Research On Indirect Signal Data Fault Diagnosis Method Based On Tensor Model And Convolutional Neural Network

The traction motor and bearing are the key equipment of the rail transit train,which are of great significance to ensure the normal operation of the traction motor and bearing for train running and passenger safety.At present,the fault diagnosis of bearing is often based on the analysis of the vibration signal collected by its sensor.In fact,the fault of bearing will be transmitted to the adjacent equipment in the interaction between the equipment,and then it is reflected in the monitoring signal of the adjacent equipment.This kind of monitoring signal is called indirect signal in this paper.Indirect signal can not only be used as the auxiliary signal for vibration signal to improve the accuracy of diagnosis,but also can be used as a main means of diagnosis when the direct sensor of equipment fails.Thus,the study of indirect signal is of great significance to the field of fault diagnosis.In this paper,the bearing driven by permanent magnet synchronous motor is taken as the research object.Aiming at the problems that the time-frequency characteristics of indirect signal are not obvious and the damage degree of bearing is difficult to identify,the algorithms of dimension and noise reduction,feature extraction,damage classification and damage degree identification suitable for direct and indirect signal data are proposed.The main work of this paper is as follows.(1)The selection of indirect signal based on periodic characteristics of signal.High order principal component analysis and correlation analysis are used to identify the current and torque signals from vibration,current,torque and radial force signals.These two indirect signals can be used for fault diagnosis.(2)The dimension reduction and noise reduction algorithm of indirect signal.To solve the problem that the conventional low-order time-frequency analysis method is not suitable for indirect signals,a third-order signal processing model based on T-HOSVD and L-curve method is proposed.The third-order tensor model of N-mode matrix signal is used to the decompose the truncated singular value of each matrix,and the L-curve method is used to obtain the truncation parameters.The noise reduction is completed in the condition of retaining their characteristics.Each truncated left singular matrix is used to form a new kernel tensor,and then the tensor will be reconstructed.The denoising performance is analyzed,and the mapping of fault characteristic frequency in indirect signal is studied by means of power spectrum.(3)The feature extraction of indirect signal.In order to solve the problem that the time-frequency characteristics of indirect signals are not obvious,a feature extraction method based on tensor T-QR-HOOI is proposed.The third-order tensor model of the original signal is decomposed into feature tensor and eigenvalue tensor by T-QR algorithm.As the initial solution,the feature tensor can be decomposed into the core tensor which can represent the global characteristics of the signal by HOOI decomposition.The vectorization of feature tensor is used as the initial feature vector.After the dual screening of distance and error weight,the final feature vector can be used as the input of the classifier for fault diagnosis.Several machine learning classifier models are used to verify the performance of indirect signals in the accuracy of damage classification of bearing.(4)The algorithm of damage degree evaluation of bearing based on indirect signal.In view of the instability of indirect signal on damage degree identification,a dual mode CNN algorithm is proposed.The 2D input end of time-frequency graph and 1D end of feature vector are dealt by independent convolution pool to extract feature,and feature fusion is carried out in the convergence layer for diagnosis and classification.The experiment shows that the indirect signal can be used as the input of model to realize the accurate diagnosis of damage classification and assessment of damage degree.