Comparison of different DC voltage supervision strategies in a local
单介质和双介质阻挡放电低温等离子体降解甲苯的比较
文章栏目：大气污染防治DOI 10.12030/j.cjee.201906002中图分类号 X511 文献标识码 A于欣, 党小庆, 李世杰, 等. 单介质和双介质阻挡放电低温等离子体降解甲苯的比较[J]. 环境工程学报，2020, 14(4): 1033-1041.YU Xin, DANG Xiaoqing, LI Shijie, et al. Comparison of single and double dielectric barrier discharge non-thermal plasma for toluene removal[J]. Chinese Journal of Environmental Engineering, 2020, 14(4): 1033-1041.单介质和双介质阻挡放电低温等离子体降解甲苯的比较于欣，党小庆*，李世杰，黄准，郭惠，张金龙西安建筑科技大学环境与市政工程学院，陕西省环境工程重点实验室，西安 710055第一作者：于欣(1993—)，女，硕士研究生。
研究方向：大气污染物治理。
E-mail ：187********@ 摘 要 为研究介质阻挡放电(DBD)反应器结构对低温等离子体降解甲苯的影响，设计了具有单层介质和双层介质的DBD 反应器。
对2种反应器的放电特征、甲苯去除率、矿化率、CO 2选择性和能量效率进行了比较，并对施加电压和初始浓度对甲苯降解效果的影响进行了分析。
结果表明：在相同电压下，双介质反应器(DDBD)具有更高的电场强度，而单介质反应器(SDBD)的输入功率更高；当甲苯浓度和电压分别为616、1 027、1 848 mg·m −3和14~24 kV 时，双介质中的甲苯去除率为9.4%~100%、7.4%~99%、5.1%~64%，单介质为67%~98%、46%~90%、26%~59%。
这说明低电压下单介质反应器的甲苯去除率更高，而高电压下则相反，并且，浓度降低、电压升高有利于甲苯的降解。
腾讯NDA限制-腾讯NDA限制-Jacinto6Eco SoC电源解决方案-DRA72 TDA2
TPS22965 + TPS51200
8
New PDN Concept
0.40 = $3.25
(3 AVS @ 1‐2.5A, Dual 1.8/3.3V IO, DDR3L)
18 + 4 + 4.6 + 4.6 + 4 + 9 = 44
$4.42 128 $2.53 74
$7.32 190 $4.64 119
5. “PDN’s AVS Capability” is the achievable power if all AVS power rails are increased to 90% of capacity while other power rails remain at typical Use Case modelled values.
(3 AVS @ 1‐3.5A, Dual 1.8/3.3V IO, DDR3L) (similar to EVM PDN #0)
49 + 9 + 4 = 62
#8.2 – LP87524 + LP5912 + TLV713 + LP5907 +
9.92
1.26 + 0.19 + 0.07 + 0.10 + 0.10 +
2. PDN Support component (Rs, Cs & Ls) pricing from Mouser Distribution website using single 4k – 10k/reel qty costs as of May 2016. Both PDN Support & PDN Total Costs have been provided for relative comparison only, individual customer volume pricing may vary.
胸部外伤英文
Blunt chest injury
Surgical requirement not so often
Late death, mainly from circulatory respiratory insufficiency
4. Principle of diagnosis and treatment
Rib
mediastinal pendelluft
Fracture Intercostal vessel —Hemothorax ——ventilation,
lung compression, blood loss
Lung tissue —Pneumothorax —ventilation, lung
4.1 Concept of time-effect
Time-effect:
• Time and outcome of trauma progress • Time and effect of diagnosis and management • Golden hour
4.2 Diagnosis
• According to variety of violence
Penetrating chest injury Blunt chest injury
• According to whether there is any communication between pleural cavity and atmosphere
Variety of trauma, mechanism of damage, time course after injury, vital signs
Evaluation of trauma
两种 DC-DC 转换器原理产生的噪声特性比较
两种 DC-DC 转换器原理产生的噪声特性比较Robert Marchetti，高级产品经理Vicor 公司，美国 Andover，MA目前，大部份的 DC-DC 转换器己普遍以高频率的开关技术为基础。
而在开关转换器中，有效的高频率开关一直被视为模块功率密度大小，性能表现优劣的关键。
开关频率高，所用的磁性元件和电容愈小，反应时间更快，噪声更低，滤波要求较小。
尽管 DC-DC 转换器所产生的噪声是比之前改善了，但是所有 DC-DC 转换器还是会产生 EMI 或者其它噪声的。
而所产生的噪声水平，不论是共模的，差模的或者是辐射噪声，会因为不同的生产厂，或者是采用不同的转换方法而产生很大的差异。
虽然市面上有上百种的 DC-DC 转换器，各有不同的设计和拓扑结构，大体可以归为两大类: 固定频率的脉宽调制式 (PWM) 和变频的准谐振零电流开关(ZCS) 两种。
设计师在选择 DC-DC 转换器时必须先了解两种原理的噪声表现。
脉宽调制式 (PWM) 模块由于不能够解决开关损耗问题，故此，它需要在工作效率和开关效率间作取舍。
开关元件在瞬时导通和关断时，使电感电流产生不连续性的状态，因而产生热量和噪声。
由于功耗来自开关损耗，它会随着脉宽调制式模块的开关频率增高而增大，直至它变为一个显著的耗损成因，达到了那一点，效率会迅速减低，开关元件所承受的热及电气应力变得无法处理。
这种开关损耗的属性，变为开关频率障碍，限制了转换器提升功率密度的能力。
除了所产生的噪声有差别，另一个脉宽调制式与零电流/零电压开关技术的分别是准谐振转换器的波形是一正弦波，而不是脉宽调制式的方波。
由于电流的波形没有近乎垂直上升和下降的尖削部份，减少电抗组件的应力，减低寄生噪声。
相反，PWM的输入电压开关频率是固定的(一般是数百kHz)，做成一连串的脉冲，利用调节脉冲的宽度来为负载提供正确的输出电压及足够的电流。
满载时，电流的波形是一个方波(图1)。
!图1 - 零电流开关和脉宽调制式架构的电流波形!很多电源工程师都以为，滤掉固定频率转换器所产生的噪声比滤掉变频率转换器的来得容易，这恰好是错的。
重组质粒构建
P1及YAC载体
酵母人工染色体，可容纳外源基因片 段长达200kb-1000kb，含有酵母第4号染色 体的着丝粒和自主复制序列CEN4、ARS1 ，作为端粒的Tel序列，选择性标志URA3 、TRP1和SUP4，能在酵母中稳primary） 工具酶：限制性内切酶
连接酶
常见的基因工程载体
载体：用于携带重组DNA，并且能够使外源DNA一起 复制与表达的运载工具。
❖ 根据来源：plasmid; phage; viral; BAC; YAC
❖ 根据用途分：克隆载体；原核生物表达载体；真核 生物表达载体
噬菌体载体
❖ λ噬菌体(λphage)
外源DNA：9~23kb 常用：EMBL 系列、 λgt 系列、charon系
列
❖ 粘性质粒(cosmid)： λDNA的 cos区+质 粒，双链环状DNA，克隆容量： 40~50kb
❖ M13噬菌体
λ噬菌体(λphage)
特点：一种能感染大肠杆菌的病毒， 野生型为双链线状DNA分子，长度48.5kb, 分子两端各有一个12bp组成的粘性末端, 感 染大肠杆菌后,线状DNA通过粘性末端互补 连接成环, 连接处称COS位点。
例：
干扰素是治疗癌症的重要物质，人血液中每升只能提取 0.05μg干扰素，因而其价格昂贵。但美国有一家公司用遗 传工程方法合成了价格低廉、药性一样的干扰素，其具体 做法是：
从人的淋巴细胞中提取能指导干扰素合成的基因，并使之 与一种叫做质粒的DNA结合，然后移植到酵母菌内，从而 用酵母菌来产生干扰素。
❖ 通常质粒含有某些染色体没有的基因，负责 编码某些功能蛋白，这些功能并不是细菌生 存所必需的，但在一定环境下，可对细菌宿 主的生存有利。由质粒产生的表型包括对抗 生素的抗性（R因子）、产生抗生素、降解复 杂有机化合物、产生大肠杆菌素（大肠杆菌 素因子col）、肠毒素及限制酶、修饰酶等。
低噪声放大器 LNA Highly linear wideband low power current mode LNA
II. WIDEBAND LNA DESIGN
A circuit schematic of the proposed differential wideband LNA is shown in Fig. 1. This novel architecture consists of a pair of NMOS and PMOS common gate amplifiers in each branch, sharing the same load. L1, L2 are off-chip inductors while CC1, CC2 are AC coupling capacitors with value much larger than Cgs.
DESIGN OF INTEGRATED CIRCUITS AND MICROSYSTEMS
output signal power can be achieved by allowing a larger current to the load. Obviously, in either case the limitation is the supply voltage VDD . So the signal power can be larger in the current mode, i.e. for a lower load resistance since
ห้องสมุดไป่ตู้
transistors either symmetrically for Class A or Class AB
operation, or asymmetrically. Fig. 2 shows the schematic of
an asymmetrically biased amplifier. In the next section we
comparison of various ion exchange resins
A comparison of various ion exchange resins for the removal of ferric ions from copper electrowinning electrolyte solutions Part II:Electrolytes containing antimony and bismuthBethan McKevitt a ,⁎,David Dreisinger ba Vale Inco Limited,CanadabUniversity of British Columbia,Canadaa b s t r a c ta r t i c l e i n f o Article history:Received 19December 2008Received in revised form 7April 2009Accepted 7April 2009Available online 16April 2009Keywords:Ion exchangeCopper electrowinningFour commercially available ion exchange resins were tested to compare their ability to remove ferric from copper electrowinning electrolytes containing antimony and bismuth.Ion exchange batch tests were per-formed in the initial phase of the study and indicated that all resins tested co-loaded some antimony with the ferric.The aminophosphonic resin tested was the only resin to load a signi ficant amount of bismuth.Further investigation was performed with the sulphonated monophosphonic resin,Purolite S957.Results were inconclusive as to whether antimony would continue to load onto the resin with time or whether the antimony would eventually reach an equilibrium loading on the resin.Antimony stripping was shown to be possible using a solution containing sulphuric acid and sodium chloride,indicating that ferric ion exchange is a technically feasible process for removal of iron from copper electrowinning electrolyte solutions containing antimony and bismuth.©2009Elsevier B.V.All rights reserved.1.IntroductionIn Part I of this work,the ability of four commercially available ion exchange resins containing a phosphonic acid group were evaluated for the removal of ferric iron from a synthetic copper electrowinning electrolyte solution.Most copper electrowinning operations utilize solvent extraction ahead of the electrowinning tankhouse to minimize the impurities present in the electrolyte solution.However,in a re finery where solvent extraction is not used,signi ficant amounts of other im-purities,such as antimony and bismuth,can be present in the elec-trolyte.For example,at the Vale Inco copper electrowinning facility in Copper Cliff,electrolyte puri fication is achieved without solvent extraction,and consequently the impurity levels in their electrolyte are fairly high (Sabau and Bech,2007).While the presence of antimony,and bismuth in a copper electro-winning electrolyte is unusual,these impurities are quite common in copper electrore fining electrolytes.In copper electrore fining,these impurities often contaminate the copper cathode product and so their levels in electrolyte must be controlled (Wang,2004).Ion exchange is one option available for the removal of these impurities.In a compre-hensive study of several ion exchange resins,Unitika's UR-3300and Rohm and Haas'DUOLITE C-467resins were found to have the best selectivity for removal of antimony and bismuth from copper electro-re fining electrolytes (Dreisinger and Leong,1992).Loading of bismuth and antimony onto the resin can be described by Eqs.(1)and (2):Bi 2ðSO 4Þ3þ6ðHR Þ→2ðBiR 3Þþ3H 2SO 4ð1ÞSb 2ðSO 4Þþ6ðHR Þ→2ðSbR 3Þþ3H 2SO 4:ð2ÞIn Dreisinger's study,elution was performed using hydrochloric acid and can be described by Eqs.(3)and (4):2ðBiR3Þþ6HCl →2BiCl3þ6ðHR Þð3Þ2ðSbR 3Þþ6HCl →2SbCl 3þ6ðHR Þ:ð4ÞAs an alternative to using hydrochloric acid,elution schemes with solutions containing sulphuric acid and sodium chloride have been proposed (Fukui et al.,2000).This approach is advantageous from a practical standpoint since many copper re fineries already have sulphuric acid and sodium chloride readily available on-site.This allows for the creation of hydrochloric acid according to Eq.(5):H 2SO 4þ2NaCl →Na 2SO 4þ2HCl :ð5ÞIt is important to note that both resins shown by Dreisinger to have excellent removal capability to remove antimony and bismuth are aminophosphonic resins.Also note that the synthetic solutions tested contained no ferric,and no sulphonated phosphonic resins were included in the study.In a later study,UR-3300and C-467were loaded with a commercial copper electrore fining electrolyte containing antimony,bismuth,andHydrometallurgy 98(2009)122–127⁎Corresponding author.E-mail address:Bethan.McKevitt@ (B.McKevitt).0304-386X/$–see front matter ©2009Elsevier B.V.All rights reserved.doi:10.1016/j.hydromet.2009.04.007Contents lists available at ScienceDirectHydrometallurgyj o u r n a l h o me p a g e :w w w.e l s ev i e r.c o m/l o c a t e /hyd ro m e tiron.Some iron loading was observed on both resins;however this could be eliminated by pre-reducing any ferric to ferrous by first passing the solution over a bed of copper wire.Loading of antimony and bismuth on both resins was improved when the ferric was pre-reduced to ferrous (Dreisinger and Scholey,1995).Since the presence of ferric is known to affect the ion exchange removal of antimony and bismuth with aminophosphonic resins,it is likely that the presence of antimony and bismuth could affect the removal of ferric by ion exchange with a resin containing the phosphoric acid group.This work focuses on the ability of four commercially available ion exchange resins containing the phosphoric acid group to remove ferric from a copper electrolyte solution containing antimony and bismuth.2.ExperimentalAs in the work in Part I,four commercially available resins were tested:Eichrom Diphonix (sulphonated diphosphonic),Purolite S957(sulphonated monophosphonic),Generic D416(sulphonated mono-phosphonic),and Lanxess Lewatit MonoPlus TP-260(aminophospho-nic).It is important to note that all these chelating resins contain the phosphoric acid group.In addition to the phosphoric acid group,the aminophosphonic resin also contains a group from the amine family;the sulphonated monophosphonic resins contain the sulphonic acid group;and the sulphonated diphosphonic resin contains both the sulphonic acid group and a carboxylic acid group.Note that the pH of the solutions tested in this work is suf ficiently low that the carboxylic acid group,typical of a weak acid exchanger,is not expected to have any impact on results.Sketches illustrating the chemistry of these chelating resins can be found in Fig.1of Part I of this work.Prior to testing,all resins were conditioned to ensure that each resin was in the acidic (H +)form and all volumes reported are based on the acidic form of the resins.To gain a qualitative comparison of the performance of the four resins,equilibrium loading experiments were performed.A test set was performed for each resin in which a series of flasks containing different electrolyte to resin volume ratios were placed in an orbital shaker at 175rpm and 55°C for 24h.For each test set,a synthetic electrolyte solution was prepared by adding the following reagents to water:sulphuric acid,copper sulphate,nickel sulphate,cobaltous sulphate,ferric sulphate,bismuth trioxide,antimony sulphate,and arsenic pentoxide.Heat was added to this solution in an attempt to dissolve all reagents;however,in all cases there was a small amount of residual solids left so the synthetic electrolyte was filtered before use.Temperature was not monitored during this procedure,and is believed to be the reason for fluctuations in feed chemistry between test sets.It should be noted that the 24h is believed to have been more than enough time to reach equilibrium since a series of tests run withVale Inco plant electrolyte reached equilibrium within 8h for a volume ratio of 14:1(mL electrolyte to mL resin).Based on the results of this first set of scoping tests,the optimum resin type was selected for further testing.The next set of tests performed involved further equilibrium batch loading tests with a volume ratio of 100:1(mL electrolyte to mL resin).In these tests the amount of ferric added to the various flasks was varied in order to produce different iron to antimony ratios in the synthetic electrolyte solution.An extended column loading test was performed to determine if antimony or iron would displace the other element off the resin.In this test,130BV of a synthetic electrolyte was passed over the resin.Note that the synthetic electrolyte for this test was prepared with a similar chemistry as used in the batch equilibrium tests,but otherwise followed the column loading procedure reported in Part I of this work.Next,antimony elution tests were performed by loading two columns,each containing 10mL of resin,with 150BV of a solution containing sulphuric acid and saturated in antimony at a rate of 15BV/h.Stripping of both columns was performed at 50°C.The first column was stripped with a solution containing only sulphuric acid while the second column was stripped with a solution containing both sulphuric acid and sodium chloride.Finally,repeated cycle testing was performed on a single sample of Purolite S957resin.In these tests,antimony was added to both the loading and stripping electrolyte,in order to observe how antimony loading affects the resin performance over a few cycles.Ten cycles of loading and stripping were performed,followed by a column wash consisting of 50BV of solution containing sulphuric acid and sodium chloride.A final cycle of loading and stripping was performed to deter-mine whether this wash had been able to restore the performance of the ion exchange resin.It should be noted that challenges were encountered obtaining accurate antimony assays in the solution samples.All samples were analyzed at the Vale Inco Central Process Technology Laboratory in Copper Cliff,Ontario.Solution metal concentrations were determined using an in-house inductive coupled plasma procedure:ICP-PMET.In order to ensure that these results were accurate,resin loadings for the 100:1volume ratio from the first set of batch tests were deter-mined by both a mass balance using the solution assays and by resin digestion.Results showed good correlation for all elements except antimony.Therefore,for subsequent tests,the amount of antimony present in solution was determined using ICP-MS,which was found to correlate very well with resin digestion results.Resin analysis involved completely dissolving a known mass of resin in nitric acid using microwave digestion.The resulting solution was then analyzed using a standard ICP analysis and results were reported in mg/g resin.This could then be correlated back to the initial volume of resin to obtain values of mg/mL resin.Note that no antimony or bismuth was detected in samples of conditioned,unloaded resin (i.e.resin blanks)and only minimal amounts of iron (b 0.2mg Fe/mL resin)were present.3.Results and discussionThe results for the various tests are discussed in following subsections.3.1.Equilibrium loading experimentsThe composition of the synthetic electrolyte feed solutions for the various tests is displayed in Table 1.It should be noted that there is some fluctuation in the concentration of the metals in the feed solution between each tests set.This is believed to be due to the fact that temperature was not monitored during electrolyte preparation.However,since the purpose of these tests was to gain a qualitative comparison between the overall performance of the various resins,and was not meant to be a quantitative analysis of resinselectivities,Fig.1.Ferric loading curves,equilibrium loading tests.123B.McKevitt,D.Dreisinger /Hydrometallurgy 98(2009)122–127these fluctuations in feed chemistry were deemed acceptable for this work.Note that the tests were all run at approximately 220g/L acid.The loading curves for ferric and bismuth are shown in Figs.1and 2,respectively.The loading curve plots the amount of an element loaded on the resin against the volume ratio of electrolyte to resin.Note that there is no loading curve for antimony since the solution samples were only run using ICP-PMET,and not by ICP-MS.All samples with a volume ratio of 100:1were also analyzed by resin digestion,and the results from resin digestion are displayed in Fig.3.Fig.1shows that the ferric loading capacity of the sulphonated monophosphonic resins (Purolite S957and Generic D416)and the aminophophonic resin (Lewatit TP 260)are very similar at ~23mg Fe/mL resin,while the capacity of the sulphonated diphosphonic resin (Diphonix)is substantially less at ~14mg Fe/mL resin.The only exception is that the calculated iron loading for the Generic resin is ~30mg Fe/mL resin at a volume ratio of 100:1.This is believed to be due to an erroneous assay result as the resin digestion results for this sample,presented in Fig.3,shows that the total iron loaded onto the resin was ~25mg Fe/mL resin.These ferric loading capacities are similar to the ferric loading obtained in the column experiments in Part I of this work.Fig.2shows that the aminophosphonic resin (Lewatit TP 260)loaded a signi ficant amount of bismuth.In fact,these results show that at a volume ratio of 100:1,more bismuth was loaded than ferric,illustrating why aminophosphonic resin are often used for bismuth removal from copper electrore fining electrolytes.The resin digestion results in Fig.3show that all resins loaded some antimony.These results also show that the sulphonated monopho-sphonic resins (Purolite S957,Generic D416)loaded similar amounts of antimony as the aminophosphonic resin (Lewatit TP 260).While the amount of antimony in the feed solution was signi ficantly lower for the aminophosphonic resin,the Fe/Sb ratio in the feed was approximately 12,while the Fe/Sb ratio for the sulphonated monophosphonic resins was between 7and 8.As will be seen in Section 3.2,the amount of antimony loaded on Purolite S957(sulphonated monophosphonic)does not appear to vary signi ficantly above an Fe/Sb ratio of about 2.This suggests that the sulphonated monophosphonic resins may be as useful as the aminophosphonic resins for antimony removal;further work should be performed to further compare these two resin types with respect to antimony removal.Additionally,Fig.3shows the resin loadings for arsenic,copper,nickel,and cobalt.All resins appear to have loaded similar amounts of arsenic.It is interesting to note that the sulphonated diphosphonic resin (Diphonix)loaded the most copper,nickel,and cobalt while the aminophosphonic resin (Lewatit)loaded the least amount of these elements.Since the sulphonic acid group allows for non-selective cation extraction at low pH,it makes sense that the aminophosphonic resin loaded the least amount of these elements since this resin is the only resin tested that does not contain the sulphonic acid group.Since the sulphonated disphosphonic resin loaded the most copper,nickel,and cobalt while loading the least amount of iron,this may suggest that the ratio of sulphonic acid groups to phosphoric acid groups is higher in this resin than for the sulphonated monophosphonic resins tested.Based on the results from these initial equilibrium batch loading tests,the sulphonated monophosphonic resins appear to be better suited than the sulphonated diphosphonic resin or the aminopho-sphonic resin for ferric removal in the presence of antimony and bismuth.Therefore,the Purolite S957resin was selected for further testing.3.2.Effect of iron to antimony ratioA second series of equilibrium batch tests were performed with the Purolite S957Resin.The composition of the feed solutions for the flasks can be found in Table 2.At the end of the test,all resin samples were digested and analyzed.Since flasks “F ”and “I ”contained approximately the same amount of iron,and only flask “F ”contained antimony,comparing the results from these two tests gives an indication of how the presence of antimony may affect the loading of iron onto the resin.The resin loading for the flask without any antimony present (“I ”)was 30.6mg Fe/mL Resin,while the flask with antimony present (“F ”)loaded slightly less iron:27.6mg Fe/mL Resin.This suggests that the loaded antimony is taking up ion exchange sites that could have been available forferric.Fig.2.Bismuth loading curves,equilibrium loading tests.Table 1Feed composition of synthetic electrolyte for equilibrium loading tests.g/L Cu Ni Co As Fe Sb Bi Diphonix 33.013.610.0 2.6 2.00.120.29Generic 35.414.410.8 2.9 2.20.290.38Lewatit 37.015.211.6 2.8 2.30.190.46Purolite38.815.612.02.92.30.290.44Fig.3.Resin digestion results,equilibrium loading tests.Table 2Feed composition of synthetic electrolyte for ratio tests.g/L Fe(III)Sb Fe /Sb Cu Ni Co As Bi “A ”b 0.010.220.0041.610.49.5 2.70.36“B ”0.130.220.5940.610.39.4 2.70.37“C ”0.230.23 1.040.610.29.3 2.70.35“D ”0.410.23 1.839.410.09.2 2.60.33“E ”0.860.24 3.641.210.59.6 2.70.38“F ” 2.100.258.440.610.29.4 2.70.37“G ” 4.280.251741.610.59.6 2.70.38“H ”8.680.243641.010.59.6 2.70.39“I ”2.10b 0.01∞39.215.211.22.70.15124 B.McKevitt,D.Dreisinger /Hydrometallurgy 98(2009)122–127Resin loadings at the various iron to antimony ratios are plotted in Fig.4.This graph shows that in a solution with no iron present,the resin loaded the most antimony.It also shows that the amount of antimony loaded onto the resin decreased at a fairly steady rate,indicating that iron was competing favourably with the antimony,until the iron/antimony ratio reached a value of 1.75.From this point on,the amount of antimony loaded onto the resin only decreased very slightly as the iron/antimony ratio continued to increase,indicating that increasing the iron to antimony ratio beyond 1.75has very little affect on the antimony loading.Note the extremely low antimony loaded at an iron/antimony ratio of 17.This is believed to be due to an assay error as there is no evidence of a lower concentration of antimony present in the solution sample at the end of this test.Unfor-tunately,there was insuf ficient resin available to re-assay this sample.3.3.Extended loading:displacement testAn extended column loading experiment,with the feed composi-tion shown in Table 3,was performed to determine whether antimony or iron would displace the other element from the resin.Displacement of an element can be seen from a loading curve that plots C/C 0of the elements in the solution exiting the column.While an element is being loaded onto the resin,the C/C 0value will be b 1.If there is no interaction between the element and the resin,the C/C 0value will remain at 1.If an element is being displaced from the resin,the C/C 0value will be N 1.The loading curve from the displacement test is plotted in Fig.5.This graph shows a single point where the iron sample appears to have a C/C 0greater than one;however,this is likely due to assay error as there is no evidence of antimony loading on the resin at this point on the curve.Note that the breakthough point for antimony is signi fi-cantly later in the loading cycle than for iron.This suggests that the resin may be selective for antimony over iron.3.4.Antimony elutionThe composition of the feed solutions for both loading and elution of antimony from the Purolite S957resin can be found in Table 4.A mass balance using the solution samples indicate a resin loading of 43.6mg Sb /mL resin in Test 1,and 41.4mg Sb /mL resin in Test 2.The detailed loading curve from test 1is shown in Fig.6.This curve shows that at the end of the loading tests,the resin was not fully loaded with antimony since C/C 0never reached 1.Using the assays from the stripping samples,the cumulative per-cent antimony eluted from the resin was calculated for each sample from the stripping test.These results are plotted in Fig.7.This figure shows that antimony is eluted slowly in the presence of sulphuric acid alone (the reverse reaction of Eq.(2),and it is eluted very quickly in the presence of a solution containing both sulphuric acid and sodium chloride (Eq.(4)).Further testing would be required to determine whether or not complete stripping could be attained with a solution containing only sulphuric acid.It should be noted that as the stripping test with the solution containing sodium chloride started,a white precipitate was observed in the strip solution.This precipitate subsequently redissolved by the time 5BV of solution had been collected.The white precipitate was not sampled,but it is possible that SbOCl may have formed,since the initial sample of strip solution collected would have had low acidity due to displacing the deionized water remaining in the column from the load wash step.This reaction could be described by Eq.(6):SbCl 3þH 20→SbOCl ðs Þþ2HCl :ð6ÞThis result suggests that flushing the ion exchange bed with water prior to a wash with sulphuric acid and sodium chloride may not be desirable as it may lead to SbOCl precipitation in the ion exchange bed.Further work would be required to test this hypothesis.3.5.Repeat column cycle testingThe composition of the feed and stripping electrolyte used in each cycle of these tests can be found in Table 5.To determine whether the iron loading performance of the resin was signi ficantly impacted by the presence of antimony in electrolyte,a solution mass balance was performed to determine the amount of iron loaded and the amount of iron stripped for each cycle.These results are plotted in Fig.8.This graph shows that the amount of iron loaded onto the resin stayed approximately constant over the course of the tests,suggesting that the effect of antimony on iron loading is not deleterious over ten cycles.Note that the total amount of iron loaded in these tests (typically 20–21mg Fe/mL resin),is approximately the same as the values obtained in Part I,for the Purolite tests performed with no antimony present (20–22mg Fe/mL resin).Also note that there is signi ficantlymoreFig.4.Iron and antimony loadings at various Fe/Sb ratios.Table 3Feed composition of electrolyte for column displacement test.g/LCu Ni Co As Fe Sb Bi Displacement test39.213.610.92.52.10.180.33Fig.5.Displacement test loading curve.Table 4Feed composition of electrolyte for antimony elution test.g/L H 2SO 4Cu Fe(III)Sb Na Load 1219.30.00.00.450.0Load 2218.50.00.00.470.0Strip 1177.80.00.00.000.0Strip2169.90.00.00.0079.2125B.McKevitt,D.Dreisinger /Hydrometallurgy 98(2009)122–127scatter in the mg Fe/mL resin calculated from the loading mass balance than the stripping mass balance,likely indicating sampling error of the fully loaded solution.To more precisely determine the effect of antimony on iron loading,the stripping time required was plotted against cycle time.As in Part I,stripping time was de fined as the time required to reach an ORP of 500mV SHE .The plot of stripping time against cycle number is found in Fig.9.In this graph,the stripping time tends to decrease with cycle number,for cycles 1through 10.This suggests that the antimony is loading onto the resin,thus reducing the number of sites available for iron to load.Although there is only one data point after the salt wash,the stripping time did increase in cycle 11,suggesting that after the antimony is removed,ion exchange sites may once again be freed up to load more iron.It is interesting to note that the results from cycle 3and cycle 6fall signi ficantly off the general trend line.During cycle 3,the loading cycle was not stopped after 8h (approximately 80BV),and 110BV of loading solution were passed over the resin.The fact that this cycle took longer to strip may be an indication that the resin was not fully loaded after 8h.The reason for cycle 6stripping much sooner than expected is unknown.It may be that some of the cuprous from the previous stripping cycle formed copper powder and found its way into the ion exchange column.The mass balance for this test does not give any indication that this occurred;however if this were to have happened,then less iron would have loaded onto the resin since some of the ferric would have been reduced to ferrous during the loading cycle.During the loading portion of cycle 10,samples were taken to generate a loading curve.This loading curve is plotted in Fig.10.It is interesting to note that in cycle 10,the breakthrough of antimonyappears to occur slightly before the breakthrough of ferric.This is the complete opposite of what was observed during the displacement test (Fig.5).Also note that the final C/C 0value for antimony is greater than unity,which would indicate that some antimony is being displaced from the resin as the ferric loads.However,given that the stripping time is continually decreasing (Fig.9),it is possible that this may simply be due to a low initial concentration value for antimony (i.e.sampling/assay error for the feed sample).Alternatively,it may be an indication that once the resin has loaded an appreciable amount of antimony,ferric is able to displace antimony from the resin.Results are inconclusive as to whether or not the antimony will continue to load onto the resin to the point where the loaded antimony would signi ficantly impede iron loading.On the one hand,the iron loading capacity does not appear to have decreased over the ten cycles,and the shape of the antimony breakthrough curve for the tenth cycle is signi ficantly different than for the first cycle,suggesting thattheFig.6.Antimony loadingcurve.Fig.7.Antimony elution curves.Table 5Feed composition of electrolyte for repeat cycle tests.LoadStrip Cycle g/L Cu g/L Fe g/L Sb g/L Cu g/L Fe g/L Sb 131.8 1.390.20041.0b 0.010.234231.1 1.350.24142.5b 0.010.263331.5 1.370.25241.6b 0.010.260431.9 1.390.24541.6b 0.010.268531.6 1.380.23242.4b 0.010.296632.6 1.400.24742.2b 0.010.206733.6 1.420.17745.40.030.260834.2 1.450.20845.80.040.236934.0 1.430.23746.60.020.2681034.1 1.430.23946.20.020.2741132.91.400.26144.2b 0.010.324Fig.9.Stripping time vs.cyclenumber.Fig.8.Iron loaded vs.cycle number.126 B.McKevitt,D.Dreisinger /Hydrometallurgy 98(2009)122–127resin could reach an equilibrium loading of antimony on the resin.On the other hand,the required stripping time appears to have decreased steadily as the number of cycles increased,indicating the possibility of continuous antimony loading on the resin.Pilot testing is recom-mended to determine the long-term behaviour of the resin in a solution containing both antimony and iron.The salt and acid wash appears to be able to restore resin perfor-mance,indicating that iron ion exchange is technically feasible in electrolyte solutions containing antimony,even if antimony continu-ously loads on the resin.4.ConclusionsResults from the equilibrium loading tests indicate that in an application where it is desired to remove ferric from an electrolyte that also contains antimony and bismuth,all resins tested will co-load antimony with the ferric.The investigations carried out with the Purolite S957resin (sulphonated monophosphonic)have shown that ferric removal from a copper electrowinning electrolyte containing antimony and bismuth is a technically feasible process.Although antimony will load onto the resin,it should be possible to elute it with a solution containing sulphuric acid and sodium chloride.Further testing is required in order to determine whether or not the antimonywill load continuously onto the resin,or whether it will reach an equilibrium loading level.In an application where antimony removal is desired,both the sulphonated monophosphonic resins and the aminophosphonic resin appear to load approximately the same amount of antimony under the batch test conditions.Therefore,it may be worthwhile investigat-ing the performance of the sulphonated monophosphonic resins for antimony removal in the absence of ferric to determine whether or not these resins may be useful for antimony removal.Note that the loading test performed during the antimony elution tests showed that the resin was not fully loaded when it contained more than 40mg Sb/mL resin.In an application where bismuth removal is desired,the amino-phosphonic resin was clearly superior to all other resins tested.AcknowledgementsThis work was done as part of a Vale Inco sponsored MASc degree at the University of British Columbia.Vale Inco is gratefully acknowl-edged for providing the funding for this work and for performing the assay analyses.Thanks should also be given to Fenix Hydromet for providing the ion exchange resins for testing and to Dr.B.Wassink of the University of British Columbia for his guidance with the develop-ment of the experimental procedures.ReferencesDreisinger,D.B.,Leong,B.J.Y.,1992.The solvent extraction and ion exchange removal ofarsenic,antimony,and bismuth from copper re finery electrolytes:final report.The University of British Columbia,Department of Metals and Materials Engineering.Dreisinger,D.B.,Scholey,B.J.Y.,1995.In:Cooper,W.C.(Ed.),Ion exchange removal ofantimony and bismuth from copper re finery electrolytes.Copper 95—Cobre 95,vol.III.Canadian Institute of Mining,Metallurgy and Petroleum,Montreal,Quebec,Canada,pp.305–314.Fukui,A.,Tsuchida,N.,Ando,K.,2000.Method of recovering antimony and bismuthfrom copper electrolyte.United States Patent #6,153,081.Sabau,M.,Bech,K.,2007.In:Houlachi,G.E.,Edwards,J.D.,Robinson,T.G.(Eds.),Statusand improvement plans in Inco's electrowinning tankhouse.Copper —Cobre 2007,vol.V.Canadian Institute of Mining,Metallurgy and Petroleum,Montreal,Quebec,Canada,pp.439–450.Wang,S.,2004.Impurity control and removal in copper tankhouse operations.Journalof Materials 56(7),34.Fig.10.Cycle ten loading curve.127B.McKevitt,D.Dreisinger /Hydrometallurgy 98(2009)122–127。
A comparison between several circuits
Abstract This paper presents a comparison between four vibration-powered generators designed to power standalone systems, such as wireless transducers. Ambient vibrations are converted into electrical energy using piezoelectric materials. The originality of the proposed approaches is based on a particular processing of the voltage delivered by the piezoelectric material, which enhances the electromechanical conversion. The principle of each processing circuit is detailed. Experimental results confirm the predictions given by an electromechanical model: compared to usual generators, the proposed approaches dramatically increase the power of the generators. © 2005 Elsevier B.V. All rights reserved.
Sensors and Actuators A 126 (2006) 405–416
A comparison between several vibration-powered piezoelectric generators for standalone systems
A CONVERTER AND A METHOD FOR CONTROLLING A CONVER
专利名称：A CONVERTER AND A METHOD FOR CONTROLLING A CONVERTER发明人：BLIDBERG, Ingemar,SVENSSON,Kjell,SILJESTRÖM, Roland申请号：SE2003001556申请日：20031007公开号：WO04/034559P1公开日：20040422专利内容由知识产权出版社提供摘要：In a method for control of a converter for conversion of dc voltage into ac voltage or dc voltage and vice versa, in which an output of the converter may alternatively be connected to a positive pole (7), a negative pole (8) or a centre (23) of a dc voltage side of the converter in the form of different so-called main states, there is carried out, when changing between main states via a so-called minor commutation loop, an extra sequence in the form of a delayed turn-on of semiconductor elements (13, 16) in one unit of the converter relative to semiconductor elements (14, 15) in another unit in a pair of units of the converter.申请人：BLIDBERG, Ingemar,SVENSSON, Kjell,SILJESTRÖM, Roland地址：SE,SE,SE,SE国籍：SE,SE,SE,SE代理机构：ABB AB更多信息请下载全文后查看。
不同剂量甲硝唑栓剂与克林霉素联用治疗老年性阴道炎的临床效果与
【Ａｂｓｔｒａｃｔ】 Ｏｂｊｅｃｔｉｖｅ Ｔｏｉｎｖｅｓｔｉｇａｔｅｔｈｅｃｌｉｎｉｃａｌｅｆｆｅｃｔａｎｄｓａｆｅｔｙｏｆｄｉｆｆｅｒｅｎｔｄｏｓｅｍｅｔｒｏｎｉｄａｚｏｌｅｓｕｐｐｏｓｉｔｏ ｒｙｃｏｍｂｉｎｅｄｗｉｔｈｃｌｉｎｄａｍｙｃｉｎｉｎｔｈｅｔｒｅａｔｍｅｎｔｏｆｓｅｎｉｌｅｖａｇｉｎｉｔｉｓ．Ｍｅｔｈｏｄｓ １８０ｐａｔｉｅｎｔｓｗｉｔｈｓｅｎｉｌｅｖａｇｉｎｉｔｉｓｔｒｅａ ｔｅｄｉｎｏｕｒｈｏｓｐｉｔａｌｆｒｏｍＪａｎｕａｒｙ２０１５ｔｏＯｃｔｏｂｅｒ２０１７ｗｅｒｅｓｅｌｅｃｔｅｄａｎｄｒａｎｄｏｍｌｙｄｉｖｉｄｅｄｉｎｔｏｇｒｏｕｐＡ（４８ｃａｓｅｓ）， ｇｒｏｕｐＢ（５０ｃａｓｅｓ），ｇｒｏｕｐＣ（５５ｃａｓｅｓ）ａｎｄｇｒｏｕｐＤ（５５ｃａｓｅｓ）．ＧｒｏｕｐＡｗａｓｇｉｖｅｎｍｉｄｄｌｅｄｏｓｅｍｅｔｒｏｎｉｄａｚｏｌｅ ｓｕｐｐｏｓｉｔｏｒｙ；ｇｒｏｕｐＢｗａｓｇｉｖｅｎｌｏｗｄｏｓｅｍｅｔｒｏｎｉｄａｚｏｌｅｓｕｐｐｏｓｉｔｏｒｙａｎｄｃｌｉｎｄａｍｙｃｉｎ；ｇｒｏｕｐＣｗａｓｇｉｖｅｎｍｉｄｄｌｅｄｏｓｅ ｍｅｔｒｏｎｉｄａｚｏｌｅｓｕｐｐｏｓｉｔｏｒｙａｎｄｃｌｉｎｄａｍｙｃｉｎ；ｇｒｏｕｐＤｗａｓｇｉｖｅｎｈｉｇｈｄｏｓｅｍｅｔｒｏｎｉｄａｚｏｌｅｓｕｐｐｏｓｉｔｏｒｙａｎｄｃｌｉｎｄａｍｙ ｃｉｎ．Ｔｈｅｃｌｉｎｉｃａｌｔｒｅａｔｍｅｎｔｅｆｆｅｃｔａｎｄｔｈｅｉｎｃｉｄｅｎｃｅｏｆａｄｖｅｒｓｅｒｅａｃｔｉｏｎｗｅｒｅｃｏｍｐａｒｅｄ，ａｎｄｔｈｅｌｅｖｅｌｓｏｆｅｓｔｒａｄｉｏｌ （Ｅ２）ａｎｄｆｏｌｌｉｃｌｅｓｔｉｍｕｌａｔｉｎｇｈｏｒｍｏｎｅ（ＦＳＨ） ｗｅｒｅｔｅｓｔｅｄｂｙｅｎｚｙｍｅｌｉｎｋｅｄｉｍｍｕｎｏｓｏｒｂｅｎｔａｓｓａｙ（ＥＬＩＳＡ）． Ｒｅｓｕｌｔｓ ＣｏｍｐａｒｅｄｗｉｔｈｇｒｏｕｐＡａｎｄｇｒｏｕｐＢ，ｔｈｅｔｏｔａｌｅｆｆｅｃｔｉｖｅｒａｔｅｉｎｇｒｏｕｐＣａｎｄｇｒｏｕｐＤｗａｓｓｉｇｎｉｆｉｃａｎｔｌｙｉｎ ｃｒｅａｓｅｄ，ｗｉｔｈｓｔａｔｉｓｔｉｃａｌｌｙｓｉｇｎｉｆｉｃａｎｔｄｉｆｆｅｒｅｎｃｅｓ（Ｐ＜００５）．Ｂｅｆｏｒｅｔｒｅａｔｍｅｎｔ，ｔｈｅｒｅｗａｓｎｏｓｔａｔｉｓｔｉｃａｌｌｙｓｉｇｎｉｆｉｃａｎｔ ｄｉｆｆｅｒｅｎｃｅｓｉｎｔｈｅｌｅｖｅｌｓｏｆＥ２ａｎｄＦＳＨ（Ｐ＞００５）．Ａｆｔｅｒｔｒｅａｔｍｅｎｔ，ｃｏｍｐａｒｅｄｗｉｔｈｇｒｏｕｐＡ，ＢａｎｄＣ，ｔｈｅｌｅｖｅｌｏｆ Ｅ２ｉｎｇｒｏｕｐＤｗａｓｓｉｇｎｉｆｉｃａｎｔｌｙｄｅｃｒｅａｓｅｄ，ｂｕｔｔｈｅＦＳＨ ｌｅｖｅｌｓｉｇｎｉｆｉｃａｎｔｌｙｉｎｃｒｅａｓｅｄ，ｗｉｔｈｓｔａｔｉｓｔｉｃａｌｌｙｓｉｇｎｉｆｉｃａｎｔ ｄｉｆｆｅｒｅｎｃｅｓ（Ｐ＜００５）．Ｔｈｅｒｅｗｅｒｅｎｏｓｔａｔｉｓｔｉｃａｌｌｙｓｉｇｎｉｆｉｃａｎｔｄｉｆｆｅｒｅｎｃｅｓｉｎｔｈｅｉｎｃｉｄｅｎｃｅｓｏｆａｄｖｅｒｓｅｒｅａｃｔｉｏｎｓｓｕｃｈ ｎａｕｓｅａ，ｖｏｍｉｔｉｎｇ，ｈｅａｄａｃｈｅａｎｄｄｉａｒｒｈｅａａｍｏｎｇｔｈｅｆｏｕｒｇｒｏｕｐｓ（ａｌｌＰ＞００５）．Ｃｏｎｃｌｕｓｉｏｎｓ Ｈｉｇｈｄｏｓｅｍｅｔｒｏｎ ｉｄａｚｏｌｅｓｕｐｐｏｓｉｔｏｒｙｃｏｍｂｉｎｅｄｗｉｔｈｃｌｉｎｄａｍｙｃｉｎｉｓｓａｆｅａｎｄｅｆｆｅｃｔｉｖｅｉｎｔｈｅｔｒｅａｔｍｅｎｔｏｆｓｅｎｉｌｅｖａｇｉｎｉｔｉｓ，ｗｈｉｃｈｃａｎ ｓｉｇｎｉｆｉｃａｎｔｌｙｉｍｐｒｏｖｅｔｈｅｌｅｖｅｌｏｆｅｓｔｒｏｇｅｎ．
基于GAN和特征选择技术的入侵检测数据增强
第１４卷㊀第３期Ｖｏｌ．１４Ｎｏ．３㊀㊀智㊀能㊀计㊀算㊀机㊀与㊀应㊀用ＩｎｔｅｌｌｉｇｅｎｔＣｏｍｐｕｔｅｒａｎｄＡｐｐｌｉｃａｔｉｏｎｓ㊀㊀２０２４年３月㊀Ｍａｒ．２０２４㊀㊀㊀㊀㊀㊀文章编号：２０９５－２１６３（２０２４）０３－０１７４－０７中图分类号：ＴＰ３９３．０８文献标志码：Ａ基于ＧＡＮ和特征选择技术的入侵检测数据增强崔子才，钟伯成，赵欣阳（上海工程技术大学电子电气工程学院，上海２０１６２０）摘㊀要：为了解决传统ＧＡＮ模型的缺陷，更好地扩展网络入侵数据和缓解数据高维性问题，本文提出了ＧＡＮ－ＣＳ数据增强模型㊂对数据进行预处理后，使用改进后的ＷＧＡＮ－ＧＰ对攻击数据进行增强，生成额外的攻击样本后，使用卡方检验方法选择最能够代表数据集的特征，生成用于分类器训练平衡后的数据集，最后使用多种不同的分类器对数据集进行分类，评估模型效果㊂本文基于ＵＮＳＷ－ＮＢ１５分别进行了数据增强数据量选择实验㊁模型可行性实验㊁模型优越性比较等３个维度的实验㊂结果表明，在多个分类器下，本文提出的模型均表现出比同类模型更好的效果，可以有效提高入侵检测模型的检测性能㊂关键词：入侵检测；数据增强；ＷＧＡＮ－ＧＰ算法；ＵＮＳＷ－ＮＢ１５数据集ＩｎｔｒｕｓｉｏｎｄｅｔｅｃｔｉｏｎｄａｔａａｕｇｍｅｎｔａｔｉｏｎｂａｓｅｄｏｎＧＡＮａｎｄｆｅａｔｕｒｅｓｅｌｅｃｔｉｏｎｔｅｃｈｎｉｑｕｅＣＵＩＺｉｃａｉ，ＺＨＯＮＧＢｏｃｈｅｎｇ，ＺＨＡＯＸｉｎｙａｎｇ（ＳｃｈｏｏｌｏｆＥｌｅｃｔｒｏｎｉｃａｎｄＥｌｅｃｔｒｉｃａｌＥｎｇｉｎｅｅｒｉｎｇ，ＳｈａｎｇｈａｉＵｎｉｖｅｒｓｉｔｙｏｆＥｎｇｉｎｅｅｒｉｎｇＳｃｉｅｎｃｅ，Ｓｈａｎｇｈａｉ２０１６２０，Ｃｈｉｎａ）Ａｂｓｔｒａｃｔ：ＩｎｏｒｄｅｒｔｏａｄｄｒｅｓｓｔｈｅｓｈｏｒｔｃｏｍｉｎｇｓｏｆｔｈｅｔｒａｄｉｔｉｏｎａｌＧＡＮｍｏｄｅｌ，ｂｅｔｔｅｒｅｘｔｅｎｄｔｈｅｎｅｔｗｏｒｋｉｎｔｒｕｓｉｏｎｄａｔａａｎｄａｌｌｅｖｉａｔｅｔｈｅｐｒｏｂｌｅｍｏｆｈｉｇｈｄｉｍｅｎｓｉｏｎａｌｉｔｙｏｆｄａｔａ，ｔｈｉｓｐａｐｅｒｐｒｏｐｏｓｅｓｔｈｅＧＡＮ－ＣＳｄａｔａｅｎｈａｎｃｅｍｅｎｔｍｏｄｅｌ．ＴｈｅｄａｔａｉｓｐｒｅｐｒｏｃｅｓｓｅｄａｎｄｔｈｅｎａｕｇｍｅｎｔｅｄｗｉｔｈａｔｔａｃｋｄａｔａｕｓｉｎｇｔｈｅｉｍｐｒｏｖｅｄＷＧＡＮ－ＧＰｔｏｇｅｎｅｒａｔｅａｄｄｉｔｉｏｎａｌａｔｔａｃｋｓａｍｐｌｅｓ．Ｔｈｅｎ，ｔｈｅｆｅａｔｕｒｅｓｔｈａｔｂｅｓｔｒｅｐｒｅｓｅｎｔｔｈｅｄａｔａｓｅｔａｒｅｓｅｌｅｃｔｅｄｕｓｉｎｇａＣｈｉ－Ｓｑｕａｒｅｔｅｓｔｍｅｔｈｏｄ，ａｎｄａｂａｌａｎｃｅｄｄａｔａｓｅｔｉｓｇｅｎｅｒａｔｅｄｆｏｒｃｌａｓｓｉｆｉｅｒｔｒａｉｎｉｎｇ，ｆｉｎａｌｌｙｔｈｅｄａｔａｓｅｔｉｓｃｌａｓｓｉｆｉｅｄｕｓｉｎｇａｖａｒｉｅｔｙｏｆｄｉｆｆｅｒｅｎｔｃｌａｓｓｉｆｉｅｒｓｔｏｅｖａｌｕａｔｅｔｈｅｍｏｄｅｌｅｆｆｅｃｔ．Ｉｎｔｈｉｓｐａｐｅｒ，ｅｘｐｅｒｉｍｅｎｔｓｂａｓｅｄｏｎＵＮＳＷ－ＮＢ１５ａｒｅｃｏｎｄｕｃｔｅｄｔｏｐｅｒｆｏｒｍｔｈｒｅｅｄｉｍｅｎｓｉｏｎｓｅｘｐｅｒｉｍｅｎｔｓｓｕｃｈａｓｄａｔａｅｎｈａｎｃｅｍｅｎｔｄａｔａｖｏｌｕｍｅｓｅｌｅｃｔｉｏｎｅｘｐｅｒｉｍｅｎｔｓ，ｍｏｄｅｌｆｅａｓｉｂｉｌｉｔｙｅｘｐｅｒｉｍｅｎｔｓ，ａｎｄｍｏｄｅｌｓｕｐｅｒｉｏｒｉｔｙｃｏｍｐａｒｉｓｏｎ，ｒｅｓｐｅｃｔｉｖｅｌｙ，ａｎｄｔｈｅｒｅｓｕｌｔｓｓｈｏｗｔｈａｔｔｈｅｍｏｄｅｌｓｐｒｏｐｏｓｅｄｉｎｔｈｉｓｐａｐｅｒａｌｌｓｈｏｗｂｅｔｔｅｒｒｅｓｕｌｔｓｔｈａｎｓｉｍｉｌａｒｍｏｄｅｌｓｕｎｄｅｒｍｕｌｔｉｐｌｅｃｌａｓｓｉｆｉｅｒｓ，ｗｈｉｃｈｃａｎｅｆｆｅｃｔｉｖｅｌｙｉｍｐｒｏｖｅｔｈｅｄｅｔｅｃｔｉｏｎｐｅｒｆｏｒｍａｎｃｅｏｆｉｎｔｒｕｓｉｏｎｄｅｔｅｃｔｉｏｎｍｏｄｅｌｓ．Ｋｅｙｗｏｒｄｓ：ｉｎｔｒｕｓｉｏｎｄｅｔｅｃｔｉｏｎ；ｄａｔａｅｎｈａｎｃｅｍｅｎｔ；ＷＧＡＮ－ＧＰａｌｇｏｒｉｔｈｍ；ＵＮＳＷ－ＮＢ１５ｄａｔａｓｅｔ基金项目：国家自然基金青年科学基金项目（６２１０２２４１）㊂作者简介：崔子才（１９９８－），女，硕士研究生，主要研究方向：网络安全；赵欣阳（１９９６－），男，硕士研究生，主要研究方向：网络安全㊂通讯作者：钟伯成（１９６４－），男，博士，教授，主要研究方向：计算机网络安全㊂Ｅｍａｉｌ：ｂｃｚｈｏｎｇ＠ｓｕｅｓ．ｅｄｕ．ｃｎ收稿日期：２０２３－０３－０８０㊀引㊀言随着信息技术的不断发展，互联网已在政治㊁军事㊁经济㊁交通等领域发挥着重大作用㊂与此同时，各种网络攻击行为始终存在于互联网中，不仅可能造成巨大的经济损失，严重时甚至还会威胁到国家安全和社会的稳定发展㊂入侵检测系统（ＩｎｔｒｕｓｉｏｎＤｅｔｅｃｔｉｏｎＳｙｓｔｅｍ，ＩＤＳ）是一种积极主动的安全防护技术，通过对网络进行实时监控，能够有效感知网络攻击行为，为安全管理人员提供相应决策㊂近年来，入侵检测系统已广泛地利用机器学习算法监控恶意活动，如贝叶斯网络㊁支持向量机㊁决策树等㊂此外，随着深度学习的快速发展，卷积神经网络㊁循环神经网络等陆续在入侵检测系统中得到广泛应用㊂然而，基于深度学习的检测系统高度依赖数据集，数据类间不平衡将严重影响检测的准确率㊂研究中发现，异常流量数据远小于征程流量数据，并且获取所需供给流量的途径有限㊂主要包括３方面：（１）采用真实攻击手段对网络用户进行入侵；（２）根据各种攻击代码的行为对其进行建模后，根据模型生成所需的攻击流量；（３）从真实存在的攻击事件中获取攻击流量［１］㊂第１种方法的网络攻击代码难以获取，第２种方法模型的准确率会影响攻击流量的可靠性，第３种方法不适合进行大规模的采集工作㊂数据增强（ＤａｔａＡｕｇｍｅｎｔａｔｉｏｎ）技术通过某些技术手段让有限的数据产生更多的等价数据，实现数据更加复杂的表征，能够一定程度缓解数据缺乏和类间不平衡等问题㊂在已有的研究中，成本函数等算法级别上的解决方案，在ＩＤＳ数据不平衡问题上的研究较少，部分研究是通过欠采样㊁过采样等方法解决该问题，但欠采样的方式缩小了整体的样本数量，而过采样的方式又容易引发过拟合问题，并不能较好地处理数据不平衡问题㊂２０１４年，Ｇｏｏｄｆｅｌｌｏｗ等学者［２］提出了一种新的生成模型生成对抗网络（ＧｅｎｅｒａｔｉｖｅＡｄｖｅｒｓａｒｉａｌＮｅｔｗｏｒｋｓ，ＧＡＮ），通过生成模型和判别模型的相互博弈学习生成高质量样本，能够很好地处理数据类别不平衡问题㊂然而，传统的ＧＡＮ模型对于网络入侵中的离散数据生成效果较差，生成的离散数据不能够以均匀的概率分布㊂为了解决传统ＧＡＮ模型的缺陷㊁更好地扩展网络入侵数据，本文对ＷＧＡＮ－ＧＰ模型的网络结构进行改进，结合特征选择算法，提出了ＧＡＮ－ＣＳ模型㊂将预处理后的数据输入ＧＡＮ－ＣＳ模型，生成的样本数据更加逼真和详细，整体效果更稳定，增强后数据多分类的准确性得到了提高㊂１㊀相关研究传统的数据增强方法（如：过采样），通过随机过采样和合成过采样来生成新的少数类样本，以均衡数据集中各类别的数量［３］㊂为了缓解随机过采样生成的新样本与原样本相似度高的问题，Ｃｈａｗｌａ等学者［４］提出ＳＭＯＴＥ算法，通过随机选择少数类样本点作为采样种子点，并使用线性插值的方法生成新的少数类样本㊂但是，当少数类样本由多个子群组成，并且多数类样本分布在子群中间时，线性插值法会生成与多数类重叠的样本，导致分类性能下降㊂Ｂａｒｕａ等学者［５］提出基于多数类加权的少数类样本过采样技术，用来确定边界的少数类样本㊂但该方法的性能在很大程度上取决于如何对少数类样本进行分区以及加权，并且所选样本可能存在冗余信息㊂Ｂｅｊ等学者［６］提出ＬｏＲＡＳ方法，通过基于样本点的凸集生成新的少数类样本㊂传统的过采样方法更适于处理低维数据，难以处理高维数据㊂随着深度学习的不断发展，在不均衡图像数据分类中会经常用到深度生成模型㊂其中变分自动编码器（ＶａｒｉａｔｉｏｎａｌＡｕｔｏｅｎｃｏｄｅｒ，ＶＡＥ）［７］，以及生成对抗网络（ＧＡＮ）被广泛应用㊂ＶＡＥ应用于给定的不均衡数据来捕捉特征维度间的关联性，进而获得样本在隐空间上的分布，最后通过解码器获得原始空间上的扩充数据集［８］㊂研究是利用最小平方误差，衡量生成样本分布与原始样本分布之间的距离㊂然而，基于元素点之间的误差无法很好地捕捉数据的真实分布㊂为了提升ＶＡＥ的性能，ＧＵＯ等学者［９］通过２个高斯分布，分别对多数类和少数类的隐空间变量进行建模，该模型适用于多分类数据㊂作为一种数据生成策略，ＧＡＮ能有效地学习隐空间到原始空间的映射函数㊂研究者提出了条件ＧＡＮ（ＣｏｎｄｉｔｉｏｎａｌＧＡＮ，ｃＧＡＮ）［１０］，用来生成特定类别的样本㊂基于ＧＡＮ的数据生成方法中，生成器的输入通常是随机噪声，可能会导致特征高度纠缠并破坏方向相关的特征［１１］㊂为了缓解该问题，研究人员提出ＢＡＧＡＮ［１２］，是将ＡＥ（Ａｕｔｏｅｎｃｏｄｅｒ）和ｃＧＡＮ集成在一起，将新的隐空间编码作为ｃＧＡＮ的输入㊂但该方法中，ＧＡＮ模式崩溃以及梯度消失和爆炸的问题仍无法避免，生成的数据甚至可能导致类别边界变形［１３］㊂Ｓａｌｅｍ等学者［１４］使用Ｃｙｃｌｅ－ＧＡＮ将ＡＤＦＡ－ＬＤ数据集转换为图像，再使用Ｃｙｃｌｅ－ＧＡＮ学习正常数据的图像来创建异常数据的图像，将生成的综合异常数据与原始数据一起用于模型的训练㊂实验结果表明，该方法优于综合采样技术，显示了生成对抗网络在异常生成中的潜力㊂Ｙｉｎ等学者［１５］提出Ｂｏｔ－ＧＡＮ模型㊂该模型根据各种流量的异常行为提取出相关特征，并结合ＧＡＮ生成假样本，经判别器判别为真样本后，可继续细分为正常或异常样本㊂实验结果表明，相较于原始检测模型，Ｂｏｔ－ＧＡＮ模型在测试集上有较高的检测准确率㊂然而，入侵检测数据属于非图像类数据，在将其转换成图像处理的过程中可能会带来精度损失，因此本文提出ＧＡＮ－ＣＳ模型解决ＩＤＳ数据的不平衡问题㊂２㊀关键技术２．１㊀生成对抗网络生成对抗网络（ＧＡＮ）中包含２个网络㊂一是生成网络Ｇ，用于生成假样本；另一个是判别网络Ｄ，用于判别样本的真假㊂这２个目标相反的网络不断地进行交替训练，当最后收敛时，如果判别网络再也无法判断出一个样本的来源，也就等价于生成网５７１第３期崔子才，等：基于ＧＡＮ和特征选择技术的入侵检测数据增强络可以生成符合真实数据分布的样本㊂ＧＡＮ的对抗博弈可以通过判别函数Ｄ（Ｘ）：Ｒｎ➝［０，１］和生成函数Ｇ：Ｒｄ➝Ｒｎ之间目标函数的极大极小值来进行数学化表示㊂生成器Ｇ将随机样本ｚɪＲｄ分布γ，转化为生成样本Ｇ（ｚ）㊂判别器Ｄ试图将其与来自分布μ的训练样本区分开来，而Ｇ试图使生成的样本在分布上与训练样本相似㊂ＧＡＮ解决的极小极大值的描述如下所示：ｍｉｎＧｍａｘＤＶ（Ｄ，Ｇ）：＝ｍｉｎＧｍａｘＤ（Ｅｘ μ［ｌｏｇＤ（ｘ）］＋㊀㊀㊀㊀㊀Ｅｚ γ［ｌｏｇ（１－Ｄ（Ｇ（ｚ）））］）（１）其中，Ｅ表示关于下标中指定分布的期望值㊂对于给定的生成器Ｇ，ｍａｘＤＶ（Ｄ，Ｇ）优化判别器Ｄ，以区分生成的样本Ｇ（ｚ）㊂其原理是尝试将高值分配给来自分布μ的真实样本，并将低值分配给生成的样本Ｇ（ｚ）㊂相反，对于给定的判别器Ｄ，ｍｉｎＧＶ（Ｄ，Ｇ）优化Ｇ，使得生成的样本Ｇ（ｚ）将试图误导判别器Ｄ以分配高值㊂ＧＡＮ算法描述如下㊂算法１㊀生成对抗网络的训练过程㊀㊀输入㊀训练集Ｄ，对抗训练迭代次数Ｔ，每次判别网络的训练迭代次数Ｋ，小批量样本数量Ｍ㊀㊀输出㊀生成对抗网络Ｇ（ｚ，θ）１．随机初始化θ和φ２．㊀Ｆｏｒｔ １ｔｏＴｄｏ３．㊀／／训练判别器网络Ｄ（ｘ，φ）４．㊀Ｆｏｒｋ－１ｔｏＫｄｏ５．㊀㊀／／采集小批量训练成本６．㊀㊀从训练集合Ｄ中采集Ｍ个样本｛ｘ（ｍ）｝，１ɤｍɤＭ７．㊀㊀从分布Ｎ（０，１）中采集Ｍ个样本｛ｚ（ｍ）｝，这里１ɤｍɤＭ８．㊀㊀使用随机梯度上升更新φ，梯度为：９．㊀㊀∂∂φ［１ＭðＭｍ＝１（ｌｏｇＤ（ｘ（ｍ），φ）＋ｌｏｇ（１－Ｄ（Ｇ（ｚ（ｍ），θ），φ）））］１０．Ｅｎｄｆｏｒ１１．／／训练判别器网络Ｄ（ｚ，θ）１２．从分布Ｎ（０，１）中采集Ｍ个样本｛ｚ（ｍ）｝，１ɤｍɤＭ１３．使用随机梯度上升更新θ，梯度为：１４．∂∂θ［１ＭðＭｍ＝１（Ｄ（Ｇ（ｚ（ｍ），θ），φ））］１５．Ｅｎｄｆｏｒ２．２㊀ＷＧＡＮ－ＧＰ在实际训练中，ＧＡＮ经常遇到的问题：一是模式崩溃，生成器生成非常窄的分布，仅覆盖数据分布中的单一模式，即生成器只能生成非常相似的样本；二是没有指标可以表征收敛情况㊂总之，判别器越好，生成器梯度消失越严重㊂在判别器最优的前提下，把原始ＧＡＮ定义的生成器ｌｏｓｓ等价变换为最小化真实分布与生成分布之间的ＪＳ散度，最小化生成器的ｌｏｓｓ即近似于最小化真实分布与生成分布之间的ＪＳ散度㊂若希望２个分布之间的ＪＳ散度越小，通过优化ＪＳ散度就能将生成分布转化为真实分布，最终实现以假乱真；若２个分布完全没有重叠的部分，或者相互重叠的部分可忽略，则两者之间的ＪＳ散度就一直是ｌｏｇ２㊂因此，原始ＧＡＮ问题的根源可以归结为２点：等价优化的距离衡量（ＪＳ散度）不合理，以及生成器随机初始化后的生成分布很难与真实分布有不可忽略的重叠㊂基于以上问题，研究者提出了Ｗａｓｓｅｒｓｔｅｉｎ距离，即Ｅａｒｔｈ－Ｍｏｖｅ距离衡量２个分布之间的距离［１６］㊂其优越性在于，即使２个分布没有任何重叠，也可以反映两者之间的距离，公式如下㊂Ｗ（Ｐｒ，Ｐｇ）＝１ＫＳＵＰｆ ＬɤＫＥｘ Ｐｒ［ｆ（ｘ）］－Ｅｘ Ｐｇ［ｆ（ｘ）］（２）其中，Ｐｒ为最小化真实分布，Ｐｇ为生成分布㊂ＷＧＡＮ算法描述如下㊂算法２㊀ＷＧＡＮ的训练过程输入㊀学习率α，裁剪参数ｃ，对抗训练迭代次数Ｔ，每次判别网络的训练迭代次数Ｋ㊂初始临界参数ｗ０，初始生成器参数θ０１．㊀ｗｈｉｌｅθ未收敛ｄｏ２．㊀ｆｏｒｔ＝０，１， ，Ｋｄｏ３．㊀㊀ｘ（ｉ）{}Ｔｉ＝１ Ｐｒａｂａｔｃｈｆｒｏｍｒｅａｌｄａｔａ４．㊀㊀ｚ（ｉ）{}Ｔｉ＝１ ｐ（ｚ）ａｂａｔｃｈｏｆｐｒｉｏｒｓ５．㊀㊀ｇｗѳÑｗ［１ＴðＴｉ＝１ｆｗ（ｘ（ｉ））－１ＴðＴｉ＝１ｆｗ（ｇθ（ｚ（ｉ）））］６．㊀㊀ｗѳｗ＋α㊃ＲＭＳＰｒｏｐ（ｗ，ｇｗ）７．㊀㊀ｗѳｃｌｉｐ（ｗ，－ｃ，ｃ）８．㊀Ｅｎｄｆｏｒ９．㊀ｚ（ｉ）{}Ｔｉ＝１ ｐ（ｚ）ａｂａｔｃｈｏｆｐｒｉｏｒｓａｍｐｌｅｓ１０．㊀ｇθѳ－Ñθ１ＴðＴｉ＝１ｆｗ（ｇθ（ｚ（ｉ）））１１．㊀θѳθ－α㊃ＲＭＳＰｒｏｐ（θ，ｇθ）１２．㊀㊀∂∂θ［１ＭðＭｍ＝１（Ｄ（Ｇ（ｚ（ｍ），θ），φ））］６７１智㊀能㊀计㊀算㊀机㊀与㊀应㊀用㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀第１４卷㊀１３．㊀Ｅｎｄｗｈｉｌｅ㊀㊀然而，ＷＧＡＮ中权重裁剪的实现方式存在２个重要问题：（１）判别器的ｌｏｓｓ希望尽可能拉大真假样本的分数差，实验发现基本上最终权重集中在两端，这样参数的多样性会减少，使判别器得到的神经网络学习一个简单的映射函数，造成巨大的浪费㊂（２）容易导致梯度消失或者梯度爆炸，若把裁剪阈值设得较小，每经过一个网络，梯度就会变小，多级之后会成为指数衰减；反之则会导致指数爆炸㊂为此，ＷＧＡＮ－ＧＰ引入了梯度惩罚项（ＧｒａｄｉｅｎｔＰｅｎａｌｔｙ）［１７］㊂当且仅当一个可微函数的梯度范数（ＧｒａｄｉｅｎｔＮｏｒｍ）在任意处都不超过１时，该函数满足１－Ｌｉｐｓｃｈｉｔｚ条件㊂损失函数公式如下：Ｌ＝Ｅｘ Ｐｇ［Ｄ（ｘ ）］－Ｅｘ Ｐｒ［Ｄ（ｘ）］＋λＥ＾ｘ Ｐ¥（Ñ＾ｘ ２－１）２[]（３）３㊀ＧＡＮ－ＣＳ模型本文提出ＧＡＮ－ＣＳ模型对不平衡数据进行增强，并对模型的性能进行评估㊂模型框架如图１所示㊂由图１可知，模型包括数据预处理㊁数据增强㊁特征选择及性能评估四部分㊂为验证模型生成的数据质量，分别使用处理后的数据集和混合数据集训练相同的分类器，通过比较２种数据集下的多分类结果，评估本文所提模型的性能㊂原始数据集生成数据混合数据集数值化子数据集训练&测试M L模型特征选择数据增强归一化图１㊀模型框架Ｆｉｇ．１㊀Ｍｏｄｅｌｆｒａｍｅｗｏｒｋｄｉａｇｒａｍ３．１㊀数据预处理入侵检测数据通常包含非数字特征，例如协议和状态等㊂为了能更好地被计算机识别和处理，需先通过ｏｎｅ－ｈｏｔ编码方法将入侵检测数据集中存在的离散型数据进行数值化，然后将所有数字特征进行归一化，以保证消除数据的可读性和消除异常值㊂这些非数字特征需要转换为数字特征以适合本文模型㊂非数字特征映射到０到Ｓ－１之间的整数值，其中Ｓ是符号数㊂不同维度的数据特征尺度不一致会影响入侵检测的结果㊂需要对数据进行归一化处理，以消除指标之间的维度影响㊂除了攻击类型标签，将所有特征缩放到［０，１］㊂ｍｉｎ－ｍａｘ归一化用于线性缩放数据值，如下所示：ｘᶄ＝ｘ－ｘｍｉｎｘｍａｘ－ｘｍｉｎ（４）㊀㊀其中，ｘ为归一化前的值；ｘᶄ为归一化后的值；ｘｍａｘ为样本数据的最大值；ｘｍｉｎ为样本数据的最小值㊂将预处理后的数据集通过改进后的ＷＧＡＮ－ＧＰ模型进行扩充，与预处理数据集混合，形成混合数据集㊂３．２㊀特征选择特征选择是一种数据降维方法，常用于处理高维㊁复杂的数据㊂从所有的特征中，选择出有意义㊁对模型有帮助的特征，以避免必须将所有特征都导入模型去训练的情况，从而提升模型的训练速度和效率，并提升准确率㊂卡方检验是以χ２分布为基础的一种常用假设检验方法㊂方法的无效假设Ｈ０是：观察频数与期望频数没有差别［１８］㊂该检验的基本思想是：首先假设Ｈ０成立，基于此前提计算出χ２值，则表示观察值与７７１第３期崔子才，等：基于ＧＡＮ和特征选择技术的入侵检测数据增强理论值之间的偏离程度㊂根据χ２分布及自由度可以确定在Ｈ０假设成立的情况下获得当前统计量及更极端情况的概率Ｐ㊂如果当前统计量大于Ｐ值，说明观察值与理论值偏离程度太大，应当拒绝无效假设，表示比较资料之间有显著差异；否则就不能拒绝无效假设，尚不能认为样本所代表的实际情况和理论假设有差别㊂其公式如下：χ２＝ð（Ａ－Ｅ）２Ｅ（５）㊀㊀其中，Ａ为实际值，Ｅ为理论值㊂４㊀实验结果与分析４．１㊀ＵＮＳＷ－ＮＢ１５数据集ＵＮＳＷ－ＮＢ１５数据集［１９］由澳大利亚网络安全中心的网络靶场实验室创建㊂该数据集包含各种新颖的攻击，因此已广泛用于入侵检测㊂其中包含９种类型的攻击来模拟真实网络环境，即Ｆｕｚｚｅｒｓ㊁Ａｎａｌｙｓｉｓ㊁Ｂａｃｋｄｏｏｒ㊁ＤｏＳ㊁Ｅｘｐｌｏｉｔｓ㊁Ｇｅｎｅｒｉｃ㊁Ｒｅｃｏｎｎａｉｓｓａｎｃｅ㊁Ｓｈｅｌｌｃｏｄｅ和Ｗｏｒｍｓ㊂ＵＮＳＷ－ＮＢ１５数据集包含一个训练集和一个测试集㊂训练集有８２３３２条记录，测试集有１７５３４１条记录㊂ＵＮＳＷ－ＮＢ１５数据集构成如图２所示㊂由图２可知，不同种类流量数据分布严重不均，且数据间存在数量的差异㊂如Ｗｏｒｍｓ在数据集中的分布仅占不足０．１％，而Ｎｏｍａｌ的数量超过整体数据集中的１／３㊂N o r m a lR e c o n n a i s s a n c eS h e l l c o d eWo r m sB a c k d o o r D o sE x p l o i t sF u z z e r sG e n e r i cA n a l y s i s图２㊀ＵＮＳＷ－ＮＢ１５数据集构成Ｆｉｇ．２㊀ＵＮＳＷ－ＮＢ１５ｄａｔａｓｅｔ４．２㊀评价指标数据在二分类问题中可被分为正样本和负样本，并将数据按照真实类别和预测类别划分为４种类型：当样本预测为正且实际为正时的真阳性（ＴＰ），当样本预测为负且实际为负时的真阴性（ＴＮ），当样本预测为正但实际为负时的假阳性（ＦＰ），当样本预测为负但实际上为正时的假阴性（ＦＮ）㊂为了评价本文所提模型的性能，采用识别准确率（Ａｃｃｕｒａｒｙ）㊁精确率（Ｐｒｅｃｉｓｉｏｎ）㊁召回率（Ｒｅｃａｌｌ）㊁Ｆ值（Ｆ－Ｍｅａｓｕｒｅ）作为评价指标㊂（１）准确率（Ａｃｃｕｒａｃｙ）㊂研究推得的定义公式为：Ａ＝ＴＰ＋ＴＮＴＰ＋ＦＰ＋ＴＮ＋ＦＮ（６）㊀㊀（２）精确率（Ｐｒｅｃｉｓｉｏｎ）㊂精确率又称查准率，是常用的评价指标，就是计算所有被预测为正的样本中实际为正样本的概率，公式如下：Ｐ＝ＴＰＴＰ＋ＦＰ（７）㊀㊀（３）召回率（Ｒｅｃａｌｌ）㊂召回率又称查全率，就是计算实际为正的样本中被预测为正样本的概率，公式如下：Ｒ＝ＴＰＴＰ＋ＦＮ（８）㊀㊀其中，ＦＮ（ＦａｌｓｅＮｅｇａｔｉｖｅ）表示将正样本预测为负样本的数量㊂（４）Ｆ１值（Ｆ１－ｖａｌｕｅ）㊂为避免精确率和召回率相矛盾的情况，需要将两者进行综合考虑，最常见的方法为Ｆ值，就是计算精确率和召回率的加权调和平均，公式如下：Ｆ＝α２＋１()ˑＰˑＲα２ˑＰ＋Ｒ()（９）㊀㊀其中，当参数α＝１时，即为Ｆ１值㊂４．３㊀实验结果与分析本文所有实验结果均由１０次实验后取平均值得到㊂研究内容分述如下㊂（１）实验一：数据增强数据量选择实验㊂为更好地体现本文提出模型的数据增强的效果，在使用同一分类器（实验选择决策树）的前提下，分别将不同攻击类别的数据由原来的数据量增加１００００㊁增加至１００００㊁２００００和３７０００（Ｎｏｒｍａｌ的数据量为３７０００）（见表１），并对混合后数据集中的准确率和原数据集的准确率做了比较（见表２）㊂由此可见，当攻击样本数据增加至２００００时，分类器的准确率最高㊂８７１智㊀能㊀计㊀算㊀机㊀与㊀应㊀用㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀第１４卷㊀表１㊀数据增强前后数据量对比Ｔａｂｌｅ１㊀Ｃｏｍｐａｒｉｓｏｎｏｆｄａｔａｖｏｌｕｍｅｂｅｆｏｒｅａｎｄａｆｔｅｒｄａｔａａｕｇｍｅｎｔａｔｉｏｎ攻击类别增强前数量增强１００００增强至１００００增强至２００００增强至３７０００Ａｎａｌｙｓｉｓ６７７１０６７７１００００２００００３７０００Ｂａｃｋｄｏｏｒ５８３１０５８３１００００２００００３７０００Ｄｏｓ４０８９１４０８９１００００２００００３７０００Ｅｘｐｌｏｉｔｓ１１１３２１１１３２１１１３２２００００３７０００Ｆｕｚｚｅｒｓ６０６２１６０６２１００００２００００３７０００Ｇｅｎｅｒｉｃ１８８７１１８８７１１８８７１２００００３７０００Ｎｏｒｍａｌ３７０００３７０００３７０００３７０００３７０００Ｒｅｃｏｎｎａｉｓｓａｎｃｅ３４９６１３４９６１００００２００００３７０００Ｓｈｅｌｌｃｏｄｅ３７８１０３７８１００００２００００３７０００Ｗｏｒｍｓ４４１００４４１００００２００００３７０００表２㊀数据增强前后准确率对比（使用决策树分类器）Ｔａｂｌｅ２㊀Ｃｏｍｐａｒｉｓｏｎｏｆａｃｃｕｒａｃｙｂｅｆｏｒｅａｎｄａｆｔｅｒｄａｔａａｕｇｍｅｎｔａｔｉｏｎ（ｕｓｉｎｇｄｅｃｉｓｉｏｎｔｒｅｅｃｌａｓｓｉｆｉｅｒ）数据量准确率增强前０．４９８３增强１０００００．６９６５增强至１０００００．４６５７增强至２０００００．７０２５增强至３７００００．６９９８㊀㊀（２）实验二：模型有效性实验㊂在实验一的基础上，使用决策树㊁随机森林㊁ＫＮＮ和神经网络对原数据集和混合数据集进行多分类，比较结果如图３所示㊂由图３可知，每个分类器的准确率都有不同程度的提高，表明本文所提模型可以提高分类器的整体性能，数据增强模型是有效的，且效果明显㊂所有分类器的召回率也都有不同程度的提高，表明本文生成的样本提高了攻击样本的多样性，从而增强了分类器知识学习的泛化性㊂由于不同的检测器具有不同的学习能力，因此影响程度存在一定差异㊂最后，分类器的Ｆ１值表明分类器的整体性能得到了有效的提升㊂（a ）准确率（b ）查准率（c ）召回率（d ）F 1值N B D TR FK N N 神经网络N B D T R F K N N神经网络N BD T R F K N N 神经网络N B D TR FK N N 神经网络增强前增强后增强前增强后增强前增强后增强前增强后0.80.70.60.50.40.30.20.100.80.70.60.50.40.30.20.100.90.80.70.60.50.40.30.20.100.80.70.60.50.40.30.20.10.57250.60730.49830.73590.72190.79630.71540.72540.63730.73680.52580.70640.50430.75310.70210.81260.69940.73600.59210.75110.57250.56330.49630.68480.72190.72170.71540.70090.63730.69920.53060.64190.44270.73160.67660.79720.68150.72910.54830.7559图３㊀增强前后对比Ｆｉｇ．３㊀Ｃｏｍｐａｒｉｓｏｎｂｅｆｏｒｅａｎｄａｆｔｅｒｅｎｈａｎｃｅｍｅｎｔ㊀㊀（３）实验三：模型优越性比较实验㊂为了证明本文提出的数据增强模型ＧＡＮ－ＣＳ优于同类模型，本实验将ＷＧＡＮ－ＧＰ和ＧＡＮ－ＦＳ作为对比对象㊂实验结果见表３㊂９７１第３期崔子才，等：基于ＧＡＮ和特征选择技术的入侵检测数据增强表３㊀本文模型与其他模型在ＵＮＳＷ－ＮＢ１５数据增强效果的比较Ｔａｂｌｅ３㊀ＣｏｍｐａｒｉｓｏｎｏｆｔｈｅｅｎｈａｎｃｅｍｅｎｔｅｆｆｅｃｔｏｆｔｈｉｓｍｏｄｅｌｗｉｔｈｏｔｈｅｒｍｏｄｅｌｓｉｎＵＮＳＷ－ＮＢ１５ｄａｔａ分类器ＮＢＡｃｃｕｒａｃｙＦ１ＤＴＡｃｃｕｒａｃｙＦ１ＲＦＡｃｃｕｒａｃｙＦ１ＫＮＮＡｃｃｕｒａｃｙＦ１神经网络ＡｃｃｕｒａｃｙＦ１ＷＧＡＮ－ＧＰ［１７］０．５６３３０．５２３６０．７０１５０．６８４８０．７２１７０．６８３５０．７００９０．６７４８０．６７９２０．５８６６ＧＡＮ－ＦＳ［２０］０．５９７００．５３８１０．６８６００．６５９７０．７２０６０．６８２８０．７００９０．６７４８０．６８３２０．５８９７ＧＡＮ－ＣＳ０．６０７３０．５４１９０．７３５９０．７０１６０．７９６３０．７３７２０．７２５４０．６８９１０．７３６８０．６１５９㊀㊀从表３中数据可以看出，基于本文提出的模型的分类器性能高于其他方案，且ＧＡＮ－ＣＳ在ＤＴ和ＲＦ上表现更好㊂ＤＴ模型的思想是使用信息熵作为度量来构建熵下降最快的树，ＷＧＡＮ－ＧＰ算法基于原始分布生成样本，增加了样本的多样性，并通过特征选择去除了不必要的特征，因此，经过过采样后，决策树可以更好地对样本进行分类㊂与ＤＴ相比，ＲＦ是综合学习算法，学习能力更优异，可以提高分类性能㊂试验结果还表明ＲＦ的性能普遍高于ＤＴ㊂５㊀结束语为缓解数据缺乏和类间不平衡等问题，本文提出了ＧＡＮ－ＣＳ数据增强模型㊂将预处理过的数据集通过改进后的ＷＧＡＮ－ＧＰ模型中进行增强，生成后的数据和原数据集混合后得到的混合数据集经特征选择后用于训练入侵检测多分类器㊂经过３种不同维度的实验可以得出，本文方法提高了入侵检测模型的性能，并且优于其他同类方案㊂虽然好的数据增强模型能够提升入侵检测的性能，但分类器又影响了入侵检测模型性能的提升，因此设计一个好的分类器将成为下一步的研究方向㊂参考文献［１］陈家浩，王轶骏，吕诚．一种基于Ｐｙｔｈｏｎ符号执行的自动化网络攻击流量获取方法［Ｊ］．计算机应用与软件，２０１９，３６（２）：２９４－３０７．［２］ＧＯＯＤＦＥＬＬＯＷＩＪ，ＰＯＵＧＥＴ－ＡＢＡＤＩＥＪ，ＭＩＲＺＡＭ，ｅｔａｌ．Ｇｅｎｅｒａｔｉｖｅａｄｖｅｒｓａｒｉａｌｎｅｔｓ［Ｃ］／／Ｐｒｏｃｅｅｄｉｎｇｓｏｆｔｈｅ２７ｔｈＩｎｔｅｒｎａｔｉｏｎａｌＣｏｎｆｅｒｅｎｃｅｏｎＮｅｕｒａｌＩｎｆｏｒｍａｔｉｏｎＰｒｏｃｅｓｓｉｎｇＳｙｓｔｅｍｓ．Ｃａｍｂｒｉｄｇｅ：ＭＩＴＰｒｅｓｓ，２０１４：２６７２２６８０．［３］王馨月．生成式数据增强的不均衡数据分类方法研究［Ｄ］．北京：北京交通大学，２０２１．［４］ＣＨＡＷＬＡＮＶ，ＢＯＷＹＥＲＫＷ，ＨＡＬＬＬＯ，ｅｔａｌ．ＳＭＯＴＥ：ｓｙｎｔｈｅｔｉｃｍｉｎｏｒｉｔｙｏｖｅｒ－ｓａｍｐｌｉｎｇｔｅｃｈｎｉｑｕｅ［Ｊ］．ＪｏｕｒｎａｌｏｆＡｒｔｉｆｉｃｉａｌＩｎｔｅｌｌｉｇｅｎｃｅＲｅｓｅａｒｃｈ，２００２，１６：３２１－３５７．［５］ＢＡＲＵＡＳ，ＩＳＩＡＭＭＭ，ＹＡＯＸｉｎ，ｅｔａｌ．ＭＷＭＯＴＥ ｍａｊｏｒｉｔｙｗｅｉｇｈｔｅｄｍｉｎｏｒｉｔｙｏｖｅｒｓａｍｐｌｉｎｇｔｅｃｈｎｉｑｕｅｆｏｒｉｍｂａｌａｎｃｅｄｄａｔａｓｅｔｌｅａｒｎｉｎｇ［Ｊ］．ＩＥＥＥＴｒａｎｓａｃｔｉｏｎｓｏｎＫｎｏｗｌｅｄｇｅａｎｄＤａｔａＥｎｇｉｎｅｅｒｉｎｇ，２０１２，２６：４０５－４２５．［６］ＢＥＪＳ，ＮＡＲＥＫＤ，ＭＡＲＫＵＳＷ，ｅｔａｌ．ＬｏＲＡＳ：Ａｎｏｖｅｒｓａｍｐｌｉｎｇａｐｐｒｏａｃｈｆｏｒｉｍｂａｌａｎｃｅｄｄａｔａｓｅｔｓ［Ｊ］．ＭａｃｈｉｎｅＬｅａｒｎｉｎｇ，２０２１，１１０：２７９－３０１．［７］ＫＩＮＧＭＡＤＰ，ＷＥＬＬＩＮＧＭ．Ａｕｔｏ－ｅｎｃｏｄｉｎｇｖａｒｉａｔｉｏｎａｌｂａｙｅｓ［Ｊ］．ａｒＸｉｖｐｒｅｐｒｉｎｔａｒＸｉｖ：１３１２．６１１４，２０１３．［８］ＷＡＮＺｈｉｑｉａｎｇ，ＺＨＡＮＧＹａｚｈｏｕ，ＨＥＨａｉｂｏ．Ｖａｒｉａｔｉｏｎａｌａｕｔｏｅｎｃｏｄｅｒｂａｓｅｄｓｙｎｔｈｅｔｉｃｄａｔａｇｅｎｅｒａｔｉｏｎｆｏｒｉｍｂａｌａｎｃｅｄｌｅａｒｎｉｎｇ［Ｃ］／／２０１７ＩＥＥＥＳｙｍｐｏｓｉｕｍＳｅｒｉｅｓｏｎＣｏｍｐｕｔａｔｉｏｎａｌＩｎｔｅｌｌｉｇｅｎｃｅ（ＳＳＣＩ）．Ｈｏｎｏｌｕｌｕ，ＵＳＡ：ＩＥＥＥ，２０１７：１－７．［９］ＧＵＯＴｉｎｇ，ＺＨＵＸｉｎｇｑｕａｎ，ＷＡＮＧＹａｎｇ，ｅｔａｌ．Ｄｉｓｃｒｉｍｉｎａｔｉｖｅｓａｍｐｌｅｇｅｎｅｒａｔｉｏｎｆｏｒｄｅｅｐｉｍｂａｌａｎｃｅｄｌｅａｒｎｉｎｇ［Ｃ］／／Ｔｗｅｎｔｙ－ＥｉｇｈｔｈＩｎｔｅｒｎａｔｉｏｎａｌＪｏｉｎｔＣｏｎｆｅｒｅｎｃｅｏｎＡｒｔｉｆｉｃｉａｌＩｎｔｅｌｌｉｇｅｎｃｅ．Ｍａｃａｏ，Ｃｈｉｎａ：ＩｎｔｅｒｎａｔｉｏｎａｌＪｏｉｎｔＣｏｎｆｅｒｅｎｃｅｓｏｎＡｒｔｉｆｉｃｉａｌＩｎｔｅｌｌｉｇｅｎｃｅＯｒｇａｎｉｚａｔｉｏｎ，２０１９：２４０６－２４１２．［１０］ＧＡＵＴＨＩＥＲＪ．Ｃｏｎｄｉｔｉｏｎａｌｇｅｎｅｒａｔｉｖｅａｄｖｅｒｓａｒｉａｌｎｅｔｓｆｏｒｃｏｎｖｏｌｕｔｉｏｎａｌｆａｃｅｇｅｎｅｒａｔｉｏｎ［Ｊ］．ＣｌａｓｓＰｒｏｊｅｃｔｆｏｒＳｔａｎｆｏｒｄＣＳ２３１Ｎ：ＣｏｎｖｏｌｕｔｉｏｎａｌＮｅｕｒａｌＮｅｔｗｏｒｋｓｆｏｒＶｉｓｕａｌＲｅｃｏｇｎｉｔｉｏｎ，２０１４（５）：２．［１１］ＮＡＺＲＵＬＨ，ＢＨＡＴＴＡＣＨＡＲＹＹＡＤＫ，ＫＡＬＩＴＡＪＫ．ＢｏｔｎｅｔｉｎＤＤｏＳａｔｔａｃｋｓ：Ｔｒｅｎｄｓａｎｄｃｈａｌｌｅｎｇｅｓ［Ｊ］．ＩＥＥＥＣｏｍｍｕｎｉｃａｔｉｏｎｓＳｕｒｖｅｙｓ＆Ｔｕｔｏｒｉａｌｓ，２０１５，１７（４）：２２４２－２２７０．［１２］ＭＡＲＩＡＮＩＧ，ＳＣＨＥＩＤＥＧＧＥＲＦ，ＩＳＴＲＡＴＥＲ，ｅｔａｌ．ＢＡＧＡＮ：ＤａｔａａｕｇｍｅｎｔａｔｉｏｎｗｉｔｈｂａｌａｎｃｉｎｇＧＡＮ［Ｊ］．ａｒＸｉｖｐｒｅｐｒｉｎｔａｒＸｉｖ：１８０３．０９６５５，２０１８．［１３］ＳＡＮＴＵＲＫＡＲＳ，ＳＣＨＭＩＤＴＬ，ＭＡＤＲＹＡ．Ａｃｌａｓｓｉｆｉｃａｔｉｏｎ－ｂａｓｅｄｓｔｕｄｙｏｆｃｏｖａｒｉａｔｅｓｈｉｆｔｉｎＧＡＮｄｉｓｔｒｉｂｕｔｉｏｎｓ［Ｃ］／／ＩｎｔｅｒｎａｔｉｏｎａｌＣｏｎｆｅｒｅｎｃｅｏｎＭａｃｈｉｎｅＬｅａｒｎｉｎｇ．Ｓｔｏｃｋｈｏｌｍ，Ｓｗｅｄｅｎ：ＰＭＬＲ，２０１８：４４８０－４４８９．［１４］ＳＡＬＥＭＭ，ＴＡＨＥＲＩＳ，ＹＵＡＮＪＳ．Ａｎｏｍａｌｙｇｅｎｅｒａｔｉｏｎｕｓｉｎｇｇｅｎｅｒａｔｉｖｅａｄｖｅｒｓａｒｉａｌｎｅｔｗｏｒｋｓｉｎｈｏｓｔ－ｂａｓｅｄｉｎｔｒｕｓｉｏｎｄｅｔｅｃｔｉｏｎ［Ｃ］／／２０１８９ｔｈＩＥＥＥＡｎｎｕａｌＵｂｉｑｕｉｔｏｕｓＣｏｍｐｕｔｉｎｇ，Ｅｌｅｃｔｒｏｎｉｃｓ＆ＭｏｂｉｌｅＣｏｍｍｕｎｉｃａｔｉｏｎＣｏｎｆｅｒｅｎｃｅ（ＵＥＭＣＯＮ）．ＮｅｗＹｏｒｋ，ＵＳＡ：ＩＥＥＥ，２０１８：６８３－６８７．［１５］ＹＩＮＣｈｕａｎｌｏｎｇ，ＺＨＵＹｕｅｌｅｉ，ＬＩＵＳｈｅｎｇｌｉ．Ａｎｅｎｈａｎｃｉｎｇｆｒａｍｅｗｏｒｋｆｏｒｂｏｔｎｅｔｄｅｔｅｃｔｉｏｎｕｓｉｎｇｇｅｎｅｒａｔｉｖｅａｄｖｅｒｓａｒｉａｌｎｅｔｗｏｒｋｓ［Ｃ］／／２０１８ＩｎｔｅｒｎａｔｉｏｎａｌＣｏｎｆｅｒｅｎｃｅｏｎＡｒｔｉｆｉｃｉａｌＩｎｔｅｌｌｉｇｅｎｃｅａｎｄＢｉｇＤａｔａ（ＩＣＡＩＢＤ）．Ｃｈｅｎｇｄｕ，Ｃｈｉｎａ：ＩＥＥＥ，２０１８：２２８－２３４．［１６］ＡＲＪＯＶＳＫＹＭ，ＣＨＩＮＴＡＬＡＳ，ＢＯＴＴＯＵＬ．Ｗａｓｓｅｒｓｔｅｉｎｇｅｎｅｒａｔｉｖｅａｄｖｅｒｓａｒｉａｌｎｅｔｗｏｒｋｓ［Ｃ］／／ＩｎｔｅｒｎａｔｉｏｎａｌＣｏｎｆｅｒｅｎｃｅｏｎＭａｃｈｉｎｅＬｅａｒｎｉｎｇ．Ｓｙｄｎｅｙ：ＰＭＬＲ，２０１７：２１４－２２３．［１７］ＧＵＬＲＡＪＡＮＩＩ，ＡＨＭＥＤＦ，ＡＲＪＯＶＳＫＹＭ，ｅｔａｌ．ＩｍｐｒｏｖｅｄｔｒａｉｎｉｎｇｏｆＷａｓｓｅｒｓｔｅｉｎＧＡＮｓ［Ｃ］／／Ｐｒｏｃｅｅｄｉｎｇｓｏｆｔｈｅ３１ｓｔＩｎｔｅｒｎａｔｉｏｎａｌＣｏｎｆｅｒｅｎｃｅｏｎＮｅｕｒａｌＩｎｆｏｒｍａｔｉｏｎＰｒｏｃｅｓｓｉｎｇＳｙｓｔｅｍｓ．ＬｏｎｇＢｅａｃｈ：ＣｕｒｒａｎＡｓｓｏｃｉａｔｅｓＩｎｃ．，２０１７：５７６９－５７７９．［１８］陈谌，梁雪春．基于基尼指标和卡方检验的特征选择方法［Ｊ］．计算机工程与设计，２０１９，４０（８）：２３４２－２３４５，２３６０．［１９］ＺＯＧＨＩＺ，ＳＥＲＰＥＮＧ．ＵＮＳＷ－ＮＢ１５ｃｏｍｐｕｔｅｒｓｅｃｕｒｉｔｙｄａｔａｓｅｔ：Ａｎａｌｙｓｉｓｔｈｒｏｕｇｈｖｉｓｕａｌｉｚａｔｉｏｎ［Ｊ］．ａｒＸｉｖｐｒｅｐｒｉｎｔａｒＸｉｖ：２１０１．０５０６７，２０２１．［２０］ＬＩＵＸｉａｏｄｏｎｇ，ＬＩＴ，ＺｈａｎｇＲｕｎｚｉ，ｅｔａｌ．ＡＧＡＮａｎｄｆｅａｔｕｒｅｓｅｌｅｃｔｉｏｎ－ｂａｓｅｄｏｖｅｒｓａｍｐｌｉｎｇｔｅｃｈｎｉｑｕｅｆｏｒｉｎｔｒｕｓｉｏｎｄｅｔｅｃｔｉｏｎ［Ｊ］．ＳｅｃｕｒｉｔｙａｎｄＣｏｍｍｕｎｉｃａｔｉｏｎＮｅｔｗｏｒｋｓ，２０２１（１）：１－１５．０８１智㊀能㊀计㊀算㊀机㊀与㊀应㊀用㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀㊀第１４卷㊀。
电压源的英语单词
Voltage
Electrical potential
The voltage difference between two points in an electric field
The electrical potential difference between two points in a circuit
02
Electrical potential difference
The electrical potential difference is the difference in potential energy between two points in a circuit
03
Electrical potential
03 WorkinБайду номын сангаас principle of voltage s
Voltage generation
Voltage is generated in a closed loop circuit when there is a difference in potential energy between two points This potential difference is caused by the presence of a resistor or conductor in the circuit
01 Overview of voltage sources
Definition of voltage source
Voltage source
An electronic device that converts energy into electrical energy and outputs a preset voltage
IGBT瞬态短路失效分析及其有限元热电耦合模型研究
IGBT瞬态短路失效分析及其有限元热电耦合模型研究窦智峰;晋玉祥;郭新飞【摘要】针对传统热网络模型不适用于IGBT短路情况下结温测量的问题,通过现场瞬态短路破坏性试验,在分析现场瞬态短路情况下IGBT失效机理的基础上,界定了IGBT短路失效时的临界能量值,发现了临界能量值在不同初始温度和不同母线电压情况下的演化规律,并建立了有限元热电耦合模型.结果表明,随着直流母线电压、初始温度和电流密度的增加,IGBT的短路维持时间和临界能量值会大幅度减小,而IGBT发生失效的临界温度点与初始温度的大小无关;在短路过程中IGBT失效前热量的传递仅达到焊料层,且短路瞬间其结温最高点分布在耗尽层的边界处.【期刊名称】《轻工学报》【年(卷),期】2018(033)006【总页数】8页(P101-108)【关键词】绝缘栅型晶闸管;瞬态短路;结温测量;临界能量值;有限元热电耦合模型【作者】窦智峰;晋玉祥;郭新飞【作者单位】[1]郑州轻工业学院、电气信息工程学院,河南郑州450002;[1]郑州轻工业学院、电气信息工程学院,河南郑州450002;[1]郑州轻工业学院、电气信息工程学院,河南郑州450002;【正文语种】中文【中图分类】TM460 引言绝缘栅型晶闸管IGBT(insulated gate bipolar transistor)具有驱动简单、功率密度大、功耗小、热稳定性好等性能，被广泛应用于各种中、大功率电能变换装置，是目前应用较为广泛的全控型电力电子器件[1-2].作为各种电能变换的主功率开关器件，在某些工作环境恶劣的应用场合，如海上风力发电、航空航天电能转换等，电能转换器内外部工作条件的巨大差异，常导致功率器件的可靠性降低，使失效机率大幅增加，最终有可能造成严重损失或灾难[3-4].根据IGBT产生失效历经时间的长短，可将IGBT的失效分为老化失效和现场失效.造成IGBT老化失效的主要原因是IGBT模块在导通和关断的过程中受到高频率热冲击的影响，加之不同封装工艺下IGBT各层异质结构的热膨胀系数不均衡，使IGBT在受到长时间过应力作用时表现出不同的失效模式[5-6].现场失效是IGBT在经受瞬间的过应力冲击之后的失效.IGBT的现场失效又可分为开路失效和短路失效.从系统可靠性的角度分析，开路失效故障的影响可以被隔离并限定在一定范围内，因此，开路失效对系统本身安全运行的影响并不是致命的[5].造成短路失效的原因主要包括过电压、闩锁效应、二次击穿和过能量冲击.由于IGBT在各种短路失效模式下都伴随着很高的电流上升率和瞬间高温的现象，因此这种致命的失效模式对IGBT的损坏是不可逆转的[7-8]，如果不加以保护和控制，严重时会导致器件烧毁、爆炸、系统停机，甚至造成更大的损失.根据失效的部位不同，又可将IGBT失效分为封装失效和芯片失效.现有IGBT均采取叠层封装技术，该技术有效提高了封装密度和器件的运行效率.但在功率循环或温度循环中，由于器件各层材料的热膨胀系数不同，在温度变化时热应变不同，因此连接层之间会产生热应力疲劳损耗，即封装失效.因此，IGBT的封装失效实质是器件各层异质结构导热系数不同导致的.导致IGBT芯片失效的原因有很多，如电源或负载波动、控制或驱动电路故障、散热装置故障、线路短路等，但最终都可归结为电击穿或热击穿，其中电击穿失效的本质也是温度过高的热击穿失效[2].IGBT结温瞬态变化机理直接影响功率器件寿命预测理论的建立.传统的结温测量方法是在模块底部安装热敏电阻或者热电偶，此方法响应速度非常慢，一般在s级[9]；直接测量法一般采用光导纤维和红外测温仪，此方法的响应速度一般在ms 级，而IGBT在短路的情况下维持时间非常短，通常是几十μs甚至几μs，因此传统的热网络模型并不适用于短路情况下的结温测量[10-14].鉴于此，本文首先对IGBT进行短路破坏性试验，分析IGBT在短路情况下的失效机理，界定导致IGBT失效的临界能量值，分析临界能量值在不同情况下的变化规律.然后利用由试验得到的短路数据，建立有限元热电耦合模型，对IGBT在短路情况下的瞬态结温变化趋势和IGBT内部的温度分布进行分析，从而为IGBT的老化失效分析、器件状态健康管理和功率器件寿命预测提供参考，也为高可靠IGBT模块的设计提供数据积累和实验参考.1 短路试验1.1 短路试验平台IGBT瞬态短路失效试验是通过搭建多重安全保护功率器件失效测试平台实现的.此试验平台通过不同能量等级的预设，对功率器件进行定量(能量)短路失效试验，可实现雪崩击穿、二次击穿、器件爆炸等由不同失效机理引起的失效模式试验.失效试验的电路原理如图1所示.通过对单个IGBT的控制，将8个460 μF/600 V 薄膜电容预存储的能量完全释放，能够实现最高能量值为330 J，瞬间额定电流值超过17 kA2·s的破坏性试验.此装置回路总阻抗和总感抗分别为180 μΩ，120 nH.IGBT短路失效电流的测量是通过CWT3/6/30/150R型柔性电流探头实现的.IGBT的基板下面安装有一个电加热盘，用于控制IGBT的温度.图1 短路失效测试电路原理图Fig.1 Schematic diagram of circuit testingshort-circuit failure1.2 短路失效分析本试验选用英飞凌PG-TO247-3封装型 AIKW30N60CT 600V-30A的单IGBT-Diode半桥臂结构模块，对此模块施加300 V的直流母线电压，保持基板温度为25 ℃，驱动电压为+15 V，得到其瞬态短路波形图如图2所示.其中Vgs为驱动电压，Id为集电极电流，Vdc为直流母线电压，Vpt为测试平台的保护信号.整个短路过程可以分为4个阶段.图2 在直流母线电压300 V，基板温度25 ℃条件下IGBT短路波形图Fig.2 Short-circuit waveform diagram of IGBT with Vdc=300 V and Tcase=25 ℃阶段1(t1～t2)：当短路发生在t1时刻时，主回路中仅存在较小的线路阻抗，集电极的电流迅速上升.此时，IGBT从正向阻断区进入饱和区，集射极两端的饱和电压几乎等于直流母线电压.在此阶段，由于MOS沟道的自由度与温度呈现正反馈关系(直到600 K)，因此电流呈现不断上升的趋势.阶段2(t2～t3)：在此阶段，IGBT一直处于饱和区，将会产生相当大的功率损耗，开关管的自热效应使IGBT的结温迅速上升.当温度上升到一定程度，将会导致MOS沟道(漂移区)载流子的自由度降低，短路电流会略微下降.此时，如果IGBT 的结温在安全范围内，它仍然能够通过控制驱动信号将IGBT关断.阶段3(t3～t4)：随着结温的继续上升，短路电流与温度呈正反馈关系.这是由MOS沟道电子电流下降率低于由热辅助的碰撞电离所引起的泄漏电流上升率所导致的.因此，短路电流又呈现出略微的上升趋势.阶段4(t4～)：t4时刻IGBT被关断，关断之后存在非常大的拖尾泄漏电流，使IGBT结温瞬间以指数规律上升，最终导致IGBT发生热失效.这种失效模式发生在IGBT被关断延迟一段时间之后.图3和图4，分别为基板温度150 ℃，直流母线电压分别为300 V和375 V时IGBT的瞬态短路波形，其过程与上述4个阶段的分析一致.图3 在直流母线电压300 V，基板温度150 ℃条件下IGBT短路波形图Fig.3 Short-circuit waveform diagram of IGBT with Vdc=300 V and Tcase=150 ℃图4 在直流母线电压375 V，基板温度150 ℃条件下IGBT短路波形图Fig.4 Short-circuit waveform diagram of IGBT with Vdc=375 V and Tcase=150 ℃1.3 临界能量值界定IGBT在承受不同能量值的冲击过程中，当能量值低于临界能量值时，其失效之前能够承受的冲击次数至少在万次以上；当能量值大于临界能量值时，经受一次能量冲击，器件就会发生失效[7].因此，在短路过程中，界定临界能量值非常重要.临界能量值Ec定义为Ec=Vds·Iddt在不同的试验条件下，IGBT的临界能量值根据基板温度、直流母线电压的变化趋势如图5和图6所示.从图5可以看出，在同一直流母线电压下，低温时需要更大的能量损耗才能使其达到临界失效温度点.随着温度的上升，IGBT失效临界能量值呈线性略微下降的趋势.从图6可以看出，在同一温度下，随着直流母线电压的升高，引起IGBT失效的临界能量值也略微降低. 因此，IGBT的临界能量值与温度、直流母线电压均呈负反馈的关系.图5 当直流母线电压为330 V时，不同温度下临界能量值的比较Fig.5 Comparison of critical energy value under different temperature conditions at Vdc=330 V图6 当基板温度150 ℃时，不同直流母线电压下临界能量值的比较Fig.6 Comparison of critical energy value under different DC bus voltage conditions at Tcase=150 ℃2 IGBT的有限元热电耦合模型为了得到短路情况下IGBT的温度分布，本文基于短路试验的数据建立了IGBT的有限元热电耦合模型，如图7所示.IGBT各层材料的物理参数如表1所示[15].图7 IGBT有限元热电耦合模型Fig.7 Finite element thermoelectric coupling model of IGBT表1 IGBT各层材料的物理参数Table 1 Each layer of physical parameters of IGBT名称(材料)比热容/(J·kg-1·K-1)密度/(kg·m-3)热传导系数/(W·m-1·K-1)IGBT 芯片层(硅)7002329131二极管(硅)7002329131覆铜层(铜)3858700400焊料层200736048陶瓷(Al2O3)200736035基板(铜)3858700400在短路瞬间，当直流母线电压施加在功率器件上时，将会形成一个宽度为xp的P 沟道和一个宽度为xn的N漂移区的耗尽层，即①②其中，εs为硅材料的电介质常数，q为电子电荷，Na和Nd分别代表P沟道和N 漂移区的掺杂浓度.在此模型中，根据击穿电压的大小可以确定N漂移区的厚度为20 μm，掺杂浓度为2×1015cm-3.在不同的短路条件下，IGBT的温度分布可以通过在笛卡尔坐标下求解热传递公式而得到③其中，kp，ρ，cp分别为硅材料的热传导系数、材料密度和比热容；Q为IGBT 在短路过程中的功率损耗所产生的热量.由于热量的流动方向是从IGBT芯片层向基板传递，因此式③可以简化为④内部热源Q可以由下式计算得到：⑤其中，J(t)为短路电流密度；S为IGBT的有效面积；I(t)为试验所测短路电流；E(x)为空间电荷区的电场分布，可以由式⑥⑦计算得到⑥⑦将式①②⑤⑥⑦带入式④中可以得出⑧⑨根据式⑧和⑨，得出以下结论：1)随着直流母线电压的升高，短路电流也升高，将会导致结温迅速上升.对于给定的临界失效温度，从短路发生到器件失效之间的维持时间将会大大缩短.2)在同样的直流母线电压下，结温的上升率与短路电流密度成正比.在此模型中，基板下表面(x=xc)被设定为恒定温度(从25 ℃到200 ℃)，假设IGBT内部热量流通是单向的，且由于绝热硅胶芯片层上表面(x=-xs)是绝热的，因此短路前IGBT的温度等于基板温度.由此可得模型中的边界条件和初始值如下：T(x=xc,t)=TcT(x,t=0)=Tc3 仿真结果与分析根据试验短路电流波形和IGBT的有限元热电耦合模型，通过数值仿真的方法可以得到IGBT的温度分布.图8为直流母线电压300 V，基板温度25 ℃的情况下IGBT发生短路后关断时刻垂直方向温度分布图.从图8可以看出，由于短路时间非常短，在整个短路过程中，IGBT的热量传递仅仅传递到焊料层且最高温度在耗尽区的边界(x=0)，这是由于此边界处的电场强度最高.如此短的热量传递距离反映了在短路过程中温度的分布与IGBT封装技术和外部冷却条件无关.此外，在耗尽区产生的集中热量将会造成栅极氧化层和金属层的退化甚至破坏.图9为IGBT在不同的直流母线电压和基板温度下，耗尽层边界处(x=0)的瞬态功率、温度曲线对比图.图8 IGBT短路关断时刻垂直方向温度分布图Fig.8 Vertical temperature distribution diagram of IGBT at short-circuit shutdown time图9 不同直流母线电压和基板温度下，耗尽层边界处不同瞬态功率、温度曲线图Fig.9 Curves of different transient power and temperature at the boundary of depletion layer under different DC bus voltage and base temperature 从图9可以看出，发生短路时IGBT温度的变化对失效临界温度点影响不大，仅仅缩短了短路过程所维持的时间.而在不同的直流母线电压下，直流母线电压的升高会使失效临界温度点有所升高，并且短路过程所维持的时间几乎缩短了一半.相比较于较高的温度，由于短路过程所维持的时间更短，较高的直流母线电压对IGBT 短路所造成的损害更加严重，这也对整个系统的保护装置提出了更高的要求.4 结语本文对IGBT进行了破坏性短路试验，在分析IGBT短路过程中的失效机理、IGBT 失效临界能量值及其在不同情况下演化规律的基础上，建立了IGBT的有限元热电耦合模型，通过仿真得出了IGBT在瞬态短路情况下的内部温度分布和直流母线电压与初始温度对内部温度分布的影响.结果如下：1)随着直流母线电压、初始温度、电流密度的增加，IGBT的短路维持时间和临界能量值都将会大幅度减小，而IGBT发生失效的临界温度点与初始温度的大小无关，与直流母线电压、电流密度有关，且呈现出正反馈关系.2)IGBT发生短路瞬间，结温迅速上升且结温的最高点分布在耗尽层的边界处(x=0)；短路电流密度越大将会导致结温上升的速度越快，且造成较大的温度梯度.3)在短路过程中，IGBT失效前热量的传递仅仅达到焊料层，这说明IGBT的短路能力与其封装、散热条件无关.对于电力电子功率电路的设计者来说，在设计电路的过程中必须考虑一定的安全裕度，本文所得试验数据和仿真数据具有一定参考价值，可为高可靠性IGBT模块的设计提供数据积累和实验参考，也可为电力电子故障诊断和容错提供新思路.参考文献:【相关文献】[1] 李武华,陈玉香,罗皓泽,等.大容量电力电子器件结温提取原理综述及展望[J].中国电机工程学报,2016,36(13):3546.[2] 汪波,罗毅飞,张烁,等.IGBT极限功耗与热失效机理分析[J].电工技术学报,2016,31(12):135.[3] SONG Y,WANG B.Survey on reliability of power electronic systems[J].IEEE Transactions on Power Electronics,2013,28(1):591.[4] 徐帅,杨欢,王田刚,等.电力电子变换器可靠性研究[J].北京交通大学学报(自然科学版),2015,39(5):125.[5] WU R,BLAABJERG F,WANG H,et al.Catastrophic failure and fault-tolerant design ofIGBT power electronic converters-an overview[C]∥IECON 2013—39th Annual Conference of the IEEE Industrial Electronics Society,Vienna:IEEE,2013:507-513.[6] 赖伟,陈民铀,冉立,等.老化实验条件下的IGBT寿命预测模型[J].电工技术学报,2016,31(24):173.[7] LEFEBVRE S,KHATIR Z,SAINT-EVE F,et al.Experimental behavior of single-chip IGBT and CoolMOS devices under repetitive short-circuit conditions[J].IEEE Transactions on Electron Devices,2005,52(2):2763.[8] 汪波,胡安,唐勇,等.IGBT电压击穿特性分析[J].电工技术学报,2011,26(8):145.[9] 窦智峰,翟朝伟,崔光照,等.IGBT现场失效短路结温测量方法研究[J].轻工学报,2017,32(4):73.[10]AMMOUS A,AMMOUS K,MOREL H,et al.Electro-thermal modeling of IGBTs:application to short-circuit conditions[J].IEEE Transactions on Power Electronics,2000,15(4):778. [11]RACITI A,MUSUMECI S,CRISTALDI D.Modeling and simulation of IGBT thermal behavior during a short circuit power pulse[C]∥2015 International Conference on Clean Electrical Power (ICCEP),Taormina:IEEE,2015:542-547.[12]CLEMENTE S.Transient thermal response of power semiconductors to short power pulses[J].IEEE Transactions on Power Electronics,1993,8(4):337.[13]唐云宇,林燎源,马皓.一种改进的并联IGBT模块瞬态电热模型[J].电工技术学报,2017,32(12):88.[14]唐勇,汪波,陈明.IGBT开关瞬态的温度特性与电热仿真模型[J].电工技术学报,2012,27(12):146.[15]SANO K,HAYASHI K,KAWAFUJI H ,et al.Ultra-small compact transfer molded package for power modules[C]∥2008 58th Electronic Components and Technology Conference,Lake Buena Vista:IEEE,2008:1832.。
DC参数测试
.DC参数测试摘要本章节我们来说说DC参数测试，大致有以下内容，⏹欧姆定律等基础知识⏹DC测试的各种方法⏹各种DC测试的实现⏹各类测试方法的优缺点基本术语在大家看DC测试部分之前，有几个术语大家还是应该知道的，如下：Hot Switching 热切换，即我们常说的带电操作，在这里和relay（继电器）有关，指在有电流的情况下断开relay或闭合relay的瞬间就有电流流过（如：闭合前relay两端的电位不等）。
热切换会减少relay的使用寿命，甚至直接损坏relay，好的程序应避免使用热切换。
Latch-up 闩锁效应，由于在信号、电源或地等管脚上施加了错误的电压，在CMOS器件内部引起了大电流，造成局部电路受损甚至烧毁，导致器件寿命缩短或潜在失效等灾难性的后果。
BinningBinning（我很苦恼这玩意汉语怎么说——译者）是一个按照芯片测试结果进行自动分类的过程。
在测试程序中，通常有两种Binning的方式——hard binning 和soft binning. Hard binning控制物理硬件实体（如机械手）将测试后的芯片放到实际的位置中去，这些位置通常放着包装管或者托盘。
Soft binning控制软件计数器记录良品的种类和不良品的类型，便于测试中确定芯片的失效类别。
Hard binning的数目受到外部自动设备的制约，而Soft binning的数目原则上没有限制。
下面是一个Binning的例子：Bin# 类别01 100MHz下良品02 75MHz下良品10 Open-Short测试不良品11 整体IDD测试不良品12 整体功能测试不良品13 75MHz功能测试不良品14 功能测试VIL/VIH不良品15 DC测试VOL/VOH不良品16 动态/静态IDD测试不良品17 IIL/IIH漏电流测试不良品从上面简单的例子中我们可以看到，Hard bin 0，Soft bin 01-02是良品，是我们常说的GoodBin；而Hard bin 1,Soft bin 10-17是不良品，也就是我们常说的FailedBin。
燃料电池车用双向DC_DC变换器的建模分析
燃料电池车用双向DC /DC 变换器的建模分析陈玮山 赵景辉 陈 沛 (上海汽车集团股份有限公司燃料电池事业部)【摘要】 针对目前还无法给出双向DC /DC 变换器的精确数学模型的研究现状,文章通过小信号分析法分析了双向DC /DC 变换器的拓扑结构,建立了双向DC /DC 变换器的传递函数方程,并利用此模型对双向DC /DC 变换器进行优化设计。
该小信号模型准确描述了双向DC /DC 变换器的静态和动态特性,对于燃料电池车用双向DC /DC 变换器的设计制造过程有较高的理论指导依据和工程应用价值。
【主题词】 燃料电池 汽车 变换器0 引言收稿日期:2007-07-10燃料电池车用双向DC /DC 变换器主要是针对燃料电池输出电压波动较大、负载特性较软而设计。
在燃料电池汽车混合动力系统中,燃料电池和镍氢蓄电池是整车所需能量的来源,双向DC /DC 变换器是整个动力系统能量流动的重要环节。
双向DC /DC 变换器是燃料电池和蓄电池之间的一个周期性通断的开关控制装置,具有双向升降压功能。
燃料电池客车在启动、加速时,蓄电池放电,供给电机电能,驱动电机加速,这时双向DC /DC 变换器工作在升压状态。
当燃料电池客车减速或匀速时,要求电机处在能量再生制动状态,双向DC /DC 变换器给作为燃料电池的储能装置蓄电池充电,实现再生能量回收,这时双向DC /DC 变换器工作在降压状态。
可以说,双向DC /DC 变换器在燃料电池客车的能量系统中起到了能量转换中枢的作用。
为了便于对双向DC /DC 变换器进行系统设计、仿真研究和稳定性分析,希望能建立双向DC /DC 变换器的数学模型。
但是由于开关电源在本质上是一个离散的非线性系统,所以给出精确的数学模型是非常困难的。
本文将通过小信号分析方法来建立系统传递函数。
1 双向DC /DC 变换器的建模现有分析方法一般有数字仿真法和解析建模法,本文主要是通过小信号分析方法来分析双向DC /DC 变换器。
- 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)。
Available online at Comparison of different DC voltage supervision strategies in a local Power Distribution System of More Electric AircraftHe Zhang a,∗,Christophe Saudemont a,Benoît Robyns a,Régis Meuret ba Ecole des Hautes Etudes d’Ingénieur(HEI),Laboratory of Electrotechnics and Power Electronics of Lille(L2EP),13,rue de Toul,59046Lille,Franceb Hispano-Suiza Group SAFRAN,Moissy Cramayel,FranceReceived1October2008;received in revised form29April2010;accepted10May2010Available online16May2010AbstractThe local DC Power Distribution System of More Electric Aircraft is one of the cores of the electric power transmission.Because of some unidirectional devices(diode bridge and anti-run-back diodes),the system is not totally reversible.Some storage and dissipation systems are necessary to store or/and dissipate the returned energy to maintain the DC bus voltage and to increase the energy efficiency.This paper introduces four supervision strategies in two levels to maintain the DC bus voltage and compares these strategies with the help of two criteria:the voltage performance and efficiency.©2010IMACS.Published by Elsevier B.V.All rights reserved.Keywords:On-board network;Fuzzy logic supervision;DC power supply;Power distribution1.IntroductionThe More Electric Aircraft(MEA)is the trend of the future aircraft because of ecological,economic and sustainable development reasons.The replacement of the mechanical,hydraulic and pneumatic power transfers by an electrical one is one of the most important innovations[5,9]of MEA.In the European Power Optimized Aircraft project(POA),it has been proven that the use of a local DC Power Distribution System(PDS)makes it possible to reach the subsystems optimum in order to get some benefits,like a weight reduction for example.Such a PDS is a local sub-network supplying loads such as,for example,flight control loads,landing system,etc.There are a lot of storage system applications in hybrid electric vehicles[7],power distribution systems[8]and wind powerfields[1,6].Because of the bidirectional characteristics of the loads and unidirectional characteristic of source and different dynamic characteristic between the two,the consumed power and supplied power will be unbalanced in some cases.It will consequently cause some instability of voltage and frequency.The main function of the storage system in these applications is to ensure the power balance between the loads and the supplier.It takes the role of load or supplier according to the case in order to control the voltage[1,7,8]and/or frequency[2],also to have some economic effects[6].Some control and supervision strategies,with the help of fuzzy logic or PI controller,are developed to control the DC bus voltage and improve the performance of power systems in these applications.But the dynamic∗Corresponding author.E-mail address:he.zhang@hei.fr(H.Zhang).0378-4754/$36.00©2010IMACS.Published by Elsevier B.V.All rights reserved.doi:10.1016/j.matcom.2010.05.009264H.Zhang et al./Mathematics and Computers in Simulation81(2010)263–276Fig.1.Definition of local DC power distribution system.( P/ t)of the load profile in the aeronautic domain is faster than in thesefields(hybrid electric vehicle,power distribution system and wind power)[10].In this context,a hybrid storage and dissipation system is proposed for the reason of dynamic compensation and security.In order to achieve the aim of voltage control and,at the same time, energetic optimization in the fast dynamic application,the network supervision becomes a synthesis problem.This paper presents some control strategies for this hybrid storage and dissipation system to ensure that the DC bus voltage is in the range of acceptable tolerance and to optimize the energy efficiency.In Section2,a local DC PDS of MEA is defined.The different supervision strategies with the help of fuzzy logic or PI controller are proposed in Section 3.In Section4,these strategies are tested in different conditions with the help of simulation.Two criteria(efficiency and voltage performance)are proposed to compare the simulation results.The conclusion is given in Section5.2.The definition of a local PDS and the problem facingThe DC local power distribution network under study is defined in Fig.1.This network includes2independent sources and3loads in which2are bidirectional.Among both sources(PMG or AC aircraft network),just one of them can be connected to the DC network at the same time.The rated voltage of DC bus is270V,but the permissible tolerance is between250V and280V as shown in Fig.2according to the standard MIL-STD-704[3].Because of some unidirectional devices(diode bridge and anti-run back diodes),the electrical energy cannot return to the source Permanent Magnet Generator(PMG)or AC aircraft network.When the sum of returned power of load1and load2is greater than the consumed power of load3,the diodes are turned off and the returned energy is stored in the capacitors. This will cause an over-voltage of DC bus.A hybrid storage and dissipation system(shown in Fig.3)can be used to resolve this problem.A resistance associated with a chopper can compensate the dynamic weakness of the storage system.Also,it can be the backup of the storage system when the storage system is disabled.3.Supervision strategiesTo control this hybrid system,two levels of supervision strategies have to be considered individually.First level,to determine the total absorbed or generated power reference of the hybrid system,P total.Second level,to split the P total into two parts:the part to store,P stor and the part to dissipate,P resi.For each supervision level,different strategies are proposed.Fig.4shows the global structure of the different supervision strategies.The detail of each strategy will be presented in the following parts.Table1summarizes the objectives,constraints,means of action and tools used for each proposed strategy.3.1.Level I:determination of the total power referenceAs mentioned above,the aim of thefirst supervision level is to determine the total power reference in order to maintain the DC bus voltage.H.Zhang et al./Mathematics and Computers in Simulation81(2010)263–276265Fig.2.V oltage standard of MIL-STD-704.Fig.3.Hybrid storage and dissipation system.Fig.4.The global structure of two level supervision strategies.266H.Zhang et al./Mathematics and Computers in Simulation81(2010)263–276Table1Objectives,constraints and action means of each power management strategy.Strategy Objectives Constraints Means of action ToolsLevel IPI DC bus standardTo maintain the DC bus voltage Storage limit P total PI controllerThe principle source is unidirectionalMP To maintain the DC bus voltage DC bus standard Fuzzy logic To maintain the PMG working Storage limit P totalTo manage the capacity limit of the storage The principle source is unidirectionalLevel IIVB To maximize the utilization of storage DC bus standard P stor Fuzzy logicStorage dynamic P resiFC To compensate the storage dynamic DC bus standard P stor Power balanceStorage dynamic P resi3.1.1.Proportional Integral controller(PI)When the total load power is negative,the diodes are turned off and the voltage controller,Generator Control Unit (GCU),cannot control the bus voltage.From this moment,the hybrid system takes the role of the bus voltage controller. The bus voltage is measured to calculate how much power should be consumed in order to balance the exceeding power. In this case,a PI controller is used to calculate this reference power value P total.Fig.5shows an example of PI controller structure to control a system.The upper part represents the system to control.The typicalfirst order system is used where K sys represent the gain of the system andτsys represent the time constant of the system.The bottom part represents the PI controller.In our case,the system to control is shown in Fig.6.C equi represent the capacitor equivalent of these capacitors at the port of the loads and R c represent the equivalent leakage resistance of the capacitors.As shown in Fig.6and(1) and compared with Fig.5,we can choose the bus voltage V bus to be the variable to control.V bus I total =R cC equi·R c·s+1(1)Fig.5.Proportional Integral controller.Fig.6.The system to control.H.Zhang et al./Mathematics and Computers in Simulation81(2010)263–276267Fig.7.The strategy of Minimum supplied Power.3.1.2.Minimum supplied Power(MP)In the previous control strategy,when the recovered power is greater than the consumed power,the power supplied by the generator becomes null because the anti-run back diodes are turned off.In the Minimum supplied Power strategy, the hybrid storage system is controlled to be a power consumer when the loads return power.If the current is always nonzero,the GCU will always take in charge the voltage control.The description of the main function of this strategy is shown in Fig.7.This strategy has three goals.Firstly,this strategy should slow the tendency of the supplied power decreasing in order to avoid turning off the diodes.Secondly,sometimes,the dynamic of load is faster than the storage system.The storage system cannot achieve the reference power instantly.In this case,the current of the generator will be zero and the GCU will not take in charge the voltage control.The second goal is to control the bus voltage by using the storage system when the diodes are turned off.The third goal is to manage the storage system to ensure that it is always available to absorb and supply the energy.According to the goals mentioned above,the informations of the generator current I gen,the bus voltage V bus and the storage level store level are the inputs of this strategy and the total power reference P total(the output variable)will be sent to the hybrid storage system.In Fig.4the block MP shows the structure of this strategy.The power convention is defined as follows:consumed power is positive;generated power is negative.So,in this strategy,the power supplied by the generator,the DC bus voltage and the storage level should be supervised.The supervisor must ensure to provide the power reference P total.Fuzzy logic is well adapted to develop a multi variable supervisor to achieve power management[1,4,6].Fig.8(a)–(c)shows the input membership functions.Fig.8(d)shows the output membership function.According to the goals mentioned above,the fuzzy rules can be described as follows:if the generator supplies a high power and the storage level is high,then the hybrid system supplies power;if the generator supplies a low power and the storage level is low,then hybrid system absorbs power,etc.Because there are three inputs and one output,a 2D table is not sufficient to represent the function of the four variables.So the rules are studied in two cases:I gen is nonzero and I gen is zero.Tables2and3show the fuzzy rules of the MP strategy.Figs.9–11show the evolution of P total vs.the storage level and generator current under different voltage inputs. We can note that when the generator current is close to zero,the total power reference depends on the bus voltage value.At this moment,the storage system acts as a voltage controller.When the current is out of the zero zone, there is no difference between different voltage levels,because in this moment,the GCU take the role of voltage controller.268H.Zhang et al./Mathematics and Computers in Simulation81(2010)263–276Fig.8.Membership functions of Minimum supplied Power strategy.(a)Input1I gen,(b)input2V bus,(c)input3store level,and(d)output P total.3.2.Level II:determination of the power reference partitionIn the level II,the strategy to split the P total is discussed.Obviously,we could use only the resistance to dissipate all returned energy from loads.But in the aspect of energetic optimization and temperature constraint,this could not be considered as a satisfactory solution.The objective of the power partition is to use the storage system as much as possible to maintain the bus voltage.3.2.1.Voltage Band(VB)The rated voltage of the bus is270V,but the permissible tolerance is between250V and280V as shown in Fig.2 according to the standard MIL-STD-704[3].The objective of this strategy is to use the resistance as little as possible.A fuzzy logic supervisor is used in this strategy to determine a partition coefficient k in order to create P stor and P resi reference as expressed in(2).The bus voltage and its evolution are considered as the inputs of the supervisor.Table2Fuzzy rules:total power reference P total when I gen is nonzero.stor level I genBig Medium SmallBig Negative Negative Zero Medium Negative Zero Positive Small Zero Positive PositiveH.Zhang et al./Mathematics and Computers in Simulation81(2010)263–276269 Table3Fuzzy rules:total power reference P total when I gen is zero.stor level V busBig Medium SmallBig Positive Zero Negative Medium Positive Zero ZeroSmall Positive Zero Zero Fig.12shows the membership functions.Fig.13shows the variation of the partition coefficient k vs.the bus voltage and V/ t.P resi=k×P total(2)P stor=(1−k)×P totalAccording to the objective mentioned above,the fuzzy rules are shown in Table4Fig.9.Generated surface of MP strategy for P total when V bus is200V.Fig.10.Generated surface of MP strategy for P total when V bus is270V.270H.Zhang et al./Mathematics and Computers in Simulation81(2010)263–276Fig.11.Generated surface of MP strategy for P total when V bus is300V.3.2.2.Full Compensation(FC)Because of the storage system dynamics,its output power,P stor out,cannot reach the power reference instantly.In this case,the dissipation system is used to dissipate the exceeding power P resi.This strategy is expressed in(3) P stor=P total(3)P resi=P total−P stor outFig.12.The membership functions of V oltage Band strategy.(a)Input1V bus,(b)input2 V bus/ t,and(c)output the partition coefficient.H.Zhang et al./Mathematics and Computers in Simulation81(2010)263–276271Fig.13.Generated surface for partition coefficient of strategy VB.Fig.14.simplified model of storage.4.Simulation results and comparison of the strategiesIn this section,the four different strategies are tested.And the results are compared according to two criteria:An efficiency criterion in order to express the performance of energetic optimization,and a voltage criterion in order to express the voltage control performance.4.1.Simulation resultsTwo supervisor levels,with respectively two variants lead to four different strategies.The abbreviations PI+VB, PI+FC,MP+VB and MP+FC are used to identify the different strategies.To test the strategies,a simplified model of the storage system is used[2],as shown in Fig.14where P max is the output power limit;τstor is the time constant of the storage system and W stor is the energy stored by storage system. The cumulated total profile of the three loads is shown in Fig.15.From0to0.1s,the loads start up.From0.1s to 0.15s,the load power is constant.From0.15s,the bidirectional loads begin to regenerate power and the unidirectional load continues to consume power.Figs.16and17show respectively for the MP+VB and PI+FC strategies evolution of the DC bus voltage and gener-ator current(Figs.16(a)and17(a)),the stored power and the power dissipated in the resistance(Figs.16(b)and17(b)), the stored energy and the energy dissipated in the resistance(Figs.16(c)and17(c)).Because the comparison focuses on the strategy of voltage supervision by using the hybrid storage system,an ideal DC source is used to replace the Table4Fuzzy rules:The partition coefficient k.V bus/ t V busBig Medium Small Negative Small Small Small Zero Big Small Small Positive Big Small Small272H.Zhang et al./Mathematics and Computers in Simulation81(2010)263–276Fig.15.Load profile.source PMG in Fig.1in order to simplify the study.Because the PI strategy does not take into account the storage level,the capacity limit of storage system W max is set to a value great enough and the initial value of storage level is set to medium,in order to eliminate the saturation effect of storage system.In Fig.16,we observe that as long as the generator current,I gen,is different from zero,the DC voltage V bus is strictly controlled by this source at270V.ButFig.16.Result of strategy MP+VB withτstor=0.05s and P max=20kW.(a)Bus voltage and current of generator,(b)power dissipated in the resistance and stored,and(c)energy dissipated in the resistance and stored.Fig.17.Result of strategy PI+FC withτstor=0.05s and P max=20kW.(a)Bus voltage and current of generator,(b)power dissipated in the resistance and stored,and(c)energy dissipated in the resistance and stored.when I gen is null,because of returned power,the hybrid system becomes in charge of controlling the DC bus voltage instead of the generator.Then,if the storage device can not balance exceeded power because of its dynamic weakness, the VB strategy permits V bus to increase,in the limit described in Fig.2.If the voltage is over this limit,the dissipation system takes part in to dissipate exceeded power.When the hybrid system absorbs more than the returned power,I gen is different from zero,and then the generator is in charge of controlling V bus again.So,the succession of these different configurations is the reason of the bus voltage oscillation between0.15s and0.18s in Fig.16.The length of this period depends on the dynamic of storage system.4.2.Definition of criteria4.2.1.EfficiencyWe define the efficiency by(4),with stored energy variations, W stor,supplied energy variations, W gen and returned loads energy variations, W return.efficiency(%)=W storW gen+ W return×100%(4)The loads begin to return power at0.15s and the criteria will be calculated from this moment.4.2.2.Voltage performanceThe area surrounded by the rated voltage270V and measured voltage is calculated in order to evaluate the voltage performance.The criteria is computed by(5)and described in Fig.18.According to a condition of a voltage value inFig.18.Definition of criterion voltage performance.the acceptable tolerance range,the less area is,the better performance is.A =|V measur −270|dt(5)Same as the efficiency performance,the voltage area will be calculated from 0.15s.4.3.Efficiency and voltage performance comparisonFour series of tests have been carried out by using different dynamic parameter (time constant τstor and maximum power P max )of the storage system.In this comparison,the capacity of storage system W max has not been taken into account.The test results according to both criteria are shown in Table 5.If we place the results in an area–efficiency space,we can get Fig.19.As shown in Fig.19,for each series,the better voltage performance we obtain,the less efficiency we get.We connect the average point of each series with the origin point.We can find that the slope of the line identifies the dynamic characteristic of each series.Gentler the slope is,faster the dynamic is.We find that between the two dynamic parameters of storage system (the maximum power and the time constant),the time constant is the more important factor of efficiency and voltage performance.Table 6compares the slopes of each dynamic series.Table 5Result of efficiency and voltage performance (area).Level IILevel I MP PI P maxτstor P maxτstor 0.5s 0.05s 0.5s 0.05s Area (V s)Efficiency Area (V s)Efficiency Area (V s)Efficiency Area (V s)Efficiency VB 20kW 0.99021.4%0.12793.3%20kW 0.54040.4%0.19891.3%5kW 1.06417.3%0.44379.1%5kW 0.62019.7%0.3079.3%FC20kW 0.00913.2%0.00966.5%20kW 0.15313.1%0.15266.1%5kW0.0099.2%0.00946.7%5kW0.15313.1%0.15266.1%Table 6The slope of different dynamic series.P maxτstor 0.5s0.05s 20kW 1.920.155kW3.120.34Fig.19.The comparison of different strategies on different condition.Table7Comparison of the strategies.Level II Level IMP PIAdvantage Drawback Advantage DrawbackVB Possible to supervise the storage level,high efficiency performance Complex implementation High efficiency performance Impossible to supervise thestorage levelFC The highest voltage performance Low efficiency performance Simple implementation,highvoltage performanceLow efficiency performanceIn another side,different strategies have different performance in the same dynamic series.Wefind that the VB strategy of level II can always obtain the best efficiency compared with FC strategy,because the strategy VB makes use of the voltage tolerance to utilize the storage as much as possible in order to increase the efficiency.For the level I,the choice of the strategy influences less the result of efficiency and voltage performance.The strategy MP has the capacity to control the storage level.But the strategy PI cannot take into account this aspect.Table7compares the advantage and drawback of the four strategies.5.ConclusionThe storage system can take the role of power supplier or load to keep the power balance in many applicationfields. Because of the fast dynamic of the load profile in aeronautic domain,the configuration of hybrid storage and dissipation system can be a new feasible solution.In order to maintain the DC bus voltage and,at the same time,consider the energy efficiency in the fast dynamic application,some supervision strategies are introduced in this paper.These strategies are separated into two independent levels.For each level,there are two different strategies.The strategies in different levels can be combined freely,therefore there are four supervision strategies.To compare the simulation results of the four supervision strategies,two criteria are introduced in this paper:the voltage and efficiency performance.The comparison shows that in this fast dynamic application,not only can the dynamic of the storage system affect the results,but also the supervision strategy can change the voltage-efficiency performance.Fuzzy logic was used in these strategies.It is well adapted to develop the multi objective supervision.The comparison shows the fuzzy logic supervision strategies have the better performance.AcknowledgmentsThis work was supported by afinancing from the regional Council Nord-Pas de Calais,HISPANO-SUIZA Company, and HEI.References[1]G.O.Cimuca,C.Saudemont,B.Robyns,M.M.Radulescu,Control and performance evaluation of aflywheel energy storage system associatedto a variable speed wind generator,IEEE Transactions on Industrial Electronics53(4)(2006)1074–1085.[2]V.Courtecuisse,M.El Mokadem,B.Robyns,B.Francois,M.Petit,J.Deuse,Association of wind turbine based dispersed generators andstorage systems to participate in primary frequency control,in:12th European Conference on Power Electronics and Applications,EPE, Aalborg,Denmark,2007.[3]Department of Defense of United States of America,Department of Defense Interface Standard–Aircraft Electric Power Characteristics,Military Standard of USA,2004.[4]L.Karunarathne,J.T.Economou,K.Knowles,Fuzzy logic control strategy for fuel cell/battery aerospace propulsion system,in:IEEE VehiclePower and Propulsion Conference(VPPC),Harbin,China,2008.[5]nglois,E.Foch,X.Roboam,H.Piquet,De l’avion plusélectriqueàl’avion toutélectrique:état de l’art et prospective sur les réseaux debord,J3EA,4(1)(2005),/more.php?rub=j3ea&vol=vol4HS1&cont=tdm.[6]L.Leclercq,B.Robyns,J.Grave,Control based on fuzzy logic of aflywheel energy storage system associated with wind and diesel generators,Mathematics and Computers in Simulation63(2003)271–280.[7]W.Lhomme,P.Delarue,A.Bouscayrol,P.Le Moigne,Control Strategies for Maximum Storage of Supercapacitor System,ESSCAP’06,Lausanne,Switzerland,2006,CD-ROM.[8]P.Thounthong,S.Raël,B.Davat,Control strategy of fuel cell and supercapacitors association for a distributed generation system,IEEETransactions on Industrial Electronics54(6)(2007)3225–3233.[9]D.Van Den Bossche,More electric control surface actuation,in:10th European Conference on Power Electronics and Applications,EPE,Toulouse,France,2003.[10]H.Zhang,C.Saudemont,B.Robyns,M.Petit,Electrical features comparison between more electric aircrafts and hybrid electric vehicles,Electromotion Journal16(2009)111–119.。