Rolling bearing fault diagnosis method based on local reservation auto-encoder

Abstract

The invention discloses a rolling bearing fault diagnosis method based on a local reservation auto-encoder, and the method comprises the steps: 1, carrying out the fragmentation and whitening of original vibration signal samples, collected at different shaft speeds, of rolling bearings in various health states, so as to construct a global training sample set; 2, randomly selecting a certain proportion of samples from the global training sample set to form a local training sample set; 3, constructing a weight map by using a complexity-invariant distance measurement method; 4, using the global and local training sample sets and the weight map to train a local reservation auto-encoder model, and using trained model network parameters to extract feature vectors of original vibration signal samples so as to form a classifier training data set; and 5, training a classifier for bearing fault diagnosis by using the classifier training data set. According to the method, the data local structure information of the original space is reserved in the reconstruction space, so that the characteristics with rich information and identifiability are obtained, and the fault identification effect is effectively improved.

Description

Technical Field

[0001] The invention relates to the field of mechanical fault diagnosis, in particular to a rolling bearing fault diagnosis method based on a local retention autoencoder. Background Art

[0002] With the growing demand for modernization and automation in the manufacturing industry, mechanical equipment is rapidly developing in the direction of large-scale, high-speed and high-precision. Timely and rapid diagnosis of faults in complex systems can effectively avoid unnecessary downtime, serious economic losses and even casualties. Therefore, accurate fault diagnosis is crucial to the reliable operation of mechanical equipment.

[0003] Intelligent fault diagnosis can adaptively learn fault identification features from a large amount of raw monitoring data and accurately diagnose the health status of machinery. Feature extraction is a key part of intelligent fault diagnosis and can directly affect the diagnosis effect. Unlike traditional feature extraction methods that rely too much on experience and prior knowledge and the limitation that extracted features cannot be used in multi-task scenarios, unsupervised feature learning can directly learn features for fault diagnosis from diagnostic data. Autoencoders are one of the most widely used unsupervised feature learning methods, but they only learn approximate expressions of identity functions and cannot fully mine the structural information of the data. Summary of the invention

[0004] The present invention aims to solve the deficiencies of the above-mentioned prior art and proposes a rolling bearing fault diagnosis method based on a local preservation autoencoder, in order to utilize an unsupervised feature learning method to mine the fault information hidden in the original data, thereby improving the accuracy of rolling bearing fault diagnosis.

[0005] In order to achieve the above-mentioned purpose, the present invention adopts the following technical scheme:

[0006] The rolling bearing fault diagnosis method based on the local preservation autoencoder of the present invention is characterized in that it comprises the following steps:

[0007] Step 1: Slice and whiten the original vibration signal samples of rolling bearings in various health states collected at different shaft speeds to construct a global training sample set. ;

[0008] Obtaining a sample set of vibration signals of rolling bearings ,in, For the vibration signal samples, P represents the total number of sampling points, yes The corresponding health status; C represents the type of health status;

[0009] right The overlap rate is , the fragment length is After the fragmentation processing, whitening is performed to obtain the tth local sample fragment set , thus obtaining the global training sample set ,in, express The i-th local sample segment in , M represents the total number of local sample segments of each vibration signal sample, ;

[0010] Step 2: From the global training sample set Randomly select a certain proportion The local sample fragments constitute the local training sample set ,in, express The kth local sample fragment extracted in, represents the total number of local sample fragments extracted, and ;

[0011] Step 3: Based on local training sample set , using the complexity-invariant distance measurement method to construct a weighted graph ,in, is the edge set of the weighted graph;

[0012] Step 4: Construct a local preservation autoencoder, including encoder h and decoder d, and and After processing, we get The corresponding global reconstruction results and The corresponding local reconstruction results ;

[0013] Step 5: Use formula (4) to construct the loss function Loss of the local preservation autoencoder:

[0014] (4)

[0015] In formula (4), The similarity matrix and the identity matrix The combination matrix of is the local preservation penalty parameter, T represents transposition, represents the trace of a matrix;

[0016] Step 6: Taking the minimum loss function Loss as the goal, use the L-BFGS algorithm to optimize the parameters of the local preservation autoencoder to obtain the trained local preservation autoencoder;

[0017] Step 7: Use the trained local preservation autoencoder to extract the vibration signal sample set The feature vector of ;

[0018] Will Input the trained local preservation autoencoder model for processing to obtain the tth local feature representation set , and After average pooling, we get The corresponding eigenvector , thus obtaining the training data set for the classifier ,in, express Local feature representation of

[0019] Step 8: The training data set of the classifier The input is processed in the classifier, and the loss function of the classifier is constructed using formula (5) , thereby training the classifier and obtaining the optimal classifier after training;

[0020] (5)

[0021] In formula (5), is an indicator function. When the internal condition is met, is 1, otherwise, is 0; is the weight decay factor, Indicates the cth parameter corresponding to the classifier;

[0022] Step 9: Give any new vibration signal sample Input the trained local preservation autoencoder model for processing and obtain The corresponding eigenvector , and then the feature vector Input into the trained optimal classifier, and then use formula (4) to get Belong to Probability of healthy state , so as to realize the classification of bearing faults;

[0023] (6)

[0024] In formula (4), is the cth parameter of the optimal classifier.

[0025] The rolling bearing fault diagnosis method based on the local preservation autoencoder described in the present invention is also characterized in that the step 3 includes the following steps:

[0026] Step 3.1: Use the distance measurement formula with constant complexity to obtain of Nearest neighbor sample set ;

[0027] Step 3.2: Use formula (1) to obtain the similarity matrix :

[0028] (1)

[0029] In formula (1), express and The local sample fragments extracted The relationship between season , otherwise, let ;

[0030] Step 3.3: Each local sample fragment in is used as the vertex of the weight graph, and the similarity between the vertices, that is, the connection weight, is used as the edge of the weight graph to construct the weight graph .

[0031] Further, the step 4 comprises the following steps:

[0032] Step 4.1: Global training sample set Composing the global data matrix , and input into the encoder h for processing, so as to obtain its corresponding hidden feature representation using formula (2) ;

[0033] (2)

[0034] In formula (2), , are the weight matrix and bias of encoder h, respectively. Indicates that copy The operator of the second is the activation function;

[0035] Step 4.2: Decoder Using formula (3) to represent the hidden features Perform decoding operation and get The corresponding global reconstruction results ;

[0036] (3)

[0037] In formula (3), , are the weight matrix and bias of decoder d respectively;

[0038] Step 4.3: From the local training sample set Composing local data matrix , and after encoding and decoding according to step 4.1-step 4.2, we get The corresponding local reconstruction results .

[0039] The electronic device of the present invention includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the rolling bearing fault diagnosis method, and the processor is configured to execute the program stored in the memory.

[0040] The present invention provides a computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and the computer program executes the steps of the rolling bearing fault diagnosis method when the computer program is executed by a processor.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] 1. Since bearings usually work under complex working conditions of variable speed and variable load, their monitoring signals are highly nonlinear, non-stationary, noisy, and have weak fault characteristics. The present invention achieves accurate reconstruction and local feature retention of data from the original space to the reconstructed space by jointly optimizing the local retention penalty term of the cost function of the standard autoencoder, effectively improving the learning ability of discriminative features under an unsupervised framework and enhancing the ability to extract complex signal features of bearings.

[0043] 2. Since the monitoring signal of the bearing is usually time series data, the Euclidean distance as a common similarity measurement method is not suitable for measuring the similarity between two time series data. The present invention uses the complexity invariant distance measurement to replace the traditional Euclidean distance measurement, thereby accurately evaluating the similarity between two time series data. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a flow chart of the rolling bearing fault diagnosis method of the present invention;

[0045] Figure 2 The figure is a flow chart of fault feature extraction of the method of the present invention. DETAILED DESCRIPTION

[0046] In this embodiment, Figure 1 As shown, a rolling bearing fault diagnosis method based on a local retention autoencoder includes the following steps:

[0047] Step 1: Slice and whiten the original vibration signal samples of rolling bearings in various health states collected at different shaft speeds to construct a global training sample set. ;

[0048] Obtaining a sample set of vibration signals of rolling bearings ,in, For the vibration signal samples, P represents the total number of sampling points, yes The corresponding health status; C represents the type of health status;

[0049] right The overlap rate is , the fragment length is After the fragmentation processing, whitening is performed to obtain the tth local sample fragment set , thus obtaining the global training sample set ,in, express The i-th local sample segment in , M represents the total number of local sample segments of each vibration signal sample, .

[0050] Step 2: From the global training sample set Randomly select a certain proportion The local sample fragments constitute the local training sample set ,in, express The kth local sample fragment extracted in, represents the total number of local sample fragments extracted, and .

[0051] Step 3: Based on local training sample set ,The weight graph is constructed using the complexity invariant distance measurement method to obtain the local geometric structure of the vibration signal samples;

[0052] Step 3.1: Use the distance measurement formula with constant complexity to obtain of Nearest neighbor sample set ; Each sample It is a time series signal (i.e. time series data). Due to the time shift characteristics of time series data, Euclidean distance is not suitable for processing time series data. Complexity invariant distance has higher measurement accuracy in evaluating the complexity information of two time series data segments. Therefore, complexity invariant distance is used to measure the similarity of two time series data segments.

[0053] Step 3.2: Use formula (1) to obtain the similarity matrix :

[0054] (1)

[0055] In formula (1), express and The local sample fragments extracted The relationship between season , otherwise, let ; Formula (1) is a standard constrained quadratic programming problem, which consists of N sub-optimization problems. In each sub-problem, the optimization parameter is a K-dimensional vector, and the constraints include 1 equality constraint and K inequality constraints;

[0056] Step 3.3: Each local sample fragment in is used as the vertex of the weight graph, and the similarity between the vertices (i.e., the connection weight) is used as the edge of the weight graph to construct the weight graph ; is the set of vertices of the graph, is the edge set of the weighted graph.

[0057] Step 4: Construct a local preservation autoencoder, including the encoder and decoder , and and After processing, we get The corresponding global reconstruction results and The corresponding local reconstruction results ;

[0058] Step 4.1: Global training sample set Each sample in is stacked together in columns to form a column of size The global data matrix , and input the encoder In order to process it, we can use formula (2) to Encode and get its corresponding hidden feature representation ;

[0059] (2)

[0060] In formula (2), , are the weight matrix and bias of encoder h, respectively, and the operator Indicates that copy The operator of the second It is an activation function; the commonly used Sigmoid activation function has the advantages of nonlinearity and being differentiable everywhere;

[0061] Step 4.2: Decoder Using formula (3) to represent the hidden features Perform decoding operation and get The corresponding global reconstruction results ;

[0062] (3)

[0063] In formula (3), , are the weight matrix and bias of decoder d respectively;

[0064] Step 4.3: From the local training sample set Each sample in is stacked together in columns to form a column of size The local data matrix , and after encoding and decoding according to step 4.1-step 4.2, we get The corresponding local reconstruction results ; and All sizes are The data matrix is ​​encoded and decoded in the same way as in equation (2) and equation (3). and The purpose is to construct a local structure of the original training samples, and to mine the discriminative feature information contained in the local structure by constructing a loss function to train the network.

[0065] Step 5: Use formula (4) to construct the loss function Loss of the local preservation autoencoder:

[0066] (4)

[0067] In formula (4), is composed of the similarity matrix and the identity matrix The combination matrix of is the local retention penalty parameter; Represents the trace of the matrix.

[0068] Step 6: Taking the minimum loss function Loss as the goal, use the L-BFGS algorithm to optimize the parameters of the local preservation autoencoder to obtain the trained local preservation autoencoder; L-BFGS is a commonly used quasi-Newton method with the advantages of high computational efficiency and adaptive selection of learning rate. In each iteration, only the gradient and loss function value of the previous step need to be provided to achieve the current iteration;

[0069] Step 7: Use the trained local preservation autoencoder to extract the vibration signal sample set The feature vector of , the specific steps are as follows Figure 2 As shown, Input the trained local preservation autoencoder model for processing to obtain the tth local feature representation set , and After average pooling, we get The corresponding eigenvector , thus obtaining the training data set for the classifier ,in, express local feature representation.

[0070] Step 8: The training data set of the classifier The input is processed in the classifier, and the loss function of the classifier is constructed using formula (5) , thereby training the classifier and obtaining the optimal classifier after training;

[0071] (5)

[0072] In formula (5), is an indicator function. When the internal condition is met, is 1, otherwise, is 0; is the weight decay factor, represents the cth parameter corresponding to the classifier; T represents transposition; Formula (5) can also be optimized using the L-BFGS algorithm.

[0073] Step 9: Give any new vibration signal sample Input the trained local preservation autoencoder model for processing and obtain The corresponding eigenvector , and then the feature vector Input into the trained optimal classifier, and then use formula (4) to get Belong to Probability of healthy state , so as to realize the classification of bearing faults;

[0074] (6)

[0075] In formula (4), is the cth parameter of the optimal classifier, There are C elements in Corresponding to labels respectively, each label corresponds to a bearing fault category, according to The label corresponding to the largest element in can determine the current input signal When the bearing is faulty, the fault type of the bearing can be determined, thereby realizing the fault diagnosis of the bearing.

[0076] In this embodiment, an electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the above method, and the processor is configured to execute the program stored in the memory.

[0077] In this embodiment, a computer-readable storage medium stores a computer program on the computer-readable storage medium, and the computer program executes the steps of the above method when executed by a processor.

Claims

1. A rolling bearing fault diagnosis method based on local preservation autoencoder, characterized in that: The following steps are involved: Step 1: Slice and whiten the original vibration signal samples of rolling bearings in various health states collected at different shaft speeds to construct a global training sample set. ; Obtaining a sample set of vibration signals of rolling bearings ,in, For the vibration signal samples, P represents the total number of sampling points, yes The corresponding health status; C represents the type of health status; right The overlap rate is , the fragment length is After the fragmentation processing, whitening is performed to obtain the tth local sample fragment set , thus obtaining the global training sample set ,in, express The i-th local sample segment in , M represents the total number of local sample segments of each vibration signal sample, ; Step 2: From the global training sample set Randomly select a certain proportion The local sample fragments constitute the local training sample set ,in, express The kth local sample fragment extracted in, represents the total number of local sample fragments extracted, and ; Step 3: Based on local training sample set , using the complexity-invariant distance measurement method to construct a weighted graph ,in, is the edge set of the weighted graph; Step 4: Construct a local preservation autoencoder, including encoder h and decoder d, and and After processing, we get The corresponding global reconstruction results and The corresponding local reconstruction results ; Step 5: Use formula (4) to construct the loss function Loss of the local preservation autoencoder: (4) In formula (4), The similarity matrix and the identity matrix The combination matrix of is the local preservation penalty parameter, T represents transposition, represents the trace of a matrix; Step 6: Taking the minimum loss function Loss as the goal, use the L-BFGS algorithm to optimize the parameters of the local preservation autoencoder to obtain the trained local preservation autoencoder; Step 7: Use the trained local preservation autoencoder to extract the vibration signal sample set The feature vector of ; Will Input the trained local preservation autoencoder model for processing to obtain the tth local feature representation set , and After average pooling, we get The corresponding eigenvector , thus obtaining the training data set for the classifier ,in, express Local feature representation of Step 8: The training data set of the classifier The input is processed in the classifier, and the loss function of the classifier is constructed using formula (5) , thereby training the classifier and obtaining the optimal classifier after training; (5) In formula (5), is an indicator function. When the internal condition is met, is 1, otherwise, is 0; is the weight decay factor, Indicates the cth parameter corresponding to the classifier; Step 9: Give any new vibration signal sample Input the trained local preservation autoencoder model for processing and obtain The corresponding eigenvector , and then the feature vector Input into the trained optimal classifier, and then use formula (4) to get Belong to Probability of healthy state , so as to realize the classification of bearing faults; (6) In formula (4), is the cth parameter of the optimal classifier.

2. A rolling bearing fault diagnosis method based on local preservation autoencoder according to claim 1, characterized in that: The step 3 comprises the following steps: Step 3.1: Use the distance measurement formula with constant complexity to obtain of Nearest neighbor sample set ; Step 3.2: Use formula (1) to obtain the similarity matrix : (1) In formula (1), express and The extracted local sample fragments The relationship between season , otherwise, let ; Step 3.3: Each local sample fragment in is used as the vertex of the weight graph, and the similarity between the vertices, that is, the connection weight, is used as the edge of the weight graph to construct the weight graph .

3. A rolling bearing fault diagnosis method based on local preservation autoencoder according to claim 2, characterized in that: The step 4 comprises the following steps: Step 4.1: Global training sample set Composing the global data matrix , and input into the encoder h for processing, so as to obtain its corresponding hidden feature representation using formula (2) ; (2) In formula (2), , are the weight matrix and bias of encoder h, respectively. Indicates that copy The operator of the second is the activation function; Step 4.2, Decoder Using formula (3) to represent the hidden features Perform decoding operation and get The corresponding global reconstruction results ; (3) In formula (3), , are the weight matrix and bias of decoder d respectively; Step 4.3: From the local training sample set Composing local data matrix , and after encoding and decoding according to step 4.1-step 4.2, we get The corresponding local reconstruction results .

4. An electronic device, comprising a memory and a processor, characterized in that: The memory is used to store a program that supports the processor to execute the rolling bearing fault diagnosis method described in any one of claims 1 to 3, and the processor is configured to execute the program stored in the memory.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the rolling bearing fault diagnosis method according to any one of claims 1 to 3 are executed.

Images