 1，LI Bo 2（1.Bijie Power Supply Bureau ，Guizhou Power Grid Co.，Ltd.，Bijie ，Guizhou 551700，China ；2.Bijie Suburban Power Supply Bureau ，Guizhou Power Grid Co.，Ltd.，Bijie ，Guizhou 551700，China ）Abstract ：Aiming at the fault condition of the power distribution network system ，a fault diagnosis method for the grounding faultof the distribution network based on the FTU acquisition unit and the traveling wave positioning method is proposed.The FTU acquisi -tion unit collects the fault signals in the distribution network line ，and obtains the transient voltage and transient current data of differ -ent monitoring nodes in the distribution network system in real time ，and collects the original faults through the A/D conversion unit.The wave voltage and current analog signals are converted into digital signals ，and the computer processing system analyzes and calcu -lates the received digital signals by using the traveling wave positioning method.The EMD algorithm is used to analyze the modal aliasing phenomenon and the endpoint effect in the signal.The VMD algorithm is used to decompose the acquired fault signal.This method greatly reduces the pseudo component in the fault signal of the distribution network ，effectively removes the signal noise ，andreuses it.The traveling wave positioning formula calculates the fault location of the distribution network and obtains the fault informa -tion.The test data indicates that the error of the distribution network fault diagnosis method designed is small..Key words ：power distribution network system ；FTU acquisition unit ；A/D conversion unit ；EMD algorithm ；VMD algorithm收稿日期：2019-07-05作者简介：顾健（1986—），男，贵州毕节人，本科，工程师，研究方向：电力调度运行管理技术。
循环冲击作用下缺陷飞机液压管路微泄漏故障诊断
doi：10.11832/j.issn.1000-4858.2020.07.005循环冲击作用下缺陷飞机液压管路微泄漏故障诊断郭长虹1,高静1,王阔强1,李涛3,权凌霄1!2(1.燕山大学机械工程学院，河北秦皇岛066004；2.燕山大学河北省重型机械流体动力传输与控制实验室，河北秦皇岛066004；3.中国商飞上海飞机设计研究院，上海201210)摘要：带有表面微缺陷的飞机液压管路在循环压力冲击作用下会逐渐形成非贯穿裂纹，但是随着裂纹处应力的循环作用，会扩展成贯穿性微裂纹，导致管路产生微泄漏。
基于管路内瞬变流数学模型，采用AMESim软件对循环压力冲击载荷下飞机液压管路泄漏故障进行模拟，分析不同泄漏情况对管路压力信号波形的影响；然后根据管路压力信号波形的仿真结果，通过Daubechies小波系中5类小波函数分别对其进行小波分解，分析并对比小波函数的分解结果，选取最优小波函数分解结果定位信号中奇Ct位置；最后开展管路泄漏故障实验，通过最优小波函数和负压波定位法检测及定位管路泄漏位置。
研究成果为飞机液压管路泄漏故障诊断提供新的方法，同时，为我国飞机液压管路视情维修模式进行探索。
关键词：循环压力冲击；管路泄漏故障；小波变换；负压波法中图分类号:TH137文献标志码:B文章编号：1000-4858(2020)07-0028-08Fault Diagnosis of Micro-leakaar in Hydraulic Pipeline of DefectedAircraft Under Cyclic ImpvrtGUO Chang-hong1,GAO Jing1,WANG Kuo-qiang1,L【Tao3,QUAN Ling-xiao1，2(1.School or Mechanicol Engineering,Yanshan University,Qinhuangdao,Hebri066004；2.Hebri Provincial Key Laboratoro of Heavy Machinero FluiO Power Transmission and Control,Yanshan University,Qinhuangdao,Hebri066004；3.China COMAC Shanghai Aircraft Design and Research Institute,Shanghai201210)Abstract:Mcro-defects on the surfacc of the aircraft hydrauUy lines con gradua l y form non-penetrating cracks undeoehecDclscpoe s uoe,buewseh ehecDclscaceson otseoe s,coacksmaDetpand sneopeneeoaesngmscoo-coacks, resulting in micro--eakage of pipelines.In this paper,AMESim soawarv is used to simulate the leakage failure of aircraft hydraulir pineline under the cyclic possuo impact load,based on the mathematicol modd of transient flow in pineline.In addition,the influencc of diieet leakage conditions on the pipeline pressure signal waveform is analyzed.Then,according to the simulation results of the pineline pressure signal waveform,the waveler decomposition is coirmd out using5kinds of waveler functions in Daubechies waveler system.The decomposition results of the wavelel function are analyzed and compared.The optimal wavelel function decomposition results are selected to locale the singulaoty points in the signl.Finlly,the pipeline leakage fault expeomenl is coirmd out to收稿日期：2019-08--0基金项目：国家自然科学基金青年项目(51505410)作者简介:郭长虹(1977—),女，河北秦皇岛人,副教授，博士，主要从事材料学、流体传动与控制的科研和教学工作&detect and locate tha pipeline leakage1x0X0-through the optimal wavelel Onction and the h—Lie pressure wave location method.Results p rovide a new method for tha fault diaanosis for tha aircraL hydraulic pipeline leakage. Key words:cyclic pressure shock,pipeline leakaaa failure,simulation analysis,neyative pre s s ura wave method引言飞机液压管有多、散、乱、长、特[1],多管路和接头多，以C919飞机为例，其液压管路有1186；管路总，C919管总为884m；乱管间，且多;管遍布机身、发机吊挂和平尾等各个部位；工（大、高压、强振和多变）。
基于录波数据解析模型的电力系统故障元件诊断方法
Telecom Power Technology运营维护技术基于录波数据解析模型的电力系统故障元件诊断方法刘 荣（泰州三新供电服务有限公司泰兴分公司，江苏能够准确定位电力系统中的故障元件。
频率等确定故障发生的位置，并指导维修人员进行修复，从而缩短故障处理时间，增强电力系统的可靠性和稳定性。
通过有效识别疑似故障区域，构建故障诊断解析模型。
在此基础上，分别对实际故障特征以及期望故障特征进行判电网故障诊断；故障特征；解析模型Fault Component Diagnosis Method of Power System Based on Wave Recording DataAnalysis ModelLIU Rong(Taizhou Sanxin Power Supply Service Co., Ltd., Taixing Branch, Taizhou 故障录波器和电网元件映射表故障区域 自动识别模块故障分析 计算模块故障诊断 求解模块元件-元件 关联矩阵故障录波数据自动上传故障区域自动识别故障区域内元件拓扑结构识别录波电气量判据计算建立解析模型智能优化算法求解解析模型确定故障元件自动推送故障诊断报告图1 故障诊断算法整体结构3 故障诊断解析模型电网发生故障时，故障特征判据与电网元件的状态有着密切的关联[2]。
2023年11月10日第40卷第21期261 Telecom Power TechnologyNov. 10, 2023, Vol.40 No.21刘 荣：基于录波数据解析模型的电力系统故障元件诊断方法针对疑似故障区内的元器件，研究人员用S =[s 1,s 2,…,s n ]表示其真实运行状态，若S j =1表示该元器件处于故障状态，若S j =0表示该元件处于正常状态。
该系统中，使用C rh 表示h 类故障波特征判据，其范围为C r =[C ri 1,C ri 2,…,C riz ]，其中变量z 表示同类故障元件个数。
Fault_Diagnosis_Overview
CauseCause-Effect Diagnosis
Behavior Signature
010001010100010101010 …
Defective Circuit Tests
Comparison & Conclusion
Nodes X and Y bridged:
0 1 1 1
X
0
Y
1/0
Node X forces Y to a value of 0
Some Diagnostic Fault Models
Gate Fault Net Fault
Bridging Fault
Path Fault
Fault Simulators
©2005 David Lavo
Fault Diagnosis Overview
19
The Pass-Fail Dictionary Pass• For each fault, store only the test vector responses • One bit per vector, pass ( 0 ) or fail ( 1 ) • Total storage requirement: f × v bits • Much smaller than full-response, and often practical for even very large circuits
• A fault simulator can simulate instances of a particular fault model • Inputs: – Circuit (netlist) – Test set – Faultlist (list of fault instances) • Output: circuit response • Usually, simulates the presence of a single fault instance (“single-fault assumption”)
修正的Helmholtz方程柯西问题的四阶修正法
It is easy to know that the solution of problem (2.5)–(2.7) is u ˆ(x, ξ ) = ϕ ˆ(ξ ) cosh( ξ 2 + k 2 x). (2.8)
Note that the function cosh( ξ 2 + k 2 x) in (2.8) is unbounded as |ξ | tends to infinity for 0 < x ≤ 1. If let u ˆ(x, ξ ) is a function in L2 (R) with respect to ξ , the exact data function ϕ ˆ(ξ ) must decay rapidly as |ξ | → ∞. Small errors in high-frequency components can blow up and completely destroy the solution for 0 < x ≤ 1. That means, the solution u ˆ(x, ξ ) does not depend continuously on the data ϕ ˆ(ξ ) and the Cauchy problem (1.1)–(1.3) is ill-posed in the Hadamard sense. Now we will modify the equation (1.1) by adding a fourth-order mixed derivative term and consider the following problem − v + k 2 v + µ2 vxxyy = 0, y ∈ R, x ∈ (0, 1), y ∈ R, (2.9) (2.10) (2.11)
故障诊断技术的回顾与展望
2. Some Problems of FD Technique
2.3 Basic Tasks of Fault Diagnosis
故障诊断是一门综合性技术，其研究涉及到 多门学科，如控制理论 （经典、现代、鲁棒、 自适应）、可靠性理论、数理统计、模糊集 理论、信息处理、模式识别人工智能等学科 理论
FD任务分两步完成 残差（征兆）生成 (Residual/Symptom Generation) 残差（征兆）评价 (Residual /Symptom Evaluation)
硬件冗余（hardware redundancy）：用同样的硬件重构过程 的元部件。特点是可靠性高、故障分离直接，但成本过高
解析冗余(analytical redundancy)：与硬件冗余相对应，指通过 用解析方式表示的系统数学模型来产生冗余信号
2.1 Basic Concepts of Fault Diagnosis Technique
2.5 Classification of Fault Diagnosis Methods
国际故障诊断权威，德国的P.M.Frank教授认为 故障诊断方法可以分为
※ 基于模型的方法 (model-based) ※ 基于知识的方法 (knowledge-based) ※ 基于信号处理的方法 (Signal-processing-based)
2.5 Classification of Fault Diagnosis Methods
诊断方法的研究在于：寻找征兆与故障之间的有效对应关系 最简单的故障检测方法就是所谓界限判别法 也即判别两类过程状态(正常和异常状态)
如使用一个传感器信号x，可按如下条件描述： 如果x＜Hth，那么状态正常，否则状态异常
机电一体化英文论文
机电⼀体化英⽂论⽂An innovative digital method for the dynamic simulation ofDCelectromechanical systemsChen Chaoying a,*, P. Di Barbab, A SavinibaDepartment of Electrical Engineering, Tianjin University, 300072 Tianjin, People'sRepublic of ChinabDepartment of Electrical Engineering, University of Pavia, 27100 Pavia, Italy AbstractIn this paper, an innovative digital simulation method, named `R-K-T' method, is presented. The new methodology combines Runge±Kutta and trapezoidal methods and possesses the advantages of both of them. The errors featuring the proposed method are analysed and their correction is worked out. As a case study, the circuit model of a small DC motor, acting as the engine starter of a road vehicle, is considered;the proposed methodology is applied to carry out the dynamic simulation of the electromechanical device. The results are obtained ef?ciently and with a good degree of accuracy; in particular, the numerical oscillations are suppressed.q1998 Elsevier Science Ltd. All rights reserved.Keywords:Numerical methods; Time integration; Dynamicsystems;Electromechanics; DC motor1. IntroductionSeveral digital methods, such as Euler, trapezoidal,Runge±Kutta and linear multistep methods are generallyused to carry out numerical integration and differentiation.The Euler method is simple, but with low accuracy; its cutoff error isO(h2), whereas that of the trapezoidal methoddecrease asO(h3). The Runge±Kutta method has relatively high accuracy but requires large amount of computational work; finally, the multistep method has high accuracy, but it can not be self-started [1]. Therefore, the trapezoidal method finds widespread applications in transient digital simulations. However, in DC system simulations, the trapezoidal method often introduces numerical oscillations with equal amplitudes, so that its application in this case is critical. Since the backward Euler method can avoid such oscillations, in the literature [2], a damped trapezoidal method was proposed; this method introduces a damping factor into the trapezoidal method which effectively decreases the numerical oscillations but at the sacrifice of accuracy.After analysing trapezoidal and Runge±Kutta methods carefully, this paper presents an innovative simulation method, called `R-K-T', which combines Runge±Kutta and trapezoidal methods ingeniously. The advantages of the new method are: the Runge±Kutta method can be expressed by the companion model just like the trapezoidal method does; the numerical oscillations can be attenuated efficiently.According to frequency spectrum analysis, the errors of the method are calculated and corrected. It makes it possible to simulate DC systems accurately and efficiently.2. Numerical oscillations of trapezoidal method in DC systemsConsidering the inductive circuit shown in Fig. 1(a) the governing equation iswhere currenti is the unknown. Using the trapezoidal method for time integration, one can get:Where h is the time step of calculation.LetthenFig. 1. Inductive impedance (a) and its companion models (b) and (c).whereIts companion model is shown in Fig. 1(c).Suppose, when a DC current ¯ows through the inductive impedance.From Eq. (3) the voltage response of the inductive branch can be calculated asIt can be seen that the oscillation of voltage is undepressed.Otherwise assume, whenn?k, the current is switched off, i.e.from Eq. (3) one can get:that isThe voltage response is also an undepressed oscillation.It can be proved that the backward Euler method can avoid such an oscillation. For inductive impedance it gives:It can be seen thatun11is not dependent onun, so this makes it possible to avoid numerical oscillations but greatly reduces the accuracy of backward Euler method. To solve this contradiction, the literature [2] proposes a trapezoidal method with damping. For the differential equationit givesFor the inductive impedance shown in Fig. 1 it gives:Wherea is the damping factor (0This method turns into the trapezoidal one when a=0,and becomes the backward Euler method when a=1. From Eq. (9), it can be seen that the coeficient of un isaccording to experience: its optimum value is dif?cult to be determined.3. The R-K-T methodThe Runge±Kutta method has higher accuracy and better stability in DC systems, but it requires the calculation of the values of a function many times during asingle step; it cannot be expressed by a companion model like the trapezoidal method. If one can combine the Runge±Kutta method and the trapezoidal method to form a new method, then it will possess the advantages of both two methods. Take the 3rd order Runge±Kutta method for example, to deduce the new method. For the differential equationby the 3rd order Runge±Kutta method, one has [3]For the inductive impedance, one has:whereFrom Eq. (10) it followsWhere is the voltage at the midpoint of the step, which can be found by solving the equations of the system. But we calculate it by trapezoidal method. It can be done in two different ways (A) and (B):(A) Take the average values of un and un11 and letSubstitutingun11/2 from Eq. (14) into Eq. (13) Eq. (10) gives:Substituting the above formula into Eq. (10), one can get:whereIt is obvious that the 3rd order Runge±Kutta method with un11/2 substituted by Eq. (14) may be expressed by the companion model shown in Fig. 1(b), as for the trapezoidal method; the parameters of the model are:The distinguishing feature of this method is that the coef?-cients ofun11andunare not equal; their ratio A may be used to attenuate the numerical oscillation with equal amplitudes of trapezoidal method. It turns to the trapezoidal method when R=0,(B) TakeUsing the trapezoidal method, one has:By substituting Eq. (19) into eq, (13), one can get:By substituting the above equations into Eq. (10), it followsWhereFormula (20) may be expressed by a companion model of inductive impedance as Fig. 1(b), whereFormula (20) has also the function of attenuating the numerical oscillations like Eq. (15), and it also turns to the trapezoidal method for pure inductance when R=0.For the 4th order Runge±Kutta method, it gives:Similarly one can obtain the companion model for the 4th order Runge±Kutta method as follows [4].(A) Taking one has:WhereIts companion model in Fig. 1(b) is(B) Taking one can getwhereIts companion model for Fig. 1(b) is:Both of the 4th order models introduced above also turn to trapezoidal method for pure inductance when R?0. Thus,the Runge±Kutta method is combined with trapezoidal method to form a new `R-K-T' method, which exhibits the advantages of these two methods.4. Analysis and calculation of error for the R-K-T methodIn a real-life system, voltages and currents, whatever wave forms they may have, can be analyzed by the method of frequency spectrum. The error of simulation can be analyzed for every frequency component; the components are then added together according to the theory of superposition to obtain the total errors.Where w any one frequency component. Let us rewrite the 3rd R-K-T method (20) as follows:Substituting Eq. (27) into Eq. (28), one can deduce:From the formula of inductive impedance, one hasThe difference of the two sides of Eq. (29) represents the error of R-K-T method for frequency component, so that the error function can be de?ned as:If the exciting sources contain a number of frequency components,e(v) should be computed for every frequency component and added together. The summation of all the e(v) gives the total error function of the 3rd order R-K-T method.5. Correction of error for the R-K-T methodFrom Eq. (29) it is clear that there is unbalance in the formula of the R-K-T method for angular frequency; it is due to the method itself and not related to the exciting sources. If one would match the two sides of Eq. (29) by adding some items, then it could give the accurate result for frequency component. Letv0be the main angular frequency of the exciting source, in order to conduct accurate calculation forv0, it is necessary to transform Eq. (29) as follows:The coeficients of the two sides of Eq. (32) are equal w=w0. It means that it gives accurate results for w=w0.Restoring Eq. (32), one can deduce:whereEq. (33) is the formula of the R-K-T method after correction. If the exciting source of the system has a single frequency v0 , then correction can be made for this frequency. If the system has a multifrequency exciting source, then correction may be made for one of the dominant lower frequency which has higher amplitude.6. Numerical resultsTo check the accuracy of the method presented, the circuit shown in Fig. 2 has been considered. Its parameters are:The accurate expression of currenti is:The test circuit shown in Fig. 2 has been solved by each of the models stated above with time step h=0.1 ms, as well as by means of formula (34) giving a more accurate result.In each case error is defined as the maximum absoluteFig. 3. Error curves for each model (T: cycle) (see Table 1).Fig. 5. Error curves for each model (T: cycle) (see Table 1).Fig. 6. Error curves for each model (T: cycle) (see Table 1).Fig. 7. Error curves for each model (T: cycle) (see Table 1).Fig. 8. Error curves for each model (T: cycle) (see Table 1).Fig. 9. Error curves for each model (T: cycle) (see Table 1).Fig. 10. Error curves for each model (T: cycle) (see Table 1).difference between the result of the model considered and that obtained by Eq.(34).From Table 1 and error curves shown in Figs 3±10, the following remarks can be drawn.1. The error of the trapezoidal method is small, that of the damped trapezoidal method is bigger while the backward Euler method proves to be very inaccurate. This means that the trapezoidal method is a simple method with high accuracy.2. For either the 3rd or 4th R-K-T method, the errors are as small as for the trapezoidal method. This means that the accuracy of the R-K-T method is comparable to that of the trapezoidal method.3. The error of the corrected R-K-T method is more than halved with respect to that of the trapezoidal method.4. The steady errors of the Euler, trapezoidal and uncorrected R-K-T methods arethe sinusoidal functions, but the error of the corrected R-K-T method is the only one tending towards zero within a few cycles. This means that the corrected method can be used to eliminate ¯uctuations and to calculate with high accuracy.7. Dynamic simulation of a DC motorFor the sake of application, let us consider the circuit shown in Fig. 11. It can represent [5] the lumped parameters model of a DC motor with series excitation, that acts as the engine starter in the electromechanical equipment of a road vehicle. The inductance parametersLa, Lf, Ghave been previously identified by means of a field simulation based on two-dimensional finite elements. From the designer viewpoint it is of great concern to predict the dynamic behaviour of the device and, in particular, to estimate the peak current.When the commutator is switched on the voltage source, the governing equation of the circuit can be written as:Where Re is the source resistance, Rf is the field winding resistance,Lf is the field winding self inductance, Ra is the armature resistance,La is the armature self inductance,Gis the speed coefficient and vis the angular speed of the armature.Fig. 11. DC motor with series excitation.Where J is the moment of inertia, Kis the damping coefficient, and T0 is Coulomb damping. The initial state is zero. Eq. (35) Eq. (36) are a non-linear coupled system; for calculating them, the method of prediction and correction for speed is introduced.7.1. Simulation of the motor with linear parametersLinear parameters of the motor in Fig. 11 are listed in Table 2.Several digital simulation methods are used to compare each other. Table 3 lists some simulation results of motor current computed by Runge±Kutta±Merson [6], Trapezoidal, R-K-T and Damping trapezoidal methods, respectively (t=0.01 ms). In the Runge±Kutta±Merson method, a tolerance parameter is introduced to control the error in the integration; the time step is changeable according to the error estimation. Here the tolerance parameter is 0.1 *10^12.Being of higher accuracy, the Runge±Kutta±Merson method can be taken as the reference to compare other methods. Table 4 lists the maximum and minimum errors, respectively, of each method with respect to the results obtained via the Runge±Kutta±Merson approach.It can be seen that the maximum error of the R-K-T ,method is smaller than for the damping trapezoidal method and greater than for the trapezoidal method; but the minimum error of the R-K-T method is the smallest of the three methods. So the conclusions obtained from the previoussection have been confirmed.Fig. 12 shows some simulation results by the R-K-TFig. 12. DC motor simulation by the R-K-T method (linear case).method: (a) is the armature voltage; (b) is motor current; (c) is motor speed; (d) is the relationship between motor speed and torque. It can be seen that the results give a clear transient process and no oscillations occurred. Fig. 13 representssome results simulated by the trapezoidal method. It shows that the armature voltage of the motor exhibits some undepressed numerical oscillations.7.2. Simulation of the motor with non-linear parametersConsidering saturation of the iron of the motor, the parametersLfandGin Table 1 will be non-linearly dependent on current. To simulate these non-linear parameters, a curve fitting method has been used. Fig. 14 shows the parameterFig. 14. Non-linear behaviour of parameterLf.Fig. 15. Non-linear behaviour of parameterGLf and its curve ?tting result. The curve ?tting model is:Fig. 15 shows the non-linear parameterGand the curve fitting result. Its mathematical model is:To avoid the crest values ofLf andGproduced by curve fitting shown in the beginning of Fig. 14 Fig. 15, a piecewise linear case in the ?rst section of Fig. 14 Fig.15 shown as a dashed line is introduced in the program.Lf being nonlinear, Eq. (35) must be changed as:When each step of calculation is completed, the parameters Lf,dLf /di andGin Eq.(39) must be modified according to the new motor current; then the calculation can move on to the next step. The main effect of non-linearity can be noted during the transient process. Fig. 16 shows some simulation results obtained by the R-K-T method.It can be seen that the peak value of the motor current in Fig. 15 is higher than in the linear case shown in Fig. 12; on the other hand, the armature voltage in the non-linear model is characterized by a bit longer transient process.7.3. Simulation of the non-linear case with speed controlFor controlling the motor speed, a simple control system shown in Fig. 17 is simulated. It includes a speed measuring link with time constant of current measuring T1, a comparing link and a switch controlling link.In the control system, speed measured value is compared with speed reference to get the speed error. When the error is over the limit, it turns the switch off; otherwise, it turns the switch on. So Eq. (39) will be changed to:with switch function:Fig. 16. DC motor simulation by the R-K-T method (non-linear case).The simulation results of the non-linear case with speed control (?12 002 rad/s) by the R-K-T method are shown innFig. 18, where (a) is the voltage of the source; (b) is the voltage of the armature; (c) is the motor current; (d) is the motor speed.The results show that the R-K-T method gives clear dynamic process in DC motor simulations. This method is simple and with high accuracy. Furthermore, it caneliminate the numerical oscillations ef?ciently.8. ConclusionA digital method, called R-K-T, for the simulation of DC systems has been presented and discussed. It results from combining, in an innovative way, the Runge ±Kutta and trapezoidal methods, the positive features of which are kept. The proposed methodologyDwhose code can be implemented on a desktop personal computerDsimplifies calculations and saves runtime without loosing accuracy. In particular, the procedure of error correction worked out can actually suppress spurious numerical oscillations. The application of the method to the circuit model of a small DC motor has given good results in the simulation of the dynamic behaviour of the device.Fig. 18. Non-linear case with speed control (v?1200^2 rad/s) by the R-K-T method (T1=0.002 s).References[1] Dommell HW. Electromagnetic transients program reference manual,2001.[2] Alvarado RL, Lasseter RH, Sanchez JJ. Testing of trapezoidal integration with damping for the solution of power transient problems. IEEE Trans PAS- 1993;102(12):3783±3790.[3] Bao Xuesong. Digital method of solving ordinary differential equations. In:。
- 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)。
1 INTRODUCTIONThe reciprocating pumps are the important equipments of petroleum mine. Pump valve module usually lose efficacy as a key part of its liquid end for its bad work condition, and it will cause the severe economy loss, so it has important meaning to carry on state monitor and the fault diagnosis [1-3]. In recent years, the vibrant test technique is gradually applied in pump valve of the fault diagnosis, and accumulate much successful experience [4-6]. Its fault information mainly comes from the vibration acceleration signal which collect form pump valve. But while the pump valve of reciprocating pumps close by turn will knock on the machine body, and also exist lots of prompting vibration source, such as reciprocating vibration source and revolving vibration source, liquid knocks on machine body and the interference of various inertial unbalance force etc., these reasons cause it has a bigger difficulty to identify fault. Therefore, this text puts forward the new method which extracts fault feature vectors by the pressure signal of single pump as a system feature signal in a creative way and makes use of the intelligence fault diagnosis technique to diagnose real time and accurately.2 THE ANALYSIS OF MANY PUMP VALVES WHICH APPEAR FAULTSYNCHRONOUSLYWhen the reciprocating pump works, the instantaneousflux and non-uniform speed of the liquid in the discharged pipeline produce the acceleration and the force of inertia, increase the suction and the exhaust resistance of the pump. the discharged pipeline’s and the system’s back pressure generally can satisfy the request of the normal work and guarantee certain over measure, meanwhile canThis work is supported by National Nature Science Foundation under Grant 10550325adopt the measure which can eliminate pulse and reduce the force of inertia, for example: The actual object is the triplex list function reciprocating pump, oneself can reduce the pulse function of the instantaneous flux; reduce pipeline length, increase the pipeline inside diameter, the reduction of reciprocation number reduces the tube performance head; Establish the air chamber on the nearness import and export pipeline of the pump, in order to reduce the liquid pulse on the pipeline. Moreover, if the pump has the air chamber , the liquid non-steady flow only occurs in the section of distances between pump work room and corresponding air chamber ,and the fluid flow of the discharged pipeline and the suction pipeline outside the air chamber is stable.t guarantees theoretically when a pump valve breaks down cannot change other fluid cylinder internal pressure, in other words, when many pump valves break down synchronistically, each fluid cylinder’s pressure variation is independent, doesn’t disturb mutually.Therefore, this article needs to solve the main question. when the pump valves have the single failure, how to find the effective fault feature information from the pressure signal of pump cylinder and diagnose for it.3 EXTRACT WAVELET PACKET’SANALYSIS CHARACTERISTIC Moyal put forward the integral of wavelet transformation scope square and the signal energy is proportional,2+22-(,)()f W a b dbda a C x t dt ψ∞∞+∞∞−∞=³³³(1)So when the pump valve breaks down, the frequency component of pump cylinder internal pressure change and the corresponding energy distribution also change along with it. I t contains the rich fault information in various frequency component’s signal energy, some kind of or some several kind of frequency component’s energy change namely has represented some kind of faultThe New Fault Diagnosis Method of Wavelet Packet Neural Network onPump Valves of Reciprocating PumpsDuan Yu-bo; Wang Xing-zhu; Han Xue-songElectricity and Information Engineering College, Daqing Petroleum Institute, Daqing, 163318, P.R.ChinaE-mail: xujj@Abstract: Two key issues of fault diagnosis for the pump valves of reciprocating pump are extracting the fault featureinformation of nonstationary time variation process efficiently from system feature signals and classifying the faults feature correctly. A new method of fault feature is proposed by ordinary pressure signal (pressure in pump cylinder) as system feature signals. A diagnosis method based on “frequency-energy-fault identification” pattern recognition diagnosis approach isintroduced to the fault detection on pump valves of reciprocating pumps. The improved BP neural network is used to diagnose various faults of pump valves by the feature vectors constructed above. This approach deals with the primitive pressure signal simply and acquires fault feature vectors easily. And the pressures in different valve boxes have no influence each other. Key Words: Reciprocating Pumps, Pressure Signal, Wavelet Packet, Neural Network, Fault Diagnosis3285978-1-4244-2723-9/09/$25.00c2009IEEEsituation. Therefore, this article introduces monitortechnology of pump valve work status by the recognitionmethod of fault diagnosis pattern based on Frequency-Energy-Fault. This method doesn’t need the model structure of the system but directly uses the changeof various frequency components’ energy to manifest the fault of the equipment. Using this kind of relation establishes the mapping relations between the charge of energy and all kind of pump valve’s fault, and obtains the fault feature vectors. The steps of this method which extracts system’s fault feature vectors by wavelet packet transformation are asfollows:Step1: First carries on three wavelet packet’s decomposition to the A/D sampling signal, extracts separately eight signal characteristic of frequencyingredient from low frequency to the high frequency in the third layer. Its decomposition structure is shown in Figure 1.In Figure 1, (i,j) expresses the jth crunodes in the ith layer,thereinto, i=0,1,2,…,7,each crunodes all represents certainsignal’s characteristic. And the (0,0) crunodes representsthe primary signal S; the(1,0) crunodes represents first lowfrequency coefficient X10 which is from the waveletpacket‘s decomposition; the (1,1) crunodes represents first highfrequency coefficient X11 which is also from the wavelet packet ‘s decomposition; the (3,0) crunodes represents thecoefficient of the 0th crunodes in the 3rd layer, and the restmay be deduced by analogy.Fig1.3-layer decomposition tree structure of wavelet packetStep2˖Restructure the decomposition coefficient of wavelet packet and extract signals of all frequency bands scope. Use S30 to represent the restructuring signal of X30, S31 represent the restructuring signal of X31, and the rest may be deduced by analogy. In this article, we only carry on the analysis to all crunodes of the 3rd layer, then the general signal can express ˖3736353433323130S S S S S S S S S +++++++= (1)On the assumption that the lowest frequency ingredient is 0 and the highest frequency ingredient is 1 in the primary signal S. Then the eight frequency ingredient which we extract represent the frequency range are shown in table 1.Table1.Frequency dimensions of every factsSignal S 30 S 31 S 32 S 33Frequency range0-0.125 0.125-0.250 0.250-0.375 0.375-0.500Signal S 34 S 35 S 36 S 37 Frequencyrange 0.500-0.625 0.625-0.750 0.750-0.875 0.875-1.000Step3: Calculate the total energy of all frequency bands signal. Because the input signal is a random signal, its output is also a random signal. On the assumption that the corresponding energy for 3j S is 3=0,1, 7(j )j E ⋅⋅⋅, thenwe can get:()21233¦³===n k jk t j j x d t S E (2)There, 0,1,71,2,(=,n)jk x j k ⋅⋅⋅⋅=⋅⋅represent the discrete points’ amplitude value of the restructuring signal S3j. Step4: Structure feature vectors. Because when the system appears to break down, it will have bigger influence to the energy of signal in each frequency band, therefore, taking energy as element to construct a feature vector. The feature vector T is: 3031323334353637,,,,,,,T E E E E E E E E ªº«»¬¼= (3) When energy is bigger, 3j 0,1, 7E (j=)⋅⋅⋅is often a bigger number, it will bring some inconvenient problems in thedata analysis. Therefore we can carry on an improvement to the feature vector T, namely processing the feature vector by normalization, order212703¸¸¹·¨¨©§=¦=j jE E (4)The feature vector T ′which is got after normalization.'3031323334353637[,,,,,,,,]T E E E E E E E E E E E E E E E E = (5)Step5: Ascertain the characteristic value of feature vector and the tolerance scope when the system is normal and in various faults .we can get them by the method of mechanism analysis and the method of experiment Statistic. The method of mechanism analysis is based on the model of the system, but when the model of system is more complicated or don't know at all, this kind of method seems to be obvious deficiency; The method of experiment Statistic isn’t based on the math model of the system, so it has extremely important meaning in the engineering application, in this article ,we use the method of experiment Statistic to ascertain the characteristic value and the tolerance scope. Suppose the first element of vector’s (30E E ) characteristic value is 0C and the tolerance scope is 0C Δ, the second element of vector’s (31E E )characteristic value is 1C and the tolerance scope is 1C Δ, the rest may be deduced by analogy. So the eighth element of vector’s (37E E ) characteristic value is 7C and32862009Chinese Control and Decision Conference (CCDC 2009)the tolerance scope is 7C Δ.j C and 0,1,7()j C j Δ=" can be calculated bynxC nk jkj ¦==1(6)n is the number of the experimentIf the value of j C is bigger, we can carry on normalization to the characteristic value and order21702¸¸¹·¨¨©§=¦=j j C C (7)The characteristic value after normalization is'01235746[,,,,,,,,]T cv C C C C C C C C C C C C C C C C = (8)The tolerance scope (0,1,7)j C j Δ="is()21121¸¹·¨©§−==Δ¦=nk j jk j C x n k k C σ 5~3=K (9)There, n is the number of experiment, the tolerance scope is 3-5 times as variance.f the characteristic value of feature vector has made normalized processing, then the tolerance scope also should make the corresponding change, namely each element of tolerance scope vector should correspondingly eliminates C, so the tolerance scope TS C ′Δ is'01235746[,,,,,,,]TS C C C C C C C C C C C C C C C C C Δ=ΔΔΔΔΔΔΔΔ (10)The request for value of n: If the repeatability (or stability) of experiment data are bigger, and then the experimental number may be less; I f the repeatability (or stability) of experiment data are less, and then the experimental number may be bigger.4 THE NEURAL NETWORK IN FAULTDIAGNOSISChart of fault classified diagnosis with neural network isfollowing:Fig.2 Chart of fault classified diagnosis with neural networkSpeaking of the fault diagnosis process of reciprocating pump, the neural network input vectors are eight dimensions characteristic vector of the system which is got by using wavelet packet analysis to the pressure signal of valve box; But the output vectors are the abstract serial number which is got in different faults. This article discusses three fault of pump valve, therefore the serial number use 001,010,100.Structures training sample data based on the fault feature vectors which is got in last section, takes the initial weight separately between (-0.5,+0.5) and (-1,+1) and conducts the contrast research, validates the influence of initial weight toward the training performance of network . Sample data for training neural network is following:Table2.Sample data for training neural networkFault categoryfault feature vectors Output patternSeal packing collar fault0.0000 0.6290 0.6451 0.6276 0 0.4709 0.5163 0.54190010.0000 1.2730 1.3195 1.2865 0 0.5247 0.5324 0.55890.0000 1.8722 1.7781 1.8798 0 0.4359 0.4387 0.45140.0532 0.6268 0.4502 0.4942 0 0.4804 0.4663 0.5058Pump valve attrition fault0.0000 12.1791 1.1485 0.9190 0 0.1105 0.1333 0.31820100.0000 10.4451 1.3423 1.0895 0 0.0392 0.0655 0.31420.0000 6.9534 1.6012 1.3753 0 0.0245 0.0347 0.24790.0000 4.5924 1.0789 1.0792 0 0.0970 0.1120 0.1033Spring break fault0.0052 4.2419 0.6519 0.7022 0 0.1367 0.0646 0.08791000.0294 5.0306 1.1451 1.3974 0 0.8621 0.8318 0.56510.0106 4.0416 0.6069 1.4195 0 1.3354 1.2738 0.49070.0061 6.4665 0.7928 0.5089 0 0.1100 0.0755 0.0967Adopt the conjugate gradient BP algorithm to train the neural network. The curve of error are different in the training initial period. The big starting value’s error rate of descent is quick, and it shows the initial weight has certain influence to calculate t the error; We may see that the erroneous change curve tendency is approximately same and all can achieve convergence along with increasing the iterative number from the erroneous drop process. But, in the training time it costs a little time to train with the small initial weight. Initial weight selection in (- 1, 1) between trains The train steps is 1108 if the initial weight is in ˄-1ˈ+1˅, but the train steps is 777 in ˄-0.5ˈ0.5˅. The selection of initial weight has certain influence to the neural network system’s performance from above analysis.5 DESIGN THE FAULT DIAGNOSIS OFTHE SYSTEMThe system equipment include the 3DJ-1.5/16 triplex jar single function pour and gather pump, the Y180L-8 asynchronous motor, the 2635 electric charge amplifier, the ZY-D pressure transmitter, and the ADAM4017and 4520 data acquisition module and the VSD2000 frequency changer, and also a computer which is used to achieve data and analyse and so on.. In the experiment, The electricity eddy current sensor was installed on the cylinder body as the time signal to examine the dead center where the2009Chinese Control and Decision Conference (CCDC 2009)3287plunger approach to the power end and used its signal as the external trigger signal to gather the pressure signal in the fluid cylinder; The pressure transmitter was installed nearby the pump body to measure the change of the pump valve cylinder internal pressure and select a point in the fluid cylinder as measured point.The dimension of the fault feature is 8 and the corresponding fault is 3 according to the wavelet packet analysis ,so the input vector dimension of neural network structure is 8 and the output vector dimension is 3. According to above analysis, the neural network layer determination is 3, namely adopt three layers conjugate gradient BP algorithm and use the batch processing to train the network. The network starting value has been determined through the experimental confirmation method, namely elect 1.5 as the starting value and 0.6 as the inertia factor; The network structure is three BP network and the number of node is 19 in the concealment level, namely three network structure is 8-19-3.The main data when training neural network is shown in table 3.Table3.several data when training neural network The training step sum of the squares of errors Study speed Inertiafactor204 0.02 1.5 0.66 CONCLUSION At present this system has used Yu note in the real-timefault diagnosis system of the oil field station. The mainresearch in this article is following:1. Takes by the pressure of valve box as the characteristicsignal of the system and extracts the fault characteristicinformation of the system by the analysis method ofwavelet packet. This method has made up the insufficiencywhich can’t extract the fault information by using the vibration signal as the system’s characteristic signal when the multi-pump valves breakdown synchronously.2. Carries on the analysis and processing to the initial pressure signal by using the time domain, the frequency range and the wavelet packet analysis separately and so on, extracts the system‘s fault characteristic, and compares the effect of each method. We can see from the final result, the fault characteristic information which is got by using the method of wavelet packet analysis in this article to process the pressure signal of the valve box is effective.REFERENCES[1] Bai yun dong 㧘Tu liang yao 㧘Yang chun bao. Theapplication of time domain radial direction primary function network diagnosis method in reciprocating pump fault diagnosis[J].Vibration Engineering Journal, 2008,15(2): 162㨪165.[2] Zhang ping 㧘Wang gui zeng 㧘Zhou dong hua. Dynamicsystem fault diagnosis method[J]. The control theory and applies, 2005, 17(2):153㨪158.[3] Zhou dong hua , Wang gui zeng. The summary of faultdiagnosis technology[J]. Chemical automation and measuring appliance, 2008, 25(1):58㨪62. [4] Huang Qun-gu, Ren zhen,et al. Fault signal Analysis in Power System Based on Similar Wavelet Transforms.Journal of South China University of Technology, 2004,29(5): 5㨪9.[5] Mallat S G, Hwang W L, Singularity detection andprocessing with wavelet. IEEE Trans on Information Theory, 2002, 38(2): 617㨪643. [6] Kumamaru K, Hu J, Inoue K and Soderstrom T. Robust fault detection using index of Kullback discrimination[C]. Proc. of IFAC World Congress, San Francisco, USA,2006, 205㨪210.32882009Chinese Control and Decision Conference (CCDC 2009)。