A bearing fault diagnosis method based on multi-scale residual subfield adaptation

Abstract

The application provides a bearing fault diagnosis method based on multi-scale residual subfield adaptation, and belongs to the technical field of industrial bearing fault diagnosis. The method solves the problems of large distribution difference of bearing monitoring data under different working conditions and the scarcity of labeled data. The method comprises the following steps: under a working condition, collecting bearing vibration signals with known fault information as a source domain data set; under other working conditions, collecting bearing vibration signals with unknown fault information as a target domain data set; constructing a multi-scale residual network to extract transferable features of the source domain samples and the target domain samples; using a local center distance difference algorithm to perform subfield adaptation processing on the transferable features of the source domain samples and the target domain samples; taking the local center distance difference loss and the source domain classifier loss as a target optimization function, and updating the training parameters in reverse; saving the model training parameters to realize state recognition under the condition of no labeled target domain data.

Description

Technical Field

[0001] This invention relates to the field of industrial bearing fault diagnosis and deep transfer learning model construction technology, and in particular to a bearing fault diagnosis method based on multi-scale residual subdomain adaptation. Background Technology

[0002] With the rapid development of modern industry, mechanical equipment is becoming increasingly complex, large-scale, and intelligent. As a key supporting component of mechanical equipment, bearing failure can lead to major safety accidents and property losses; therefore, effective diagnosis of bearing failures is extremely important.

[0003] With the rapid development of technologies such as the Industrial Internet of Things, bearing fault diagnosis technology based on machine learning has become a research hotspot, mainly including BP neural network, probabilistic neural network (PNN), singular value decomposition (SVD), support vector machine (SVM) and artificial neural network (ANN). These methods have achieved relatively good results, but due to the difficulty of extracting small features of fault data from shallow network structures, the accuracy of this type of diagnostic method is difficult to further improve. Deep learning models utilize multi-layer neural networks, which can further extract bearing signal features. They have the characteristics of not relying on prior information and strong discrimination ability, avoiding the uncertainty brought about by shallow training, and solving the above problems well. For example, the article [Guo Lu et al. Rolling bearing fault diagnosis based on Teager energy spectrum and convolutional neural network [J]. Computer Applications and Software, 2021, 38(09):111]

[119] proposed a method to convert one-dimensional bearing vibration signals into two-dimensional images, and to realize bearing fault diagnosis based on convolutional neural networks (CNN) by extracting features from the converted two-dimensional images; [Zhang et al., Kangshuo Xing, Ruxue Bai, et al. An enhanced convolutional neural network for bearing fault diagnosis based on time–frequency image[J]. Measurement, 2020, 157:107667] proposed an intelligent bearing diagnosis model based on multi-scale CNN, which uses different kernel sizes for multi-scale convolution sampling and performs diagnosis through feature information from different receptive fields; [Li et al. A Novel Bearing Fault Diagnosis of RawSignals Based on 1D Residual Convolution Neural Network[J]. IEEE, 2020, 32:1] [6] A residual convolutional neural network is proposed. By introducing CNN, residual blocks, global average pooling layers, and batch normalization, bearing features can be effectively identified and extracted. The above method utilizes powerful feature learning capabilities and has achieved great progress in the field of bearing fault diagnosis.

[0004] However, the good diagnostic performance of deep learning is limited by two basic assumptions: 1) the training set and the test set follow independent and identically distributed (i.i.d.) and 2) a large amount of labeled data is required. However, under actual working conditions, the collected data usually runs under different speeds and loads, making it difficult to meet the assumption of independent and identically distributed (i.i.d.) conditions; at the same time, it is also difficult to obtain a large amount of labeled data in the field of bearing fault diagnosis. Transfer learning effectively solves the above problems and further improves the performance of the target task by mining the feature information of the inconsistent source domain (labeled) and target domain (few or no labels) and continuously narrowing the difference in feature information distribution between the two domains. For example, the article [Mao Wentao et al. Multi-scale residual attention deep domain adaptation model for early bearing fault detection [J] , Computer Applications, 2020 ,40(10):2890

[2898] proposes a domain adaptation transfer learning method, which further reduces the distribution difference between the source and target domains in the mapping space through maximum mean difference (MMD); Article [Li et al. Central moment discrepancy-based domain adaptation for intelligent bearing fault diagnosis[J]. Neurocomputing ,2021 ,429:12

[14] proposes a domain adaptive bearing fault diagnosis method based on central moment difference (CMD). It uses CNN to extract features from vibration signals in two domains and minimizes the high-order matrix of features through the CMD criterion, which effectively avoids the disadvantages of complex MMD kernel matrix calculation and the influence of kernel parameter selection on domain adaptability.

[0005] The methods described above have all achieved good results, but they mainly align distributions from the perspective of global domain adaptation, neglecting the distribution relationships of related subdomains within the same category. This can easily confuse diagnostic knowledge for different bearing fault categories. Furthermore, transfer learning is based on deep learning, and traditional deep learning models do not sufficiently extract bearing signal features, which can also affect the transfer learning effect. Summary of the Invention

[0006] To address the above problems, the present invention aims to provide a bearing fault diagnosis method based on multi-scale residual subdomain adaptation. This method can fully extract the transferable features of the source domain and the target domain, and further improve the accuracy of bearing fault diagnosis by reducing the distribution differences of the same category.

[0007] This invention provides a bearing fault diagnosis method based on multi-scale residual subdomain adaptation, comprising the following steps:

[0008] S1: Under one working condition, the bearing vibration signal with known fault information is collected as the source domain dataset. Under other working conditions (variable load and variable frequency), the bearing vibration signal with unknown fault information is collected as the target domain dataset. The target domain dataset includes a test set and a training set.

[0009] S2: Construct a multi-scale residual network to fully extract the deep transferable features of the training sets in the source domain dataset and the target domain dataset;

[0010] S3: Utilize the deep transferable features of the source and target domains, the true labels in the source domain dataset, and the pseudo labels in the target domain dataset to calculate the difference in local central distance, which measures the difference in the distribution of related subdomains of the same category between the source and target domains.

[0011] S4: Use the local center distance difference loss and the classifier loss as the objective optimization function, and update the parameters in reverse through the Adam optimizer to complete the model training;

[0012] S5: Save the model training parameters, take the test set in the target domain dataset as input, and realize state recognition in the case of unlabeled target domain data.

[0013] Furthermore, in step S2, a multi-scale residual network is used to fully extract the deep transferable features of the training sets in both the source and target domain datasets, including the following steps:

[0014] 1 1) Global feature extraction is performed on the bearing vibration signal using a 1*16 large convolutional kernel, pooling layer, and ReLU activation function, as shown in the following formula: in, and represents the feature vectors of the input and output of the convolutional layer, respectively, and w and b represent the weights and biases during the training process of this layer, respectively. Represents the convolution operation. This represents a pooling operation, where m and n are... The corresponding pool kernel covers the region, where s is the length of the non-overlapping segment. This represents the feature vector after pooling. This represents the activated feature vector. This represents an activation operation;

[0015] 1 2) Simplify the above operations by defining convolution, pooling, and ReLU activation functions as convolution modules, as shown in the following equation:

[0016] in, This represents the feature vector after passing through the convolution module. Represents the input to the convolutional layer. This represents the overall operation of the convolution module. Represents the convolution operation. Represents pooling operations. This represents an activation operation;

[0017] 1 3) Local feature extraction is performed on the bearing signal. A multi-scale residual block is designed, with four channels inside. Each channel uses four convolutional modules of different sizes (1*1, 1*3, 1*5, and 1*7 small convolutional kernels). This allows for the learning of local bearing features at different scales. Additionally, 1*1 convolutional kernels are added to the branches to increase the dimensionality of the features, ensuring that the channel dimensions are consistent when the main path and branches are added within the residual block, as shown in the following formula:

[0018] Where R∈{1,2,3,4} represents a multi-scale residual module (multi The scale residual block (MRB) has 4 channels, where C represents... The number of feature channels, where K represents a 1×C tensor. Represents residual calculation. Represents branch convolution operation, Feature vectors representing the fusion of feature information at different scales. The feature vector representing the local feature extraction of the multi-scale residual block. This indicates that the operation has been performed by two convolutional modules. Represents the convolution operation;

[0019] 14) After global and local feature extraction, the feature vector M is obtained. l Flatten it into a one-dimensional vector, as shown in the following equation:

[0020] in, Represents the tiling function; represents the eigenvector of the tiling layer. This represents the feature vector extracted from global and local features;

[0021] 1 5) After applying the tiling function, the feature vector is obtained as follows: Perform a full connection operation on it, as shown in the following equation:

[0022] in, Represents a fully connected mapping operation. The feature vector representing the tiling layer, where w and b represent the weights and biases during the training process of this layer, respectively. The feature vector representing the fully connected layer;

[0023] 1 6) After the fully connected operation, the Softmax function is used as the classifier to predict the classification result of the input sample, as shown in the following formula:

[0024] Where Q represents the total number of fault categories, q represents a specific fault category, and F 3 P represents the output feature vector of the classifier, and P represents the probability distribution of the classification layer. The input data is represented by w and b, which represent the weights and biases during the training process of this layer, respectively.

[0025] Furthermore, in step S3, a local center distance difference algorithm is used to perform sub-domain adaptation processing on the transferable features of source domain samples and target domain samples to reduce the distribution differences of related sub-domains of the same category, including the following steps:

[0026] 2 1) Before the subdomain adaptation process, the original center distance difference algorithm is as follows:

[0027] in, and Represents the interval Features of the source and target domains; It is the expected experience of the source domain; The central distance vector of order k for the source domain sample; and Similarly, this represents the empirical expectation of the target domain and the k-th order central distance vector;

[0028] 2 2) To align the distribution of related subdomains within the same category in the source and target domains, the center distance difference algorithm is improved into a local center distance difference algorithm, as shown in the following formula:

[0029] Where 'c' represents the fault category. and Representing source domain samples respectively With target domain samples The probability weight of belonging to category c, Represents the source and target domain intervals. Represents the expected experience of the source domain; The central distance vector of order k for the source domain sample; and Similarly, this represents the empirical expectation of the target domain and the k-th order central distance vector;

[0030] 2 3) Calculate source domain samples The probability weights for belonging to category c are shown in the following formula:

[0031] in, express The corresponding one hot vector tags The cth item, This represents the sum of all samples belonging to class c;

[0032] 2 4) Considering the target domain samples There are no tags in the middle. Uncomputable formula It can be achieved by calculating the target domain samples From the probability distribution, we can obtain the probability of belonging to the c-th item. Further pseudo-labels for the target domain As shown in the following formula:

[0033] in, P represents a pseudo-label for the target domain. j The probability distribution representing the output layer vector of the target domain;

[0034] 2 5) During the subdomain adaptation process, samples from the source domain dataset are extracted from the multi-scale residual network. and target domain dataset samples Deeply transferable features: The source domain contains tags. The feature vector of the fully connected layer, This indicates that the target domain contains pseudo-labels. The feature vector of the fully connected layer, and For those belonging to type c, the local center difference is calculated as follows:

[0035] Where 'c' represents the fault category. and Representing source domain samples respectively With target domain samples The probability weight of belonging to category c, Represents the feature intervals of the source and target domains. Empirical expectations representing the characteristics of the source domain; The central distance vector of order k for the source domain sample; and Similarly, this represents the empirical expectation of the target domain features and the k-th order center distance vector.

[0036] Furthermore, in step S6, the local center distance difference loss and the classifier loss are used as the objective optimization function, and the parameters are updated in reverse through the Adam optimizer to complete the model training. The specific formula is as follows:

[0037] in, Let be the predicted label for the i-th source domain sample. Represents the cross-entropy loss of the source domain samples. The subdomain distribution difference alignment loss represents the loss. The source domain contains tags. The feature vector of the fully connected layer, This indicates that the target domain contains pseudo-labels. The feature vector of the fully connected layer, where λ represents the regularization penalty factor.

[0038] The beneficial effects of adopting the above technical solution are as follows:

[0039] This invention proposes a bearing fault diagnosis method based on multi-scale residual subdomain adaptation. scale Residual Sub Domain Adaptation (MRSA) addresses the problem of large differences in the distribution of bearing monitoring data under different operating conditions and the scarcity of labeled data. This invention improves upon these two processes by fully extracting transferable features from the source and target domains and by reducing the distribution differences within the same category. This invention improves the residual neural network, called the Multiscale Residual Network (MRN). Compared with traditional residual networks, this network adds multi-channel convolutional layers of different sizes to the residual network structure, deeply extracting multi-scale feature information and expanding the network width. It combines the advantages of residual network identity mapping and multi-scale feature extraction, resulting in richer extracted feature information. This solves the problem that traditional residual neural network modules using single-size convolutional kernels are not accurate enough for bearing feature extraction. Inspired by the subdomain adaptation idea of ​​Local Maximum Mean Difference (LMMD), this invention aligns the distribution of related subdomains of the same category in the source and target domains and improves the center distance difference algorithm, called the Local Central Moment Discrepancy (LCMD) algorithm. This solves the shortcomings of MMD, such as the complexity of kernel matrix calculation and the influence of kernel parameter selection on domain adaptation ability, as well as the problem of ignoring the distribution relationship of related subdomains of the same category and easily confusing the diagnostic knowledge of different bearing fault categories.

[0040] Compared with traditional deep transfer learning methods, this invention: (1) avoids the inaccuracy of single-size convolutional kernels in extracting bearing features by using multi-scale residual networks to extract nonlinear and non-stationary rolling bearing signals, and effectively avoids the gradient vanishing problem of deep network structures; (2) the proposed local center distance difference algorithm can solve the problem in traditional bearing fault diagnosis domain adaptation methods that focus on reducing the global distribution difference between the source domain and the target domain, while ignoring the relationship between related subdomains, leading to confusion in the distribution of data of the same category in the two domains; (3) the use of a center distance difference algorithm that does not require expensive kernel matrix calculation effectively avoids the disadvantages of MMD kernel matrix calculation being complex and kernel parameter selection affecting the (sub)domain adaptation capability.

[0041] The proposed method was applied to the bearing fault datasets of Western Reserve University and the University of Paderborn, Germany. Through experimental analysis, the effectiveness and generalization of the proposed bearing fault diagnosis method based on multi-scale residual subdomain adaptation were verified. By comparing (1) CNN: by stacking 3 convolutional module structures (composed of 1*3 convolutional layers, 1*3 max pooling layers and ReLU activation function), 1 tiling layer, 1 fully connected layer and 1 output layer; (2) ResNet: by stacking 2 residual blocks (composed of 2 1*3 convolutional layers), 1 tiling layer, 1 fully connected layer and 1 output layer; (3) MRN; among which, models (1) to (3) are deep learning networks, without domain adaptation ability, and are only trained in the source domain and directly applied to the target domain. (4) MRN MMD; (5) MRN CMD. Models (4) and (5) both use MRN as the deep learning model to extract features, and MMD and CMD are used as global domain adaptation distribution difference criteria, respectively. (6) ResNet LCMD; (7) MRN LMMD; (8)MRSA(MRN) (LCMD). Models (6) and (8) use different deep learning models with the same subdomain adaptation distribution differences, while models (7) and (8) use the same deep learning model with different subdomain adaptation distribution differences. The bearing fault diagnosis accuracy of the method proposed in this invention is significantly improved compared with the diagnosis accuracy of the other 7 models. In addition, it can reduce the feature distribution differences in the subdomains and increase the distance between the subdomains, thereby improving the classification effect. This further verifies that the method proposed in this invention has effectiveness and generalization ability, meeting the actual needs of engineering applications. Attached Figure Description

[0042] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0043] Figure 1 This is a framework diagram of a bearing fault diagnosis method based on multi-scale residual subdomain adaptation;

[0044] Figure 2 It is a diagram of the multi-scale residual network structure;

[0045] Figure 3 It is a multi-scale residual block structure diagram;

[0046] Figure 4 This is a training process diagram of a bearing fault diagnosis method based on multi-scale residual subdomain adaptation;

[0047] Figure 5 This is a feature visualization comparison chart of the deep learning task in Case 1;

[0048] Figure 6 This is a comparison chart of the confusion matrix for the transfer learning task in Case 1;

[0049] Figure 7 This is a feature visualization comparison chart of the transfer learning task in Case 1;

[0050] Figure 8 This is the confusion matrix of the transfer task in Case 2;

[0051] Figure 9 This is a feature visualization of the transfer task in Case 2. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] The terms "comprising" and "having," and any variations thereof, used in the embodiments of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the steps or units listed, but may optionally include other steps or units not listed, or may optionally include other steps or units inherent to such processes, methods, products, or devices.

[0054] This invention uses an industrial bearing fault diagnosis and deep transfer learning model as its carrier, and employs a multi-scale residual network and a local center distance difference algorithm as its main algorithmic framework. The model is as follows: Figure 1 As shown, it includes the following steps:

[0055] S1: Under one working condition, the bearing vibration signal with known fault information is collected as the source domain dataset. Under other working conditions (variable load and variable frequency), the bearing vibration signal with unknown fault information is collected as the target domain dataset. The target domain dataset includes a test set and a training set.

[0056] S2: Constructing a multi-scale residual network, such as... Figure 2 As shown, fully extracting deep transferable features from the training sets of both the source and target domain datasets includes the following steps:

[0057] 1 1) Global feature extraction is performed on the bearing vibration signal using a 1*16 large convolutional kernel, pooling layer, and ReLU activation function, as shown in the following formula:

[0058]

[0059]

[0060] in, and represents the feature vectors of the input and output of the convolutional layer, respectively, and w and b represent the weights and biases during the training process of this layer, respectively. Represents the convolution operation. This represents a pooling operation, where m and n are... The corresponding pool kernel covers the region, where s is the length of the non-overlapping segment. This represents the feature vector after pooling. This represents the activated feature vector. This represents an activation operation;

[0061] 1 2) Simplify the above operations by defining convolution, pooling, and ReLU activation functions as convolution modules, as shown in the following equation:

[0062] in, This represents the feature vector after passing through the convolution module. Represents the input to the convolutional layer. This represents the overall operation of the convolution module. Represents the convolution operation. Represents pooling operations. This represents an activation operation;

[0063] 1 3) Local feature extraction is performed on the bearing signal. A multi-scale residual block is designed, with four channels inside. Each channel uses four convolutional modules of different sizes (1*1, 1*3, 1*5, and 1*7 small convolutional kernels). This allows for the learning of local bearing features at different scales. Additionally, 1*1 convolutional kernels are added to the branches to increase the dimensionality of the features, ensuring that the channel dimensions are consistent when the main path and branches are added within the residual block, as shown in the following formula:

[0064] in, Represents a multi-scale residual module (multi The scale residual block (MRB) has 4 channels, where C represents... The number of feature channels, where K represents a 1×C tensor. Represents residual calculation. Represents branch convolution operation, Feature vectors representing the fusion of feature information at different scales. The feature vector representing the local feature extraction of the multi-scale residual block. This indicates that the operation has been performed by two convolutional modules. Represents the convolution operation;

[0065] 1 4) After global and local feature extraction, the feature vector is obtained as follows: Flatten it into a one-dimensional vector, as shown in the following equation:

[0066] in, Represents a tiling function; The feature vector representing the tiling layer, This represents the feature vector extracted from global and local features;

[0067] 1 5) After applying the tiling function, the feature vector is obtained as follows: Perform a full connection operation on it, as shown in the following equation:

[0068] in, F represents a fully connected mapping operation. 1 The feature vector represents the tiling layer, w and b represent the weights and biases during the training process of this layer, respectively, and F 2 The feature vector representing the fully connected layer;

[0069] 1 6) After the fully connected operation, the Softmax function is used as the classifier to predict the classification result of the input sample, as shown in the following formula:

[0070]

[0071] Where Q represents the total number of fault categories, q represents a specific fault category, and F 3 The output feature vector of the classifier is represented by P, which represents the probability distribution of the classification layer. input The input data is represented by w and b, which represent the weights and biases during the training process of this layer, respectively.

[0072] The parameters of the multi-scale residual network are shown in Table 1:

[0073] Table 1 Parameters of Multi-Scale Residual Network

[0074] S3: Utilize the deep transferable features of the source and target domains, the true labels in the source domain dataset, and the pseudo labels in the target domain dataset to calculate the difference in local central distance, measuring the difference in the distribution of related subdomains of the same class between the source and target domains. This includes the following steps:

[0075] 2 1) Before the subdomain adaptation process, the original center distance difference algorithm is as follows:

[0076] Where X = (x1, x2, ..., x) s ) and Y = (y1, y2…y t () represents the interval [a, b] n Features of the source and target domains; It is the expected experience of the source domain; C k (X)=E(x E(x)) k Represents the center distance vector of the k-th order of the source domain samples; E(Y) and C k (Y) similarly represents the empirical expectation of the target domain and the k-order central distance vector;

[0077] 2 2) To align the distribution of related subdomains within the same category in the source and target domains, the center distance difference algorithm is improved into a local center distance difference algorithm, as shown in the following formula:

[0078] Where 'c' represents the fault category. and Representing source domain samples respectively With target domain samples The probability weights for belonging to category c, [a, b] n Represents the source and target domain intervals. Represents the expected experience of the source domain; The central distance vector of order k for the source domain sample; and Similarly, this represents the empirical expectation of the target domain and the k-th order central distance vector;

[0079] 2 3) Calculate source domain samples The probability weights for belonging to category c are shown in the following formula:

[0080] in, express The corresponding one hot vector tags The cth item, This represents the sum of all samples belonging to class c;

[0081] 2 4) Considering the target domain samples There are no tags in the middle. Uncomputable formula It can be achieved by calculating the target domain samples From the probability distribution, we can obtain the probability of belonging to the c-th item. Further pseudo-labels for the target domain As shown in the following formula:

[0082] in, P represents a pseudo-label for the target domain. j The probability distribution representing the output layer vector of the target domain;

[0083] 2 5) During the subdomain adaptation process, samples from the source domain dataset are extracted from the multi-scale residual network. and target domain dataset samples Deeply transferable features: The source domain contains tags. The feature vector of the fully connected layer, This indicates that the target domain contains pseudo-labels. The feature vector of the fully connected layer, and For those belonging to type c, the local center difference is calculated as follows:

[0084]

[0085] Where 'c' represents the fault category. and Representing source domain samples respectively With target domain samples The probability weights for belonging to category c, [a, b] n Represents the feature intervals of the source and target domains. Empirical expectations representing the characteristics of the source domain; The central distance vector of order k for the source domain sample; and Similarly, the expected value of the target domain features and the k-th order central distance vector are represented.

[0086] S4: The local center distance difference loss and the classifier loss are used as the objective optimization functions, and the parameters are updated in reverse through the Adam optimizer to complete the model training. The specific formula is shown below:

[0087] in, Let ε be the predicted label for the i-th source domain sample. j ε represents the cross-entropy loss of the source domain samples. d The subdomain distribution difference alignment loss represents the loss. The source domain contains tags. The feature vector of the fully connected layer, This indicates that the target domain contains pseudo-labels. The feature vector of the fully connected layer, where λ represents the regularization penalty factor.

[0088] S5: Save the model training parameters, take the test set in the target domain dataset as input, and realize state recognition in the case of unlabeled target domain data;

[0089] Based on the above steps, this invention effectively solves the problems of large differences in the distribution of bearing monitoring data under different operating conditions and the scarcity of labeled data. Firstly, addressing the issue that using a single-size convolutional kernel within the residual neural network module is insufficient for accurate bearing feature extraction, a multi-scale residual network is proposed. By adding multi-channel convolutional layers of different sizes to the residual network structure, multi-scale feature information is extracted at depth, and the network width is expanded. This combines the advantages of residual network identity mapping and multi-scale feature extraction, resulting in richer extracted feature information. Secondly, addressing the problem that traditional bearing fault diagnosis domain adaptation methods focus on reducing the global distribution difference between the source and target domains while ignoring the relationships between related subdomains, leading to confusion in the distribution of data of the same category in both domains, a local center distance difference algorithm is proposed. By reducing the feature distribution difference between the source and target domains within the same category, it achieves good matching of local and global distributions, improving fault classification accuracy. This invention significantly improves bearing fault diagnosis accuracy and effectively enhances transfer learning.

[0090] Experimental verification of the bearing fault diagnosis method based on multi-scale residual subdomain adaptation of this invention:

[0091] 1. Test Environment

[0092] The experimental platform built for this invention was mainly conducted on a Windows 10 64-bit operating system. The equipment configuration information was as follows: CPU was AMD Ryzen 7 5800H, RAM was 16.0GB, the deep learning framework used was TensorFlow+Keras, and the programming language was Python 3.7.

[0093] 2. Experimental verification

[0094] Case 1: Diagnostic Results and Analysis on the Case Western Reserve University Bearing Dataset

[0095] (1) Dataset Description

[0096] The experimental data were selected from the rolling bearing dataset of Case Western Reserve University (CWRU). The experimental setup consists of a motor, encoder, and dynamometer, and the experimental bearing is an SKF6205 motor drive end bearing.

[0097] The experimental datasets are shown in Table 2. Four datasets were constructed for different operating conditions. The transfer learning task was simulated using varying load and frequency. Each dataset contained 10 types, including normal types and combinations of three faulty parts (rolling element, inner ring, and outer ring) in three sizes (0.07, 0.14, and 0.21 inches). Each dataset contained 500 data points, and each sample had 1024 points. The training and test sets were divided in a 7:3 ratio. To simulate the scarcity of labeled data in real-world conditions, samples in dataset A were labeled, while samples in datasets B, C, and D were unlabeled. Each transfer task was denoted as A→B, where A represents the source domain dataset and B represents the target domain dataset. Knowledge was learned from the training sets of the source and target domains to predict the sample status of the target domain test set.

[0098] Table 2 describes the feature vectors of the fully connected layers in the experimental dataset, where λ represents the regularization penalty factor.

[0099] S5: Save the model training parameters, take the test set in the target domain dataset as input, and realize state recognition in the case of unlabeled target domain data;

[0100] Based on the above steps, this invention effectively solves the problems of large differences in the distribution of bearing monitoring data under different operating conditions and the scarcity of labeled data. Firstly, addressing the issue that using a single-size convolutional kernel within the residual neural network module is insufficient for accurate bearing feature extraction, a multi-scale residual network is proposed. By adding multi-channel convolutional layers of different sizes to the residual network structure, multi-scale feature information is extracted at depth, and the network width is expanded. This combines the advantages of residual network identity mapping and multi-scale feature extraction, resulting in richer extracted feature information. Secondly, addressing the problem that traditional bearing fault diagnosis domain adaptation methods focus on reducing the global distribution difference between the source and target domains while ignoring the relationships between related subdomains, leading to confusion in the distribution of data of the same category in both domains, a local center distance difference algorithm is proposed. By reducing the feature distribution difference between the source and target domains within the same category, it achieves good matching of local and global distributions, improving fault classification accuracy. This invention significantly improves bearing fault diagnosis accuracy and effectively enhances transfer learning.

[0101] Experimental verification of the bearing fault diagnosis method based on multi-scale residual subdomain adaptation of this invention:

[0102] 1. Test Environment

[0103] The experimental platform built for this invention was mainly conducted on a Windows 10 64-bit operating system. The equipment configuration information was as follows: CPU was AMD Ryzen 7 5800H, RAM was 16.0GB, the deep learning framework used was TensorFlow+Keras, and the programming language was Python 3.7.

[0104] 2. Experimental verification

[0105] Case 1: Diagnostic Results and Analysis on the Case Western Reserve University Bearing Dataset

[0106] (1) Dataset Description

[0107] The experimental data were selected from the rolling bearing dataset of Case Western Reserve University (CWRU). The experimental setup consists of a motor, encoder, and dynamometer, and the experimental bearing is an SKF6205 motor drive end bearing.

[0108] The experimental datasets are shown in Table 2. Four datasets were constructed for different operating conditions. The transfer learning task was simulated using varying load and frequency. Each dataset contained 10 types, including normal types and combinations of three faulty parts (rolling element, inner ring, and outer ring) in three sizes (0.07, 0.14, and 0.21 inches). Each dataset contained 500 data points, and each sample had 1024 points. The training and test sets were divided in a 7:3 ratio. To simulate the scarcity of labeled data in real-world conditions, samples in dataset A were labeled, while samples in datasets B, C, and D were unlabeled. Each transfer task was denoted as A→B, where A represents the source domain dataset and B represents the target domain dataset. Knowledge was learned from the training sets of the source and target domains to predict the sample status of the target domain test set.

[0109] Table 2 Description of Experimental Datasets

[0110]

[0111] (2) Comparative analysis of experimental results

[0112] To verify the effectiveness of MRSA, the following eight models were used for comparative experiments, with accuracy as an important indicator. The experimental performance of the eight models on the target domain test set is shown in Table 3.

[0113] Table 3. Comparison of experimental performance of the eight models

[0114]

[0115] The accuracy results in Table 4 lead to the following four conclusions:

[0116] 1) MRN has better feature extraction capabilities. Comparing experiments (1), (2), and (3) in Task A, the accuracy of traditional CNN and ResNet is 87.40% and 93.32%, respectively, lower than MRN's 99.47%. This is because CNN cannot fully exploit the correlation of temporal data, and ResNet's use of single-size convolutional kernels for bearing feature extraction is not precise enough, both leading to insufficient feature extraction. In contrast, the MRN structure uses multi-scale convolutions instead of the single-scale convolutions in the original network structure, expanding the network width and resulting in richer feature extraction. To more clearly demonstrate the feature extraction capabilities of MRN, we use t... Distributed random neighborhood embedding (t) The SNE (Search Engine Network) technique is mapped to a two-dimensional space to visualize the feature distribution of three experiments in the fully connected layer. Feature visualization examples include... Figure 5 As shown in the figure, different colors correspond to different bearing states. It can be seen that MRN, compared to CNN and ResNet, is able to learn more distinguishable features from complex bearing signals.

[0117] 2) Domain adaptation methods outperform deep learning methods. Comparing experiments (1)(2)(3) with (4)(5)(6)(7)(8) in tasks A→B, A→C, and A→D reveals that deep learning methods without domain adaptation generally have lower accuracy, reaching a maximum of only 75.19%. This is because the data distributions of the source and target domains differ, meaning diagnostic knowledge learned from the source domain cannot be directly applied to the target domain, resulting in suboptimal diagnostic accuracy. Domain adaptation methods, by continuously narrowing the distribution differences between the source and target domains, significantly improve accuracy, generally reaching over 90%.

[0118] 3) CMD has better domain adaptability than MMD. By comparing experiments (4)(7) and (5)(8) in tasks A→B, A→C and A→D, it can be seen that the diagnostic results of CMD (LCMD) distribution difference are better than those of MMD (LMMD). This is because the MMD method reduces the distance between the weighted sums of all original moments, while CMD does not require expensive kernel matrix calculation, which effectively avoids the disadvantages of MMD kernel matrix calculation being complex and kernel parameter selection affecting (sub)domain adaptability.

[0119] 4) Subdomain adaptation outperforms traditional global domain adaptation, as demonstrated by comparing experiments (4) in tasks A→B, A→C, and A→D.

[0120] (5) As shown in experiments (6), (7), and (8), it can be calculated that the MRN using subdomain adaptation... The average accuracy of LMMD and MRSA were 97.69% and 98.99%, respectively, both significantly higher than that of MRN using global domain adaptation. MMD and MRN CMD (Continuous Domain Adaptation) improves fault classification accuracy because subdomain adaptation reduces the difference in feature distribution between the source and target domains within the same class, adapting to both local and global distributions effectively. However, ResNet, which also uses subdomain adaptation... The LCMD results did not improve significantly because the ResNet network structure did not extract enough features, making it difficult to extract deep transferable features, which in turn affected the subdomain adaptation effect.

[0121] Based on the above four conclusions, it can be concluded that MRSA combines the advantages of subdomain adaptation and CMD, achieving the highest accuracy. To illustrate the classification results for each state in fault diagnosis in detail, the confusion matrices of experiments (3)(4)(5)(6)(7)(8) on task A→B are compared as follows: Figure 7 As shown in the diagram. The vertical axis displays the true labels of the samples, and the horizontal axis displays the predicted labels. From... Figure 6 From (a)(b)(c)(d), we can see MRN, MRN MMD, MRN CMD and ResNet Various states in LCMD can lead to classification errors; from Figure 6 MRN can be seen in (e). Only a small percentage of LMMD classification errors occur; from Figure 7 As can be seen from (f), except for the rolling element fault type which caused a diagnostic error, the MRSA achieved a 100% accuracy rate in identifying the other categories, demonstrating the best classification performance.

[0122] To further illustrate MRSA's ability to align the same category, we compared the high-dimensional features of experiments (3)(4)(5)(6)(7)(8) on task A→B using t... SNE visualizes the feature distribution of six experiments in the fully connected layer and maps it to a two-dimensional space. Feature visualization examples include... Figure 7 As shown, features starting with S represent source domain features, and features starting with T represent target domain features. From... Figure 7 As shown in (a), MRN lacks a domain adaptation process, leading to significant differences in features between the source and target domains. It can be observed that training a model solely on the source domain cannot effectively generalize to the target domain. Figure 7 From (b) and (c), we can see that MRN MMD and MRN After CMD's global domain adaptation, the source and target domains are roughly the same, but both exhibit a problem where the subdomain data distributions of BO_07, BO_14, and BO_21 types are too similar, causing classification confusion. Figure 7From (d)(e)(f), we can see that ResNet LCMD, MRN Through subdomain adaptation, LMMD and MRSA clearly demonstrate distinct boundaries among the 10 types, thanks to ResNet. The limitations of ResNet, the feature extractor in LCMD, lead to excessive differences in the distribution of the source and target domains within subdomains, which affects the subdomain adaptation effect. MRSA, compared to MRN... LMMD can reduce the difference in feature distribution within subdomains while increasing the distance between subdomains, thus improving classification performance.

[0123] Case Study 2: Diagnostic Results and Analysis on the University of Paderborn Bearing Dataset

[0124] (1) Data Description

[0125] The experimental dataset uses the rolling bearing dataset from Paderborn University (PU), Germany. The test bench consists of an electric motor, a torque measurement shaft, a rolling bearing test module, a flywheel, and a load motor.

[0126] The experimental datasets are shown in Table 5. Four bearing datasets under different operating conditions were constructed. The transfer learning task was simulated using varying speed, load torque, and radial force. Data was collected at a sampling rate of 64 kHz. Each dataset has three types: normal type, inner race, and outer race fault types, with 500 data points for each type. Each sample segment has 1024 points, and the training and test sets are divided in a 7:3 ratio. To simulate the actual operating conditions where labeled data is scarce, samples in dataset E are labeled, while samples in datasets F, G, and H are unlabeled. Each transfer task is represented as E→F, where E represents the source domain dataset and F represents the target domain dataset. Knowledge is learned from the training sets of the source and target domains to predict the sample states of the target domain test set.

[0127] Table 4 Description of the experimental dataset

[0128] (2) Comparative Experiment Results and Analysis

[0129] The PU dataset is more consistent with real-world industrial scenarios than the CWRU dataset, focusing more on the transfer of load torque and radial force. To verify the generalization ability of the bearing fault diagnosis method based on multi-scale residual subdomain adaptation, the eight models mentioned above were compared in an experiment, and the results are shown in Table 5.

[0130] Table 5. Comparison of experimental performance of the eight models

[0131]

[0132] The average values ​​of each diagnostic method show that domain adaptation outperforms deep learning methods, CMD outperforms MMD, and subdomain adaptation outperforms traditional global domain adaptation, consistent with the conclusions in Section 3.1.2. Therefore, in complex transfer tasks, the average diagnostic accuracy of MRSA is 97.56%, surpassing the other seven methods and further validating its superiority and effectiveness. Again, using the E→F transfer task as an example, the confusion matrix and feature visualization of MRSA are shown below. Figure 8 and 9 As shown. From Figure 8 It can be seen that both NORMAL and IR types have approximately 2% diagnostic errors, while the OR type has an accuracy rate of 100%, which can cause NORMAL types to be mistakenly identified as IR types. Figure 9 As can be seen, MRSA can aggregate transferable features from different operating conditions, enabling accurate classification of features of the same category and effective separation of different categories.

[0133] 3. Conclusion

[0134] To address the challenges of significant variations in bearing monitoring data distribution under different operating conditions and the scarcity of labeled data, a bearing fault diagnosis method based on multi-scale residual subdomain adaptation is proposed. First, transferable features between the source and target domains are extracted using a multi-scale residual network. Then, the difference in local center distance between the feature distributions of the source and target domains is minimized. The local center distance difference loss and the classifier loss are then used as objective optimization functions, and the parameters are continuously optimized through back-up updates. Finally, the model training parameters are saved to achieve state recognition of unlabeled target domain samples. Experimental results show that the bearing fault diagnosis method based on multi-scale residual subdomain adaptation achieves high accuracy and better stability even with unlabeled target domain data. This also verifies the effectiveness and generalization of the method, meeting the practical needs of engineering applications.

[0135] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes: USB flash drive, portable hard drive, read-only memory (ROM). ROM (ROM-only memory), RAM (random access memory), magnetic disks, optical disks, and other media that can store program code.

[0136] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the scope of the technology disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention. All should be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A bearing fault diagnosis method based on multi-scale residual subdomain adaptation, characterized in that, Includes the following steps: S1: Under one working condition, the bearing vibration signal with known fault information is collected as the source domain dataset. Under other working conditions, the bearing vibration signal with unknown fault information is collected as the target domain dataset. The target domain dataset includes a test set and a training set. S2: Construct a multi-scale residual network to fully extract the deep transferable features of the training sets in the source domain dataset and the target domain dataset; S3: Utilize the deep transferable features of the source and target domains, the true labels in the source domain dataset, and the pseudo labels in the target domain dataset to calculate the difference in local central distance, which measures the difference in the distribution of related subdomains of the same category between the source and target domains. S4: Use the local center distance difference loss and the classifier loss as the objective optimization function, and update the parameters in reverse through the Adam optimizer to complete the model training; S5: Save the model training parameters, take the test set in the target domain dataset as input, and realize state recognition in the case of unlabeled target domain data.

2. The bearing fault diagnosis method based on multi-scale residual subdomain adaptation according to claim 1, characterized in that: In step S2, a multi-scale residual network is used to fully extract the deep transferable features of the training sets in the source domain dataset and the target domain dataset, including the following steps: 1-1) Global feature extraction is performed on the bearing vibration signal using a 1*16 large convolutional kernel, pooling layer, and ReLU activation function, as shown in the following formula: C l =f(wx l-1 +b) P l =pool(max{C l |s(m-1)+1≤n≤sm}) Where, x l-1 and C l Let x and n represent the input and output feature vectors of the convolutional layer, respectively; w and b represent the weights and biases during the training process of this layer, respectively; f(·) represents the convolution operation; pool(·) represents the pooling operation; and m and n are the features of x. l-1 The corresponding pool kernel covers the region, where s is the length of the non-overlapping segment, and P l R represents the feature vector after pooling. l This represents the activated feature vector. This represents an activation operation; 1-2) Simplify the above operations by defining convolution, pooling, and ReLU activation functions as convolution modules, as shown in the following equation: Among them, V l x represents the feature vector after passing through the convolution module. l-1 φ represents the input to the convolutional layer, φ(·) represents the overall operation of the convolutional module, f(·) represents the convolution operation, and pool(·) represents the pooling operation. This represents an activation operation; 1-3) Local feature extraction of bearing signals is performed. A multi-scale residual block is designed, with four channels inside. Each channel uses four convolutional modules of different sizes to learn local bearing features from different scales. At the same time, 1*1 convolutional kernels are added to the branches to increase the dimensionality of the features and ensure that the channel dimension is consistent when the main path and branches are added inside the residual block, as shown in the following formula: Where R∈{1,2,3,4} represents the number of 4 channels within the MRB, and C represents... The number of feature channels, where K represents a 1×C tensor. Represents residual operation, f(wx) l-1 +b) represents the branch convolution operation, Z concat M represents a feature vector fused from feature information at different scales. l The feature vector representing the local feature extraction of the multi-scale residual block. This indicates that the operation has been performed by two convolutional modules, and f(·) represents the convolution operation; 1-4) After global and local feature extraction, the feature vector M is obtained. l Flatten it into a one-dimensional vector, as shown in the following equation: F 1 =flatten(M l ) Where flatten(·) represents the tiling function; F 1 M represents the feature vector of the tiling layer. l This represents the feature vector extracted from global and local features; 1-5) After applying the tiling function, the feature vector F is obtained. 1 Perform a full connection operation on it, as shown in the following equation: F 2 =σ(wF 1 +b) Where σ(·) represents the fully connected mapping operation, F 1 The feature vector represents the tiling layer, w and b represent the weights and biases during the training process of this layer, respectively, and F 2 The feature vector representing the fully connected layer; 1-6) After the fully connected operation, the Softmax function is used as the classifier to predict the classification result of the input sample, as shown in the following formula: P=[p(F 3 =1|x input )…p(F 3 =Q|x input )] T Where Q represents the total number of fault categories, q represents a specific fault category, and F 3 The output feature vector of the classifier is represented by P, which represents the probability distribution of the classification layer. input The input data is represented by w and b, which represent the weights and biases during the training process of this layer, respectively.

3. The bearing fault diagnosis method based on multi-scale residual subdomain adaptation according to claim 1, characterized in that: In step S3, a local center distance difference algorithm is used to perform sub-domain adaptation processing on the transferable features of source domain samples and target domain samples, reducing the distribution differences of related sub-domains of the same category. The specific calculation formula is as follows: 2-1) Before the subdomain adaptation layer process, the original center distance difference algorithm is as follows: Where X = (x1, x2, ..., x) s ) and Y = (y1, y2…y t () represents the interval [a, b] n Features of the source and target domains; It is the expected experience of the source domain; C k (X)=E(xE(x)) k Represents the center distance vector of the k-th order of the source domain samples; E(Y) and C k (Y) similarly represents the empirical expectation of the target domain and the k-order central distance vector; 2-2) To align the distribution of related subdomains of the same category in the source and target domains, the center distance difference algorithm is improved into a local center distance difference algorithm, as shown in the following formula: Where 'c' represents the fault category. and Representing source domain samples respectively With target domain samples The probability weights for belonging to category c, [a, b] n Represents the source and target domain intervals. Represents the expected experience of the source domain; The central distance vector of order k for the source domain sample; and Similarly, this represents the empirical expectation of the target domain and the k-th order central distance vector; 2-3) Calculate source domain samples The probability weights for belonging to category c are shown in the following formula: in, express Corresponding one-hot vector labels The cth item, This represents the sum of all samples belonging to class c; 2-4) Considering the target domain samples There are no tags in the middle. Uncomputable expression It is possible to calculate the target domain samples From the probability distribution, we can obtain the probability of belonging to the c-th item. Further pseudo-tags for the target domain As shown in the following formula: in, P represents a pseudo-label for the target domain. j The probability distribution representing the output layer vector of the target domain; 2-5) During the subdomain adaptation process, samples from the source domain dataset are extracted from the multi-scale residual network. and target domain dataset samples Deeply transferable features: The source domain contains tags. The feature vector of the fully connected layer, This indicates that the target domain contains pseudo-labels. The feature vector of the fully connected layer, and For those belonging to type c, the local center difference is calculated as follows: Where 'c' represents the fault category. and Representing source domain samples respectively With target domain samples The probability weights for belonging to category c, [a, b] n Represents the feature intervals of the source and target domains. Empirical expectations representing the characteristics of the source domain; The central distance vector of order k for the source domain sample; and Similarly, this represents the empirical expectation of the target domain features and the k-th order center distance vector.

Images