Cross-modal bearing fault diagnosis method based on self-supervised space learning

Abstract

The invention discloses a cross-modal bearing fault diagnosis method based on self-supervised space learning, which mainly constructs a unified optimization model of class labels and a cross-modal space projection direction, so that the self-supervised learning of the class labels can be realized under the condition of unsupervised fault diagnosis, and the accuracy of fault diagnosis is effectively improved. The specific implementation process is as follows: (1) constructing a cross-modal self-supervision space learning model by using cross-modal class label information and a related space theory; (2) theoretically deriving indirect representation of a cross-modal class center, and further obtaining an analytical solution of a class label in a learned cross-modal consistent space; and (3) obtaining cross-modal fault features of the fault sample data on the basis of the cross-modal self-supervised space learning projection direction, and inputting the low-dimensional discrimination unified fault features into the classifier to obtain a final bearing fault diagnosis result. Compared with the prior art, the fault diagnosis method is more effective and robust.

Description

Technical Field

[0001] The invention relates to a cross-modal bearing fault diagnosis method based on self-supervised space learning, which belongs to the field of pattern recognition and fault diagnosis. Background Art

[0002] In the field of fault diagnosis, bearings, as key components of mechanical equipment, can ensure the stable operation of mechanical components, significantly improve the operating efficiency of equipment, and reduce energy consumption by reducing friction and improving rotation stability. However, once a device fails, it will cause huge economic losses or even safety accidents. Therefore, it is a key issue to effectively diagnose the key components of mechanical equipment, extract valuable identification features from high-dimensional fault data, and improve diagnostic performance. Among the methods to solve such problems, spatial learning is an effective method. However, due to the lack of class labels for rolling bearing faults in unsupervised cross-modal spatial learning, it is impossible to extract richer feature information and it is difficult to effectively improve the class separation of the learning space. To this end, the invention of this paper constructs a unified optimization model of class labels and cross-modal spatial projection directions with the help of relevant space theory and cross-modal class label functions. This model can solve the problem of lack of class label identification information in unsupervised learning. Through the theoretical derivation of the optimization model, the analytical solution of class labels in the cross-modal self-supervised spatial learning model is obtained, and self-supervised learning of class labels in unsupervised conditions is realized. In addition, the invention also effectively solves the problem of difficulty in fault separation caused by the lack of adaptability between cross-modal fusion and downstream task clustering. Summary of the invention

[0003] In order to effectively utilize the rich identification information of self-supervised pseudo-labels, the present invention constructs a cross-modal self-supervised spatial learning model based on spatial learning theory, and theoretically derives the analytical solution of the model, thereby solving the problem of lack of class label identification information in unsupervised learning. The specific implementation steps of the present invention are as follows:

[0004] 1. Collect the operating data of mechanical equipment bearing faults through sensors, extract relevant statistical features of multiple bearing fault data from the time domain, frequency domain and time-frequency domain to construct a cross-modal fault data sample set where d i represents the characteristic dimension of the i-th data set, N represents the number of samples, and m represents the number of modes of fault data; the fault data is divided into training sample sets trainX1, trainX2 and test sample sets testX1, testX2 in proportion.

[0005] 2. Construct a cross-modal self-supervised spatial learning model to obtain a low-dimensional representation of high-dimensional fault features.

[0006] The specific steps for constructing the cross-modal self-supervised spatial learning model are as follows:

[0007] (2a) Construction of class label optimization function:

[0008] Assuming that there are K class centers in each mode, after each mode is mapped The class center is In the shared subspace, the class center learning function in the mth mode is expressed as:

[0009]

[0010] in, represents the number of samples belonging to class v in the m mode. And further define For the class label function:

[0011]

[0012] Among them, the class label is represented by F = [f1,f2,…,f n ]∈R (K×N) ,f n =[f n1 ,f n2 ,…,f nK ]∈R (K×N) , n=1,2,…,N. m=1,2 means if and only if f nv =1, belongs to class v, otherwise f nv =0.

[0013] (2b) Cross-modal neighborhood consistent class center objective optimization function:

[0014] Cross-modal data is usually high-dimensional data containing noise and redundant information. The neighborhood relationship based on high-dimensional data often deviates from the inherent neighborhood relationship, resulting in weak class separability of fused related features. To address this problem, the present invention projects high-dimensional data into the relevant fusion subspace to obtain inter-modal related features. The related features between different modalities have the same dimension. The cross-modal class center is constructed in the fused subspace, and the indirect representation of the cross-modal class center is theoretically derived:

[0015]

[0016] Among them, F represents the class label matrix, c v represents the cross-modal class center, which can be regarded as a set of discrete points within the class.

[0017] The cross-modal neighborhood consistent class center objective optimization function can effectively integrate the identification information between modalities, improve the complementarity between modalities, and weaken the impact of label distortion. The cross-modal neighborhood consistent class center objective optimization function is defined as:

[0018]

[0019] in, S represents the feature matrix.

[0020] (2c) Construction of cross-modal self-supervised spatial learning model:

[0021] Based on the solution of the above-mentioned class label optimization function and the cross-modal neighborhood consistent class center objective optimization function and related space learning theory, a cross-modal self-supervised space learning model is constructed to address the problem of lack of class label identification information in unsupervised learning:

[0022]

[0023] in, The cross-modal self-supervised spatial learning model seeks a shared low-dimensional representation between different modalities, minimizes the intra-class discreteness of each modality, and effectively enhances the class separability of related features; maximizes the correlation between modalities, and retains the complementary information between cross-modal data to a large extent.

[0024] 3. Use iterative optimization method to solve the class center matrix C and projection dimension reduction matrix H of the cross-modal self-supervised spatial learning model m and the class label matrix F.

[0025] (3a) Fix the class label matrix F and optimize the projection matrix H m And the center matrix C:

[0026] The Lagrange multiplier method is used to obtain the analytical representation of C and construct the Lagrange multiplier function L(c v ):

[0027]

[0028] Among them, λ is the Lagrange multiplier, which Substitute into the above formula and calculate c v Find partial derivatives:

[0029]

[0030] make Zero gives:

[0031]

[0032] Therefore, c vThe representation is:

[0033]

[0034] By expanding the above formula, we can deduce the cross-modal class center matrix C:

[0035]

[0036] When the Lagrange multiplier method is used to solve the problem, the solution is the PR corresponding to the first d largest eigenvalues -1 The generalized eigenvector of . is a block matrix, From S we can get the projection vector

[0037] (3b) Fixed projection matrix H m And the center matrix C, optimize the class label matrix F:

[0038] The optimization objective function solved by the cross-modal self-supervised spatial learning model is reformulated as:

[0039]

[0040] When H m and c v When fixed, the denominator in the formula is a constant, and only the numerator in the formula is minimized. v is fixed, F=[f1,f2,…,f n The optimization problem of ] is simplified to the nearest neighbor problem of the cluster center. n =[f n1 ,f n2 ,…,f nK ]∈R (K×N) The optimal solution for (n=1,2,…,N) is:

[0041]

[0042] 4. Obtain the low-dimensional training sample feature set (trainY1, trainY2) and low-dimensional test sample feature set (testY1, testY2) of bearing fault data through spatial projection; use a parallel fusion strategy to fuse low-dimensional training sample sets (trainY1, trainY2) or low-dimensional test sample feature sets (testY1, testY2) of different modes, and then obtain the fused low-dimensional training sample set and the fused low-dimensional test sample set; finally, use a classifier for classification to obtain the fault diagnosis result.

[0043] The method of the present invention has the following advantages:

[0044] (1) The present invention can effectively utilize the rich identification information of self-supervised pseudo-labels, and based on this, learn the cross-modal self-supervised space, effectively improving the class separation of cross-modal bearing fault features;

[0045] (2) The present invention further constrains the learning of the spatial projection matrix by means of the correlation space learning theory and the cross-modal class label function, which can effectively solve the problem of lack of class label identification information in unsupervised learning, realize self-supervised learning in an unsupervised environment, and significantly improve the accuracy of fault diagnosis;

[0046] (3) Through theoretical derivation, the present invention obtains an analytical solution for the cross-modal self-supervised spatial learning model, effectively utilizing class label information and capturing similar structures between different modalities. By minimizing the intra-class dispersion of each modality, the class separability of related features is effectively enhanced; by maximizing the correlation between modalities, the complementary information between cross-modal data is retained as much as possible. In this way, the cross-modal fault test features of the fault test samples can be quickly obtained, achieving more accurate fault type classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0048] Figure 1 is the flow chart of the present invention, where n is the number of fault categories

[0049] Figure 2 is the classification accuracy of the random experiment DETAILED DESCRIPTION

[0050] The specific implementation steps of the present invention are as follows:

[0051] 1. Collect the operating data of mechanical equipment bearing faults through sensors, extract relevant statistical features of multiple bearing fault data from the time domain, frequency domain and time-frequency domain to construct a cross-modal fault data sample set where d i represents the characteristic dimension of the i-th data set, N represents the number of samples, and m represents the number of modes of fault data; the fault data is divided into training sample sets trainX1, trainX2 and test sample sets testX1, testX2 in proportion.

[0052] 2. Based on the cross-modal self-supervised spatial learning model, an iterative solution method is used to continuously update the optimization class center matrix C and the projection dimension reduction matrix H m And the class label matrix F, get the low-dimensional projection direction By weakening the incompatibility between dimensionality reduction and clustering of high-dimensional data samples, the clustering performance is optimized, and the class label can be used to indicate the class center in an unsupervised environment.

[0053] 3. Obtain low-dimensional training sample feature sets and low-dimensional test sample feature sets of bearing fault data through spatial projection; use a parallel fusion strategy to fuse low-dimensional training sample sets or low-dimensional test sample feature sets of different modes, and then obtain the fused low-dimensional training sample sets and low-dimensional test sample sets; finally, use a classifier for classification to obtain the fault diagnosis results.

[0054] The effect of the present invention is further verified by the following experiments:

[0055] Two data modes were selected from the Paderborn bearing data set for experimental verification, namely the motor current signal and the vibration signal. In this experiment, the selected bearing fault data consists of 1000 samples, the sampling frequency is 64KHz, and the current and vibration signals are divided into 250 samples with a sampling length of 1024. The experiment selected two groups of man-made damage fault data, one type of accelerated life damage data, and one type of fault-free data. Among them, the inner ring fault and the outer ring fault are man-made damage faults, recorded as F1 and F2; the mixed fault of the inner ring and the outer ring is accelerated life damage, recorded as F3; the fault-free data is recorded as F4. Figure 2 The classification accuracy of bearing fault diagnosis in each random experiment is intuitively displayed. Figure 2 It can be seen that as the number of training samples gradually increases, the accuracy of the method of the present invention increases accordingly, and the average diagnostic accuracy of the best random experiment reaches 99.79%, and the stability is good. The experimental results show that the method disclosed by the present invention is an effective bearing fault diagnosis method.

Claims

1. The present invention discloses a cross-modal bearing fault diagnosis method based on self-supervised spatial learning, which is characterized by The method comprises the following steps: (1) Collect the operating data of mechanical equipment bearing faults through sensors, extract the relevant statistical features of multiple bearing fault data from the time domain, frequency domain and time-frequency domain, and construct a cross-modal fault data sample set where d i represents the characteristic dimension of the i-th data set, N represents the number of samples, and m represents the number of modes of fault data; the fault data is divided into training sample sets trainX1, trainX2 and test sample sets testX1, testX2 in proportion; (2) Construct a self-supervised spatial learning model for cross-modal fault data; (3) Use iterative optimization method to solve the class center matrix C and projection dimension reduction matrix H of the cross-modal self-supervised spatial learning model m and class label matrix F; (4) Obtain a low-dimensional training sample feature set and a low-dimensional test sample feature set of the bearing fault data through spatial projection; use a parallel fusion strategy to fuse low-dimensional training sample sets or low-dimensional test sample feature sets of different modes to obtain a fused low-dimensional training sample set and a fused low-dimensional test sample set; finally, use a classifier to perform classification to obtain a fault diagnosis result.

2. The cross-modal bearing fault diagnosis method based on self-supervised spatial learning according to claim 1 is characterized in that The steps of constructing a cross-modal self-supervised spatial learning model in step (2) to obtain a low-dimensional representation of the high-dimensional fault features are as follows: (1) Assuming that each mode has K cluster centers, each mode is mapped The class center is definition The optimization function for class labels is derived theoretically, and the optimization function for the cross-modal neighborhood consistent class center objective is obtained; (2) Combining the solution of the class label optimization function and the cross-modal neighborhood consistent class center objective optimization function and related spatial learning theory, a cross-modal self-supervised spatial learning model is constructed to address the problem of lack of class label identification information in unsupervised learning: Among them, H m is the required projection vector, S represents the feature matrix, F represents the class label matrix, c v represents the cross-modal class center, Taking the class label function as an example, construct a class label optimization function in, represents the number of samples belonging to class v in the m mode, F = [f1, f2, ..., f n ]∈R (K×N) , f n =[f n1 , f n2 , ..., f nK ]∈R (K×N) , n=1,2,...,N; It means that if and only if f nv =1, belongs to class v, otherwise f nv =0; The cross-modal neighborhood consistent class center objective optimization function can effectively integrate the identification information between modalities, improve the complementarity between modalities, and weaken the influence of label distortion. The cross-modal neighborhood consistent class center objective optimization function is defined as: The cross-modal self-supervised spatial learning model seeks a shared low-dimensional representation between different modalities, minimizes the intra-class discreteness of each modality, and effectively enhances the class separability of related features; maximizes the correlation between modalities, and retains the complementary information between cross-modal data to a large extent.

3. The cross-modal bearing fault diagnosis method based on self-supervised spatial learning according to claim 1 is characterized in that The iterative optimization method described in step (3) is used to solve the class center matrix C and the projection dimension reduction matrix H m And the class label matrix F, the steps are as follows: (3a) Fix the class label matrix F and optimize the projection matrix H m And the class center matrix C: The Lagrange multiplier method is used to obtain the analytical representation of C and construct the Lagrange multiplier function L(c v ): Among them, λ is the Lagrange multiplier, which Substitute into the above formula and calculate c v Find the partial derivative: make Zero gives: Therefore, c v The representation is: By expanding the above formula, we can deduce the cross-modal class center matrix C: When the Lagrange multiplier method is used to solve the problem, the solution is the PR- 1 The generalized eigenvector of is a block matrix, From S we can get the projection vector (3b) Fixed projection matrix H m And the center matrix C, optimize the class label matrix F: The optimization objective function solved by the cross-modal self-supervised spatial learning model is reformulated as: When H m and c v When fixed, the denominator in the formula is a constant, and only the numerator in the formula is minimized; since c v is fixed, F = [f1, f2, ..., f n ] is simplified to the nearest neighbor problem of the cluster center, f n =[f n1 , f n2 , ..., f nK ]∈R (K×N) , n=1,2,...,N, the optimal solution is: The present invention uses canonical correlation analysis theory to fuse cross-modal information, enhance the complementarity of fault samples, obtain richer identification information, and further consider cross-modal class centers to weaken the influence of class label distortion. Finally, the projection matrix is ​​obtained. The class center matrix C and class label matrix F realize self-supervised learning in an unsupervised situation and sharing of class centers.

Images