Cross-working-condition bearing fault diagnosis method of prototype comparison network based on minimum entropy optimization

Abstract

The invention provides a cross-working-condition bearing fault diagnosis method of a prototype comparison network with minimum entropy optimization, and relates to the technical field of intelligent fault diagnosis. In a pre-training stage, an auxiliary domain discriminator is constructed, DA is assisted by discrimination information of a classifier, classification difficulty is evaluated through sample entropy, and performance degradation in a DA process is inhibited; in the training stage, a prototype is found in a learning vector quantization mode; through intra-domain prototype comparative learning, sample features are prompted to closely gather around similar prototypes in a feature space, and meanwhile, the sample features are kept separated from heterogeneous prototypes; further, the intra-class consistency of the features is enhanced, the inter-class distinction degree is improved, and therefore intra-domain accurate alignment is achieved on the feature level; besides, the cross-domain instance-prototype learning aligns the semantic structure in the shared embedding space, and through a fine-grained alignment strategy, the negative migration problem is relieved, and the generalization ability of the model is improved; and the generalization performance of the model in a cross-working-condition small sample scene is improved through pseudo label generation and a weighted loss function.

Description

Technical Field

[0001] This invention relates to the field of intelligent fault diagnosis technology, and in particular to a prototype comparison network optimized with minimum entropy for small sample bearing fault diagnosis across operating conditions. Background Technology

[0002] Bearings, as key fundamental components of large-scale mechanical equipment, are widely used in power, transmission, and actuation systems. Their health directly determines the operational efficiency and reliability of the entire equipment. Bearing failures often lead to equipment downtime or even complete failure, resulting in significant economic losses. Therefore, timely and accurate diagnosis of bearing faults is of paramount importance for ensuring stable equipment operation and reducing maintenance costs. In recent years, with the rapid development of sensor and data storage technologies, data-driven intelligent fault diagnosis technology has emerged and been widely applied. However, intelligent fault diagnosis technology typically relies on large amounts of labeled data and assumes that training and test data have the same probability distribution. In practical engineering applications, the complex and variable operating conditions of mechanical equipment and the numerous limitations on fault data collection make it difficult to obtain a sufficient quantity of uniformly distributed fault data. Furthermore, differences in signal acquisition equipment and the uncertainty of bearing operating conditions further exacerbate the heterogeneity of data distribution, posing numerous challenges to traditional intelligent fault diagnosis methods in practical applications. Therefore, how to overcome existing technological bottlenecks and develop new methods suitable for bearing fault diagnosis under complex operating conditions has become a critical issue that urgently needs to be addressed in current academic research and industrial applications.

[0003] For bearing fault diagnosis under small sample conditions across various operating conditions, traditional intelligent diagnostic methods rely on the unrealizable assumption that fault samples are sufficiently identically distributed. First, the limited sample size in practical engineering applications prevents DA methods from acquiring sufficient domain-invariant features, leading to decreased model generalization ability. Second, traditional adversarial networks force feature distribution alignment through domain classifiers, but sample weight allocation depends solely on classification confidence, neglecting differences in sample classification difficulty. For example, difficult-to-classify samples exhibit high uncertainty; assigning them large weights significantly impacts model parameter updates during training, potentially causing the model to learn incorrect patterns. Finally, DA methods face unavoidable challenges: dissimilar samples within different domains may be incorrectly grouped into the same category, while samples of the same category may be incorrectly mapped to different categories across different domains, resulting in suboptimal data matching.

[0004] Therefore, how to simultaneously address the variability of bearing distribution across operating conditions and the limitations of small samples in bearing fault diagnosis, achieve feature-level domain alignment, and enhance feature discrimination has become a key technical bottleneck in improving the accuracy of bearing fault diagnosis tasks under small sample conditions across operating conditions. Summary of the Invention

[0005] This invention proposes a method for cross-condition bearing fault diagnosis using a prototype comparison network optimized by minimum entropy, aiming to address the limitations of the variable cross-condition distribution of bearing samples and the small sample size in practical industrial systems.

[0006] This invention provides a method for cross-condition bearing fault diagnosis using a minimum entropy optimized prototype comparison network, comprising the following steps:

[0007] Step 1: Collect bearing vibration signals under multiple operating conditions to construct source domain datasets and target domain datasets;

[0008] Step 2: Extract depth feature representation of bearing vibration signal using convolutional layers and transformer layers;

[0009] Step 3: Establish auxiliary information between global features and prediction information so that the model can capture the non-linear relationship between global features and prediction information;

[0010] Step 4: Calculate the sample entropy, assess the uncertainty of sample classification, and weight the results according to the difficulty of sample classification;

[0011] Step 5: Construct a four-layer fully connected domain discriminator, and combine auxiliary information and sample entropy to optimize the feature distribution alignment between the source and target domains through a domain adversarial loss function;

[0012] Step 6: Use the feature vectors of the bearing data to obtain the prototype of the data by learning vector quantization;

[0013] Step 7: Perform intra-domain prototype contrast learning, compare the similarity components between the feature vector and the prototype, and calculate the intra-domain prototype contrast loss;

[0014] Step 8: Perform cross-domain instance-prototype learning, compare the similarity components between the source domain feature vector and the target domain prototype, and calculate the cross-domain instance-prototype loss;

[0015] Step 9: Design a weighted total loss function that includes domain adversarial loss, intra-domain prototype contrast loss, cross-domain instance-prototype loss, and classification loss to optimize the model;

[0016] Step 10: Use the trained model to classify and diagnose the bearing fault status in the target domain.

[0017] Furthermore, the method for extracting the depth feature representation of the bearing vibration signal using convolutional layers and transformer layers in step S2 includes:

[0018] (1) The first CNN is performed using 16 convolutional kernels of size 3;

[0019] (2) A second CNN is performed using 32 convolutional kernels of size 3;

[0020] (3) A third CNN is performed using 64 convolutional kernels of size 3;

[0021] (4) Use Transformer layers to analyze the contextual relationships between each feature extracted by the CNN to form global features f;

[0022] (5) After passing through two fully connected layers (FC), a deep feature representation is generated;

[0023] (6) After passing through the Softmax layer, the prediction information g is obtained.

[0024] Furthermore, the method for establishing auxiliary information between global features and prediction information in step S3 to capture the nonlinear relationship between global features and prediction information includes:

[0025] (1) Combine global features f and prediction information g into auxiliary information.

[0026] (2) The combined representation is processed using an auxiliary domain discriminator consisting of four fully connected layers, with the number of neurons in the fully connected layers being 256, 128, 64, and 1, respectively.

[0027] (3) The process of capturing the nonlinear relationship between features and prediction information through the auxiliary domain discriminator is expressed as follows:

[0028]

[0029] Among them, D Auxiliary This indicates an auxiliary discriminator.

[0030] Furthermore, the method for calculating sample entropy, assessing the uncertainty of sample classification, and weighting the results according to the difficulty of sample classification in step S4 includes:

[0031]

[0032] ω(Entropy(g))=1+e -Entropy(g)

[0033] Where N represents the number of classes, g n Let represent the probability that a sample is predicted to be in the nth class; Entropy(g) represents the sample entropy; ω(Entropy(g)) represents the weights applied to the sample; e -Entropy(g) To measure the difficulty of classifying a sample, the greater the sample entropy, the smaller its weight should be.

[0034] Furthermore, the method for constructing a four-layer fully connected domain discriminator in step S5, combining auxiliary information and sample entropy, and optimizing the feature distribution alignment between the source and target domains through a domain adversarial loss function includes:

[0035] Domain adversarial loss can be expressed as:

[0036]

[0037]

[0038] Among them, D s D represents the source domain dataset; t Let i represent the target domain dataset; s represent the source domain index; j represent the target domain index; and t represent the target domain. The feature distributions of the source domain (SD) and the target domain (TD) are aligned by minimizing the weighted domain adversarial loss function.

[0039] Furthermore, the method for obtaining the prototype of the data using learning vector quantization in step S6 includes:

[0040] Randomly select k feature vectors from SD as the initial prototype vectors {p1, p2, ..., pk}. k}, whose tags serve as the corresponding preset tags. Where k represents the number of classes;

[0041] Randomly select a sample from SD. Calculate the distance between the sample and each prototype:

[0042] d ij =||f i s -p j ||2,j∈[1,k]

[0043] Find the relationship with f i s The latest prototype p j* ;if With p j* The corresponding tag t j* If they are the same, then bring p closer. j* Distance from the sample; otherwise, move p away. j* Distance from the sample:

[0044]

[0045] Where η∈(0,1) are hyperparameters; the final SD prototype is obtained after 1000 iterations. and TD prototype

[0046] Further, the method for performing intra-domain prototype contrastive learning, comparing the similarity components between the feature vector and the prototype, and calculating the intra-domain prototype contrastive loss in step S7 includes:

[0047] To accurately capture subtle differences in bearing fault signals and improve the accuracy of fault diagnosis, f is calculated. i s and The similarity components between them are:

[0048]

[0049] Where φ is the temperature value;

[0050] Intra-domain prototype contrast loss can be expressed as:

[0051]

[0052] Where, index s and index t Indicates the index of the input data; L CE (·) represents the cross-entropy loss;

[0053] By minimizing the intra-domain prototype contrast loss, each prototype represents the central trend and typical features of a class of data; by bringing samples of the same class closer together and pushing samples of different classes further apart, the model is able to learn more discriminative features.

[0054] Further, the method for performing cross-domain instance-prototype learning, comparing the similarity components between the source domain feature vector and the target domain prototype, and calculating the cross-domain instance-prototype loss in step S8 includes:

[0055] Given the eigenvectors f of the source domain i s and the prototype of the target domain Calculate similarity components:

[0056]

[0057] To identify the matching relationship between the prototype and the features of the foreign bearing, Perform entropy minimization, obtaining E(V) based on the difficulty of sample classification. i s→t Finally, the cross-domain instance-prototype loss is obtained as follows:

[0058]

[0059] in, Entropy representing the similarity component between source domain features and target domain prototypes; Entropy representing the similarity component between the target domain features and the source domain prototype; D tu This indicates that the target field label is not visible;

[0060] By minimizing the cross-domain instance-prototype loss, the model can avoid the outlier effects of bearing data.

[0061] Furthermore, the method for optimizing the model using a weighted total loss function comprising domain adversarial loss, intra-domain prototype contrast loss, cross-domain instance-prototype loss, and classification loss, as described in step S9, includes:

[0062] Classification loss, adversarial loss, and PCL loss are used to aid in the model training process, and parameters are used to control the impact of the global loss; the losses for each part are as follows:

[0063] (1) Sample reweighting module:

[0064]

[0065] Where λ1 and λ2 are hyperparameters; L represents the predicted label, y represents the true label; pre-total This represents the total loss during the pre-training phase; the training process minimizes L. pre-total Domain adversarial loss L Auxiliary Features f acting on the source and target domains s,t and prediction information g s,t Classification loss L CE It only applies to the source domain; and This represents the network parameters before and after the pre-training phase update, and the domain discriminator Θ. D Using GRL for gradient inversion, the optimization objective is:

[0066]

[0067] (2) Prototype Comparison Learning Module:

[0068]

[0069] Where λ3 and λ4 are hyperparameters; L train-total This represents the total loss during the training phase; the training process minimizes L. train-total Prototype contrast loss L In and L Cross Features f acting on the source and target domains s,t Classification loss L CE It only applies to the source domain; the optimization objective is:

[0070]

[0071] in, This represents the network parameters updated during the training phase.

[0072] Compared with the prior art, the present invention has the following advantages:

[0073] This invention proposes a Minimum Entropy Optimization Prototypical Contrastive Network (MEOPCN). In the pre-training phase, an auxiliary domain discriminator is constructed, leveraging the discriminative information from the classifier to assist in the differential identification (DA). The classification difficulty is evaluated using sample entropy, suppressing performance degradation during the DA process. In the training phase, prototypes are found using learned vector quantization. Through intra-domain prototype contrastive learning, sample features are encouraged to cluster tightly around similar prototypes in the feature space while remaining separated from dissimilar prototypes. This process significantly enhances intra-class consistency and inter-class discriminability, achieving precise intra-domain alignment at the feature level. Furthermore, cross-domain instance-prototype learning aligns semantic structures in a shared embedding space. Through fine-grained alignment strategies, it effectively mitigates the negative transfer problem, further improving the model's generalization ability. This framework requires no target domain labeled data or additional modality support; it significantly improves the model's generalization performance in small-sample scenarios across different operating conditions, advancing bearing fault classification technology towards practical application, solely through pseudo-label generation and a weighted loss function.

[0074] Based on the implementation methods provided in the above aspects, this application can be further combined to provide more implementation methods. Attached Figure Description

[0075] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent upon reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of the invention are illustrated by way of example and not limitation, with the same or corresponding reference numerals denoteing the same or corresponding parts, wherein:

[0076] Figure 1 This is a general framework diagram of the present invention;

[0077] Figure 2 This is an accuracy radar chart on the CWRU dataset of this invention;

[0078] Figure 3 This is an accuracy radar chart on the HUST dataset of this invention;

[0079] Figure 4 This is a confusion matrix diagram for tasks H0 and H8 of the present invention;

[0080] Figure 5 This is a visualization of the t-SNE features of tasks H0 and H8 of the present invention;

[0081] Figure 6 This is a visualization of the t-SNE features of tasks H3 and H9 of the present invention;

[0082] Figure 7 Box plots of fault diagnosis accuracy for different models of the present invention on C6 and H11 missions;

[0083] Figure 8 This is a diagram of the CWRU experimental platform of the present invention;

[0084] Figure 9 This is a diagram of the HUST experimental platform of the present invention.

[0085] In the diagram: 1. Load motor; 2. Torque sensor; 3. Induction motor; 4. Fan-end acceleration sensor; 5. Drive-end acceleration sensor; 6. Speed ​​controller; 7. Electric motor; 8. Bearing; 9. Data acquisition unit; 10. Acceleration sensor. Detailed Implementation

[0086] The exemplary embodiments disclosed in this application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of this application and to fully convey the scope of this application to those skilled in the art. Unless otherwise specified, the technical means used in the embodiments are conventional means well known to those skilled in the art.

[0087] This invention proposes a Minimum Entropy Optimization Prototypical Contrastive Network (MEOPCN). In the pre-training phase, an auxiliary domain discriminator is constructed, leveraging the discriminative information from the classifier to assist in the differential identification (DA). The classification difficulty is evaluated using sample entropy, suppressing performance degradation during the DA process. In the training phase, prototypes are found using learned vector quantization. Through intra-domain prototype contrastive learning, sample features are encouraged to cluster tightly around similar prototypes in the feature space while remaining separated from dissimilar prototypes. This process significantly enhances intra-class consistency and inter-class discriminability, achieving precise intra-domain alignment at the feature level. Furthermore, cross-domain instance-prototype learning aligns semantic structures in a shared embedding space. Through fine-grained alignment strategies, it effectively mitigates the negative transfer problem, further improving the model's generalization ability. This framework requires no target domain labeled data or additional modality support; it significantly improves the model's generalization performance in small-sample scenarios across different operating conditions, advancing bearing fault classification technology towards practical application, solely through pseudo-label generation and a weighted loss function.

[0088] Combination Figures 1 to 9 As shown, an optional embodiment of the present invention provides a method for cross-condition bearing fault diagnosis using a minimum entropy optimized prototype comparison network, the specific steps of which include:

[0089] Step 1: Collect bearing vibration signals under multiple operating conditions to construct source domain datasets and target domain datasets;

[0090] Step 2: Extract local features of the bearing signal using CNN. A first CNN is performed using 16 convolutional kernels of size 3, a second CNN using 32 convolutional kernels of size 3, and a third CNN using 64 convolutional kernels of size 3, interspersed with three max-pooling layers. A Transformer layer is used to analyze the contextual relationships between each feature extracted by the CNN to construct the global feature f. Then, it passes through two fully connected layers to generate a deep feature representation.

[0091] Step 3: Establish auxiliary information between global features and prediction information so that the model can capture the non-linear relationship between global features and prediction information;

[0092] Step 4: Calculate sample entropy, assess the uncertainty of sample classification, and weight the results according to the difficulty of sample classification to prevent the model from learning incorrect patterns and improve the overall classification performance of the model.

[0093] Step 5: Construct a four-layer fully connected domain discriminator, combine sample entropy, optimize feature representation through adversarial loss, enhance domain invariance and suppress negative migration;

[0094] Step 6: Utilize the feature vectors of the bearing data to obtain the prototype of the data through learning vector quantization;

[0095] Step 7: Perform intra-domain prototype contrast learning, compare the similarity components between the feature vector and the prototype, and calculate the intra-domain prototype contrast loss;

[0096] Step 8: Perform cross-domain instance-prototype learning, compare the similarity components between the source domain feature vector and the target domain prototype, and calculate the cross-domain instance-prototype loss;

[0097] Step 9: Design a weighted total loss function that includes domain adversarial loss, intra-domain prototype contrast loss, cross-domain instance-prototype loss, and classification loss to optimize the model;

[0098] Step 10: Use the trained model to classify and diagnose the bearing fault status in the target domain.

[0099] Optionally, the specific method for extracting the depth feature representation of the bearing vibration signal using convolutional layers and Transformer layers in step S2 includes:

[0100] (1) The first CNN is performed using 16 convolutional kernels of size 3;

[0101] (2) A second CNN is performed using 32 convolutional kernels of size 3;

[0102] (3) A third CNN is performed using 64 convolutional kernels of size 3;

[0103] (4) Use Transformer layers to analyze the contextual relationships between each feature extracted by the CNN to form global features.

[0104] (5) After passing through two fully connected layers (FC), a deep feature representation is generated;

[0105] (6) After passing through the Softmax layer, the prediction information g is obtained.

[0106] Optionally, the specific method for establishing auxiliary information between feature representation and prediction information in step S3 to capture the nonlinear relationship between feature representation and prediction information includes:

[0107] (1) Combine feature representation f and prediction information g into auxiliary information.

[0108] (2) The combined representation is processed using an auxiliary domain discriminator consisting of four fully connected layers, with the number of neurons in the fully connected layers being 256, 128, 64, and 1, respectively.

[0109] (3) The process of capturing the nonlinear relationship between features and prediction information through the auxiliary domain discriminator is expressed as follows:

[0110]

[0111] Among them, D Auxiliary This indicates an auxiliary discriminator.

[0112] Optionally, the specific methods for calculating sample entropy, assessing sample classification difficulty, and minimizing classification uncertainty in step S4 include:

[0113] Calculate sample entropy, assess the uncertainty of sample classification, and weight the results according to the difficulty of sample classification:

[0114]

[0115] ω(Entropy(g))=1+e -Entropy(g)

[0116] Where N represents the number of classes, g n Let represent the probability that a sample is predicted to belong to the nth class. Entropy(g) represents the sample entropy. ω(Entropy(g)) represents the weights applied to the sample. -Entropy(g) To measure the difficulty of classifying a sample, the greater the sample entropy, the smaller its weight should be.

[0117] Optionally, in step S5, the specific methods for constructing a four-layer fully connected domain discriminator and optimizing the feature distribution alignment between the source and target domains using a domain adversarial loss function, in conjunction with auxiliary information and sample entropy, include:

[0118] Domain adversarial loss can be expressed as:

[0119]

[0120] Among them, D s D represents the source domain dataset; t represents the target domain dataset; i represents the source domain index; s represents the source domain; j represents the target domain index; t represents the target domain.

[0121] By minimizing the weighted domain adversarial loss function, the feature distributions of the source domain (SD) and the target domain (TD) are aligned.

[0122] Optionally, the specific method for obtaining the prototype of the data using learning vector quantization in step S6 includes:

[0123] Given a feature vector:

[0124] Randomly select k feature vectors from SD as the initial prototype vectors {p1, p2, ..., pk}. k}, whose tags serve as the corresponding preset tags. Where k represents the number of classes.

[0125] Randomly select a sample from SD. Calculate the distance between the sample and each prototype:

[0126] d ij =||f i s -p j ||2,j∈[1,k]

[0127] Find the relationship with f i s The latest prototype p j* .if With p j* The corresponding tag t j* If they are the same, then bring p closer. j* Distance from the sample; otherwise, move p away. j* Distance from the sample:

[0128]

[0129] Here, η∈(0,1) are hyperparameters. The final SD prototype is obtained after 1000 iterations. and TD prototype

[0130] Optionally, the specific method for performing intra-domain prototype contrastive learning, calculating the similarity components between the feature vector and the prototype, and calculating the intra-domain prototype contrastive loss in step S7 includes:

[0131] To accurately capture subtle differences in bearing fault signals and improve the accuracy of fault diagnosis, f is calculated. i s and The similarity components between them are:

[0132]

[0133] Where φ is the temperature value;

[0134] Intra-domain prototype contrast loss can be expressed as:

[0135]

[0136] Where, index s (·) and index t (·) indicates the index of the input data; L CE(·) represents the cross-entropy loss;

[0137] By minimizing the intra-domain prototype contrast loss, each prototype represents the central trend and typical features of a class of data; by bringing samples of the same class closer together and pushing samples of different classes further apart, the model is able to learn more discriminative features.

[0138] Optionally, the specific methods for performing cross-domain instance-prototype learning, calculating the similarity components between the source domain feature vector and the target domain prototype, and calculating the cross-domain instance-prototype loss in step S8 include:

[0139] Given the eigenvectors f of the source domain i s and the prototype of the target domain Calculate similarity components:

[0140]

[0141] To identify the matching relationship between the prototype and the features of the foreign bearing, Perform entropy minimization, based on the difficulty of sample classification. Finally, the cross-domain instance-prototype loss is obtained as follows:

[0142]

[0143] By minimizing the cross-domain instance-prototype loss, the model can avoid the outlier effects of bearing data.

[0144] Optionally, the specific method for optimizing the model using a weighted total loss function comprising domain adversarial loss, intra-domain prototype contrast loss, cross-domain instance-prototype loss, and classification loss as described in step S9 includes:

[0145] Classification loss, adversarial loss, and PCL loss are used to aid in the model training process, and parameters are used to control the impact of the global loss; the losses for each part are as follows:

[0146] (1) Sample reweighting module:

[0147]

[0148] Where λ1 and λ2 are hyperparameters; Let y be the predicted label and y be the true label; the training process minimizes L. pre-total Domain adversarial loss L Conditional Features f acting on the source and target domains s,t and prediction information g s,t Classification loss L CE It only applies to the source domain; and These represent the network parameters before and after the update, respectively, and the domain discriminator Θ. D Using GRL for gradient inversion, the optimization objective is:

[0149]

[0150] (2) Prototype Comparison Learning Module:

[0151]

[0152] Where λ3 and λ4 are hyperparameters; the training process minimizes L train-total Prototype contrast loss L In and L Cross Features f acting on the source and target domains s,t Classification loss L CE It only applies to the source domain; These represent the updated network parameters, and the optimization objective is:

[0153]

[0154] Specifically, this invention employs a combination of pre-training and fine-tuning, the framework of which is as follows: Figure 1 As shown, the model is first pre-trained using the Sample Reweighting (SR) module. Then, it is fine-tuned using the Prototypical Contrastive Learning (PCL) module. Classification loss, adversarial loss, and PCL loss are comprehensively used to optimize model performance, and parameters are introduced to adjust the weights of the global loss, ensuring the effectiveness and efficiency of model training. The definitions of the symbols and formulas required in this invention are as follows.

[0155] First, for bearing fault diagnosis, the sample space with sufficient data for each state under a certain operating condition is taken as the source domain, and the dataset is defined as follows: Similarly, taking the sample set composed of state data under another operating condition as the target domain, the dataset is defined as follows: Secondly, the small sample fault diagnosis task T few By support set and query set Composition can be represented as Each set consists of K samples from each of the N classes. Therefore, T few This can be defined as an N-way-K-shot problem. Furthermore, this invention details key components such as the SR module and the PCL module, and designs corresponding loss functions to optimize model performance, thereby improving the model's generalization ability in the target domain.

[0156] 1. Pre-training phase

[0157] Bearing faults are diverse, each with different frequencies and amplitudes of waveforms, resulting in unique vibration signal characteristics. The SR module is used to pre-train the feature extractor to generate feature representations for different domains. Marginal alignment primarily focuses on reducing the difference in marginal distributions of SD and TD in the feature space, typically achieved by matching the global distributions of the two domains. This approach is too coarse and fails to capture differences in local feature distributions. Excessive differences between the two domains can lead to degraded model performance. Given the discrepancies in learned feature distributions caused by cross-domain marginal probability alignment biases, and the impact of the difficulty of predicting features on the DA process, this invention proposes an SR module. This module uses the distribution of labels under given input features as auxiliary information, achieving finer-grained alignment by considering the joint distribution of features and labels, thus more effectively reducing the risk of negative transfer. Furthermore, the results are weighted according to the difficulty of sample classification, and domain adversarial loss and classification loss are used to ensure the effectiveness of the pre-trained model. The pre-training phase consists of three parts: a feature extractor, a classifier, and an auxiliary domain discriminator, such as... Figure 1 As shown, the feature extractors and classifiers operating on SD and TD share parameters. To address the discrepancies between cross-domain feature distributions, this invention uses an auxiliary domain discriminator to combine the learned features f and predicted information g for domain alignment, thereby enabling D to achieve the desired result. s and D t The characteristic distributions should be as close as possible. The specific process is as follows.

[0158] The network input is a one-dimensional bearing vibration signal. First, this invention utilizes a CNN to extract local features from the bearing signal. A first CNN is performed using 16 convolutional kernels of size 3, a second CNN using 32 convolutional kernels of size 3, and a third CNN using 64 convolutional kernels of size 3, interspersed with three max-pooling layers. Then, a Transformer layer is used to analyze the contextual relationships between each feature extracted by the CNN to construct global features. Finally, the feature representation f is obtained. Then, it passes through two fully connected layers (FC) to generate a deep feature representation. Finally, it passes through a Softmax layer to obtain the prediction information g. Specific network parameters are shown in Table 1.

[0159] Table 1 Network Structure Parameters

[0160]

[0161] To reduce the distribution differences in bearing data under different operating conditions, this invention uses a four-layer FC (Functionally Connected) domain discriminator of size (256, 128, 64, 1) to classify the domain to which a feature belongs, specifically using binary classification. To better represent the differences in features during this process, this invention considers the conditional probabilities of feature representation and prediction information as auxiliary information:

[0162]

[0163] Among them, D Auxiliary This represents an auxiliary discriminator. Thus, the model can capture the non-linear relationship between features and predicted information, improving its classification performance for bearing signals. However, difficult-to-classify bearing samples often have high uncertainty; if assigned large weights, they can significantly impact model parameter updates during training, leading to the model learning incorrect patterns. To address this issue, this invention proposes using sample entropy to evaluate the difficulty of classifying a sample, weighting the results according to the magnitude of the sample entropy:

[0164]

[0165] ω(Entropy(g))=1+e -Entropy(g) (3)

[0166] Where N represents the number of classes, g n Let represent the probability that a sample is predicted to belong to the nth class. Entropy(g) represents the sample entropy. ω(Entropy(g)) represents the weights applied to the sample. -Entropy(g) To measure the difficulty of classifying a sample, the higher its entropy, the smaller its weight should be. Therefore, the domain adversarial loss can be expressed as:

[0167]

[0168] By minimizing the domain adversarial loss, the model can align with significantly different cross-domain distributions, learn knowledge from SD that can be transferred to TD, and effectively handle bearing cross-condition diagnostic tasks.

[0169] 2. Training Phase

[0170] Due to factors such as setpoint adjustments, changes in operating conditions, and mechanical wear, the actual operating conditions of bearings can vary, leading to differences in the data distribution between SD and TD. DA technology aims to align SD with TD. However, this technology faces unavoidable challenges: dissimilar samples within different domains may be incorrectly merged into the same category, while samples of the same category may be incorrectly mapped to different categories across different domains, resulting in suboptimal data matching. To address these issues, this invention proposes the PCL module. This invention interprets only concise symbols on SD and performs similar operations on TD.

[0171] 1) Intra-domain PCL: To achieve high differentiation between different patterns across classes with limited bearing data, this invention designs an intra-domain PCL to capture the inherent commonalities and differences in bearing signals, thereby enhancing the similarity of data within the same category. Simultaneously, it reduces the similarity between data across categories in the semantic space. A Gaussian mixture model is used to obtain the prototype of the data, avoiding incorrect intra-domain clustering and indiscriminate feature learning. This method helps improve the model's ability to identify subtle differences in the features of bearing signals from different categories, enhancing its discriminative power and robustness in practical applications.

[0172] Specifically, given a feature vector: The prototype is found using learned vector quantization. The specific process is as follows: k feature vectors are randomly selected from the SD (Simplified Design Model) as the initial prototype vectors {p1, p2, ..., p...}. k}, whose tags serve as the corresponding preset tags. Where k represents the number of classes. A sample is randomly selected from SD. Calculate the distance between the sample and each prototype:

[0173] d ij =||f i s -p j ||2,j∈[1,k] (5)

[0174] Find the relationship with f i s The latest prototype p j* .if With p j* The corresponding tag t j* If they are the same, then bring p closer. j* Distance from the sample; otherwise, move p away. j* Distance from the sample:

[0175]

[0176] Here, η∈(0,1) are hyperparameters. The final SD prototype is obtained after 1000 iterations. and TD prototype

[0177] To accurately capture subtle differences in bearing fault signals and improve the accuracy of fault diagnosis, this invention is inspired by intra-domain contrastive learning and calculates f. i s and The similarity components between them are:

[0178]

[0179] Where φ is the temperature value. The in-domain prototype contrast loss can be expressed as:

[0180]

[0181] Where, index s and index t Indicates the index of the input data. L CE (·) represents the cross-entropy loss. By minimizing the intra-domain prototype contrast loss, each prototype represents the central tendency and typical characteristics of a class of data. By bringing samples of the same class closer together and increasing the distance between samples of different classes, the model can learn more discriminative features, which is particularly important for bearing fault classification tasks.

[0182] 2) Cross-Domain Instance-Prototype Learning: Due to the variability of bearing operating conditions, data distribution often contains outliers. If an outlier in the SD (Structured Instance) is close to many samples in the TD (Digital Instance), this outlier may incorrectly match multiple samples in the TD, thus affecting the accuracy of the entire classification process. Prototypes are specific to each fault and discriminative between different faults. This invention employs cross-domain instance-prototype learning, eliminating the influence of outliers by matching prototypes with bearing samples from different domains, rather than matching individual samples across domains. This allows the knowledge learned by the model on SD to improve diagnostic performance on TD. Given the feature vector f of SD... i s The prototype of TD Calculate similarity components:

[0183]

[0184] To identify the matching relationship between the prototype and the features of the foreign bearing, Perform entropy minimization, according to equation (2) Finally, the cross-domain instance-prototype loss is obtained as follows:

[0185]

[0186] By minimizing the cross-domain instance-prototype loss, the model can avoid the outlier effects of bearing data.

[0187] 3. Optimize training

[0188] The MEOPCN proposed in this invention first uses the SR module to pre-train the feature extractor and classifier, and then uses the PCL module to fine-tune the network model. Classification loss, adversarial loss, and PCL loss are used to aid the model training process, and parameters are used to control the impact of the global loss. The losses for each part are as follows.

[0189] 1) Sample reweighting module:

[0190]

[0191] Where λ1 and λ2 are hyperparameters. Let y be the predicted label and y be the true label. The training process minimizes L. pre-total Domain adversarial loss L Auxiliary Features f acting on TD and SD s,t and prediction information g s,t Classification loss L CE It only applies to SD. and These represent the network parameters before and after the update, respectively, and the domain discriminator Θ. D Using GRL for gradient inversion, the optimization objective is:

[0192]

[0193] 2) Prototype Comparison Learning Module:

[0194]

[0195] Where λ3 and λ4 are hyperparameters. The training process minimizes L. train-total Prototype contrast loss L In and L Cross Features f acting on SD and TD s,t Classification loss L CE It only applies to SD. These represent the updated network parameters, and the optimization objective is:

[0196]

[0197] To gain a more thorough understanding of the present invention, the present invention will be further described in detail below with reference to specific embodiments.

[0198] Step 1: Collect bearing vibration signals under multiple operating conditions (using CWRU dataset and HUST dataset as examples).

[0199] Step 2: Extract local features of the bearing signal using CNN. A first CNN is performed with 16 convolutional kernels of size 3, a second with 32 kernels of size 3, and a third with 64 kernels of size 3, interspersed with three max-pooling layers. A Transformer layer is used to analyze the contextual relationships between each feature extracted by the CNN to construct the global feature f. Then, it passes through two fully connected layers to generate a deep feature representation.

[0200] Step 3: Establish auxiliary information between global features and prediction information so that the model can capture the non-linear relationship between feature representation and prediction information.

[0201] Step 4: Calculate sample entropy, assess the uncertainty of sample classification, and weight the results according to the difficulty of sample classification to prevent the model from learning incorrect patterns and improve the overall classification performance of the model.

[0202] Step 5: Construct a four-layer fully connected domain discriminator, combine sample entropy, optimize feature representation through adversarial loss, enhance domain invariance and suppress negative migration.

[0203] Step 6: Utilize the feature vectors of the bearing data to obtain the prototype of the data through learning vector quantization.

[0204] Step 7: Intra-domain feature matching: Calculate the similarity components between the feature vector and the prototype; Define intra-domain prototype contrast loss: Calculate the intra-domain prototype contrast loss.

[0205] Step 8: Cross-domain feature matching: Given the feature vector of the source domain and the prototype of the target domain, calculate the similarity component between the feature vector of the source domain and the prototype of the target domain; Entropy minimization and matching relationship identification: Perform entropy minimization on the prototype of the target domain, and identify the matching relationship between the prototype and the features of the bearing in the foreign domain according to the relevant formula; Define cross-domain instance-prototype loss: Calculate the cross-domain instance-prototype loss.

[0206] Step 9: Design a weighted total loss function that includes domain adversarial loss, PCL loss, and classification loss.

[0207] Step 10: Design small-sample experiments across operating conditions to validate and optimize the cross-domain model, and quantitatively analyze the contributions of the SR and PCL modules through ablation experiments.

[0208] Step 11: Use confusion matrix analysis to improve the classification of difficult-to-classify samples (such as complex faults); use t-SNE to show intra-class concentration and inter-class separation; use box plots to evaluate the robustness and adaptability of the model in small sample tasks across operating conditions.

[0209] The following is an experiment and results evaluation:

[0210] 1. Experimental Procedure

[0211] (1) Dataset

[0212] 1) CWRU Dataset: This dataset was provided by Case Western Reserve University in the United States, and its data collection is based on, for example... Figure 8 The experimental platform shown includes the following core components: a load motor 1, driven by an induction motor 3, which supports the test bearing; a torque sensor 2, used to measure the motor's output torque and speed, providing feedback and recording for setting operating conditions; an induction motor 3, acting as a power source to drive the entire transmission system, allowing for adjustment of motor speed to simulate different bearing operating conditions; a fan-end accelerometer 4, used to collect the vibration acceleration signal of the bearing at that end; and a drive-end accelerometer 5, which directly monitors the vibration response of the test bearing under various preset fault conditions. Four faults were collected under four different operating conditions: normal, rolling element fault, outer race fault, and inner race fault, as shown in Table 6. Each fault included damage of three different diameters: 7 mils, 14 mils, and 21 mils. Therefore, the dataset can be divided into 10 categories, including 9 fault categories and 1 normal category. The sampling frequency was set to 12.5 kHz, with continuous sampling for 10 seconds. This experiment selected signals sampled from the drive end.

[0213] Table 6 Fault Types in the CWRU Dataset

[0214]

[0215] 2) HUST Dataset: This dataset was proposed by Huazhong University of Science and Technology, and its data collection is based on, for example... Figure 9 The experimental platform shown mainly consists of a speed controller 6, a motor 7, a bearing under test 8, an acceleration sensor 10, and a data acquisition unit (9). The speed controller 6 is used to precisely set and adjust the rotational speed of the motor 7, which drives the spindle on which the bearing under test 8 is mounted. The acceleration sensor 10 is used to monitor the vibration signal of the bearing 8 in real time and transmits the collected vibration signal to the data acquisition unit 9 for recording and storage. Five faults were set under four different operating conditions: normal, rolling element fault, outer ring fault, inner ring fault, and compound fault (inner ring + outer ring), as shown in Table 7. Each fault includes two different degrees of damage: moderate and severe. Therefore, the dataset can be divided into nine categories. The sampling frequency was set to 25.6 kHz, with continuous sampling for 10.2 s.

[0216] Table 7 HUST Dataset Fault Types

[0217]

[0218] (2) Experimental setup

[0219] To verify the superiority of MEOPCN, this invention selected five methods for comparison: CORAL, CDAN, DWPT, TDAM, and MK-DANN. The method in this invention is implemented using the PyTorch library 2.3.1 in Python, a 12th Gen Intel(R) Core(TM) i3-12100 CPU, and an Intel(R) UHD Graphics 730 GPU. Each group of 1024 adjacent points in the acquired signal constitutes a sample. A total of 48 work condition transfer tasks were designed, including four types of few-shot tasks: 10way-5shot, 10way-10shot, 9way-5shot, and 9way-10shot. The overall workflow consists of two stages. In the pre-training stage, SD and TD come from different work conditions; in the training stage, SD and TD come from the same work condition, but TD samples are not visible in SD. As shown in Table 8, the main configuration of the model is: using the Adam optimizer, λ1 = 0.75, λ2 = 0.3, λ3 = 0.6, λ4 = 0.6. For the two datasets, the pre-training learning rate was 0.001, the ratio of TD to TD was 1:1, the batch size was 600 and 800, and the epochs were 600 and 800, respectively. The fine-tuning learning rates were 0.001 and 0.0001, φ1 was set to 40 and 200, and φ2 was set to 50 and 200, respectively. To verify the accuracy of the MEOPCN method in classification results under cross-condition migration scenarios, this invention conducted 5 experiments for each task, and the results were averaged.

[0220] Table 8 Main Configurations of the Model

[0221]

[0222]

[0223] 2. Experimental Results

[0224] To verify the performance of this invention, it was compared with different models, and the comparison results are shown in Tables 2 and 3. Figure 2 and Figure 3As shown, experimental results of MEOPCN on the CWRU and HUST datasets demonstrate that its performance significantly outperforms other comparative methods. In the 10-way-5-shot task on the CWRU dataset, the MEOPCN model achieved an average classification accuracy of 99.94%, representing significant improvements of 18.22%, 0.37%, 1.36%, 2.6%, and 0.18% compared to methods such as CORAL, CDAN, DWPT, TDAM, and MK-DANN, respectively. When the number of samples per class in the training set increased to 10, the average classification accuracy of MEOPCN further improved to 99.99%, achieving 100% accuracy in 10 out of 12 work condition delivery tasks. Furthermore, in the HUST dataset, when there are only 5 samples per fault type, MEOPCN achieves an average classification accuracy of 85.10%, which is 24.72%, 21.01%, 23.7%, and 14.64% higher than 1D-CNN, Transformer, CORAL, and CDAN, respectively. When the number of samples per class increases to 10, MEOPCN's average accuracy further improves to 95.31%, and in 7 out of 12 work condition transmission tasks, the accuracy exceeds 95%.

[0225] The MEOPCN model exhibits strong adaptability and robustness, achieving high classification accuracy even when there are significant differences in the distributions of SD and TD data. This is mainly attributed to the advantages of the MEOPCN model in feature extraction and DA, which effectively reduces the impact of domain differences and improves the model's data sensitivity under limited sample conditions. Therefore, the MEOPCN model provides an efficient and robust solution for bearing fault diagnosis tasks and has broad application prospects.

[0226] The beneficial effects of this invention are:

[0227] 1. Fine-tuned cross-domain alignment, resulting in a steady improvement in classification accuracy:

[0228] To evaluate the proposed MEOPCN method, this invention applied it to the CWRU dataset and conducted 12 cross-domain fault diagnosis experiments C0-C11, where the classification results were taken from the original paper. Based on the classification accuracy analysis of the CWRU dataset (see Table 2 and...),... Figure 2MEOPCN significantly outperforms other comparative methods in terms of performance. Specifically, in the 10-way-5-shot scenario, the method achieves an average classification accuracy of 99.94%, representing improvements of 18.22%, 0.37%, 1.36%, 2.6%, and 0.18% compared to CORAL, CDAN, DWPT, TDAM, and MK-DANN, respectively. When the number of samples per class in the training set is increased to ten during the training phase, the average classification accuracy of MEOPCN further improves to 99.99%, achieving 100% accuracy in 10 out of 12 delivery tasks. Experimental data fully demonstrate that MEOPCN can reduce the impact of domain differences and improve data sensitivity with limited samples. Furthermore, under the same TD, the method achieves higher accuracy in the 10-shot task than in the 5-shot task. This can be attributed to the fact that increasing the number of labeled source samples helps enhance the model's feature mining capabilities.

[0229] Table 2. Classification accuracy (%) on the CWRU dataset.

[0230]

[0231] To further analyze the superiority of MEOPCN, this invention applied it to the HUST dataset and conducted 12 cross-domain fault diagnosis experiments H0-H11, where each method used the same baseline network as MEOPCN. The classification accuracy analysis of the HUST dataset is detailed in Table 3. Figure 3 MEOPCN significantly outperforms other comparative methods. Specifically, with only five samples per fault type, the method achieves an average classification accuracy of 85.10%, representing improvements of 24.72%, 21.01%, 23.7%, and 14.64% compared to 1D-CNN, Transformer, CORAL, and CDAN, respectively. When the number of samples per class increases to ten, MEOPCN's average classification accuracy further improves to 95.31%. In 7 out of 12 operational condition transmission tasks, the accuracy exceeds 95%. Experimental data fully demonstrate that the MEOPCN method can significantly improve the accuracy of fault diagnosis, as well as its practicality and advanced nature in handling complex fault classification problems. Furthermore, it further proves that the performance of this method increases with the number of fault type samples in the training set.

[0232] Table 3. Classification accuracy (%) on the HUST dataset

[0233]

[0234] 2. The modules complement each other, significantly improving model accuracy:

[0235] The proposed method combines two modules: 1) the SR module learns domain-invariant feature representations by combining auxiliary information and sample entropy; 2) the PCL module performs feature alignment by capturing the classification semantic structure of intradomain data and cross-domain instance-prototype pairs. To further explain the effectiveness of MEOPCN, the following experiments were conducted to demonstrate the importance and complementarity of the SR and PCL modules.

[0236] Tables 4 and 5 show the bearing signal classification performance of models trained using different methods on the CWRU and HUST datasets under small sample cross-condition conditions, respectively. In the tables, w / o SR indicates... Figure 1 The SR module was removed from the model, and w / oPCL indicates the removal of the PCL module. During training, only Nway-10shot was used for fine-tuning. As shown in the table, the average accuracy of w / o SR-w / oPCL was only 94.63% and 68.02%, respectively. The results indicate that the variable operating conditions of bearings lead to significant differences in the data distribution between SD and TD, making direct feature extraction and custom classifier training on SD data ineffective. In w / oPCL, the average accuracy improved by 4.73% and 9.73% after cross-domain alignment, indicating that the SR preprocessing can improve the performance of deep learning models. However, the average accuracy is still lower than MEOPCN, suggesting that blindly using domain transfer may lead to negative transfer. In w / o SR, this invention observed that adding the PCL module improved the average accuracy by 0.69% and 5.1%, indicating that prototype learning can improve the model's performance with small sample sizes when applied across domains. However, the average accuracy is still lower than MEOPCN, indicating that PCL is difficult to adapt to new data distributions. The average accuracy of MEOPCN reached 99.99% and 95.31% respectively, indicating that the PCL optimization based on the SR module further improved the model performance.

[0237] Table 4. Bearing classification performance of different training modules on the CWRU dataset.

[0238]

[0239] Table 5. Bearing classification performance of different training modules on the HUST dataset.

[0240]

[0241] 1) Sample reweighting module

[0242] This invention uses a confusion matrix to conduct an in-depth analysis of the performance advantages of the SR module in cross-domain tasks. Taking tasks H0 and H8 as examples, the confusion matrix results are as follows: Figure 4As shown in the figure, in the multi-class bearing signal experiment, composite fault samples are difficult to classify due to their complexity. Experimental results show that, after introducing the SR module, the classification accuracy of difficult-to-classify samples (classes 7 and 8 in H0 and H8) is significantly improved compared with the traditional DA method, with improvements of 0.51% and 5.71% and 9.9% and 24.41%, respectively. Compared with w / o SR, the improvements are 1.18% and 3.42% and 3.69% and 18.49%, respectively. This result indicates that the SR module, by introducing sample entropy and re-weighting the classification results, effectively reduces the confusion between difficult-to-classify samples, thereby significantly improving classification performance.

[0243] The t-SNE results for tasks H0 and H8 are as follows: Figure 5 As shown, the limitations of traditional DA methods are particularly evident when handling outliers. Specifically, there is significant overlap between classification clusters, and the distance between samples within a cluster and the cluster center is generally large. This not only affects classification accuracy but also leads to obvious classification errors in some cases. This indicates that traditional DA methods struggle to achieve accurate classification when faced with complex sample distributions and outlier interference. In contrast, the SR module performs fine-tuning and optimization of samples, effectively reducing the interference of outliers on the classification process. Under the action of the SR module, the boundaries between classification clusters become clearer, and samples within a cluster are closely distributed around the cluster center, forming a more compact and more segregated classification structure.

[0244] 2) Prototype Comparison Learning Module

[0245] To further explore the effectiveness of the PCL module in feature learning, this invention employs t-SNE technology to visualize and map the learned features. Taking tasks H3 and H9 as examples, the feature mapping results are as follows: Figure 6 As shown, by comparing the feature mapping results of different methods, the influence of each module on the feature space distribution can be clearly observed.

[0246] In traditional Data Acquisition (DA) methods, due to the lack of classification aids and sample-aware weights, feature clustering performance is poor under the influence of outliers. Specifically, significant overlap exists between different category clusters, and the classification error rate is high, indicating a deficiency in the ability of traditional DA methods to partition the feature space. In contrast, methods such as... Figure 6 As shown in (b) and (e), although the model optimized by the SR module can resist the interference of outliers and effectively distinguish different clusters, overlap and dispersion still exist within each cluster. This indicates that while the SR module improves the robustness of features to some extent, there is still room for improvement in the clustering performance of features. Figure 6As shown in (c) and (f), the distribution of the feature space is significantly improved after the PCL module is introduced. The division between categories becomes clearer, the boundaries between different category clusters are more defined, and the sample points within a category are more concentrated. This phenomenon indicates that the PCL module effectively clusters samples of the same category in the feature space while separating samples of different categories, thus significantly improving the separability of features. Furthermore, Figure 6 This further validates the robustness of the SR module in feature clustering. Even under the influence of outliers, the clustering results of the SR module maintain good stability, which is consistent with... Figure 5 The results are consistent with those in the previous study, further demonstrating the effectiveness of the SR module in outlier suppression.

[0247] In summary, the PCL module, through a contrastive learning mechanism, further optimizes the distribution of the feature space and enhances the clustering performance of features, thus providing a more effective feature representation for classification problems in cross-domain tasks. Meanwhile, the robustness of the SR module has been verified in multiple experimental results, demonstrating its significant advantage in dealing with outlier interference.

[0248] 3. Improve the robustness and adaptability of the model:

[0249] In training deep learning models for bearing signals, robustness becomes a crucial consideration due to the limited dataset size and complex operating conditions. Box plot comparisons can observe the performance trends of different model combinations, thereby identifying the optimal and stable model configuration. To accurately select training models, this invention evaluates models by plotting diagnostic accuracy box plots, such as... Figure 7 As shown.

[0250] Taking tasks C6 and H11 as examples, the SR-w / o PCL model exhibits higher classification accuracy and a shorter bin compared to w / o SR-w / o PCL and other DA-w / o PCL models. This indicates that the introduction of the SR module significantly enhances the model's robustness to bearing data input, enabling it to maintain high classification accuracy even with complex data distributions. When the PCL module is introduced, the model's accuracy distribution becomes more concentrated, specifically manifested in a further reduction in bin length and a significant increase in the median. This demonstrates that the PCL module significantly enhances the model's consistency and stability in feature extraction. Through a contrastive learning mechanism, the PCL module further optimizes the distribution of the feature space, causing samples of the same class to cluster more tightly in the feature space, while samples of different classes are separated more effectively. This optimization of feature distribution is directly reflected in the model's classification accuracy, making the model's performance more stable under different operating conditions. It is particularly noteworthy that without the support of the SR module, using the PCL module alone does not improve accuracy. This result is consistent with the data in Tables 4 and 5, showing that the PCL module struggles to effectively improve the classification performance of bearing signals when data distributions differ significantly. Therefore, combining the SR module with the PCL module is crucial for improving the overall performance of the model, especially in enhancing its adaptability and stability to bearing signals under different operating conditions.

[0251] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included 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 method for cross-condition bearing fault diagnosis using a minimum entropy optimized prototype comparison network, characterized in that, Includes the following steps: Step 1: Collect bearing vibration signals under multiple operating conditions to construct source domain datasets and target domain datasets; Step 2: Extract depth feature representation of bearing vibration signal using convolutional layers and transformer layers; Step 3: Establish auxiliary information between global features and prediction information so that the model can capture the non-linear relationship between global features and prediction information; Step 4: Calculate the sample entropy, assess the uncertainty of sample classification, and weight the results according to the difficulty of sample classification; Step 5: Construct a four-layer fully connected domain discriminator, and combine auxiliary information and sample entropy to optimize the feature distribution alignment between the source and target domains through a domain adversarial loss function; Step 6: Use the feature vectors of the bearing data to obtain the prototype of the data by learning vector quantization; Step 7: Perform intra-domain prototype contrast learning, compare the similarity components between the feature vector and the prototype, and calculate the intra-domain prototype contrast loss; Step 8: Perform cross-domain instance-prototype learning, compare the similarity components between the source domain feature vector and the target domain prototype, and calculate the cross-domain instance-prototype loss; Step 9: Design a weighted total loss function that includes domain adversarial loss, intra-domain prototype contrast loss, cross-domain instance-prototype loss, and classification loss to optimize the model; Step 10: Use the trained model to classify and diagnose the bearing fault status in the target domain.

2. The method for cross-condition bearing fault diagnosis using a minimum entropy optimized prototype comparison network according to claim 1, characterized in that, The specific method for extracting the depth feature representation of the bearing vibration signal using convolutional layers and transformer layers in step S2 includes: (1) The first CNN is performed using 16 convolutional kernels of size 3; (2) A second CNN is performed using 32 convolutional kernels of size 3; (3) A third CNN is performed using 64 convolutional kernels of size 3; (4) Use Transformer layers to analyze the contextual relationships between each feature extracted by the CNN to form global features f; (5) After passing through two fully connected layers (FC), a deep feature representation is generated; (6) After passing through the Softmax layer, the prediction information g is obtained.

3. The method for cross-condition bearing fault diagnosis using the minimum entropy optimized prototype comparison network according to claim 2, characterized in that, The specific method for establishing auxiliary information between global features and predicted information in step S3 to capture the nonlinear relationship between global features and predicted information includes: (1) Combine global features f and prediction information g into auxiliary information. (2) The combined representation is processed using an auxiliary domain discriminator consisting of four fully connected layers, with the number of neurons in the fully connected layers being 256, 128, 64, and 1, respectively. (3) The process of capturing the nonlinear relationship between features and prediction information through the auxiliary domain discriminator is expressed as follows: Among them, D Auxiliary This indicates an auxiliary discriminator.

4. The method for cross-condition bearing fault diagnosis using the minimum entropy optimized prototype comparison network according to claim 3, characterized in that, The specific methods for calculating sample entropy, assessing the uncertainty of sample classification, and weighting the results according to the difficulty of sample classification in step S4 include: ω(Entropy(g))=1+e -Entropy(g) Where N represents the number of classes, g n Let represent the probability that a sample is predicted to be in the nth class; Entropy(g) represents the sample entropy; ω(Entropy(g)) represents the weights applied to the sample; e -Entropy(g) To measure the difficulty of classifying a sample, the greater the sample entropy, the smaller its weight should be.

5. The method for cross-condition bearing fault diagnosis using the minimum entropy optimized prototype comparison network according to claim 4, characterized in that, In step S5, a four-layer fully connected domain discriminator is constructed. The specific method for optimizing the feature distribution alignment between the source and target domains using a domain adversarial loss function, combining auxiliary information and sample entropy, includes: Domain adversarial loss can be expressed as: Among them, D s D represents the source domain dataset; t Let i represent the target domain dataset; s represent the source domain index; j represent the target domain index; and t represent the target domain. The feature distributions of the source domain and the target domain are aligned by minimizing the weighted domain adversarial loss function.

6. The method for cross-condition bearing fault diagnosis using the minimum entropy optimized prototype comparison network according to claim 2, characterized in that, The specific method for obtaining the prototype of the data using the learning vector quantization method described in step S6 includes: Randomly select k feature vectors from SD as the initial prototype vectors {p1, p2, ..., pk}. k }, whose tags serve as the corresponding preset tags. Where k represents the number of classes; Randomly select a sample from SD. Calculate the distance between the sample and each prototype: d ij =||f i s -p j ||2,j∈[1,k] Find the relationship with f i s The latest prototype p j* ;if With p j* The corresponding tag t j* If they are the same, then bring p closer. j* Distance from the sample; otherwise, move p away. j* Distance from the sample: Where η∈(0,1) are hyperparameters; the final SD prototype is obtained after 1000 iterations. and TD prototype 7. The method for cross-condition bearing fault diagnosis using the minimum entropy optimized prototype comparison network according to claim 6, characterized in that, The specific method for performing intra-domain prototype contrastive learning, comparing the similarity components between the feature vector and the prototype, and calculating the intra-domain prototype contrastive loss as described in step S7 includes: To accurately capture subtle differences in bearing fault signals, f is calculated. i s and The similarity components between them are: Where φ is the temperature value; Intra-domain prototype contrast loss can be expressed as: Where, index s and index t Indicates the index of the input data; L CE (·) represents the cross-entropy loss; By minimizing the intra-domain prototype contrast loss, each prototype represents the central trend and typical features of a class of data; by bringing samples of the same class closer together and pushing samples of different classes further apart, the model is able to learn more discriminative features.

8. The method for cross-condition bearing fault diagnosis using the minimum entropy optimized prototype comparison network according to claim 7, characterized in that, The specific method for performing cross-domain instance-prototype learning, comparing the similarity components between the source domain feature vector and the target domain prototype, and calculating the cross-domain instance-prototype loss as described in step S8 includes: Given the eigenvectors f of the source domain i s and the prototype of the target domain Calculate similarity components: To identify the matching relationship between the prototype and the features of the foreign bearing, Perform entropy minimization, obtaining E(V) based on the difficulty of sample classification. i s→t Finally, the cross-domain instance-prototype loss is obtained as follows: in, Entropy representing the similarity component between source domain features and target domain prototypes; Entropy representing the similarity component between the target domain features and the source domain prototype; D tu This indicates that the target field label is not visible; By minimizing the cross-domain instance-prototype loss, the model can avoid the outlier effects of bearing data.

9. The method for cross-condition bearing fault diagnosis using the minimum entropy optimized prototype comparison network according to claim 8, characterized in that, The specific methods for optimizing the model using a weighted total loss function comprising domain adversarial loss, intra-domain prototype contrast loss, cross-domain instance-prototype loss, and classification loss as described in step S9 include: Classification loss, adversarial loss, and PCL loss are used to aid in the model training process, and parameters are used to control the impact of the global loss; the losses for each part are as follows: (1) Sample reweighting module: Where λ1 and λ2 are hyperparameters; L represents the predicted label, y represents the true label; pre-total This represents the total loss during the pre-training phase; the training process minimizes L. pre-total Domain adversarial loss L Auxiliary Features f acting on the source and target domains s,t and prediction information g s ,t Classification loss L CE It only applies to the source domain; and This represents the network parameters before and after the pre-training phase update, and the domain discriminator Θ. D Using GRL for gradient inversion, the optimization objective is: (2) Prototype Comparison Learning Module: Where λ3 and λ4 are hyperparameters; L train-total This represents the total loss during the training phase; the training process minimizes L. train-total Prototype contrast loss L In and L Cross Features f acting on the source and target domains s,t Classification loss L CE It only applies to the source domain; the optimization objective is: in, This represents the network parameters updated during the training phase.

Images