A Rolling Bearing Fault Diagnosis Method and System Based on the Belief Rule Base

Abstract

The present invention discloses a rolling bearing fault diagnosis method and system based on a belief rule base. The method includes: obtaining a feature extraction model by training a self-learning network; extracting features from labeled data and training a diagnostic classifier; extracting the intrinsic features of unlabeled data as the input of the diagnostic classifier and outputting pseudo-labels; correcting the pseudo-labels through the belief rule base; retraining the diagnostic classifier with the unlabeled data and the corrected labels as the input, and updating the belief rule base based on the labels output again by the trained diagnostic classifier; repeatedly and iteratively training the diagnostic classifier and updating the belief rule base to obtain a diagnostic classifier that meets the requirements; and performing fault diagnosis on the bearing based on the feature extraction model and the diagnostic classifier. The present invention fully considers the situation of incomplete belief rule base and less labeled data, and can quickly diagnose the occurring faults according to the collected data and find out the causes of the faults.

Description

Technical Field

[0001] The present invention belongs to the field of rolling bearing fault diagnosis, and particularly relates to a rolling bearing fault diagnosis method and system based on a belief rule base. Background Art

[0002] With the rapid development of technology, mechanical equipment is becoming larger and more intelligent, and plays an increasingly important role in many industrial fields. Rolling bearings are vulnerable and critical components in mechanical equipment, and they are prone to failure when working in harsh working environments and under alternating loads. Faults in rolling bearings may cause the entire mechanical system to stop running, resulting in huge economic losses. In addition, due to the influence of various non-linear factors, the vibration signals of rolling bearings usually have the characteristics of non-linearity and non-stationarity, which indicates that it is difficult to extract fault feature information and identify fault types through traditional methods (such as fast Fourier transform). Therefore, physical models cannot effectively handle systems with high complexity and strong noise, and many data-driven fault diagnosis methods have been successfully developed currently.

[0003] Although the existing data-driven methods have obtained good results currently, most of these methods are implemented based on the way of supervised learning. Data-driven methods based on the supervised learning way have some limitations. Because in the real industrial environment, it is unrealistic to obtain enough expert-labeled data, so there is less labeled bearing fault data, and most bearing fault data are not labeled. The reduction of training data will lead to a significant decline in diagnostic performance. And there is some domain knowledge in the field of rolling bearing fault diagnosis, however, this valuable information has not been fully integrated into the diagnostic model to improve the performance and interpretability of the model. Summary of the Invention

[0004] The purpose of the present invention is to provide a rolling bearing fault diagnosis method and system based on a belief rule base, to solve the problem that the accuracy of rolling bearing fault diagnosis results is low due to less labeled bearing fault data in rolling bearings, and to quickly diagnose the occurring faults.

[0005] The technical solution for realizing the purpose of the present invention is: A rolling bearing fault diagnosis method based on a belief rule base, including the steps of:

[0006] Obtain an unlabeled data training set, and obtain a feature extraction model by training a self-learning network;

[0007] Use the labeled data set as the training set, extract features of the labeled data through the feature extraction model, and train a diagnostic classifier;

[0008] Extract the internal features of the unlabeled data through the feature extraction model, use them as the input of the diagnostic classifier, and output pseudo-labels;

[0009] Based on abductive reasoning, the pseudo-labels are corrected through the belief rule base to obtain corrected labels;

[0010] The unlabeled data and the corrected labels are used as the input of the diagnostic classifier for retraining, and the belief rule base is updated based on the labels output after retraining the diagnostic classifier;

[0011] Iteratively train the diagnostic classifier and update the belief rule base repeatedly to obtain a diagnostic classifier whose accuracy meets the set threshold;

[0012] Based on the feature extraction model and the diagnostic classifier, fault diagnosis of the bearing is carried out.

[0013] Furthermore, the self-learning network includes an encoder and a decoder. The encoder includes a neural network input layer and a hidden layer, which are used to extract the intrinsic features of the unlabeled data. The decoder includes a neural network hidden layer and an output layer, which are used to decode the output of the encoder to obtain the reconstructed data.

[0014] Furthermore, the encoding process of the encoder is as follows:

[0015] z = h(x) = σ(W 1 x + b 1 )

[0016] where h represents the encoder, x is the input, z represents the intrinsic representation of the unlabeled data, and W 1 and b 1 represent the parameters of the encoder.

[0017] Furthermore, the decoder aims to reconstruct the latent features in the new feature space into the original input, and the decoding process is as follows:

[0018] x' = g(z) = σ(W 2 z + b 2 )

[0019] where g represents the decoder, z represents the intrinsic representation of the unlabeled data, x' represents the reconstructed features, and W 2 and b 2 represent the parameters of the decoder.

[0020] Furthermore, when training the self-learning network, the objective function of self-learning is expressed as follows:

[0021]

[0022] where N represents the number of unlabeled samples, dist represents the distance between the original input x and the reconstructed input x'. When the value of the objective function meets the set threshold, the encoder is the feature extraction model.

[0023] Furthermore, the diagnostic classifier adopts a BP neural network.

[0024] Furthermore, the pseudo-labels are corrected through a belief rule base to obtain corrected labels, specifically including:

[0025] Fuzzify the unlabeled data X u through a fuzzy function to obtain fuzzy samples Unlabeled data;

[0026] Find the fault mode rule set R i ;

[0027] in the belief rule base that matches the fuzzy samples, and simplify the fault mode rule set R i ;

[0028] by deleting the fault mode rules included by other rules; i Screen the simplified fault mode rule set R

[0029] according to the belief degrees of the belief vectors, and retain the fault mode rules with the highest belief degrees for each fault category;

[0030]

[0031] Among them, β k represents the belief vector, θ k represents the rule weight, and k represents the number of belief rules in the fault rule pattern rule set with the highest belief degree;

[0032] Correct the pseudo-label y u to generate a new label y r = y u * ω c .

[0033] Furthermore, the fuzzy function is:

[0034]

[0035] Among them, x i,j represents the j-th parameter of the i-th unlabeled sample, and respectively represent the upper and lower bounds of the threshold of the j-th parameter, represents the fuzzy value of the j-th parameter of the i-th unlabeled sample, and H, N, and L represent the fuzzy values corresponding to x i,j in different ranges.

[0036] Furthermore, the belief rule base is updated based on the labels output by re-acquiring the trained diagnostic classifier, specifically including:

[0037] If the input is labeled data, select the corresponding fuzzy data from the belief rule base The matching rule set, if There is a corresponding true label, the belief degree of the matching rule for this true label will increase by α l , otherwise the belief degree of the predicted label of the diagnostic classifier will be reduced by the penalty value γ; for unlabeled data, the belief degree corresponding to the predicted label in the matching fault mode rule set increases by α u , where α u <α l , based on the newly obtained belief vector Combined with the original belief vector , update the belief vector of the belief rule base as:

[0038]

[0039] A system based on the rolling bearing fault diagnosis method, including a self-learning module, a fault diagnosis model, a belief rule base, a label correction module, and a belief rule base update module, where

[0040] The self-learning module is based on an unlabeled data training set, and obtains a feature extraction model by training a self-learning network;

[0041] The fault diagnosis model includes a feature extraction model and a diagnostic classifier, and is used to obtain the cause of rolling bearing faults;

[0042] The belief rule base is a database containing belief rules;

[0043] The label correction module is used to correct the pseudo labels through the belief rule base to obtain corrected labels;

[0044] The belief rule base update module updates the belief rule base based on the labels output by re-acquiring the trained diagnostic classifier.

[0045] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention extracts the intrinsic features of tags through a self-learning method, provides more and more accurate features for fault diagnosis for model training, and improves the accuracy of the model; the present invention corrects the tags output by the fault diagnosis model through a belief rule base, and the corrected tags are used for the training of the fault diagnosis model. The output of the trained fault diagnosis model is used to supplement and update the belief rule base, and they are continuously updated and enhanced in an iterative manner to improve the accuracy of rolling bearing fault diagnosis; when the bearing fails, the present invention can quickly diagnose the occurring fault based on the collected data and find out the cause of the fault. Description of the Drawings

[0046] Figure 1 It is a schematic module diagram of the method of the present invention.

[0047] Figure 2 It is a schematic flowchart of the method of the present invention.

[0048] Figure 3 It is a flowchart for correcting pseudo-tags based on the belief rule base of the present invention. Detailed Embodiments

[0049] The following specifically introduces the present invention in combination with the drawings and specific examples.

[0050] The key problem to be solved by the present invention is to train a stable rolling bearing fault diagnosis system based on a belief rule base containing incomplete domain knowledge in a scenario with only a small amount of labeled data (including fault classification and fault parameters). The rules in the belief rule base are used to correct the pseudo-tags output by the diagnosis model, and the corrected pseudo-tags are used to train a diagnosis model with better accuracy. The architecture diagram of the entire solution is as Figure 1 shown, including a self-learning module, a fault diagnosis model, a belief rule base, a tag correction module, and a belief rule base update module. Among them, the self-learning module is based on an unlabeled data training set, and obtains a feature extraction model by training a self-learning network; the fault diagnosis model includes a feature extraction model and a diagnosis classifier, and is used to obtain the cause of rolling bearing faults; the belief rule base includes rules matching known fault causes; the tag correction module is used to correct the pseudo-tags through the belief rule base to obtain corrected tags; the belief rule base update module updates the belief rule base based on the tags re-obtained from the trained diagnosis classifier. Based on this bearing fault diagnosis system, the bearing fault cases with unknown fault causes collected are used as the input of the diagnosis system. The diagnosis system classifies the faults according to the relevant sensor parameters of the bearing faults, and the classification result is the cause of the fault occurrence, so as to achieve the purpose of fault diagnosis. Finally, the design of the present invention ensures the authenticity of the entire mechanism. Combining Figure 2 , the specific implementation method is as follows:

[0051] Step 1: Learn Intrinsic Features

[0052] In real industrial scenarios, the number of unlabeled data is much larger than that of labeled data. Therefore, it is crucial to make full use of unlabeled data to improve the performance of the diagnostic model. Self-learning is an unsupervised feature learning method and belongs to a type of transfer learning. Its purpose is to learn high-level feature representations and provide more informative features for training the diagnostic model;

[0053] An autoencoder consists of three neural network layers: an input layer, a hidden layer, and an output layer. The encoder and the decoder are two components of the autoencoder. The encoder is composed of the input layer and the hidden layer, and the decoder is composed of the hidden layer and the output layer. To obtain the intrinsic feature representation, the encoder h transforms the input x into the latent representation z. The encoding process of the autoencoder is as follows:

[0054] z = h(x) = σ(W 1 x + b 1 )

[0055] The decoder g aims to reconstruct the latent feature representation z in the new feature space into the original input x. The decoding process is as follows:

[0056] x' = g(z) = σ(W 2 x + b 2 )

[0057] The training objective of the autoencoder is to make the reconstructed input x' as similar as possible to the original input x. The objective function of self-learning is expressed as follows:

[0058]

[0059] where N represents the number of unlabeled samples, and dist represents the distance between x and x'.

[0060] Step 2: Correct Pseudo Labels

[0061] Since the initial diagnostic classifier is trained only on a limited number of labeled samples, the diagnostic performance of the classifier is poor.

[0062] If the classifier is directly retrained using the predicted pseudo-labels, the performance of the resulting classifier may degrade, especially when the quality of the training data is poor. There is some expert experience and domain knowledge in the field of rolling bearing fault diagnosis. Utilize this valuable information to improve the performance of the diagnostic model. Incorporate domain knowledge into the training process of the diagnostic model. Specifically, use domain knowledge to correct the pseudo-labels so that the pseudo-labels conform to the knowledge or rules as much as possible. Since the confidence degree of the corrected pseudo-labels is higher than that of the original pseudo-labels, they can be used again to train a new diagnostic classifier to replace the original one. Combining Figure 3 , specifically as follows:

[0063] First, a small amount of labeled dataset (X l , y l ) is passed through an encoder to obtain latent representations, and then these latent representations are used to train an initial diagnostic classifier C. The unlabeled data X u is fuzzified through a fuzzy function to obtain fuzzy samples The fuzzy function is expressed as follows:

[0064]

[0065] where x i,j represents the j-th parameter of the i-th unlabeled sample, and represent the upper and lower bounds of the threshold of the j-th parameter respectively, represents the fuzzy value of the j-th parameter of the i-th unlabeled sample. The reference values of each parameter are represented by H, N, L

[0066] Find the set of fault mode rules R i matching the fuzzy samples in the belief rule base. Then simplify the obtained set of pattern rules by deleting the pattern rules contained in other rules. If rule 1 contains rule 2, it means that if the fuzzy sample conforms to rule 1, then it must conform to rule 2, and rule 2 is deleted. Each belief rule r i in the belief rule base has a rule weight θ i and a belief vector β i . Specifically, β i represents the degree of association between r i and each fault category. Screen the simplified set of pattern rules according to the belief vector, compare using the maximum belief degree in the belief vector, and retain the set of pattern rules with the highest belief degree for each fault category. The rule matching and screening process for bearing fault cases is as Figure 2 shown. Calculate the correction weight according to the obtained optimal set of rule patterns The calculation process is as follows:

[0067]

[0068] The calibration weight is used to calibrate the pseudo-label y predicted by the initial diagnostic classifier u to generate a new label y r , and the calibration process is as follows:

[0069] y r = y u * ω c

[0070] Using the new label y obtained after calibration r and the unlabeled data X u to train a new diagnostic classifier C * to replace the original diagnostic classifier C.

[0071] Step 3: Update the belief rule base

[0072] Because expert experience or domain knowledge is incomplete and cannot cover all bearing fault situations, the rules or knowledge contained in the initial belief rule base are insufficient. Based on the initial belief rule base, the trained classifier has better diagnostic performance than the initial classifier. Therefore, these unlabeled data and more reliable pseudo-labels can provide additional rules and their relationships regarding bearing fault categories to enhance the belief rule base. The updated belief rule base can in turn improve the performance of the diagnostic classifier, and the two continuously enhance each other in an iterative manner;

[0073] Based on the initial belief rule base, the trained diagnostic classifier has achieved better performance than the initial classifier. Therefore, these unlabeled data and more reliable pseudo-labels can provide additional rules and their relationships regarding fault categories to enhance the belief rule base. Specifically, first select the rule set that matches the fuzzy data from the belief rule base. If there is a corresponding true label, then the belief degree of the matching rule for that label will increase by α l . If the prediction of is incorrect, then the belief degree of the predicted label will be reduced by the penalty value γ. For unlabeled data, the belief degree corresponding to the predicted label in the matching set increases by α u , where α u < α l is because the unlabeled data lacks true labels, so the degree of increase in belief is less than that of labeled data. For each rule, the newly obtained belief vector (the belief vector is composed of multiple belief degrees, and the new is obtained based on the increase in belief degree, which is common sense in this field and will not be elaborated here) and the old belief vector Combined as follows:

[0074]

[0075] Based on the prediction of the diagnostic model, the belief rule base is updated. The obtained new belief rule base is used again to correct the prediction of the diagnostic model, and the update is continuously carried out in an iterative manner.

[0076] Step 4: Diagnose unknown faults

[0077] After the diagnostic model is trained, the bearings with fault causes can be diagnosed; the sensor data related to the bearings with unknown fault causes collected is first passed through the encoder of the trained autoencoder to obtain high-level features, and then the high-level features are used as the input of the diagnostic model for diagnosis to obtain the results of fault diagnosis, so as to achieve the purpose of rolling bearing fault diagnosis. In addition, rules matching the bearing fault use cases can be found in the updated belief rule base, and these matched rules can assist in explaining the output of the model and improve the interpretability of the diagnostic model.

Claims

1. A rolling bearing fault diagnosis method based on a belief rule base, characterized in that, it includes the steps of: Obtain an unlabeled data training set, and obtain a feature extraction model by training a self-learning network; Use the labeled data set as the training set, extract features from the labeled data through the feature extraction model, and train a diagnostic classifier; Extract the internal features of the unlabeled data through the feature extraction model, use them as the input of the diagnostic classifier, and output pseudo-labels; Based on abductive reasoning, correct the pseudo-labels through the belief rule base to obtain corrected labels; Use the unlabeled data and the corrected labels as the input of the diagnostic classifier to retrain, and update the belief rule base based on the labels output again by the trained diagnostic classifier; Iteratively train the diagnostic classifier and update the belief rule base repeatedly to obtain a diagnostic classifier whose accuracy meets the set threshold; Based on the feature extraction model and the diagnostic classifier, perform fault diagnosis on the bearing; Correct the pseudo-labels through the belief rule base to obtain corrected labels, specifically including: Fuzzify the unlabeled data X through the fuzzy function u to obtain the fuzzy samples Unlabeled data; Find the fault mode rule set R that matches the fuzzy sample in the belief rule base i ; The method of deleting the fault mode rules included by other rules simplifies the fault mode rule set R i ; Filter the simplified fault mode rule set R according to the magnitude of the belief degree of the belief vector i and retain the fault mode rules with the highest belief degree for each fault category; Calculate the correction weight based on the fault rule pattern rule set with the highest degree of belief where β k represents the belief vector, θ k represents the rule weight, and k represents the number of belief rules in the rule set of the fault rule pattern with the highest belief degree; For the pseudo-label y u perform calibration to generate a new label y r = y u * ω c .

2. The rolling bearing fault diagnosis method based on a belief rule base according to claim 1, characterized in that, The self-learning network includes an encoder and a decoder. The encoder includes a neural network input layer and a hidden layer, which is used to extract the internal features of the unlabeled data. The decoder includes a neural network hidden layer and an output layer, which is used to decode the output of the encoder to obtain reconstructed data.

3. The rolling bearing fault diagnosis method based on a belief rule base according to claim 2, characterized in that, The encoding process of the encoder is: z = h(x) = σ(W 1 x + b 1 ) Among them, h represents the encoder, x is the input, z represents the inherent representation of the unlabeled data, and W 1 and b 1 represent the parameters of the encoder.

4. The rolling bearing fault diagnosis method based on a belief rule base according to claim 3, characterized in that, The decoding process of the decoder is: x’ = g(z) = σ(W 2 z + b 2 ) Among them, g represents the decoder, z represents the inherent representation of unlabeled data, x' represents the reconstructed feature, W 2 and b 2 represent the parameters of the decoder.

5. The rolling bearing fault diagnosis method based on a belief rule base according to claim 4, characterized in that, When training the self-learning network, the training objective function is expressed as follows: where N represents the number of unlabeled samples, dist represents the distance between the original input x and the reconstructed input x'. When the value of the objective function meets the set threshold, the encoder is the feature extraction model.

6. The rolling bearing fault diagnosis method based on a belief rule base according to claim 1, characterized in that, The diagnostic classifier uses a BP neural network.

7. The rolling bearing fault diagnosis method based on a belief rule base according to claim 1, characterized in that, The fuzzy function is: where, x i,j represents the j-th parameter of the i-th unlabeled sample, and represent the upper and lower bounds of the threshold of the j-th parameter respectively, represents the fuzzy value of the j-th parameter of the i-th unlabeled sample, and H, N, L represent the fuzzy values corresponding to x i,j in different ranges.

8. The rolling bearing fault diagnosis method based on a belief rule base according to claim 1, characterized in that, Updating the belief rule base based on the labels output again by the trained diagnostic classifier specifically includes: If the input of the diagnosis classifier is labeled data, select the corresponding fuzzy data from the belief rule base The matching rule set, if There is a corresponding true label, the belief degree of the matching rule for this true label will increase by α l , otherwise the belief degree of the predicted label of the diagnosis classifier will be reduced by the penalty value γ; for unlabeled data, the belief degree corresponding to the predicted label in the matching fault mode rule set increases by α u , where α u <α l , based on the newly obtained belief vector Combined with the original belief vector , update the belief vector of the belief rule base To:

9. A system for a rolling bearing fault diagnosis method according to any one of claims 1 to 8, characterized in that, It includes a self-learning module, a fault diagnosis model, a belief rule base, a label correction module and a belief rule base update module, wherein, The self-learning module is based on an unlabeled data training set and obtains a feature extraction model by training a self-learning network; The fault diagnosis model includes a feature extraction model and a diagnostic classifier, and is used to obtain the cause of the rolling bearing fault; The belief rule base includes rules that match known failure causes; The belief rule base is a database containing belief rules; The label correction module is used to correct the pseudo-labels through the belief rule base to obtain corrected labels; The belief rule base update module updates the belief rule base based on the labels output again by the trained diagnostic classifier.

Images