Research On Engine Fault Diagnosis Based On Tensor Decomposition Feature Extraction

The development trend of engine fault diagnosis technology is automatic and intelligent.And analysis and feature extraction of the collected signal data is the key problem.However,the traditional data analysis and feature extraction methods are usually based on the vector model,which may lose the structural information and destroy the correlation between the data and will affect the subsequent fault pattern recognition.To solve this problem,a method of engine data feature extraction based on tensor decomposition is presented in this paper.In this paper,the virtual prototype of the engine is established in the GT-Crank software,and two kinds of faults of single cylinder misfire and misalignment of shafting are simulated by changing the relevant parameters.The data are obtained under three kinds of engine states of the normal state,the single cylinder misfire and the misalignment of the shaft,including the torque of the crankshaft,the inertia moment of the flywheel,the connecting force of the crank pin and the axial force of the connecting rod.Applied the tensor theory,the expression process from low order tensor to higher order tensor is analyzed.The above data are constructed into 32 three order tensor engine state samples are constructed which form are "signal class×crank angle×rotation speed" and size are 4×720×31 according to the three conditions,to provides the basis for the subsequent tensor analysis and decomposition.Through the analysis of the Tucker tensor problem,the basic principle and algorithm steps of HOSVD tensor Tucker decomposition algorithm and ALS tensor Tucker decomposition algorithm is given,and HOSVD-ALS simultaneous tensor Tucker decomposition algorithm is proposed.Three algorithms were used to extract the features of the samples,and 96 core samples of 4×8×2 were obtained.The sample is divided into training set and test set.Three classification models of decision tree,support vector machine and K-nearest neighbor are learned,and the data is classified by the data without feature extraction and with feature extraction based on tensor Tucker decomposition algorithm respectively.And the confusion matrix,classification accuracy and the model learning time were as evaluating indicator to compare the classification results.The results show that the classification accuracy obtained by using the HOSVD-ALS simultaneous tensor Tucker decomposition algorithm to extract the tensor data feature is 94.44%,and is higher than that obtained without construction and decomposition of tensor and by using the HOSVD tensor Tucker decomposition algorithm and the ALS tensor Tucker decomposition algorithm that were 93.06%,87.50%,and 82.64% respectively.The learning time of the data input classification model obtained by using the HOSVD-ALS simultaneous tensor Tucker decomposition algorithm,the HOSVD tensor Tucker decomposition algorithm and the ALS tensor Tucker decomposition algorithm were 0.132 s,0.128 s,0.131 s,and they are very close and significantly faster than the direct quantization method of 1.156 s.The effectiveness of the engine feature extraction method based on tensor decomposition is verified adequately.