A bearing imbalance fault diagnosis method and system guided by dual classifier collaboration

Abstract

The present invention relates to a bearing imbalance fault diagnosis method and system guided by dual classifier collaboration. The method comprises: collecting bearing vibration signals in various healthy states under different operating conditions, performing fast Fourier transforms, converting the signals into frequency domain signals, obtaining a data set, dividing the data set into source domain data and target domain data under different operating conditions, and labeling the source domain data; constructing a fault diagnosis model, the fault diagnosis model comprising a shared feature extractor, a balanced classifier, and a standard classifier; and constructing a loss function based on the source domain data and the target domain data; training the fault diagnosis model using the loss function to obtain a trained fault diagnosis model; and inputting the bearing vibration signal to be tested into the trained fault diagnosis model to obtain a fault diagnosis result. The present invention can solve the problems in the prior art where the performance of diagnostic models is limited by label scarcity, imbalanced data distribution, and the need to improve model generalization capabilities.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing fault diagnosis, and in particular to a bearing imbalance fault diagnosis method and system guided by a dual-classifier collaboration. Background Art

[0002] Intelligent fault diagnosis is a key research area for improving industrial reliability. As a critical component widely used in industrial systems, the health of bearings has a direct impact on the overall system's operating efficiency and safety. Therefore, research into intelligent fault diagnosis for bearings is not only an important means of improving the overall reliability of industrial systems but also a key step in achieving innovative breakthroughs in current technological development.

[0003] Existing deep learning-based fault diagnosis methods face numerous challenges in their application in real-world industrial scenarios. These methods rely heavily on large amounts of labeled fault data for training. However, obtaining sufficient labeled data in real-world industrial settings often requires significant human and material resources. Furthermore, the unacceptable nature of machines operating under fault conditions for extended periods of time makes collecting labeled data difficult. Furthermore, the scarcity of labeled data limits the robustness and generalization capabilities of the models, making it difficult to meet the demand for high-precision diagnosis in diverse and dynamic industrial scenarios. Furthermore, in real-world industrial scenarios, mechanical equipment operates normally most of the time, with faults occurring only occasionally. Furthermore, stringent safety requirements necessitate the rapid shutdown of equipment upon the occurrence of a fault, further limiting the opportunity to collect data on the progress of the fault. The significant differences in the probability of occurrence of different fault types result in the data collected by monitoring systems being dominated by normal operation data, while fault data is extremely scarce, resulting in a dataset with a pronounced long-tail distribution. This imbalance in data distribution severely impairs the generalization capabilities of intelligent fault diagnosis models, ultimately leading to poor diagnostic performance in real-world applications. Therefore, addressing the challenges of label scarcity, data imbalance, and improving model generalization capabilities remain key challenges that urgently need to be addressed in intelligent fault diagnosis technology. Summary of the Invention

[0004] To this end, the technical problem to be solved by the present invention is to overcome the problems in the prior art in which the performance of the diagnostic model is affected by the scarcity of labels, imbalanced data distribution, and the improvement of the model generalization ability, and to provide a bearing imbalance fault diagnosis method and system guided by dual classifiers. The fault diagnosis method includes the following steps:

[0005] S1: Collect bearing vibration signals in various health states under different working conditions and perform fast Fourier transform to convert them into frequency domain signals to obtain a data set. The data set is divided into source domain data and target domain data under different working conditions, and the source domain data is labeled.

[0006] S2: constructing a fault diagnosis model, the fault diagnosis model including a shared feature extractor, a balanced classifier, and a standard classifier; and constructing a loss function based on the source domain data and the target domain data;

[0007] S3: training the fault diagnosis model using the loss function to obtain a trained fault diagnosis model;

[0008] S4: inputting the vibration signal of the bearing to be tested into the trained fault diagnosis model to obtain a fault diagnosis result;

[0009] In step S2, the method for constructing the loss function includes:

[0010] S21: Using the shared feature extractor to extract features from the source domain data and the target domain data respectively, to obtain source domain data features and target domain data features;

[0011] S22: Inputting the source domain data features into the balanced classifier and the standard classifier respectively to obtain a source domain balanced classification loss and a source domain standard classification loss;

[0012] S23: Inputting the target domain data features into a standard classifier to generate target domain data pseudo labels for the balanced classifier; combining the original output of the standard classifier for the target domain data features with the output of the standard classifier for the target domain data features after post-logit adjustment to obtain an adjusted perception weight;

[0013] S24: obtaining a target domain classification loss of a balanced classifier using the adjusted perception weights and the target domain pseudo labels; and obtaining a target domain classification loss of a standard classifier by filtering an original output of a standard classifier for target domain data features using a confidence threshold.

[0014] S25: Perform weighted calculation on the Coral loss, the source domain balanced classification loss, the source domain standard classification loss, the target domain classification loss of the balanced classifier, and the target domain classification loss of the standard classifier to obtain the loss function L total , which is expressed as follows:

[0015]

[0016] Among them, L cor Indicates Coral loss, represents the source domain standard classification loss, represents the source domain balanced classification loss, represents the target domain classification loss of the standard classifier, represents the target domain classification loss of the balanced classifier, and μ is a hyperparameter.

[0017] In one embodiment of the present invention, step S21 further includes: aligning the feature distributions of the source domain data features and the target domain data features using the Coral loss.

[0018] In one embodiment of the present invention, the Coral loss L cor It is expressed as follows:

[0019]

[0020] Among them, d E represents the output dimension of the shared feature extractor, x s and x t represent the source domain data and the target domain data respectively, ||·|| represents the Frobenius norm, C s and C t Respectively represent the covariance matrix of the source domain data and the target domain data, n s and n t Represent the source domain data features and the target domain data features respectively, and 1 is a column vector with all elements equal to 1.

[0021] In one embodiment of the present invention, the source domain balanced classification loss It is expressed as follows:

[0022]

[0023] Among them, n s is the total number of samples of the source domain data, x s,i is the i-th sample in the source domain data, y s,i is the label of the i-th sample in the source domain data, represents the output of the balanced classifier for the i-th sample in the source domain data, represents the logit adjustment factor of the balanced classifier, π s,[k] =n s,[k] / n s represents the observation frequency of the kth category in the source domain data, n s,[k] is the total number of samples in the kth class in the source domain data, represents the temperature coefficient of the balanced classifier.

[0024] In one embodiment of the present invention, the source domain standard classification loss It is expressed as follows:

[0025]

[0026] in, represents the output of the balanced classifier for the i-th sample in the source domain data, represents the logit adjustment factor of the standard classifier, represents the temperature coefficient of the standard classifier, represents the observation frequency of the kth class in the target domain data estimated by the balanced classifier, λ is a weighting parameter, represents the class distribution term of the kth class in the target domain data in the first b-1 batches estimated by the balanced classifier, Represents the class distribution item of the kth class in the target domain data in the current bth batch.

[0027] In one embodiment of the present invention, the class distribution item of the kth class in the target domain data in the current bth batch is The calculation formula is:

[0028]

[0029] Among them, n b is the number of samples in a batch, and δ(·) represents the Softmax function.

[0030] In one embodiment of the present invention, in step S23, a pseudo label of the target domain data for the balanced classifier is generated. The calculation formula is expressed as:

[0031]

[0032] in, represents the target class distribution term estimated by the standard classifier, is the temperature coefficient;

[0033] Get the adjusted perception weight w(x t,i ) is calculated as:

[0034] w(x t,i )=α m β(x t,i )

[0035] Among them, x t,i is the i-th sample in the target domain data, β(x t,i ) represents the logit consistency factor, max(·) represents the maximum value operation, δ(·) represents the Softmax function; α m represents the adjustment of the perception factor in the mth training step, n b is the number of samples in a batch.

[0036] In one embodiment of the present invention, the target domain classification loss of the standard classifier is It is expressed as follows:

[0037]

[0038] Among them, 1{·} is an indicator function, which returns 1 if the condition is met, otherwise it returns 0, T represents the confidence threshold, is the pseudo label of the target domain data of the standard classifier, is the pseudo label of the i-th sample in the k-th class in the target domain data of the standard classifier, x t,i is the i-th sample in the target domain data, x t,i,[k] is the i-th sample in the k-th class in the target domain data, n t is the total number of samples of the target domain data, and K represents the total number of categories of the target domain data.

[0039] In one embodiment of the present invention, the target domain classification loss of the balanced classifier is It is expressed as follows:

[0040]

[0041] Among them, w(x t,i ) is to adjust the perception weight, n t is the total number of samples of the target domain data, is the pseudo label of the i-th sample in the target domain data for the balanced classifier, is the pseudo label of the i-th sample in the k-th class in the target domain data for the balanced classifier, x t,i is the i-th sample in the target domain data, x t,i,[k] is the i-th sample in the k-th class in the target domain data, n t is the total number of samples of the target domain data, and K represents the total number of categories of the target domain data.

[0042] Based on the same inventive concept, the present invention also provides a bearing imbalance fault diagnosis system guided by dual classifiers in a collaborative manner. The system is used to implement the steps of the bearing imbalance fault diagnosis method guided by dual classifiers in a collaborative manner. The bearing imbalance fault diagnosis system guided by dual classifiers in a collaborative manner includes the following modules:

[0043] A data set construction module is used to collect bearing vibration signals in various health states under different working conditions, perform fast Fourier transform, convert them into frequency domain signals, obtain a data set, divide the data set into source domain data and target domain data under different working conditions, and label the source domain data;

[0044] A model construction and loss function design module is used to construct a fault diagnosis model, the fault diagnosis model including a shared feature extractor, a balanced classifier, and a standard classifier; and to construct a loss function based on the source domain data and the target domain data;

[0045] A model training module, configured to train the fault diagnosis model using the loss function to obtain a trained fault diagnosis model;

[0046] A fault diagnosis module is used to input the vibration signal of the bearing to be tested into the trained fault diagnosis model to obtain a fault diagnosis result;

[0047] The model construction and loss function design module includes a loss function construction submodule, and the loss function construction submodule is configured as follows:

[0048] Using the shared feature extractor to perform feature extraction on the source domain data and the target domain data respectively to obtain source domain data features and target domain data features;

[0049] Inputting the source domain data features into the balanced classifier and the standard classifier respectively to obtain a source domain balanced classification loss and a source domain standard classification loss;

[0050] Inputting the target domain data features into a standard classifier to generate target domain data pseudo labels for the balanced classifier; combining the original output of the standard classifier for the target domain data features with the output of the standard classifier for the target domain data features after post-logit adjustment to obtain adjusted perception weights;

[0051] Obtaining a target domain classification loss of a balanced classifier using the adjusted perception weights and the target domain pseudo labels; and obtaining a target domain classification loss of a standard classifier by filtering an original output of a standard classifier for target domain data features using a confidence threshold.

[0052] The loss function is obtained by performing weighted calculation on the Coral loss, the source domain balanced classification loss, the source domain standard classification loss, the target domain classification loss of the balanced classifier, and the target domain classification loss of the standard classifier.

[0053] The above technical solution of the present invention has the following advantages over the prior art:

[0054] By designing a dual classifier, this invention enables the model to maintain nearly balanced classification performance across all categories. The synergistic effect of the dual classifiers allows the standard classifier to better adapt to the target domain's category distribution. Furthermore, a carefully designed mechanism for adjusting perception weights effectively suppresses the negative impact of unreliable pseudo-labels, further enhancing the model's generalization capabilities in target domains with unknown category distributions and significantly improving the accuracy of imbalanced fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to make the content of the present invention more clearly understood, the present invention is further described in detail below based on specific embodiments of the present invention in conjunction with the accompanying drawings, wherein:

[0056] Figure 1 This is a flow chart of a bearing imbalance fault diagnosis method guided by dual classifier collaboration provided in the first embodiment of the present invention;

[0057] Figure 2 is a flow chart of the loss function construction method provided in the first embodiment of the present invention;

[0058] Figure 3 This is a flowchart of a specific implementation of a bearing imbalance fault diagnosis method guided by dual classifier collaboration provided in the first embodiment of the present invention;

[0059] Figure 4 It is a test diagram of the bearing data generation test bench of the present invention;

[0060] Figure 5 This is a feature visualization diagram of the present invention under five different target domain class distributions;

[0061] Figure 6 This is a schematic structural diagram of a bearing imbalance fault diagnosis system with dual classifier collaborative guidance provided in the second embodiment of the present invention;

[0062] Explanation of the accompanying figures in the specification: 100, data set construction module; 200, model construction and loss function design module; 210, loss function construction sub-module; 300, model training module; 400, fault diagnosis module. DETAILED DESCRIPTION

[0063] The present invention will be further described below with reference to the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it. However, the embodiments are not intended to limit the present invention.

[0064] Example 1

[0065] See also Figures 1 to 3As shown, the present invention provides a bearing imbalance fault diagnosis method guided by dual classifier collaboration, the fault diagnosis method comprising the following steps:

[0066] S1: Collect bearing vibration signals in various health states under different working conditions and perform fast Fourier transform to convert them into frequency domain signals to obtain a data set. The data set is divided into source domain data and target domain data under different working conditions, and the source domain data is labeled.

[0067] S2: constructing a fault diagnosis model, the fault diagnosis model including a shared feature extractor, a balanced classifier, and a standard classifier; and constructing a loss function based on the source domain data and the target domain data;

[0068] S3: training the fault diagnosis model using the loss function to obtain a trained fault diagnosis model;

[0069] S4: inputting the vibration signal of the bearing to be tested into the trained fault diagnosis model to obtain a fault diagnosis result;

[0070] Among them, in step S2, the method of constructing the loss function is as follows Figure 2 Shown, including:

[0071] S21: Using the shared feature extractor to extract features from the source domain data and the target domain data respectively, to obtain source domain data features and target domain data features;

[0072] S22: Inputting the source domain data features into the balanced classifier and the standard classifier respectively to obtain a source domain balanced classification loss and a source domain standard classification loss;

[0073] S23: Inputting the target domain data features into a standard classifier to generate target domain data pseudo labels for the balanced classifier; combining the original output of the standard classifier for the target domain data features with the output of the standard classifier for the target domain data features after post-logit adjustment to obtain an adjusted perception weight;

[0074] S24: obtaining a target domain classification loss of a balanced classifier using the adjusted perception weights and the target domain pseudo labels; and obtaining a target domain classification loss of a standard classifier by filtering an original output of a standard classifier for target domain data features using a confidence threshold.

[0075] S25: Perform weighted calculation on the Coral loss, the source domain balanced classification loss, the source domain standard classification loss, the target domain classification loss of the balanced classifier, and the target domain classification loss of the standard classifier to obtain the loss function L total , which is expressed as follows:

[0076]

[0077] Among them, L cor Indicates Coral loss, represents the source domain standard classification loss, represents the source domain balanced classification loss, represents the target domain classification loss of the standard classifier, represents the target domain classification loss of the balanced classifier, and μ is a hyperparameter.

[0078] It can be seen from the above technical solutions that the present invention can solve the following problems existing in the prior art:

[0079] (1) Solve the label scarcity problem: During the training process, not only the labeled source domain data but also the unlabeled target domain data are used. The features of the target domain data are extracted by a shared feature extractor, and pseudo labels are generated using a standard classifier. These pseudo labels supplement the label information to a certain extent. In order to reduce the noise that may be caused by pseudo labels, the present invention designs a method for adjusting the perceptual weight. This method combines the original output of the standard classifier and the output after post-logit adjustment to more accurately evaluate the reliability of the pseudo labels, thereby giving different weights during the training process and reducing the negative impact of unreliable pseudo labels.

[0080] (2) Solving the problem of unbalanced data distribution: Two parallel classifier modules, a balanced classifier and a standard classifier, are constructed. Each of them has different initialization parameters or structures, so that more diverse feature information can be captured during training. When constructing the loss function, the present invention comprehensively considers the source domain balanced classification loss, the source domain standard classification loss, the target domain classification loss of the balanced classifier, and the target domain classification loss of the standard classifier. By weighted summing these losses, the model pays more attention to the minority class during training, thereby reducing the impact of unbalanced data distribution.

[0081] (3) Improving model generalization: A feature sharing strategy is adopted. By sharing feature extractors, the balanced classifier and the standard classifier can share some feature information. This helps the model learn more general feature representations from large-scale data, thereby enhancing the model's generalization ability. At the same time, the Coral loss is introduced into the loss function. This loss term is used to quantify the difference between the data distribution of the source domain and the target domain. By minimizing the Coral loss, the model can better adapt to the data distribution of the target domain, thereby improving the model's generalization performance in unknown environments.

[0082] Furthermore, step S21 further includes: using the Coral loss to align the feature distribution of the source domain data features and the target domain data features. The Coral loss L cor It is expressed as follows:

[0083]

[0084]

[0085] Among them, d E represents the output dimension of the shared feature extractor, x s and x t represent the source domain data and the target domain data respectively, ||·|| represents the Frobenius norm, C s and C t Respectively represent the covariance matrix of the source domain data and the target domain data, n s and n t Represent the source domain data features and the target domain data features respectively, and 1 is a column vector with all elements equal to 1.

[0086] Furthermore, the source domain balanced classification loss It is expressed as follows:

[0087]

[0088] Among them, n s is the total number of samples of the source domain data, x s,i is the i-th sample in the source domain data, y s,i is the label of the i-th sample in the source domain data, represents the output of the balanced classifier for the i-th sample in the source domain data, represents the logit adjustment factor of the balanced classifier, π s,[k] =n s,[k] / n s represents the observation frequency of the kth category in the source domain data, n s,[k] is the total number of samples in the kth class in the source domain data, represents the temperature coefficient of the balanced classifier.

[0089] Furthermore, the source domain standard classification loss It is expressed as follows:

[0090]

[0091] in, represents the output of the balanced classifier for the i-th sample in the source domain data, represents the logit adjustment factor of the standard classifier, represents the temperature coefficient of the standard classifier, represents the observation frequency of the kth class in the target domain data estimated by the balanced classifier, λ is a weighting parameter, represents the class distribution term of the kth class in the target domain data in the first b-1 batches estimated by the balanced classifier, Represents the class distribution item of the kth class in the target domain data in the current bth batch.

[0092] Furthermore, the class distribution item of the kth class in the target domain data in the current bth batch is The calculation formula is:

[0093]

[0094] Among them, n b is the number of samples in a batch, and δ(·) represents the Softmax function.

[0095] Furthermore, in step S23, a pseudo label of the target domain data for the balanced classifier is generated. The calculation formula is expressed as:

[0096]

[0097] in, represents the target class distribution term estimated by the standard classifier, is the temperature coefficient;

[0098] Get the adjusted perception weight w(x t,i ) is calculated as:

[0099] w(x t,i )=α m β(x t,i )

[0100] Among them, x t,i is the i-th sample in the target domain data, β(x t,i ) represents the logit consistency factor, max(·) represents the maximum value operation, δ(·) represents the Softmax function; α m represents the adjustment of the perception factor in the mth training step, n b is the number of samples in a batch.

[0101] Furthermore, the target domain classification loss of the standard classifier is It is expressed as follows:

[0102]

[0103] Among them, 1{·} is an indicator function, which returns 1 if the condition is met, otherwise it returns 0, T represents the confidence threshold, is the pseudo label of the target domain data of the standard classifier, is the pseudo label of the i-th sample in the k-th class in the target domain data of the standard classifier, x t,i is the i-th sample in the target domain data, x t,i,[k] is the i-th sample in the k-th class in the target domain data, n t is the total number of samples of the target domain data, and K represents the total number of categories of the target domain data.

[0104] Furthermore, the target domain classification loss of the balanced classifier is It is expressed as follows:

[0105]

[0106] Among them, w(x t,i ) is to adjust the perception weight, n t is the total number of samples of the target domain data, is the pseudo label of the i-th sample in the target domain data for the balanced classifier, is the pseudo label of the i-th sample in the k-th class in the target domain data for the balanced classifier, x t,i is the i-th sample in the target domain data, x t,i,[k] is the i-th sample in the k-th class in the target domain data, n t is the total number of samples of the target domain data, and K represents the total number of categories of the target domain data.

[0107] Based on the above embodiments, this embodiment uses Figure 4 The bearing test bench shown here collects vibration signals. Using wire cutting technology, three types of faults are created for the tested bearing types: inner race fault (IF), outer race fault (OF), and ball fault (BF). Furthermore, each fault has two fault size levels: 0.2 mm and 0.4 mm. Adding the normal condition (NC), the bearing dataset has seven health conditions. Table 1 shows the fault labels for the bearing dataset:

[0108] Table 1 Description of fault labels in the bearing dataset

[0109]

[0110] The vibration signals of the test bearings are measured by a sensor located on top of the bearing mounting bracket, with a sampling frequency of 10 kHz. The bearing dataset contains three sub-datasets: L1 (0 kN), L2 (1 kN), and L3 (2 kN). The source domain data follows a long-tailed class distribution, while the target domain data follows an unknown class distribution. To facilitate the construction of the diagnostic task, five typical class distributions are designed for the target domain.

[0111] For the five typical class distributions, the number of samples per class in the target domain was set, as shown in Table 2. With the exception of the uniform distribution, two different sample number settings were used for the other four class distributions. Furthermore, two sample number settings were designed for the long-tail source domain. After setting the sample number for the source and target domains, 18 diagnostic tasks were designed for the bearing dataset, as shown in Table 3. For example, in Task A1, the SCU dataset across three operating conditions (L0, L1, and L2) was treated as a separate domain, used as either the source or target domain. When one condition (such as L0) serves as the source domain and another condition (such as L1) serves as the target domain, a sub-transfer task (L0→L1) is created. This setup results in six sub-transfer tasks in A1 (L0→L1, L0→L2, L1→L0, L1→L2, L2→L0, and L2→L1). Therefore, the evaluation metrics for Task A1 are calculated by averaging these six sub-transfer tasks. Following this structure, the other diagnostic tasks are configured similarly.

[0112] Table 2 Sample number settings for each class under different class distributions of source and target domains

[0113]

[0114] Table 3 Diagnostic task description

[0115]

[0116] The detailed structural parameter settings of the shared feature extractor, standard classifier and balanced classifier are shown in Table 4:

[0117] Table 4 Description of structured parameters

[0118]

[0119] In Table 4, "Convolution" represents a convolutional layer, and "Max_Pooling" represents a maximum pooling layer. The "Adaptive_Pooling" layer adjusts the pooling size based on the input dimension. "Kernel_size = 15*16" specifies a kernel width of 15 with 16 output channels, and "padding = invalid" indicates that no padding is applied in the convolutional layer. "ReLU" represents the rectified linear unit activation function, and "FC" represents a fully connected layer. The main parameter settings for the model training and fine-tuning phases are shown in Table 5. In addition, the stochastic gradient descent (SGD) algorithm was selected for model training.

[0120] Table 5 Main parameter settings of the present invention

[0121]

[0122] In order to verify the superiority of the method of the present invention, five methods were selected for comparison, including two homemade methods and two latest methods, as described below:

[0123] (1) Focal: Focal loss is a mature method to solve class imbalance by increasing the loss weight of the minority class, thereby shifting the focus of the model to more difficult samples.

[0124] (2) ACIL: Adaptive class imbalance loss adaptively emphasizes difficult-to-classify classes through class-level weighting terms, enhances the inter-class margin through boundary regularization, and improves the model's detection sensitivity for minority classes.

[0125] (3) SMOTE: Synthetic Minority Oversampling Technique SMOTE addresses the class imbalance problem by generating synthetic samples for the minority class. It expands the dataset by inserting between existing samples of the minority class.

[0126] (4) ACWGAN-GP: ACWGAN-GP addresses the data imbalance problem in fault diagnosis by generating high-quality samples for the minority class, and improves the diagnosis accuracy and dataset balance through progressive enhancement with real synthetic samples.

[0127] (5) CA-SupCon: Class-aware Supervised Contrastive Learning The CA-SupCon method enhances imbalanced fault diagnosis by combining supervised contrastive learning with a class-aware sampler, obtaining a discriminative feature space with better intra-class compactness and inter-class separability for minority class faults.

[0128] Table 6 shows the comparison of detection accuracy for 18 diagnostic tasks:

[0129] Table 6 Comparison of test accuracy between this method and the comparison method

[0130]

[0131] Table 6 summarizes the diagnostic accuracy of the proposed DCS and the five compared methods. As can be seen from Table 6, Focal performs the worst, with an average accuracy of 60.9% and 61.3% on the two datasets, respectively. This shows that the focal strategy, which assigns greater weight to the minority class in the source domain, fails to address the challenges posed by the different target class distributions. Compared with Focal, ACIL significantly improves the accuracy by optimizing the margins of the classification boundary. However, it still lacks a customized strategy for different target class distributions, resulting in performance still lower than the proposed DCS. SMOTE generates samples for the minority class through linear interpolation. It assumes a linear data distribution that rarely aligns with real-world scenarios, resulting in poor sample quality and limited adaptability to changes in the target class distribution. ACWGAN-GP significantly improves the quality of generated samples by enhancing GAN. Although the source domain can be transformed into an approximately uniform distribution through data augmentation, its accuracy drops significantly when dealing with non-uniformly distributed target domains.

[0132] As shown in Table 6, ACWGAN-GP achieves significantly higher accuracy on tasks A5 and A6 than on other tasks. Compared to CA-SupCon, the proposed DCS improves the average accuracy by 10.2% and 9.4% on the two datasets, respectively, and outperforms CA-SupCon across all diagnostic tasks. This demonstrates that the proposed DCS can better handle diverse target class distributions and more effectively address the problem of class distribution bias.

[0133] Figure 5 The feature visualization results under five different target domain category distribution scenarios are shown. These five distributions are A4 (long-tail distribution), A6 (uniform distribution), A10 (reverse long-tail distribution), A14 (head-heavy and tail-light distribution), and A18 (intermediate distribution). Figure 5 It can be found that in the A4, A6, and A18 tasks where the target domain presents long-tail distribution, uniform distribution, and intermediate distribution, the method of the present invention is particularly effective in clustering samples of corresponding categories in the source domain and the target domain, and the classification boundaries between categories are very clear. However, in the A10 task with an inverse long-tail distribution, the classification performance of the two minority categories (6 and 7) in the source domain is relatively poor, and there is a certain degree of confusion between these two categories of samples and the second category samples. This phenomenon shows that when there is a large difference in the category distribution between the source domain and the target domain, the insufficient number of samples of the minority category in the source domain will have a negative impact on the classification effect of the corresponding category in the target domain. Overall, the method of the present invention can obtain satisfactory clustering effects under various category distribution conditions, which fully verifies the effectiveness and applicability of the method of the present invention.

[0134] Example 2

[0135] Based on the same inventive concept as that of the first embodiment, the present invention also provides a bearing imbalance fault diagnosis system guided by a dual classifier collaboration, which is used to implement the steps of the bearing imbalance fault diagnosis method guided by a dual classifier collaboration as described in the first embodiment. Figure 6 As shown, the bearing imbalance fault diagnosis system guided by dual classifier collaboration includes the following modules:

[0136] The data set construction module 100 is used to collect bearing vibration signals in various health states under different working conditions, perform fast Fourier transform on them, convert them into frequency domain signals, obtain a data set, divide the data set into source domain data and target domain data under different working conditions, and label the source domain data;

[0137] A model construction and loss function design module 200 is used to construct a fault diagnosis model, the fault diagnosis model including a shared feature extractor, a balanced classifier, and a standard classifier; and to construct a loss function based on the source domain data and the target domain data;

[0138] A model training module 300 is used to train the fault diagnosis model using the loss function to obtain a trained fault diagnosis model;

[0139] A fault diagnosis module 400 is used to input the vibration signal of the bearing to be tested into the trained fault diagnosis model to obtain a fault diagnosis result;

[0140] The model construction and loss function design module 200 includes a loss function construction submodule 210, and the loss function construction submodule 210 is configured as follows:

[0141] Using the shared feature extractor to perform feature extraction on the source domain data and the target domain data respectively to obtain source domain data features and target domain data features;

[0142] Inputting the source domain data features into the balanced classifier and the standard classifier respectively to obtain a source domain balanced classification loss and a source domain standard classification loss;

[0143] Inputting the target domain data features into a standard classifier to generate target domain data pseudo labels for the balanced classifier; combining the original output of the standard classifier for the target domain data features with the output of the standard classifier for the target domain data features after post-logit adjustment to obtain adjusted perception weights;

[0144] Obtaining a target domain classification loss of a balanced classifier using the adjusted perception weights and the target domain pseudo labels; and obtaining a target domain classification loss of a standard classifier by filtering an original output of a standard classifier for target domain data features using a confidence threshold.

[0145] The loss function is obtained by performing weighted calculation on the Coral loss, the source domain balanced classification loss, the source domain standard classification loss, the target domain classification loss of the balanced classifier, and the target domain classification loss of the standard classifier.

[0146] A bearing imbalance fault diagnosis system with collaborative guidance of dual classifiers proposed in this embodiment is used to implement the aforementioned bearing imbalance fault diagnosis method with collaborative guidance of dual classifiers. Therefore, the specific implementation method in the synchronous response single-point displacement monitoring system can be seen in the embodiment part of the aforementioned bearing imbalance fault diagnosis method with collaborative guidance of dual classifiers. For example, the data set construction module 100, the model construction and loss function design module 200, the model training module 300 and the fault diagnosis module 400 are respectively used to implement steps S1 to S4 in the bearing imbalance fault diagnosis method with collaborative guidance of dual classifiers in embodiment one. Accordingly, the functions implemented by the loss function construction submodule correspond to steps S21 to S25 in the bearing imbalance fault diagnosis method with collaborative guidance of dual classifiers in embodiment one. Therefore, its specific implementation method can refer to the description of the corresponding embodiments of each part. In order to avoid redundancy, it will not be repeated here.

[0147] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0148] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0149] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0150] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0151] Obviously, the above embodiments are merely examples for clarity of explanation and are not intended to limit the implementation methods. Those skilled in the art will appreciate that other variations or modifications can be made based on the above description. It is not necessary and impossible to enumerate all implementation methods here. Obvious variations or modifications arising therefrom remain within the scope of protection of the present invention.

Claims

1. A bearing imbalance fault diagnosis method guided by dual classifier collaboration, characterized in that: The following steps are involved: S1: Collect bearing vibration signals in various health states under different working conditions and perform fast Fourier transform to convert them into frequency domain signals to obtain a data set. The data set is divided into source domain data and target domain data under different working conditions, and the source domain data is labeled. S2: constructing a fault diagnosis model, the fault diagnosis model including a shared feature extractor, a balanced classifier, and a standard classifier; and constructing a loss function based on the source domain data and the target domain data; S3: training the fault diagnosis model using the loss function to obtain a trained fault diagnosis model; S4: inputting the vibration signal of the bearing to be tested into the trained fault diagnosis model to obtain a fault diagnosis result; In step S2, the method for constructing the loss function includes: S21: Using the shared feature extractor to extract features from the source domain data and the target domain data respectively, to obtain source domain data features and target domain data features; S22: Inputting the source domain data features into the balanced classifier and the standard classifier respectively to obtain a source domain balanced classification loss and a source domain standard classification loss; S23: Inputting the target domain data features into a standard classifier to generate target domain data pseudo labels for the balanced classifier; combining the original output of the standard classifier for the target domain data features with the output of the standard classifier for the target domain data features after post-logit adjustment to obtain an adjusted perception weight; S24: obtaining a target domain classification loss of a balanced classifier using the adjusted perception weights and the target domain data pseudo-labels; and obtaining a target domain classification loss of a standard classifier by filtering an original output of a standard classifier for target domain data features using a confidence threshold. S25: Perform weighted calculation on the Coral loss, the source domain balanced classification loss, the source domain standard classification loss, the target domain classification loss of the balanced classifier, and the target domain classification loss of the standard classifier to obtain the loss function L total , which is expressed as follows: Among them, L cor Indicates Coral loss, represents the source domain standard classification loss, represents the source domain balanced classification loss, represents the target domain classification loss of the standard classifier, represents the target domain classification loss of the balanced classifier, and μ is a hyperparameter.

2. The bearing imbalance fault diagnosis method guided by dual classifier collaboration according to claim 1 is characterized in that: Step S21 further includes: using the Coral loss to align the feature distributions of the source domain data features and the target domain data features.

3. The bearing imbalance fault diagnosis method guided by dual classifier collaboration according to claim 1 or 2, characterized in that: The Coral loss L cor It is expressed as follows: Among them, d E represents the output dimension of the shared feature extractor, x s and x t represent the source domain data and the target domain data respectively, ||·|| represents the Frobenius norm, C s and C t Respectively represent the covariance matrix of the source domain data and the target domain data, n s and n t Represent the source domain data features and the target domain data features respectively, and 1 is a column vector with all elements equal to 1.

4. The bearing imbalance fault diagnosis method guided by dual classifier collaboration according to claim 1, characterized in that: The source domain balanced classification loss It is expressed as follows: Among them, n s is the total number of samples of the source domain data, x s,i is the i-th sample in the source domain data, y s,i is the label of the i-th sample in the source domain data, represents the output of the balanced classifier for the i-th sample in the source domain data, represents the logit adjustment factor of the balanced classifier, π s,[k] =n s,[k] / n s represents the observation frequency of the kth category in the source domain data, n s,[k] is the total number of samples in the kth class in the source domain data, represents the temperature coefficient of the balanced classifier.

5. The bearing imbalance fault diagnosis method guided by dual classifier collaboration according to claim 1, characterized in that: The source domain standard classification loss It is expressed as follows: in, represents the output of the balanced classifier for the i-th sample in the source domain data, represents the logit adjustment factor of the standard classifier, represents the temperature coefficient of the standard classifier, represents the observation frequency of the kth class in the target domain data estimated by the balanced classifier, λ is a weighting parameter, represents the class distribution term of the kth class in the target domain data in the first b-1 batches estimated by the balanced classifier, Represents the class distribution item of the kth class in the target domain data in the current bth batch.

6. The bearing imbalance fault diagnosis method guided by dual classifier collaboration according to claim 5, characterized in that: The class distribution item of the kth class in the target domain data in the current bth batch The calculation formula is: Among them, n b is the number of samples in a batch, and δ(·) represents the Softmax function.

7. The bearing imbalance fault diagnosis method guided by dual classifier collaboration according to claim 1, characterized in that: In step S23, generate pseudo labels for the target domain data of the balanced classifier The calculation formula is expressed as: in, represents the target class distribution term estimated by the standard classifier, is the temperature coefficient; Get the adjusted perception weight w(x t,i ) is calculated as: w(x t,i )=a m ·β(x t,i ) Among them, x t,i is the i-th sample in the target domain data, β(x t,i ) represents the logit consistency factor, max(·) represents the maximum value operation, δ(·) represents the Softmax function; α m represents the adjustment of the perception factor in the mth training step, n b is the number of samples in a batch.

8. The bearing imbalance fault diagnosis method guided by dual classifier collaboration according to claim 1, characterized in that: The target domain classification loss of the standard classifier It is expressed as follows: Among them, 1{·} is an indicator function, which returns 1 if the condition is met, otherwise it returns 0, T represents the confidence threshold, is the pseudo label of the target domain data of the standard classifier, is the pseudo label of the i-th sample in the k-th class in the target domain data of the standard classifier, x t,i is the i-th sample in the target domain data, x t,i,[k] is the i-th sample in the k-th class in the target domain data, n t is the total number of samples of the target domain data, and K represents the total number of categories of the target domain data.

9. The bearing imbalance fault diagnosis method guided by dual classifier collaboration according to claim 1, characterized in that: The target domain classification loss of the balanced classifier It is expressed as follows: Among them, w(x t,i ) is to adjust the perception weight, n t is the total number of samples of the target domain data, is the pseudo label of the i-th sample in the target domain data for the balanced classifier, is the pseudo label of the i-th sample in the k-th class in the target domain data for the balanced classifier, x t,i is the i-th sample in the target domain data, x t,i,[k] is the i-th sample in the k-th class in the target domain data, n t is the total number of samples of the target domain data, and K represents the total number of categories of the target domain data.

10. A bearing imbalance fault diagnosis system guided by dual classifiers, characterized in that: The method for diagnosing a bearing imbalance fault using a dual classifier collaborative guidance method according to any one of claims 1 to 9 is implemented, wherein the dual classifier collaborative guidance bearing imbalance fault diagnosis system comprises the following modules: A data set construction module is used to collect bearing vibration signals in various health states under different working conditions, perform fast Fourier transform, convert them into frequency domain signals, obtain a data set, divide the data set into source domain data and target domain data under different working conditions, and label the source domain data; A model construction and loss function design module is used to construct a fault diagnosis model, the fault diagnosis model including a shared feature extractor, a balanced classifier, and a standard classifier; and to construct a loss function based on the source domain data and the target domain data; A model training module, configured to train the fault diagnosis model using the loss function to obtain a trained fault diagnosis model; A fault diagnosis module is used to input the vibration signal of the bearing to be tested into the trained fault diagnosis model to obtain a fault diagnosis result; The model construction and loss function design module includes a loss function construction submodule, and the loss function construction submodule is configured as follows: Using the shared feature extractor to perform feature extraction on the source domain data and the target domain data respectively to obtain source domain data features and target domain data features; Inputting the source domain data features into the balanced classifier and the standard classifier respectively to obtain a source domain balanced classification loss and a source domain standard classification loss; Inputting the target domain data features into a standard classifier to generate target domain data pseudo labels for the balanced classifier; combining the original output of the standard classifier for the target domain data features with the output of the standard classifier for the target domain data features after post-logit adjustment to obtain adjusted perception weights; Obtaining a target domain classification loss of a balanced classifier using the adjusted perception weights and the target domain data pseudo-labels; And by using the confidence threshold, the original output of the standard classifier of the target domain data features is filtered to obtain the target domain classification loss of the standard classifier; The loss function is obtained by performing weighted calculation on the Coral loss, the source domain balanced classification loss, the source domain standard classification loss, the target domain classification loss of the balanced classifier, and the target domain classification loss of the standard classifier.

Images