Rolling bearing fault diagnosis method and system

Abstract

The invention discloses a rolling bearing fault diagnosis method and system, and relates to the technical field of intelligent fault diagnosis, and the method comprises the following steps: carrying out the preprocessing of rolling bearing vibration original signals of a plurality of source domains, and obtaining a plurality of time domain data sets and a plurality of frequency spectrum data sets; constructing a fault diagnosis model, wherein the fault diagnosis model comprises a first feature extractor, a second feature extractor, a first classifier, a second classifier, a fusion feature extractor and a joint classifier; performing data enhancement on similar samples in the plurality of time domain data sets to generate a plurality of enhanced data sets; and training the fault diagnosis model through the time domain data set, the enhanced data set and the frequency spectrum data set. According to the method, the cross-domain invariant representation can be effectively extracted under the condition that the difference between the fields is large, and then the refined diagnosis performance of the model under the unknown working condition is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of rolling shaft intelligent fault diagnosis, and in particular to a rolling bearing fault diagnosis method and system. Background Art

[0002] Rotating machinery is widely used in modern industrial systems. Rolling bearings are one of the most critical components of rotating machinery, and their operating condition directly affects the overall performance of the equipment. Over long-term operation, due to repeated changes in load and speed, rolling bearings are prone to varying degrees of wear and failure, which can easily lead to overall performance degradation or even system failure. Therefore, research on rolling bearing condition identification and fault diagnosis has important theoretical and practical significance.

[0003] In recent years, intelligent rolling bearing fault diagnosis methods have been widely used and achieved excellent performance, leveraging the powerful automatic feature extraction capabilities of deep learning. The success of deep learning methods relies on two essential assumptions: sufficient labeled training data and the training and test data following the same distribution. However, in real industrial scenarios, bearing fault data is scarce, and collecting sufficient data and manually annotating it is very time-consuming and expensive. Therefore, there is a strong incentive to use richly labeled data from similar tasks under other operating conditions to assist in diagnosis. However, bearing operating scenarios are complex and varied, and differences in factors such as operating conditions and equipment types often lead to changes in data distribution. This difference in data distribution weakens the generalization ability of diagnostic models. This can cause models that perform well in the source domain to perform significantly worse in the target domain with a different distribution.

[0004] To address the challenges of cross-domain bearing fault diagnosis caused by varying operating conditions, transfer learning methods based on domain adaptation and domain generalization have rapidly developed in the field of intelligent fault diagnosis. The success of domain adaptation methods relies on the prerequisite of having access to the prior distribution of target domain data during the training phase. However, in practical transfer diagnosis tasks, target domain test data is often unavailable, which severely limits their applicability. Intelligent fault diagnosis methods based on domain generalization learn cross-domain-invariant general diagnostic knowledge from multiple source domains. Their significant advantage is that they do not require any target domain data for training. However, limitations remain. First, feature learning is insufficient. Many existing methods forcibly push all domains into a single latent feature space to form shared cross-domain-invariant features, which are then used to diagnose faults. In the process of forming shared cross-domain-invariant features, a large number of domain-private features are discarded, which may carry fault-sensitive information that is beneficial for diagnosis. On the other hand, these discarded features also result in limited feature diversity, increasing the risk of overfitting to the source domain. Second, when there is significant disparity between domains, effective cross-domain-invariant representations are difficult to extract.

[0005] To address these challenges, several recent domain-generalized fault diagnosis methods have been proposed. While these recent improvements extract cross-domain invariant features, they also rely on extracting auxiliary invariant features directly from data forms that are less susceptible to domain shift (such as phase information). While these simple techniques can be effective in some cases, they often fail to bridge the distributional differences between source domains. Summary of the Invention

[0006] Based on the above-mentioned defects in the prior art, the present invention provides a rolling bearing fault diagnosis method and system to solve the existing problems.

[0007] The present invention adopts the following technical solutions:

[0008] In a first aspect, the present invention provides a rolling bearing fault diagnosis method, comprising the following steps:

[0009] Preprocessing the original rolling bearing vibration signals from multiple source domains to obtain multiple time domain data sets and multiple spectrum data sets;

[0010] Constructing a fault diagnosis model, wherein the fault diagnosis model includes a first feature extractor, a second feature extractor, a first classifier, a second classifier, a fusion feature extractor, and a joint classifier;

[0011] Data enhancement is performed on similar samples in multiple time domain data sets to generate multiple enhanced data sets; a fault diagnosis model is trained using the time domain data set, the enhanced data set and the spectrum data set, wherein a first feature extractor is used to extract features from the time domain data set and the enhanced data set respectively to obtain first class-level invariant features and second class-level invariant features; the first class-level invariant features and the second class-level invariant features are input into a first classifier to obtain a first predicted probability value and a second predicted probability value; a second feature extractor is used to extract features from the spectrum data set to obtain a cross-domain invariant feature; the cross-domain invariant feature is input into a second classifier to obtain a third predicted probability value; a first confidence level and a second confidence level of the first predicted probability value and the third predicted probability value are obtained, and based on the first confidence level and the second confidence level, the first class-level invariant features and the cross-domain invariant features are weighted and fused by a fusion feature extractor to obtain a fusion feature; the fusion feature is classified by a joint classifier to obtain a fault diagnosis result;

[0012] The time domain data and spectrum data of the original vibration signal of the rolling bearing in the unknown target domain are obtained and input into the trained fault diagnosis model to obtain the fault diagnosis results.

[0013] Preferably, data enhancement is performed on similar samples in multiple time domain data sets, and the data enhancement process is specifically as follows:

[0014]

[0015] Where μ is a set of mixing factors randomly drawn from the Dirichlet distribution, is the i-th enhanced sample in the c-th class, M(·) is the weighted mixing operation, is the time domain data of the i-th sample in the c-th class in the k-th source domain dataset, K is the number of multiple source domains, μ k is the factor assigned to the k-th source domain, is the label information of the i-th enhanced sample in the c-th class, is the label information of the i-th sample in the c-th class in the K-th source domain dataset.

[0016] Preferably, the first feature extractor comprises five one-dimensional convolutional layers, each of which is followed by a batch normalization layer, an activation layer, and a pooling layer;

[0017] The second feature extractor includes five two-dimensional convolutional layers, each of which is followed by an instance normalization layer and an activation layer;

[0018] The first classifier and the second classifier include three fully connected layers, the number of neurons in the last fully connected layer is consistent with the total number of health states of the rolling bearing, and batch normalization layers, activation layers and Dropout (0.5) operations are introduced between the fully connected layers;

[0019] The fusion feature extractor includes two fully connected layers, and a batch normalization layer, an activation layer and a Dropout (0.5) operation are introduced between the fully connected layers;

[0020] The joint classifier includes two fully connected layers, the number of neurons in the last fully connected layer is consistent with the total number of health states of the rolling bearing, and batch normalization layers, activation layers and Dropout (0.5) operations are introduced between the fully connected layers.

[0021] Preferably, when the fault diagnosis model is trained using the time domain dataset, the enhanced dataset, and the spectrum dataset, the model parameters of the fault diagnosis model are optimized by constructing a first optimization objective and a second optimization objective, specifically comprising the following steps:

[0022] Divide the training process into the first and second stages;

[0023] In the first stage, a first optimization objective is constructed. The first optimization objective includes a first optimization sub-objective and a second optimization sub-objective. The parameters of the first feature extractor and the first classifier are optimized by the first optimization sub-objective, and the parameters of the second feature extractor and the second classifier are optimized by the second optimization sub-objective. The first optimization sub-objective and the second optimization sub-objective are specifically as follows:

[0024]

[0025] Where, J ci and J di are the first optimization sub-goal and the second optimization sub-goal, are the trainable parameters of the first feature extractor F1 and the first classifier C1, β is a hyperparameter, are the trainable parameters of the second feature extractor F2 and the second classifier C2, is the cross entropy loss between the first feature extractor and the first classifier, L E is the entropy loss of the first feature extractor and the first classifier, L MMD is the MMD constraint loss of the second feature extractor, is the cross entropy loss between the second feature extractor and the second classifier;

[0026] In the second stage, a second optimization objective is constructed to optimize the parameters of the fusion feature extractor and the joint classifier. The second optimization objective is specifically as follows:

[0027]

[0028] Where, J joint is the second optimization objective, They are fusion feature extractors F J and joint classifier C J The trainable parameters of is the cross entropy loss of the fused feature extractor and the joint classifier.

[0029] Preferably, the weighting and concatenation of the first-class invariant features and the cross-domain invariant features by the fusion feature extractor specifically includes the following steps:

[0030] The corresponding weights are obtained through the first confidence level and the second confidence level, as shown below:

[0031]

[0032] Where, Represents class-level invariant features The confidence level, Represents cross-domain invariant features The confidence level, is the weight of the class-level invariant feature generated by the i-th sample in the k-th source domain, The weight of the cross-domain invariant feature generated by the i-th sample in the k-th source domain;

[0033] The first-level invariant features and cross-domain invariant features are weighted and concatenated, as shown below:

[0034]

[0035] Where, is the fusion feature, F J To fuse feature extractors, Concat(·,·) represents the concatenation operation along the feature dimension. is the concatenated feature of the i-th sample in the k-th source domain, is a class-level invariant feature, It is a cross-domain invariant feature.

[0036] Preferably, the rolling bearing vibration original signals of multiple source domains are preprocessed, and the preprocessing process includes equal-length segmentation, time domain Z-score normalization processing, spectrum Z-score normalization processing and two-dimensional transformation operation.

[0037] In a second aspect, the present invention provides a rolling bearing fault diagnosis system, comprising:

[0038] An acquisition module is used to pre-process the original vibration signals of rolling bearings in multiple source domains to obtain multiple time domain data sets and multiple spectrum data sets;

[0039] A construction module, configured to construct a fault diagnosis model, wherein the fault diagnosis model includes a first feature extractor, a second feature extractor, a first classifier, a second classifier, a fusion feature extractor, and a joint classifier;

[0040] A training module is used to perform data enhancement on similar samples in multiple time domain data sets to generate multiple enhanced data sets; a fault diagnosis model is trained using the time domain data set, the enhanced data set and the spectrum data set, wherein a first feature extractor is used to perform feature extraction on the time domain data set and the enhanced data set respectively to obtain a first class-level invariant feature and a second class-level invariant feature; the first class-level invariant feature and the second class-level invariant feature are input into a first classifier to obtain a first predicted probability value and a second predicted probability value; a second feature extractor is used to perform feature extraction on the spectrum data set to obtain a cross-domain invariant feature; the cross-domain invariant feature is input into a second classifier to obtain a third predicted probability value; a first confidence level and a second confidence level of the first predicted probability value and the third predicted probability value are obtained, and based on the first confidence level and the second confidence level, the first class-level invariant feature and the cross-domain invariant feature are weighted and concatenated by a fusion feature extractor to obtain a fusion feature; the fusion feature is classified by a joint classifier to obtain a fault diagnosis result;

[0041] The prediction module is used to obtain the time domain data and spectrum data of the original vibration signal of the rolling bearing in the unknown target domain, and input it into the trained fault diagnosis model to obtain the fault diagnosis result.

[0042] Compared with the prior art, the at least one technical solution adopted by the present invention can achieve the following beneficial effects:

[0043] The present invention constructs a fault diagnosis model based on dual-perspective joint domain generalization of class-level invariance and cross-domain invariance, which can effectively extract cross-domain invariant representations when there are large differences between domains. First, the original vibration signals of rolling bearings in multiple source domains are preprocessed to obtain multiple time domain data sets and multiple spectrum data sets. Before extracting the class-level invariant features of multiple source domain data, data enhancement is performed on similar samples carrying similar fault information in multiple time domain data sets to generate multiple enhanced data sets, and the non-smooth initial distribution between similar samples in multiple source domains is filled with a smoother distribution, thereby bridging the distribution differences between the source domains.

[0044] The fault diagnosis model is trained using a time domain data set, an enhanced data set, and a spectrum data set, wherein the first feature extractor is used to extract class-level invariant features with classification and discrimination capabilities, the second feature extractor is used to extract cross-domain invariant features, the first classifier and the second classifier are used to output the predicted probability values ​​of the two features, obtain the first confidence and the second confidence of the two predicted probabilities, and based on the first confidence and the second confidence, the first class-level invariant features and the cross-domain invariant features are weighted and spliced ​​and fused by the fusion feature extractor to obtain fusion features. The present invention introduces a weighting strategy based on sample confidence to evaluate the strength relationship of feature generalizability under different perspectives (class-level invariance / cross-domain invariance), thereby realizing feature adaptive weighting and further ensuring the effectiveness of fusion features. Finally, the fusion features are classified by a joint classifier to obtain the fault diagnosis results. The present invention combines the dual-perspective joint domain generalization of class-level invariance and cross-domain invariance to improve the refined diagnosis performance of the model under unknown working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0046] Figure 1 This is a flow chart of a rolling bearing fault diagnosis method of the present invention;

[0047] Figure 2 Schematic diagram of the network model training process of the present invention;

[0048] Figure 3 It is a flow chart of the testing process of the present invention. DETAILED DESCRIPTION

[0049] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0050] The present invention provides a rolling bearing fault diagnosis method, referring to Figure 1-Figure 2 , specifically including the following steps:

[0051] Step 1: The collected K source domain rolling bearing vibration raw signals are first segmented into equal lengths, and then subjected to time domain Z-score normalization processing, spectrum Z-score normalization processing, and two-dimensional transformation operations to obtain the preprocessed data set: time domain data set and spectrum datasets

[0052] in, and Represents the time domain and spectrum form of the i-th sample in the k-th source domain dataset; represents the label information of the i-th sample in the k-th source dataset, C represents the total number of categories of bearing health status, st represents the time domain format of the source dataset, and sf represents the frequency domain format of the source dataset.

[0053] Step 2: Establish an intelligent fault diagnosis model based on dual-view joint domain generalization, which includes branch feature extractor modules F1 and F2 (first feature extractor and second feature extractor), branch classifier modules C1 and C2 (first classifier and second classifier), fusion feature extractor module F J , Joint Classifier Module C J .

[0054] Step 2.1: The branched feature extractor module F1 consists of five one-dimensional convolutional layers. In F1, batch normalization, activation (all activation functions use LeakyReLU), and pooling layers are placed after each convolutional layer (except the first one). Only 50% of the features obtained after the first convolutional layer are batch normalized, while the remaining 50% are instance normalized. This enhances the convolutional network's modeling and generalization capabilities.

[0055] Step 2.2: The branched feature extractor module F2 consists of five two-dimensional convolutional layers. In F2, instance normalization layers and activation layers (all using the LeakyReLU activation function) are placed after each convolutional layer. Pooling layers are placed as needed to ensure that the flattened output features of F2 match the size of F1.

[0056] Step 2.3: Branch classifiers C1 and C2 have identical structures, each consisting of three fully connected layers. The number of neurons in the final fully connected layer matches the total number of bearing health states, C. Batch normalization layers, activation layers (all using the LeakyReLU activation function), and Dropout (0.5) are introduced between the fully connected layers. The softmax function converts the output of the final fully connected layer into a probability distribution summing to 1. This means that the branch classifier output is directly converted into the probabilities of different fault categories, enabling bearing fault classification.

[0057] Step 2.4: Fusion feature extractor module F J It contains two fully connected layers, and batch normalization layers, activation layers (activation functions are all LeakyReLU functions) and Dropout (0.5) operations are introduced between the fully connected layers.

[0058] Step 2.5: Joint Classifier Module C J The model consists of two fully connected layers, with the number of neurons in the final fully connected layer matching the total number of bearing health states, C. Batch normalization, activation layers (all using the LeakyReLU activation function), and a Dropout (0.5) operation are introduced between the fully connected layers. The softmax function converts the output of the final fully connected layer into a probability distribution summing to 1. This means that the network model output is directly converted into the probabilities of different fault categories, enabling bearing fault classification.

[0059] Step 3: A semantically consistent multi-source data augmentation method is used to generate an auxiliary augmentation domain. A branch feature extractor F1 is used to extract features from each source domain and the augmentation domain (in the time domain). Minimizing supervised classification and minimizing classification entropy constraints are introduced into the branch classifier C1 to ensure the discriminability and class-level invariance of this branch feature.

[0060] Step 3.1: Use the multivariate Dirichlet distribution similar to the beta distribution to sample the mixup factor instead of the beta distribution to generate samples with more information and higher diversity based on multi-domain data. To further ensure semantic consistency, directional data enhancement is achieved by fusing similar samples with similar fault information from different source domains. Represents samples randomly drawn from the cth class in different source domains (time domain form). Generate semantic enhancement samples The process can be expressed as:

[0061]

[0062] Where μ is a set of mixing factors randomly drawn from the Dirichlet distribution μ~Dirichlet(α), and α is a K-tuple array, all elements of which are set to the same value in this invention. M(·) represents the weighted mixing operation, K is the number of multi-source domains, and A is the abbreviation for augmented sample.

[0063] Step 3.2: The auxiliary enhancement domain constructed from the aforementioned enhancement samples fills the non-smooth initial distribution of similar samples in multiple source domains with a smoother distribution. By adding a minimum supervised classification constraint to the branch classifier C1, the branch feature extractor F1 can extract class-level invariant features with classification and discrimination capabilities from each source domain and the enhancement domain (in the time domain). The calculation process is as follows:

[0064]

[0065] in, is the cross entropy loss between the first feature extractor and the first classifier, Representative time domain form samples Features obtained after passing through the branch feature extractor F1; Representative time domain form enhanced samples The features obtained after the branch feature extractor F1; L CE (·,·) represents the cross entropy function; N A Represents the total number of all enhanced samples generated, N s，k is the total number of samples in the k-th source domain, is the label of the j-th augmented sample.

[0066] Step 3.3: Introduce the concept of entropy to further enhance the network's discriminative and generalization capabilities by minimizing the entropy constraint to encourage the decision boundary to cross low-density areas. The calculation process is as follows:

[0067]

[0068] Among them, L E is the entropy loss of the first feature extractor and the first classifier, and E(·) represents the entropy function; Represents the feature f ci The predicted probability of the cth class obtained after inputting into C1, Represents the characteristics The predicted probability of class c obtained after inputting into C1.

[0069] Step 4: Use the branch feature extractor F2 to extract features from each source domain (spectral form) respectively, introduce the constraint of minimizing the maximum mean difference to align the feature distributions from different source domains, and introduce the minimization of the supervised classification loss in the branch classifier C2 to ensure the discriminability and cross-domain invariance of this branch feature.

[0070] Step 4.1: Use the branch feature extractor F2 to extract features from each source domain (spectral form) and use the maximum mean difference (MMD) to measure the difference in feature distribution between different domains. By minimizing the MMD distance between the distributions of each source domain, the inter-domain distribution is aligned to extract cross-domain invariant features. The calculation process is as follows:

[0071]

[0072] Among them, L MMD is the MMD constraint loss of the second feature extractor, Representative spectrum form sample The features obtained after the feature extractor F2, is the mean of the cross-domain invariant features of the k1th source domain, is the mean of the cross-domain invariant features of the k2th source domain.

[0073] Step 4.2: To further enable the cross-domain invariant features to have the ability to make category decisions, the classifier C2 is introduced to minimize the supervised classification loss to train the classifier parameters. The calculation formula is as follows:

[0074]

[0075] Where, is the cross entropy loss between the second feature extractor and the second classifier.

[0076] Step 5: Introduce a weighting strategy based on sample confidence, weight the above two types of branch features separately, concatenate and fuse them, and then send them to the fusion feature extractor F J , in the joint classifier C J The method of minimizing supervised classification loss is introduced to maximize the optimization effect of the joint classifier.

[0077] Step 5.1: Introduce a weighting strategy based on sample confidence to evaluate the strength of feature generalizability under different perspectives (class-level invariance / cross-domain invariance). The model's probability prediction value for the correct label category is used as the confidence level. A higher confidence level indicates that the network is confident in making accurate predictions based on this feature. This indicates that the feature has good generalizability; conversely, if the feature is not generalizable enough, that is, lacks sufficient discriminability, the confidence level will be lower. Therefore, a weighting strategy is designed according to the following guidelines to achieve feature adaptive weighting:

[0078]

[0079] in: Represents class-level invariant features Confidence (the category label of the sample is c); Represents cross-domain invariant features Confidence (the category label of the sample is c), is the weight of the class-level invariant feature generated by the i-th sample in the k-th source domain, The weight of the cross-domain invariant feature generated by the i-th sample in the k-th source domain.

[0080] Step 5.2: Adaptively weight the above two types of features, concatenate and fuse them, and then send them to the fusion feature extractor F J , get fusion features

[0081]

[0082] Among them, Concat(·,·) represents the concatenation operation along the feature dimension. is the concatenated feature of the i-th sample in the k-th source domain.

[0083] Step 5.3: In the joint classifier C J Introducing the minimization of supervised classification loss to maximize the optimization effect of the joint classifier;

[0084]

[0085] Where, is the cross entropy loss of the fused feature extractor and the joint classifier.

[0086] Step 6: Construct an overall optimization goal. Under the guidance of the overall optimization goal, input the two types of source domain data sets into the intelligent fault diagnosis model based on dual-perspective joint domain generalization for training. After the training is completed, the model parameters are fixed.

[0087] Step 6.1: The first stage extracts dual-view invariant features in an independent and parallel manner. The optimization goal of this stage can be expressed as:

[0088]

[0089] in, Represent the trainable parameters of the in-domain feature extractor F1 and the auxiliary classifier C1, Similarly; β is a hyperparameter.

[0090] Step 6.2: In the second phase, weighted fusion of the pre-extracted features from both views is performed for joint training. It should be noted that the modules involved in the first phase remain frozen in the second phase and can only be used for forward reasoning weight calculations. The optimization objective of this phase can be expressed as:

[0091]

[0092] in, Represents the fusion feature extractor F J and joint classifier C J The trainable parameters of

[0093] Step 6.3: Restore the first-stage module to a trainable state.

[0094] Step 6.4: Repeat steps 6.1-6.3 until the whole model training is completed. After the training is completed, the modules F1, F2, F J and C J The parameters are used for subsequent testing of the model.

[0095] Step 7: Reference Figure 3 , collect the original time domain signal of bearing vibration in the unknown target domain, obtain the target domain samples in the form of time domain and spectrum after the same preprocessing as the source domain, input them into the trained branch feature extractors F1 and F2 respectively to obtain the dual-view branch features and splice and fuse them, and input the fused features into F J 、C J Then obtain the rolling bearing fault diagnosis results.

[0096] The present invention proposes a dual-perspective joint-domain generalization framework that combines class-level invariance and cross-domain invariance, which can effectively improve the diagnostic performance of the model under unknown working conditions.

[0097] This paper designs a semantically consistent multi-source data augmentation method. Compared to traditional mixup methods, which can only generate samples between two domains and randomly combine samples from unspecified categories, this method uses a multivariate Dirichlet distribution, similar to the beta distribution, to sample the mixup factor. This fully utilizes multi-domain data to generate samples with greater information content and diversity. Furthermore, by fusing similar samples from different source domains that carry similar fault information, targeted data augmentation is achieved, further ensuring semantic consistency between the generated samples and the original samples.

[0098] The present invention introduces a weighting strategy based on sample confidence to evaluate the strength relationship of feature generalizability under different perspectives (class-level invariance / cross-domain invariance), thereby realizing feature adaptive weighting and further ensuring the effectiveness of fused features.

[0099] Based on the same concept, the present invention also provides a rolling bearing fault diagnosis system, which includes an acquisition module, a construction module, a training module and a prediction module.

[0100] The acquisition module is used to preprocess the original vibration signals of rolling bearings in multiple source domains to obtain multiple time domain data sets and multiple spectrum data sets.

[0101] The construction module is used to construct a fault diagnosis model, which includes a first feature extractor, a second feature extractor, a first classifier, a second classifier, a fusion feature extractor and a joint classifier;

[0102] The training module is used to perform data enhancement on similar samples in multiple time domain data sets to generate multiple enhanced data sets; the fault diagnosis model is trained using the time domain data set, the enhanced data set and the spectrum data set, wherein the first feature extractor is used to extract features from the time domain data set and the enhanced data set respectively to obtain first class-level invariant features and second class-level invariant features; the first class-level invariant features and the second class-level invariant features are input into the first classifier to obtain first predicted probability values ​​and second predicted probability values; the second feature extractor is used to extract features from the spectrum data set to obtain cross-domain invariant features; the cross-domain invariant features are input into the second classifier to obtain a third predicted probability value; the first confidence level and the second confidence level of the first predicted probability value and the third predicted probability value are obtained, and based on the first confidence level and the second confidence level, the first class-level invariant features and the cross-domain invariant features are weighted and spliced ​​and fused by the fusion feature extractor to obtain fusion features; the fusion features are classified by the joint classifier to obtain the fault diagnosis results.

[0103] The prediction module is used to obtain the time domain data and spectrum data of the original vibration signal of the rolling bearing in the unknown target domain, and input it into the trained fault diagnosis model to obtain the fault diagnosis results.

[0104] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0105] Obviously, those skilled in the art may make various modifications and variations to the present invention without departing from the spirit and scope of the present invention. Thus, if such modifications and variations fall within the scope of the claims and their equivalents, the present invention is intended to include such modifications and variations.

Claims

1. A rolling bearing fault diagnosis method, characterized in that: The following steps are involved: Preprocessing the original rolling bearing vibration signals from multiple source domains to obtain multiple time domain data sets and multiple spectrum data sets; Constructing a fault diagnosis model, wherein the fault diagnosis model includes a first feature extractor, a second feature extractor, a first classifier, a second classifier, a fusion feature extractor, and a joint classifier; Perform data enhancement on similar samples in multiple time domain datasets to generate multiple enhanced datasets; The fault diagnosis model is trained using a time domain dataset, an enhanced dataset, and a spectrum dataset, wherein a first feature extractor is used to extract features from the time domain dataset and the enhanced dataset respectively to obtain first class-level invariant features and second class-level invariant features; the first class-level invariant features and the second class-level invariant features are input into a first classifier to obtain a first prediction probability value and a second prediction probability value; a second feature extractor is used to extract features from the spectrum dataset to obtain a cross-domain invariant feature; the cross-domain invariant feature is input into a second classifier to obtain a third prediction probability value; a first confidence level and a second confidence level of the first prediction probability value and the third prediction probability value are obtained, and based on the first confidence level and the second confidence level, the first class-level invariant features and the cross-domain invariant features are weighted and concatenated by a fusion feature extractor to obtain a fusion feature; the fusion feature is classified by a joint classifier to obtain a fault diagnosis result; The time domain data and spectrum data of the original vibration signal of the rolling bearing in the unknown target domain are obtained and input into the trained fault diagnosis model to obtain the fault diagnosis results.

2. A rolling bearing fault diagnosis method according to claim 1, characterized in that: The data enhancement process for similar samples in multiple time domain data sets is as follows: Where μ is a set of mixing factors randomly drawn from the Dirichlet distribution, is the i-th enhanced sample in the c-th class, M(·) is the weighted mixing operation, is the time domain data of the i-th sample in the c-th class in the k-th source domain dataset, K is the number of multiple source domains, μ k′ is the factor assigned to the k-th source domain, is the label information of the i-th enhanced sample in the c-th class, is the label information of the i-th sample in the c-th class in the K-th source domain dataset.

3. A rolling bearing fault diagnosis method according to claim 1, characterized in that: The first feature extractor includes five one-dimensional convolutional layers, each of which is followed by a batch normalization layer, an activation layer, and a pooling layer; The second feature extractor includes five two-dimensional convolutional layers, each of which is followed by an instance normalization layer and an activation layer; The first classifier and the second classifier include three fully connected layers, the number of neurons in the last fully connected layer is consistent with the total number of health states of the rolling bearing, and batch normalization layers, activation layers and Dropout (0.5) operations are introduced between the fully connected layers; The fusion feature extractor includes two fully connected layers, and a batch normalization layer, an activation layer and a Dropout (0.5) operation are introduced between the fully connected layers; The joint classifier includes two fully connected layers, the number of neurons in the last fully connected layer is consistent with the total number of health states of the rolling bearing, and batch normalization layers, activation layers and Dropout (0.5) operations are introduced between the fully connected layers.

4. A rolling bearing fault diagnosis method according to claim 1, characterized in that: When the fault diagnosis model is trained using the time domain dataset, the enhanced dataset, and the spectrum dataset, the model parameters of the fault diagnosis model are optimized by constructing a first optimization objective and a second optimization objective, specifically including the following steps: Divide the training process into the first and second stages; In the first stage, a first optimization objective is constructed. The first optimization objective includes a first optimization sub-objective and a second optimization sub-objective. The parameters of the first feature extractor and the first classifier are optimized by the first optimization sub-objective, and the parameters of the second feature extractor and the second classifier are optimized by the second optimization sub-objective. The first optimization sub-objective and the second optimization sub-objective are specifically as follows: Where, J ci and J di are the first optimization sub-goal and the second optimization sub-goal, are the trainable parameters of the first feature extractor F1 and the first classifier C1, β is a hyperparameter, are the trainable parameters of the second feature extractor F2 and the second classifier C2, is the cross entropy loss between the first feature extractor and the first classifier, L E is the entropy loss of the first feature extractor and the first classifier, L MMD is the MMD constraint loss of the second feature extractor, is the cross entropy loss between the second feature extractor and the second classifier; In the second stage, a second optimization objective is constructed to optimize the parameters of the fusion feature extractor and the joint classifier. The second optimization objective is specifically as follows: Where, J joint is the second optimization objective, They are fusion feature extractors F J and joint classifier C J The trainable parameters of is the cross entropy loss of the fused feature extractor and the joint classifier.

5. A rolling bearing fault diagnosis method according to claim 1, characterized in that: The weighting and splicing of the first-class invariant features and the cross-domain invariant features by the fusion feature extractor specifically includes the following steps: The corresponding weights are obtained through the first confidence level and the second confidence level, as shown below: Where, Represents the class-level invariant feature f i ci,k The confidence level, Represents the cross-domain invariant feature f i di,k The confidence level, is the weight of the class-level invariant feature generated by the i-th sample in the k-th source domain, The weight of the cross-domain invariant feature generated by the i-th sample in the k-th source domain; The first-level invariant features and cross-domain invariant features are weighted and concatenated, as shown below: Where, f i joint,k is the fusion feature, F J To fuse feature extractors, Concat(·,·) represents the concatenation operation along the feature dimension. is the concatenated feature of the i-th sample in the k-th source domain, f i ci,k is the class-level invariant feature, f i di,k It is a cross-domain invariant feature.

6. A rolling bearing fault diagnosis method according to claim 1, characterized in that: The rolling bearing vibration original signals of multiple source domains are preprocessed, and the preprocessing process includes equal length segmentation, time domain Z-score normalization processing, spectrum Z-score normalization processing and two-dimensional transformation operation.

7. A rolling bearing fault diagnosis system, characterized in that: include: An acquisition module is used to pre-process the original vibration signals of rolling bearings in multiple source domains to obtain multiple time domain data sets and multiple spectrum data sets; A construction module, configured to construct a fault diagnosis model, wherein the fault diagnosis model includes a first feature extractor, a second feature extractor, a first classifier, a second classifier, a fusion feature extractor, and a joint classifier; The training module is used to perform data enhancement on similar samples in multiple time domain datasets to generate multiple enhanced datasets; The fault diagnosis model is trained using a time domain dataset, an enhanced dataset, and a spectrum dataset, wherein a first feature extractor is used to extract features from the time domain dataset and the enhanced dataset respectively to obtain first class-level invariant features and second class-level invariant features; the first class-level invariant features and the second class-level invariant features are input into a first classifier to obtain a first prediction probability value and a second prediction probability value; a second feature extractor is used to extract features from the spectrum dataset to obtain a cross-domain invariant feature; the cross-domain invariant feature is input into a second classifier to obtain a third prediction probability value; a first confidence level and a second confidence level of the first prediction probability value and the third prediction probability value are obtained, and based on the first confidence level and the second confidence level, the first class-level invariant features and the cross-domain invariant features are weighted and concatenated by a fusion feature extractor to obtain a fusion feature; the fusion feature is classified by a joint classifier to obtain a fault diagnosis result; The prediction module is used to obtain the time domain data and spectrum data of the original vibration signal of the rolling bearing in the unknown target domain, and input it into the trained fault diagnosis model to obtain the fault diagnosis result.

Images