外文资料翻译---基于自联想神经网络的发动机控制系统传感器故障诊断与重构(编辑修改稿)

外文资料翻译---基于自联想神经网络的发动机控制系统传感器故障诊断与重构(编辑修改稿)内容摘要：

to output layer through bottleneck layer, demapping layer and output layer, and reconstruct the input data. The weights W and biases B of the AANN are op timized in order to reserve the information of input data in bottleneck layer entirely. Fig. 1 Topology architecture of AANN The key point of an AANN is the bottleneck layer, whose nodes are the smallest in dimension. The bottleneck forces an internal encoding and pression of the input, with a subsequent decoding and depression after the bottleneck, and the work output is produced. The bottleneck layer prevents the work from a simple onetoone or “ straight through” mapping during training the work. Internal restricted fact is included in the bottleneck layer of AANN, which can make AANN learn the internal relationship between all inputs rather than simple unit function. AANN realizes NLPCA, whereas normal feed forward NN can not extract feature and reconstruct data because it has no the ability of coding and decoding. AANN has the property of noise filtering, which depends on the ability of the work to produce a model of the measurements that fits the systematic correlations in the data, yet excludes random variations due to measurement noise. Noise filtering in AANN also depends on redundancy. Redundancy reduces variance in the similar way as that using samples containing multiple items reduces variance in statistical quality control. 2 Selection of Network Nodes In the bined work, there are m nodes in the input and output layers and f nodes in the bottleneck layer. The number of mapping and demapping nodes ( M1 and M2) must be selected properly in order to ensure adequate representational capacity without overfitting : M1+ M2n,where n is the number of observations. Crossvalidation (training on a subset of the training examples, reserving other examples for testing generalization ability) can also be used to select an appropriate number of mapping and demapping nodes,and to limit the intensity of training. For aircraft engine, selfrelationship can be drawn from the thermodynamics and pneumatics relation among different measurements, and then the relationship among all variables can be decided,and the number of independent variables can also be decided, and also the number of bottleneck those nonlinear objects the relationship of which is very plicated, since it is difficult to obtain their mathematical models, the selfrelationship can be determined by analysis of the covariance matrix of the training set. For a set of m measured variables, the covariant matrix R = [ Rij ] m x m is defined as 1 i=j rij= i≠j where rij reflects the dependent relation between xi and xj . If an element rij in the covariance matrix is zero ( or statistically indistinguishable from zero) , then the corresponding measurements xi and xj are independent. Rearrange R in block diagonal form by reordering the measured variable results in a matrix that reveals the dependency structure of the measurements. Each square block of nonzero elements in R represents a set of mutually correlated variables. A block (or set of blocks) not sharing variables with other blocks indicates the independence of the variables in the block (or set of blocks) from the remaining variables and can not introduce two independent groups of variables into a single AANN. Overlapping sets of blocks represent a subsystem of related variables. The number of blocks in an overlapping set of blocks is a lower bound on the number of independent variables (bottleneck nodes). 3 Sensor Fault Detection and Reconstruction Based on AANN When work is trained abundantly, it can be used for sensor fault detection because there exists redundant information among input variables and when a sensor fails or even several sensors fail,the other sensors can still provide good estimation to replace the failed sensor. Estimation Returning Scheme ( ERS) is developed to diagnosis sensor fault, by paring the output of work and the corresponding sensor output to detect sensor。