外文資料翻譯---基于自聯(lián)想神經(jīng)網(wǎng)絡(luò)的發(fā)動(dòng)機(jī)控制系統(tǒng)傳感器故障診斷與重構(gòu)（英文）

1、Received date: 2003 ?05?23; Revision received date: 2003 ?11 ?02 Foundation item: Aeronautical Science Foundation of China ( 01C52015)Sensor Fault Diagnosis and Reconstruction of EngineControl System Based on Autoassocia

2、tive Neural NetworkHUANG Xiang?hua(Power and Energy College, Nanj ing University o f Aeronautics andAstronautics, Nanj ing ? 210016 , China)Abstract: ? The topology and property of Autoassociative Neural Netw orks( AANN)

3、 and the AANN? sapplication to sensor fault diagnosis and reconstruct ion of engine control system are studied. T he key fea?ture of AANN is feature extract and noise filtering. Sensor fault detection is accomplished by

4、integratingt he optimal estimation and fault detection logic. Digital simulation shows that the scheme can detect hardand soft failures of sensors at the absence of models for engines w hich have performance deteriorate

5、in theservice life, and can provide good analytical redundancy.Key words: ? Autoassociative Neural Network; engine sensor; fault diagnosis; analytical redundancy基于自聯(lián)想神經(jīng)網(wǎng)絡(luò)的發(fā)動(dòng)機(jī)控制系統(tǒng)傳感器故障診斷與重構(gòu). 黃向華. 中國(guó)航空學(xué)報(bào)( 英文版) , 2004, 17(

6、1) : 23- 27.摘? 要: 研究自聯(lián)想神經(jīng)網(wǎng)絡(luò)及其在發(fā)動(dòng)機(jī)控制系統(tǒng)傳感器故障診斷及重構(gòu)中的應(yīng)用。自聯(lián)想神經(jīng)網(wǎng)絡(luò)關(guān)鍵在于特征提取和噪聲濾波。綜合自聯(lián)想網(wǎng)絡(luò)的最優(yōu)估計(jì)與故障診斷, 自動(dòng)區(qū)分估計(jì)誤差和傳感器故障。仿真結(jié)果表明這種方法不需要模型, 能診斷傳感器硬、 軟故障, 當(dāng)發(fā)動(dòng)機(jī)性能蛻化時(shí)也能提供很好的解析余度。關(guān)鍵詞: 自聯(lián)想網(wǎng)絡(luò); 發(fā)動(dòng)機(jī)傳感器; 故障診斷; 解析余度文章編號(hào): 1000 ?9361( 2004) 01 ?002

7、3 ?05? ? ? 中圖分類號(hào): V233. 7; V263. 6? ? ? 文獻(xiàn)標(biāo)識(shí)碼: A? ? Sensor fault diagnosis and reconstruction arerequired to achieve adequate reliability in enginecontrol system. Robustness requirement offerschallenges to the design of

8、a fault diagnosis system.T he approach using intelligent algorithms is apromising one[ 1] . For a sensor set w hich has re ?dundant information, it is possible to reconstructone or more lost sensor data if the relationsh

9、ip a ?mong the sensors is known. Usually, the relation?ship can be described as mathematical equationsw ith sensor measurements as input variables. T hemethod provided in this paper is based on Autoas?sociative Neural Ne

10、twork( AANN ) and can realizethe relationship and reconstruct failed sensors.1 ? Topology Architecture of AANN[ 2, 3]Nonlinear Principal Component Analysis( NLPCA ) is the basis of AANN. NLPCA is atechnique for mapping n

11、onlinear multidimensionaldata into low er dimensions w ith minimal loss of in?formation. Let Y= [ y 1 ? y 2 ?? y m] representa n ! m table of data ( n = number of observa?tions, m= number of variables) . T he mapping int

12、ofeature space can be represented byT = G( Y) ( 1)where T= [ t 1 ? t 2 ? ? tf ] is the principal com?ponent matrix ( n ! f ) ; f is the number of princi?pal components ( f < m ) ; G is a nonlinear vectorfunction. Res

13、toring the original dimensionality ofthe data is implemented by another nonlinear vectorfunctionY?= H( T) ( 2)? ? T he loss of information is measured by residu?al E = Y - Y?, and E consists of minor compo?nents w hich i

14、nvolve noise or unimportant variance.Functions G and H are selected to minimize #E #in order to draw principal components.Function G and H can be represented by 2? Vol. 17? No. 1 CH INESE JOURNAL OF AERONAUTICS February

15、2004ally?correlated variables. A block ( or set of blocks)not sharing variables w ith other blocks indicates theindependence of the variables in the block ( or set ofblocks) from the remaining variables and can notintrod

16、uce tw o independent groups of variables intoa single AANN. Overlapping sets of blocks repre ?sent a subsystem of related variables. T he numberof blocks in an overlapping set of blocks is a lowerbound on the number of i

17、ndependent variables( bottleneck nodes) .3 ? Sensor Fault Detection andReconstruction Based on AANNWhen netw ork is trained abundantly, it canbe used for sensor fault detection because there ex?ists redundant information

18、 among input variablesand w hen a sensor fails or even several sensors fail,the other sensors can still provide good estimationto replace the failed sensor. Estimation ReturningScheme ( ERS ) is developed to diagnosis se

19、nsorfault, by comparing the output of netw ork and thecorresponding sensor output to detect sensor faults.If the difference betw een a sensor measurement andits estimation exceeds the threshold w hile the dif?ferences of

20、 other sensors w ith their correspondingestimation ( e. g. relatively low ) , then a sensorfault is declared to happen. Once a faulty sensormeasurement is detected, it w ill be disconnectedfrom the input layer of network

21、. How ever, theneural network w ill continue to function by usingthe most recent corresponding output of the NN asinput instead of the faulty sensor measurement, be ?cause the most recent output is a good estimation ofth

22、e faulty sensor measurement w hen there are e ?nough information on input variables. And AANNhas the ability of fault tolerance for the fact thatthe disturbance from input nodes can be distributedto the network and has l

23、ittle impact on output.T he controller w ill be sw itched to the estimatedvalue to continue the system operation. U nder thisscheme, the system can remain operable even w ithmultiple sensors faults as long as normal sens

24、ors arenot less than bottleneck nodes. The ability to com?bine detection, isolation and accommodation in onestep is the key advantage of AANN based sensorvalidation scheme. This ability is based on the di?mensionality re

25、duction property of AANN.T here w ill be performance degeneration or in?stallation and manufacture tolerance w hich are thesources of uncertainties and will cause estimationerror in the optimal estimation of AANN. T hese

26、uncertainties may be taken for sensor fault or viceverse. If the degeneration is taken for sensor fault,fault w ill be w rongly w arned, causing incorrectfault accommodation. And if sensor fault is takenfor degeneration,

27、 it will cause incorrect networkcompensation. Fault control gain together with softfault detection logic[ 5] is developed to distinguishoptimal estimation error from sensor faults in thispaper. Axial directional fault si

28、gnature is used to i?dentify the cause of optimal estimation error. If theresidual is caused by optimal estimation error, thenthe weights and biases of AANN w ill be compen?sated on?line. If the residual is caused by sen

29、sorfault, then corresponding estimation is used to re?place the failed sensor, providing analytical redun?dancy. In the fault accommodation logic the faultcontrol gains are used to provide a smooth transi?tion from the f

30、ailing sensor to its corresponding es?timation.4 ? Example of Digital SimulationLet? s take a turboshaft engine for example[ 5] .Fig. 2 show s the closed loop control system of a en?gine system consisting of the engine,

31、controller andAANN?based sensor fault diagnosis. The primaryvariables of interest are, n g, n p, T t 45, P s3, M land W fB npg is the given speed of pow er turbine,and M l and n pg are inputs. The control feedbackvariabl

32、es are ng, np, T t45, Ps3.Only when the input variables of AANN arecorrelative, the valid feature of the variables can beextraced from the bottleneck layer. The covariancematrix of 6 variables, ng, np, T t45, P s3, W fB,