Cross-working-condition fault diagnosis method based on uncertainty estimation denoising

Abstract

The invention belongs to the technical field of industrial system fault diagnosis, and particularly relates to a cross-working-condition fault diagnosis method based on uncertainty estimation denoising. The method comprises the following specific steps: 1, respectively obtaining labeled source domain data and unlabeled target domain data, constructing a feature extractor based on a one-dimensional convolutional neural network and a Transform layer, and carrying out feature extraction on a fault signal; 2, performing self-supervised learning on the target domain data to obtain a pseudo tag; 3, constructing a pseudo-label redefinition selection module, dynamically selecting high-confidence pseudo-labels, and reducing noise pseudo-labels; 4, constructing a class prototype alignment strategy based on Dirichlet uncertainty estimation, and achieving the accurate recognition of a target domain sample; 5, loading the source domain data and the high-confidence target domain data to the dual-classifier adversarial network model for training to obtain a trained neural network model; and 6, realizing fault diagnosis of the rotating mechanical system by using the trained neural network model.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial system fault diagnosis, and particularly relates to a cross-condition fault diagnosis method based on uncertainty estimation and denoising. Background Technique

[0002] With the continuous deepening of industrial development, the wave of industrial intelligence is sweeping in. As an indispensable component in industrial applications, rotating machinery usually operates in non-stationary and harsh working environments and is prone to various faults. Therefore, the health monitoring and fault diagnosis of rotating machinery have become a research hotspot in industrial intelligent operation and maintenance.

[0003] In recent years, deep learning technology has dominated the field of rotating machinery fault diagnosis due to its excellent feature extraction ability. However, in actual industrial applications, rotating machinery operates for a long time under various conditions (such as speed changes, load changes, etc.), resulting in significant differences in data under different conditions, which seriously hinders the application of deep learning models in actual industries. To address these challenges and improve diagnostic performance, unsupervised domain adaptation fault diagnosis methods have emerged and received extensive attention.

[0004] Unsupervised domain adaptation technology for cross-domain fault diagnosis is mainly based on methods such as statistical metrics, domain adversarial, and pseudo-label learning. Although traditional adaptation methods have shown good performance, these methods mainly measure the similarity between samples through differences or extract domain-invariant features through adversarial learning. However, this cannot guarantee the accuracy and diversity of classifying target samples, because fuzzy target samples may be misclassified due to unreasonable decision boundaries, resulting in inaccurate alignment of class distributions in subsequent learning processes. The main reason for this problem is that previous domain adaptation methods did not fully consider the accuracy of target samples and the impact of label noise in pseudo-labels on the model generalization performance. Summary of the Invention

[0005] In view of the above problems, the present invention proposes a cross-condition fault diagnosis method based on uncertainty estimation and denoising to solve the problem of poor generalization caused by insufficient consideration of the accuracy of target domain samples and noisy pseudo-labels in previous methods.

[0006] The technical solution of the present invention is as follows:

[0007] A cross-condition fault diagnosis method based on uncertainty estimation and denoising, comprising the following steps:

[0008] Step 1: Obtain source domain data and target domain data, construct a feature extractor based on a one-dimensional convolutional neural network and a Transformer layer, and load the source domain and target domain data into the feature extractor to extract features from the fault signal;

[0009] Step 2: Perform self-supervised clustering on the target domain data to obtain pseudo-labels for the target domain samples;

[0010] Step 3: Construct a pseudo-label redefinition selection module to dynamically select high-confidence pseudo-labels and reduce noise data and labels;

[0011] Step 4: Construct a class prototype alignment strategy based on Dirichlet uncertainty estimation to achieve accurate identification of target domain samples;

[0012] Step 5: Load the source domain data and the high-confidence target domain data into a dual-classifier adversarial network model for training to obtain a trained neural network model;

[0013] Step 6: Use the trained neural network model to diagnose mechanical faults.

[0014] Among them, Step 1 is specifically implemented according to the following steps:

[0015] Obtain the source domain data and the target domain data , where represents the th data in the source domain, is its corresponding label, represents the th data in the target domain, and represent the number of samples in the source domain and the target domain respectively; construct a feature extractor based on a one-dimensional convolutional neural network and a Transformer layer , and input the source domain and target domain data into the feature extractor simultaneously to effectively capture the global and local features of the fault signal; after the feature extractor, two classifiers and are connected in parallel for adversarial optimization training.

[0016] Step 2 is specifically implemented according to the following steps:

[0017] In the initial stage of training, the model quickly fits the source domain data. Subsequently, the two softmax function outputs of the dual-classifier adversarial network are used to weight the target domain data to obtain the centroid of each class, the distance between the target samples and the centroid is quantified through the cosine distance function, and the corresponding pseudo-labels are obtained through the nearest centroid strategy.

[0018] Step 3 is specifically implemented according to the following steps:

[0019] Construct a pseudo-label redefinition selection module to redefine the pseudo-labels obtained in Step 2 by aggregating neighbor knowledge, and update the temporary pseudo-labels of all target samples during the iteration process. Through iterative updates, perceive the density of surrounding pseudo-labels and obtain a more accurate uncertainty estimate, and derive the confidence score for each target domain sample. In the initial stage of training, simple samples in the target domain are more likely to be assigned high confidence. Therefore, by setting a dynamic threshold , dynamically select high-confidence target domain samples from easy to difficult, specifically expressed as:

[0020]

[0021] wherein, represents the low threshold, represents the high threshold, represents the current iteration number, represents the total number of iterations.

[0022] Step 4 is specifically implemented according to the following steps:

[0023] Construct a class prototype alignment strategy based on Dirichlet uncertainty estimation, model the class probability distribution using the variational Dirichlet distribution. After obtaining the Dirichlet distribution, introduce subjective logic to assign a belief mass to each class, and assign an uncertainty estimate value , and further introduce this uncertainty into class prototype alignment. Specifically, the class prototype of each training batch is calculated as follows:

[0024]

[0025] wherein, is the domain index, represents the source domain, represents the target domain, represents the th sample from the source domain or the target domain, represents the feature representation calculated after the th sample from the source domain or the target domain passes through the feature extractor , is the indicator function, is the uncertainty value of the th sample, is the uncertainty threshold, represents the true label in the source domain and the reliable pseudo-labels obtained in the target domain through Step 3, represents if the uncertainty value of the th sample in the th class is less than Its value is 1 if so, otherwise 0;

[0026] Through a dynamic dictionary to store the class prototypes from the previous batch and update them in a moving average manner, specifically expressed as:

[0027]

[0028] In the formula, is a trade-off parameter; after obtaining the class prototypes of the source domain and the target domain, calculate the alignment loss of samples in the execution domain to the class prototypes and the alignment loss of class prototypes between domains , specifically expressed as:

[0029]

[0030]

[0031] In the formula, represents the number of samples in the th class in the source domain or the target domain, represents the class prototype of the th class in the source domain, represents the class prototype of the th class in the target domain.

[0032] The alignment loss of class prototypes based on Dirichlet uncertainty estimation combines the alignment loss of source domain samples to source domain class prototypes , the alignment loss of target domain samples to target domain class prototypes and the alignment loss of class prototypes between domains , specifically expressed as follows:

[0033] .

[0034] Step 5 is specifically implemented according to the following steps:

[0035] Load the source domain data and the denoised target domain data into the dual-classifier adversarial network model for adversarial training, specifically:

[0036] Step 5.1 Use the labeled source domain data and the reliable samples and pseudo-labels obtained in the target domain to train the entire network:

[0037]

[0038] In the formula, is the trade-off parameter, , and respectively represent the feature extractor and the two classifiers , parameters, represents the source domain sample set, represents the label set corresponding to the source domain samples, represents the high-confidence sample set of the target domain, represents the pseudo-label set corresponding to the high-confidence samples in the target domain. Cross-entropy loss is used for supervised training, and the classification loss of the source domain samples is calculated as follows:

[0039]

[0040] To utilize the unlabeled target domain samples, a self-supervised learning mechanism is used to strengthen the dual-classification adversarial learning. The weighted cross-entropy loss of self-supervised learning is calculated as follows:

[0041]

[0042] In the formula, is the weighted cross-entropy loss function, and its formula is as follows:

[0043]

[0044]

[0045] In the formula, represents the standard information entropy, represents the softmax function, represents the th reliable pseudo-label of the sample obtained in the target domain through step 3.

[0046] Step 5.2 Fix the feature extractor , and at the same time update the two classifiers and , to minimize the classification loss of the two classifiers in the source domain, and at the same time maximize the discrimination of the target samples, which is expressed as:

[0047]

[0048] In the formula, represents the difference loss, which is generally measured by norm, represents the discrimination difference of the classifier for the target sample ;

[0049] Step 5.3 Combine the class prototype alignment loss based on the Dirichlet uncertainty estimation between the two domains , thus obtaining the final optimization objective, which is expressed as:

[0050]

[0051] wherein is a trade-off parameter.

[0052] The beneficial effects of the present invention are as follows:

[0053] (1) The present invention proposes a cross-condition fault diagnosis method based on uncertainty estimation denoising. By constructing a feature extractor based on a one-dimensional convolutional neural network and a Transformer layer, the global and local features of the fault signal can be effectively extracted;

[0054] (2) The present invention proposes a pseudo-label redefinition selection module based on uncertainty. This strategy selects high-confidence samples as clean samples, reduces the impact of noise pseudo-labels on the model generalization performance, and ensures the separability of features.

[0055] (3) The present invention proposes a class prototype alignment strategy based on Dirichlet uncertainty estimation. By performing in-domain sample to class prototype alignment and inter-domain class prototype alignment, the fault recognition accuracy and generalization ability of the target domain samples are effectively improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 Overall framework schematic diagram, i.e., the abstract drawing;

[0057] Figure 2 Algorithm schematic diagram;

[0058] Figure 3 Pseudo-label accuracy curve of the cross-domain fault diagnosis task ;

[0059] Figure 4 Pseudo-label accuracy curve of the cross-domain fault diagnosis task ;

[0060] Figure 5 Pseudo-label accuracy curve of the cross-domain fault diagnosis task , and Feature visualization diagrams of different methods under

[0061] Figure 6 Pseudo-label accuracy curve of the cross-domain fault diagnosis task Comparison of feature visualization diagrams between DASAN and the present invention under

[0062] Figure 7 Pseudo-label accuracy curve of the cross-domain fault diagnosis task Feature visualization diagram of the present invention under

[0063] Figure 8 Pseudo-label accuracy curve of the cross-domain fault diagnosis task Visualization diagram of the features of the present invention

[0064] Figure 9 Cross-domain fault diagnosis task Visualization diagram of the features of the present invention Detailed implementation manners

[0065] The cross-condition fault diagnosis method based on uncertainty estimation denoising proposed by the present invention consists of a feature extraction module and a pseudo-label redefinition and selection module. The specific flow chart is as Figure 1 shown. The function of the feature extractor is to extract fault features from the original signals of the source domain and the target domain. After the feature extractor module, a prototype alignment term is introduced to further align the distributions and improve the accuracy of the target samples. The pseudo-label redefinition module aims to select high-confidence target domain samples and dynamically reduce the noisy labels, and finally realizes the accurate identification of fault types under variable working conditions. To enable those skilled in the art to better understand the solution of the present invention, the exemplary embodiments or examples of the present invention will be described below with reference to the accompanying drawings.

[0066] 1. Model establishment

[0067] 1.1 Feature extractor

[0068] The feature extractor aims to extract high-level feature representations of mechanical health conditions from the data of the source domain and the target domain. Convolutional neural networks are widely used as feature extractors for mechanical fault diagnosis because they have significant advantages in capturing local features. However, due to their lack of the ability to capture global features and the obvious context relationships in fault signals, it is necessary to fully consider the global and local features of fault signals. The present invention designs a feature extractor based on one-dimensional convolutional neural networks and Transformer to take into account both the global and local features of fault signals and improve the diagnostic performance of the model under variable working conditions. Figure 1 shows the schematic diagram of the algorithm proposed by the present invention. The first convolutional layer receives the one-dimensional original vibration signal as the model input. After the fourth convolutional module, two Transformer layers are embedded to better capture global features. The feature extraction process of the feature extractor can be expressed as:

[0069] ,

[0070] where represents the extracted features, is the batch input matrix.

[0071] 1.2 Self-supervised learning of target domain samples

[0072] In the initial stage of training, the model quickly fits the source domain data. Subsequently, the two softmax functions of the dual-classifier adversarial network weight the target domain data to obtain the centroid of each class , specifically expressed as:

[0073]

[0074] In the formula, represents the feature extractor, and F represents the classifier, where represents the th element corresponding in the softmax output. The target sample quantifies the distance between itself and the clustering center of each fault mode using the cosine distance function and obtains the corresponding pseudo-label through the nearest centroid strategy, specifically expressed as:

[0075] .

[0076] 1.3 Pseudo-label Redefinition Selection Module Based on Uncertainty

[0077] Construct a pseudo-label redefinition selection module to redefine the pseudo-labels obtained in step 2 by aggregating neighbor knowledge, specifically expressed as:

[0078]

[0079] In the formula, represents the one-dimensional vector of the temporary pseudo-label . The temporary pseudo-label is initialized by the pseudo-label obtained through self-supervised learning, and the temporary pseudo-labels of all target samples are updated during the iteration, specifically expressed as:

[0080]

[0081] By updating to iteratively execute estimation to perceive the density of surrounding pseudo-labels and obtain a more accurate uncertainty estimate. After the uncertainty estimate is completed, the confidence score is obtained, specifically expressed as:

[0082]

[0083] For the sample , only its nearest neighbor sharing the same label as will contribute to .

[0084] In the initial stage of training, simple samples in the target domain are more likely to be assigned high confidence and learned by the model. Therefore, a dynamic threshold is set , and high-confidence target domain samples are selected from easy to difficult, specifically expressed as:

[0085]

[0086] In the formula, represents the low threshold, represents the high threshold, represents the current iteration number, represents the total number of iterations.

[0087] To utilize unlabeled samples, a self-supervised learning mechanism is combined to strengthen the dual-class adversarial learning. The weighted classification loss of self-supervised learning is expressed as follows:

[0088]

[0089] In the formula, is the weighted cross-entropy loss function, and its calculation formula is as follows:

[0090] ,

[0091]

[0092] In the formula, represents the standard information entropy, represents the softmax function.

[0093] 1.4 Class Prototype Alignment Strategy Based on Dirichlet Uncertainty Estimation

[0094] Construct a class prototype alignment strategy based on Dirichlet uncertainty estimation, and use the variational Dirichlet distribution to model the class probability distribution. Specifically, for the classification problem of class, assume that the estimated classification probability is , where . The Dirichlet distribution is parameterized by its concentration parameters , and the probability density function of this distribution is specifically expressed as:

[0095]

[0096] In the formula, is the multivariate polynomial beta function of dimension is the simplex of dimension

[0097]

[0098] After obtaining the Dirichlet distribution, to quantify its uncertainty, subjective logic is introduced to associate the parameters of the Dirichlet distribution with belief mass and uncertainty. For a multi-class classification problem, subjective logic assigns belief mass to each class based on the Dirichlet distribution and assigns an overall uncertainty mass to all classes , that is .

[0099] Given the Dirichlet distribution, subjective logic, and evidence classification, the evidence is associated with the concentration parameter of the Dirichlet distribution. Evidence refers to the metrics obtained from the input to support classification, and it is closely related to the expected concentration parameter of the Dirichlet distribution. Specifically, let represent the evidence for the -th class, be the -th concentration parameter of the Dirichlet distribution, the relationship between and is , then the belief and uncertainty

[0100]

[0101] are calculated as follows: where denotes the Dirichlet strength. Since samples with low uncertainty have better generalization ability, this uncertainty is introduced into class prototype alignment. Therefore, the present invention proposes an uncertainty class prototype alignment strategy based on Dirichlet uncertainty estimation for aligning class-level features. Specifically, the class prototype of each training batch

[0102]

[0103] is calculated as follows: where is the domain index, denotes the source domain, denotes the target domain, denotes the -th sample from the source domain or the target domain, denotes the feature representation calculated after the -th sample from the source domain or the target domain passes through the feature extractor is the indicator function, is the -th sample uncertainty value, is the uncertainty threshold, denotes the true label in the source domain and the reliable pseudo-label obtained in the target domain through step 3. denotes that if the th sample uncertainty value in the th class is less than then its value is 1, otherwise 0;

[0104] By constructing a dynamic dictionary to store the class prototypes from the previous batch and update them in a moving average manner, which is specifically expressed as:

[0105]

[0106] where is a trade-off parameter. By calculating the class prototypes in both domains and performing the in-domain sample to class prototype alignment loss and the inter-domain class prototype alignment loss, to improve the model's recognition ability for target domain samples, the specific loss function is expressed as:

[0107]

[0108]

[0109] where denotes the number of samples in the th class in the source domain or target domain, denotes the class prototype of the th class in the source domain, denotes the class prototype of the th class in the target domain.

[0110] The class prototype alignment loss based on Dirichlet uncertainty estimation combines the in-source domain sample to in-source domain class prototype alignment loss , the in-target domain sample to in-target domain class prototype alignment loss and the inter-domain class prototype alignment loss , which is specifically expressed as follows:

[0111]

[0112] 1.5 Optimization Process

[0113] Load the source domain data and the denoised target domain data into the dual-classifier adversarial network model for adversarial training, specifically:

[0114] Use the labeled source domain data and the high-confidence samples obtained in the target domain to train the entire network:

[0115]

[0116] In the formula, is a trade-off parameter, , and respectively represent the parameters of the feature extractor and two classifiers , . represents the source domain sample set, represents the label set corresponding to the source domain samples, represents the high-confidence sample set of the target domain, represents the pseudo-label set corresponding to the high-confidence samples of the target domain. The cross-entropy loss is used for supervised training, and the classification loss of the source domain samples is calculated as follows:

[0117]

[0118] Fix the feature extractor , and at the same time update the two classifiers and to minimize the classification loss of the two classifiers in the source domain and at the same time maximize the judgment on the target samples, which is expressed as:

[0119]

[0120] In the formula, represents the difference loss, which is generally measured by the norm, represents the discriminant difference of the classifier for the target sample ;

[0121] Combine the class prototype alignment loss based on the Dirichlet uncertainty estimation between the two domains, so as to obtain the final optimization objective:

[0122]

[0123] In the formula, is a trade-off parameter.

[0124] 2. Experimental verification

[0125] To illustrate the superiority of the method proposed in the present invention, specific experiments are carried out for verification below. During the experimental verification process, seven models, namely, CORrelation ALignment (CORAL), Domain Adversarial Neural Network (DANN), Deep Convolutional Transfer Learning Network (DCTLN), Joint Maximum Mean Discrepancy (JAN), Multi-Kernel Maximum Mean Discrepancy (MKMMD), Maximum Classifier Discrepancy (MCD), and Deep Adversarial Subdomain Adaptation Network (DASAN), are selected to be compared with the present invention to obtain more persuasive verification results. To ensure a fair comparison, the feature extractors, health status classifiers, and optimizers of these models are consistent with those in the present invention. The network parameters are shown in Table 1 for details.

[0126] Table 1

[0127]

[0128] 2.1 Data Description

[0129] The present invention uses the bearing dataset of Paderborn University (PU) in Germany for experimental verification. The dataset contains artificial damages and real damages caused by accelerated life. The test bench consists of a motor and a bearing housing with four 6203-type bearings. By controlling three parameters, namely, the rotational speed, the load torque, and the radial force of the bearing, the acquisition of bearing vibration signals under four working conditions is realized. The signal sampling frequency is 64 kHz, and the working condition parameters are shown in Table 2. In this embodiment, is used to represent that the transfer task migrates from working condition 0 to working condition 1.

[0130] Table 2

[0131]

[0132] The present invention uses bearings with real damages caused by accelerated life tests to study transfer learning tasks under different conditions to better conform to the actual industrial operation scenarios. The real damage types are divided into three types: electrical machining damage, electrical engraving damage, and drilling damage. Combining the bearing type, the fault occurrence location, the severity of the fault, and the damage type, twelve fault types are combined and formed. The damage characteristics are summarized in Table 3. Plus a healthy state, finally, fault diagnosis of thirteen categories is realized.

[0133] Table 3

[0134]

[0135] First, the method proposed in the present invention is compared with several domain adaptation methods, including DANN, MKMMD, CORAL, DCTLN, MCD, and DASAN. Each model is evaluated through 10 independent trials to obtain reliable results. The experimental results are shown in Table 4. The following arguments can be drawn from these results: 1) The model proposed in the present invention achieves an average diagnostic accuracy of 93.19%, which is 6.01% higher than that of MCD. This indicates that the method proposed in the present invention effectively alleviates the distribution differences between working conditions and improves the diagnostic accuracy.

[0136] Table 4

[0137]

[0138] To verify the effectiveness of the proposed method for redefining and dynamically selecting pseudo-labels based on uncertainty, the method proposed in the present invention is compared with self-supervised clustering and the prediction results based on the softmax function in the cross-domain fault diagnosis task and The experimental results are as shown in Figure 3 and Figure 4 It can be seen from the pseudo-label selection accuracy graph that the algorithm proposed in the present invention reaches the maximum accuracy faster and exhibits excellent stability. This result provides strong evidence for the denoising effect of the present invention, and it more effectively promotes cross-domain alignment.

[0139] To further prove the effectiveness of the model proposed in the present invention and provide an intuitive understanding, the feature distributions learned by the model are visualized using the t-SNE technique in the cross-domain fault diagnosis task , and The visualization of the feature distributions of DANN, MCD, DASAN, and the method of the present invention is shown. It can be seen that, compared with other methods, the present invention achieves effective alignment at the source domain and target misclassification levels, and shows fewer misclassifications, demonstrating better diagnostic accuracy and generalization performance. Figure 5 The enlarged visualization graphs of the present invention and DASAN are shown. The results show that the class prototype alignment strategy based on Dirichlet uncertainty estimation of the present invention selects samples with low uncertainty and representative class features as class prototypes. Compared with DASAN, the method of the present invention can better align target samples with source samples at the class level and construct clearer class boundaries, while DASAN shows greater aliasing. Figure 6 The enlarged visualization graphs of the present invention and DASAN are shown. The results show that the class prototype alignment strategy based on Dirichlet uncertainty estimation of the present invention selects samples with low uncertainty and representative class features as class prototypes. Compared with DASAN, the method of the present invention can better align target samples with source samples at the class level and construct clearer class boundaries, while DASAN shows greater aliasing. Figure 7 , Figure 8 and Figure 9It is shown that the method proposed by the present invention better aligns the distribution of target samples with that of source samples at the class level by reducing the noise in the target domain, resulting in clearer classification boundaries and higher diagnostic accuracy.

[0140] As described above, the foregoing are only preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. It should be understood that the present invention is not limited to the implementation solutions described herein, and the purpose of describing these implementation solutions is to help those skilled in the art practice the present invention. Any person skilled in the art can easily make further improvements and refinements without departing from the spirit and scope of the present invention. Therefore, the present invention is only limited by the content and scope of the claims of the present invention, and it is intended to cover all alternative solutions and equivalent solutions included in the spirit and scope of the present invention defined by the appended claims.

Claims

1. A cross-operating condition fault diagnosis method based on uncertainty estimation and denoising, characterized in that: The specific steps are as follows: Step 1: Obtain source domain data and target domain data respectively, build a feature extractor based on a one-dimensional convolutional neural network and a Transformer layer, load the source domain and target domain data into the feature extractor to extract features of the fault signal; Step 2: Perform self-supervised learning clustering on the target domain data to obtain pseudo labels of the target domain samples; Step 3: Build a pseudo-label redefinition selection module to dynamically select high-confidence pseudo-labels and reduce noisy data and labels; Step 4: Construct a class prototype alignment strategy based on Dirichlet uncertainty estimation to achieve accurate recognition of target domain samples; Step 5: Load the source domain data and high-confidence target domain data into the dual classifier adversarial network model for training to obtain a trained neural network model; Step 6: Use the trained neural network model to implement fault diagnosis of rotating machinery systems.

2. The cross-operating condition fault diagnosis method based on uncertainty estimation and denoising according to claim 1 is characterized in that: In step 1, the source domain data is obtained separately and target domain data ,in, Represented as the first data, is its corresponding label, Represented as the first data, and Represents the number of samples in the source domain and the target domain respectively; constructs a feature extractor based on a one-dimensional convolutional neural network and a Transformer layer , the source domain and target domain data are simultaneously input into the feature extractor to effectively capture the global and local features of the fault signal; after the feature extractor, two classifiers are connected in parallel and Conduct adversarial optimization training.

3. The cross-operating condition fault diagnosis method based on uncertainty estimation and denoising according to claim 1 is characterized in that: In step 2, at the initial stage of training, the model quickly fits the source domain data; then, the target domain data is weighted using the two softmax function outputs of the dual classifier adversarial network to obtain the centroid of each class, and the distance between the target sample and the centroid is quantified by the cosine distance function and the corresponding pseudo label is obtained by the nearest centroid strategy.

4. The cross-operating condition fault diagnosis method based on uncertainty estimation and denoising according to claim 1 is characterized in that: In step 3, a pseudo-label redefinition selection module is constructed to redefine the pseudo-labels obtained in step 2 by aggregating neighbor knowledge, and to update the temporary pseudo-labels of all target samples in the iterative process. The density of the surrounding pseudo-labels is perceived through the update iteration and a more accurate uncertainty estimate is obtained to obtain the confidence score of each target domain sample. In the early stage of training, simple samples in the target domain are more likely to be learned by the model with high confidence, so a dynamic threshold is set. , dynamically select high-confidence target domain samples from easy to difficult, specifically expressed as: ; In the formula, Represents the low threshold, Represents the high threshold, Represents the current iteration number, Represented as the total number of iterations.

5. The cross-operating condition fault diagnosis method based on uncertainty estimation and denoising according to claim 1 is characterized in that: In step 4, a class prototype alignment strategy based on Dirichlet uncertainty estimation is constructed, and the class probability distribution is modeled using variational Dirichlet distribution. After obtaining the Dirichlet distribution, subjective logic and evidence classification are introduced to assign a belief quality to each class. , and assign an uncertainty estimate , and further introduce this uncertainty into the class prototype alignment. Specifically, each training batch The class prototype is calculated as follows: ; In the formula, is the domain indicator, represents the source domain, represents the target domain, Indicates samples from the source domain or the target domain, Indicates Samples from the source domain or target domain are passed through the feature extractor The feature representation calculated later is is the indicator function, It is The sample uncertainty value, is the uncertainty threshold, represents the true label in the source domain and the reliable pseudo label obtained in the target domain through step 3, If The first in the category The sample uncertainty value is less than If yes, its value is 1, otherwise it is 0; Through a dynamic dictionary To store the class prototypes from the previous batch and update them in a moving average manner, specifically expressed as: ; In the formula, Is a trade-off parameter; after obtaining the class prototypes of the source domain and the target domain, the loss of alignment from the sample in the execution domain to the class prototype is calculated. and inter-domain class prototype alignment loss , specifically expressed as: ; ; In the formula, Indicates the source or target domain. The number of class samples, Indicates the source domain The class prototype of the class, Indicates the target domain The class prototype of the class; Class Prototype Alignment Loss Based on Dirichlet Uncertainty Estimation Combining source domain samples to source domain class prototype alignment loss , the target domain sample to the target domain class prototype alignment loss and inter-domain class prototype alignment loss , specifically expressed as follows: 。 6. The cross-operating condition fault diagnosis method based on uncertainty estimation and denoising according to claim 1 is characterized in that: In step 5, the source domain data and the denoised target domain data are loaded into the dual classifier adversarial network model for adversarial training, specifically: Step 5.1 trains the entire network using labeled source domain samples and high confidence samples in the target domain: ; In the formula, is a trade-off parameter, , and They represent feature extractors And two classifiers , Parameters, represents the source domain sample set, represents the label set corresponding to the source domain samples, Represents a high confidence sample set of the target domain, Represents the pseudo label set corresponding to the high-confidence samples in the target domain, and the cross entropy loss Classification loss of source domain samples for supervised training The calculation is as follows: ; In order to utilize unlabeled target domain samples, a self-supervised learning mechanism is used to strengthen dual-classification adversarial learning. The weighted cross entropy loss of self-supervised learning The calculation is as follows: ; In the formula, is the weighted cross entropy loss function, and its formula is as follows: ; ; In the formula, represents the standard information entropy, represents the softmax function, represents the first Reliable pseudo-labeling of samples; Step 5.2 Fix the feature extractor , update both classifiers simultaneously and , in order to minimize the classification loss of the two classifiers in the source domain, while maximizing the judgment of the target sample is expressed as: ; In the formula, Indicates the difference loss, generally used The norm is used to measure, Represents the classifier for the target sample The difference in discrimination; Step 5.3 Combine the class prototype alignment loss based on Dirichlet uncertainty estimation between the two domains , thus obtaining the final optimization goal, expressed as: ; In the formula, is a trade-off parameter.

Images