A dual-view false label correction rolling bearing fault targeting migration diagnosis method

Abstract

A dual-view error label correction method for targeted transfer learning in rolling bearing fault diagnosis is proposed. First, deep features of the source and target domains are extracted from two perspectives using a parameter-shared feature extraction module with residual connections. Then, by comparing the feature similarity and gradient direction of incorrectly labeled samples with other purely labeled samples, an anomaly index is constructed, and incorrect labels are further modified. Finally, using the corrected samples, the original samples, and the target sample, an objective function is constructed to train a model that can correctly label the target domain even under sample anomalies. This invention achieves high diagnostic accuracy for incorrect labeling, significantly improves the precision of transfer learning in bearing fault diagnosis, and solves the problem of label anomalies affecting deep transfer learning.

Description

Technical Field

[0001] This invention belongs to the field of bearing fault diagnosis technology, specifically relating to a method for targeted migration diagnosis of rolling bearing faults with dual-view error label correction. Background Technology

[0002] The application of deep transfer learning in bearing fault diagnosis has attracted widespread attention from researchers and engineers. Deep transfer learning can apply the fault diagnosis knowledge of well-studied bearings (source domain) to other types of bearings (target domain). In this way, users do not need to train the diagnostic model through a large amount of labeled data in the target domain, but can reuse the model from source to target.

[0003] Existing deep transfer learning ([1] M.Ghorvei, M.Kavianpour, MTBeheshti, andA.Ramezani, "An unsupervised bearing fault diagnosis based on deep subdomainadaptation under noise and variable load condition," Measurement Science andTechnology, vol.33, no.2, p.025901, Dec.2021. [2] M.Ghorvei, M.Kavianpour, MTBeheshti, and A.Ramezani, "Spatial graph convolutional neural network via structured subdomain adaptation and domain adversarial learning for bearingfault diagnosis,"Neurocomputing,vol.517,pp.44-61,Jan.2023.[3]M.Kavianpour,A.Ramezani,and MTBeheshti,"A class alignment method based on graphconvolution neural network for bearing fault diagnosis in presence of missingdata and changing working conditions,"Measurement,vol.199,p.111536,Aug.2022.[4]B.Yang,Y.Lei,X.Li,and C.Roberts,Deep targeted transfer learning alongdesignable adaptation trajectory for fault diagnosis across different machines,"IEEE Transactions on Industrial Electronics,vol.70,no.9,pp.9463-9473,Sep.2023.) The success of this study is largely attributed to the collection of high-quality labeled data from the source domain.Pure labels enable diagnostic models to learn precise diagnostic knowledge, namely the relationship between monitoring data and machine health status. However, in engineering scenarios, it is difficult to label large amounts of source domain data with high-quality labels. Influenced by factors such as human error in labeling, incomplete expert knowledge, and insufficient training data for intelligent labeling software, it is challenging to collect source domain data with pure labels. Therefore, due to various factors under actual working conditions, existing transfer diagnostic methods cannot effectively implement transfer diagnostics in engineering applications. Summary of the Invention

[0004] To overcome the shortcomings of the prior art, the present invention aims to propose a dual-view error label correction method for targeted migration diagnosis of rolling bearing faults, which reduces the complexity of the model diagnosis decision surface and achieves high-precision transfer of diagnostic knowledge under data labeling anomalies.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A method for targeted migration diagnosis of rolling bearing faults with dual-view error label correction includes the following steps:

[0007] Step 1: Obtain the source domain dataset, including a clean dataset with labels. Compared with anomaly labeled datasets Furthermore, each state class in the source domain contains at least one label of pure data; then, the target domain dataset is obtained. By N t Composed of unlabeled samples;

[0008] In the above formula, This represents the m-th data point in the labeled pure dataset of the source domain dataset. The corresponding tag is M. sc Indicates the number of pure data items labeled; This represents the m-th data point in the labeled anomaly dataset within the source domain dataset. It is the corresponding exception label, M sa Indicates the number of anomaly-labeled data; This represents the nth data point in the target domain dataset.

[0009] Step 2: Utilize a domain-shared deep residual network to construct a parameter-shared feature extraction module with dual-view residual connections; simultaneously extract features from the labeled clean dataset X of the source domain dataset. sc Anomaly labeled dataset X s and target domain dataset X t Deep fault features are extracted from them. The feature extraction process for viewpoint 1 and viewpoint 2 is as follows:

[0010]

[0011] In the above formula, θ is the operation function of the feature extraction module. Feature This refers to the set of parameters to be optimized in the feature extraction module.

[0012] Step 3: Use the Softmax function σ Softmax To predict the probability of each input feature associated with a health state, we obtain predicted labels from two perspectives:

[0013]

[0014] in and Represent the training parameters for viewpoint 1 and viewpoint 2; calculate the objective function L of the dual-view transfer diagnostic model. c And update the model parameter θ in step 2. Feature :

[0015]

[0016] In the formula, K represents the total number of categories in a sample set, and I(·) represents a binary indicator function, which is 1 if the condition is true and 0 otherwise. This represents the probability that a sample from the m-th labeled pure dataset is associated with a sample from the n-th labeled normal dataset.

[0017] The optimization objective is:

[0018]

[0019] Step 4: When the number of iterations exceeds the set number ξ, calculate the distance between the m-th clean sample and the n-th labeled outlier sample in the source domain from both perspectives. and and gradient and

[0020]

[0021]

[0022]

[0023] In the above formula, These represent the m-th and n-th features of the anomaly-labeled dataset, respectively. This indicates that the calculation is performed in viewpoint 1. This indicates that the calculation is performed in viewpoint 2; L(·,·) is the cross-entropy loss function. and Let m represent the m-th predicted label and the m-th original label in the pure label dataset, respectively. and Let represent the nth predicted label in the pure label dataset and the nth original label in the anomaly label dataset, respectively. Combining the distance and gradient calculation formulas, we obtain the label purity factor of the nth anomaly label sample from both perspectives:

[0024]

[0025] Step 5: Fit the label purity factor from Step 4 using a binarized Gaussian mixture model. Based on the mean and variance in the model parameters, determine whether label correction is needed.

[0026]

[0027] In the formula, |μ v | and |σ v | These are the mean and standard deviation of the clusters with larger label purity factors; This indicates that the nth labeled sample is pure; otherwise, it indicates an anomaly.

[0028] Step 6: Combining the method for identifying anomalous labeled datasets in Step 5, correct the source domain dataset. The label of the nth sample after correction is:

[0029]

[0030] In the formula, and These represent the nth predicted label under viewpoint 1 and the nth predicted label under viewpoint 2, respectively.

[0031] Step 7: From the source domain dataset, select the anomaly-labeled dataset X. s and target domain dataset X t The minimum batch size N is randomly sampled.

[0032] Step 8: Calculate the objective function:

[0033]

[0034] In the formula, This indicates the nth predicted label of the anomaly dataset under viewpoint 1 or viewpoint 2. Let μ represent the j-th corrected label of the source domain dataset under viewpoint 1 or viewpoint 2, μ be the difference penalty coefficient of the dual-view classifier, and λ be the penalty coefficient for feature distribution adaptation.

[0035] The optimization objective is:

[0036]

[0037] Then, the Adam optimization algorithm is used to update the parameter set θ in the target. FeatureTo minimize the objective function;

[0038] Step 9: Repeat steps 2-8 to iteratively optimize the feature extraction module for parameter sharing of residual connections, and output the health labels corresponding to the unlabeled samples in the target domain.

[0039] The construction process of the feature extraction module for parameter sharing in residual connections in step 2 is as follows:

[0040] 2.1) Construct a feature input layer, whose structure consists of a wide-kernel convolutional layer and a max-pooling layer in sequence;

[0041] 2.2) Construct residual blocks and stack several of them;

[0042] 2.3) After the residual unit, the extracted deep fault features are flattened into a one-dimensional vector through a tiling layer and then normalized to map the features onto the unit hypersphere.

[0043] 2.4) After the tiling is completed, proceed to the diagnostic decision-making level.

[0044] Step 8 describes the construction of the objective function L. da The specific process is as follows:

[0045] 8.1) Based on the corrected labels obtained in step 6, the loss function under source domain anomaly labeling is obtained by exchanging the judgment results under the dual-view perspective:

[0046]

[0047] In the formula, v includes two perspectives, v1 and v2, μ is the differential penalty coefficient of the dual-view classifier, and KL(·) represents the calculation of KL divergence;

[0048] 8.2) Initialize the relation matrix The specific format is as follows:

[0049]

[0050] For each a in the matrix i,j If the i-th clean label is associated with the j-th target domain label, then a i,j =1 otherwise 0;

[0051] 8.3) For two perspectives v∈{v1,v2} and The adaptive trajectory between the i-th source domain sample and the j-th target domain sample is:

[0052]

[0053] In the formula, ε is a constant that controls the cost of cross-domain adaptation, and ε → +∞ when the i-th source domain sample is mislabeled. is Represents the i-th feature of the source domain. This represents the j-th feature of the target domain;

[0054] 8.4) Calculate the OT divergence of cross-domain features, in the following form:

[0055]

[0056]

[0057] In the formula, T t→s M represents the transformation matrix from the target domain distribution to the source domain distribution. t→s Let a represent the sample distance matrix. i Let b represent the probability quality of the i-th source domain sample. k The probability quality of the k-th target domain sample is given by the above formula, where k represents the probability quality of the target domain sample. The optimal transmission distance for the target domain sample to migrate to the source domain sample is:

[0058]

[0059] Therefore, the objective function in step 8 can be calculated.

[0060] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0061] This invention proposes a dual-view error label correction method for targeted transfer diagnosis of rolling bearing faults. A parameter-shared feature extraction module with dual-view residual connections is constructed to obtain deep fault features from the data. By using Gaussian binarization to construct a function from two perspectives to identify and reconstruct labeled anomalous data, and then combining this data with the features extracted from both perspectives to construct the objective function, the model can find and correct labeled anomalous data in the source domain, resulting in purer source domain data and better training performance. This method overcomes the problem of existing transfer diagnosis methods that do not consider the quality of the source domain dataset, ultimately improving the accuracy of bearing fault diagnosis. Attached Figure Description

[0062] Figure 1 This is a flowchart of an embodiment of the present invention.

[0063] Figure 2 This is a label conversion diagram according to an embodiment of the present invention, wherein (a) is a schematic diagram of abnormal labels; and (b) is a schematic diagram of reconstructed labels. Detailed Implementation

[0064] The present invention will now be described in further detail with reference to the accompanying drawings.

[0065] Reference Figure 1 A method for targeted migration diagnosis of rolling bearing faults with dual-view error label correction includes the following steps:

[0066] Step 1: Obtain the source domain dataset, including a clean dataset with labels. Compared with anomaly labeled datasets Furthermore, each state class in the source domain contains at least one label of pure data; obtain the target domain dataset. By N t It consists of unlabeled samples; when considering the pure labels of the source domain dataset, it also considers the abnormal labels, which is more realistic and prepares for subsequent repair.

[0067] In the above formula, This represents the m-th data point in the labeled pure dataset of the source domain dataset. The corresponding tag is M. sc Indicates the number of pure data items labeled; This represents the m-th data point in the labeled anomaly dataset within the source domain dataset. It is the corresponding exception label, M sa Indicates the number of anomaly-labeled data; This represents the nth data point in the target domain dataset.

[0068] Step 2: Utilize a domain-shared deep residual network to construct a parameter-shared feature extraction module with dual-view residual connections; simultaneously extract features from the labeled clean dataset X of the source domain dataset. sc Anomaly labeled dataset X s and target domain dataset X t Deep fault features are extracted from them. The feature extraction process for viewpoint 1 and viewpoint 2 is as follows:

[0069]

[0070] In the above formula, θ is the operation function of the feature extraction module. Feature This refers to the set of parameters to be optimized in the feature extraction module.

[0071] The construction process of the feature extraction module for parameter sharing in residual connections in step 2 is as follows:

[0072] 2.1) Construct a feature input layer, whose structure consists of a wide-kernel convolutional layer and a max-pooling layer in sequence;

[0073] 2.2) Construct residual blocks and stack several of them. This multi-layered structure is beneficial for distribution adaptation.

[0074] 2.3) After the residual unit, the extracted deep fault features are flattened into a one-dimensional vector by the tiling layer and then normalized to eliminate the risk of invalid measures in the feature distribution difference ([5] MKAfzal et al., "Discriminative feature abstraction by deep L-2 hypersphere embedding for 3Dmesh CNNs", "Information Sciences, vol.607, pp.1158-1173, Aug.2022.);

[0075] 2.4) After the tiling is completed, proceed to the diagnostic decision-making level;

[0076] Step 3: Use the Softmax function σ Softmax To predict the probability of each input feature associated with a health state, we obtain predicted labels from two perspectives:

[0077]

[0078] in and Let L represent the training parameters for viewpoint 1 and viewpoint 2. Since the source domain contains mislabeled samples, direct training with these samples would lead to overfitting. To address this, the network is first pre-trained using a very limited number of purely labeled samples. The objective function L of the dual-view transfer diagnostic model is then calculated. c And update the model parameter θ in step 2. Feature :

[0079]

[0080] In the formula, K represents the total number of categories in a sample set, and I(·) represents a binary indicator function, which is 1 if the condition is true and 0 otherwise. This represents the probability that a sample from the m-th labeled pure dataset is associated with a sample from the n-th labeled normal dataset.

[0081] The optimization objective is:

[0082]

[0083] Step 4: Samples labeled as anomalous are similar to samples labeled as pure; they have the same health status but opposite gradients. When the number of iterations exceeds a set number ζ, calculate the distance between the m-th pure sample and the n-th labeled anomalous sample in the source domain from both perspectives. and and gradient and

[0084]

[0085]

[0086]

[0087] In the above formula, These represent the m-th and n-th features of the anomaly-labeled dataset, respectively. This indicates that the calculation is performed in viewpoint 1. This indicates that the calculation is performed in viewpoint 2; L(·,·) is the cross-entropy loss function. and Let m represent the m-th predicted label and the m-th original label in the pure label dataset, respectively. and Let represent the nth predicted label of the pure dataset and the nth original label of the anomaly dataset, respectively. This helps distinguish between anomaly and pure labels. Combining the distance and gradient calculation formulas, we obtain the label purity factor of the nth anomaly sample from two perspectives:

[0088]

[0089] Step 5: Fit the label purity factor from Step 4 using a binarized Gaussian mixture model. Based on the mean and variance in the model parameters, determine whether label correction is needed.

[0090]

[0091] In the formula, |μ v | and |σ v | These are the mean and standard deviation of the clusters with larger label purity factors; This indicates that the nth labeled sample is pure; otherwise, it indicates an anomaly.

[0092] Step 6: Combining the method for identifying anomalous labeled datasets in Step 5, construct a classifier. To enhance the capabilities of both classifiers, exchange their judgments on mislabeled samples. This process corrects the source domain dataset. The label of the nth sample after correction is:

[0093]

[0094] In the formula, and These represent the nth predicted label under viewpoint 1 and the nth predicted label under viewpoint 2, respectively.

[0095] Step 7: From the source domain dataset, select the anomaly-labeled dataset X. s and target domain dataset X tThe minimum batch size N is randomly sampled.

[0096] Step 8: After processing the exception labels, calculate the objective function again:

[0097]

[0098] In the formula, This indicates the nth predicted label of the anomaly dataset under viewpoint 1 or viewpoint 2. Let μ represent the j-th corrected label of the source domain dataset under viewpoint 1 or viewpoint 2, μ be the difference penalty coefficient of the dual-view classifier, and λ be the penalty coefficient for feature distribution adaptation.

[0099] The optimization objective is:

[0100]

[0101] Then, the Adam optimization algorithm is used to update the parameter set θ in the target. Feature To minimize the objective function;

[0102] Step 8 describes the construction of the objective function L. da The specific process is as follows:

[0103] 8.1) Based on the corrected labels obtained in step 6, the loss function under source domain anomaly labeling is obtained by exchanging the judgment results under the dual-view perspective:

[0104]

[0105] In the formula, v includes two perspectives, v1 and v2, μ is the differential penalty coefficient of the dual-view classifier, and KL(·) represents the calculation of KL divergence;

[0106] 8.2) Initialize the relation matrix The specific format is as follows:

[0107]

[0108] For each a in the matrix i,j If the i-th clean label is associated with the j-th target domain label, then a i,j =1 otherwise 0;

[0109] 8.3) The smaller the distance between cross-domain samples, the lower the cost of distribution adaptation. Therefore, we should first find the distribution adaptation trajectory with lower cost; for dual-view v∈{v1,v2} and The adaptive trajectory between the i-th source domain sample and the j-th target domain sample is:

[0110]

[0111] In the formula, ε is a constant that controls the cost of cross-domain adaptation, and ε → +∞ when the i-th source domain sample is mislabeled. i s Represents the i-th feature of the source domain. This represents the j-th feature of the target domain;

[0112] 8.4) Calculate the OT divergence of cross-domain features. The OT divergence is used to find the optimal strategy with the lowest cost, and it takes the following form:

[0113]

[0114]

[0115] In the formula, T t→s M represents the transformation matrix from the target domain distribution to the source domain distribution. t→s Let a represent the sample distance matrix. i Let b represent the probability quality of the i-th source domain sample. k The probability quality of the k-th target domain sample is given by the above formula, where k represents the probability quality of the target domain sample. The optimal transmission distance for the target domain sample to migrate to the source domain sample is:

[0116]

[0117] Therefore, the objective function in step 8 can be calculated;

[0118] Step 9: Repeat steps 2-8 to iteratively optimize the feature extraction module for parameter sharing of residual connections, and output the health labels corresponding to the unlabeled samples in the target domain.

[0119] This embodiment uses two rolling bearing datasets to verify the method of the present invention, as detailed in Table 1.

[0120] Table 1 Migration Diagnosis Dataset

[0121]

[0122] The RTS dataset was collected from a bearing-rotor test bench. The tested bearings included four health states: normal, inner ring fault, outer ring fault, and rolling element fault. During the experiment, the motor speed was set to 2100 r / min and the sampling frequency was set to 12.8 kHz. The dataset contains a total of 2048 samples, with 512 samples for each health state.

[0123] The SYMotor dataset was collected from the Suntech motor fault test bench. The tested motor bearings included: normal, inner ring fault, outer ring fault, and rolling element fault. During the experiment, the motor speed was set to 1500 r / min, the friction torque applied by the magnetic powder brake was 10 N·m, the sampling frequency was 25.6 kHz, and the dataset included 2048 samples.

[0124] The transfer diagnostic task RTS→SYMotor is constructed using the aforementioned bearing dataset to verify the feasibility of this invention. It is assumed that the anomaly rate in the RTS dataset is ζ, which follows a symmetric label probability transformation matrix:

[0125] Table 2 Tag Probability Transformation Matrix

[0126]

[0127]

[0128] To construct a clean, labeled dataset, one sample from each class is randomly selected from the RTS dataset. Then, the anomaly rate ζ is set to 30%, and the labels of the remaining samples are transformed, such as... Figure 2 As shown in (a), the sample labels reconstructed by the method of the present invention are as follows: Figure 2 As shown in Figure (b), the method of the present invention can effectively correct abnormal labels in the source domain data and reconstruct labels that are close to their true labels, thus verifying the feasibility of the present invention in solving the labeling problem in the field of deep learning.

[0129] Furthermore, the source domain data annotation anomaly rates were set to 10%, 20%, and 30%. The diagnostic knowledge from the source domain data was transferred using the method of this invention to identify the health status of the target domain samples. The diagnostic accuracy is shown in Table 3. As can be seen from the results in Table 3, the method of this invention has high diagnostic accuracy under different data annotation anomaly rates.

[0130] Table 3 Comparison of Diagnostic Results

[0131]

[0132] Meanwhile, two methods were selected and compared with the diagnostic accuracy of the method proposed in this invention. Method 1 is a deep intelligent diagnostic model based on ResNet, which is trained only using source domain data and then tested on target domain samples. It can be found that as the labeling anomaly rate increases, the recognition accuracy of the deep intelligent diagnostic model for target domain samples gradually decreases, indicating that the presence of labeled anomaly data increases the complexity of the source domain model's diagnostic decision surface, thereby reducing the model's transfer accuracy. Method 2 is a feature distribution adaptation model based on optimal transfer theory. During the ResNet model training process, a regularization constraint that minimizes the optimal transfer divergence of deep features is added. Although this method reduces the distribution difference of cross-domain data features compared to Method 1, its transfer diagnostic accuracy is lower than that of the method proposed in this invention because it ignores the influence of source domain data labeling anomalies. By comparing this invention with the other two label repair methods, it is shown that this invention effectively overcomes the influence of labeling problems and improves the diagnostic accuracy of the transfer diagnostic model.

Claims

1. A method for targeted migration diagnosis of rolling bearing faults with dual-view error label correction, characterized in that, Includes the following steps: Step 1: Obtain the source domain dataset, including a clean dataset with labels. Compared with anomaly labeled datasets Furthermore, each state class in the source domain contains at least one label of pure data; then, the target domain dataset is obtained. By N t Composed of unlabeled samples; In the above formula, This represents the m-th data point in the labeled pure dataset of the source domain dataset. The corresponding tag is M. sc Indicates the number of pure data items labeled; This represents the m-th data point in the anomaly-labeled dataset within the source domain dataset. It is the corresponding exception label, M sa Indicates the number of anomaly-labeled data; This represents the nth data point in the target domain dataset. Step 2: Utilize a domain-shared deep residual network to construct a parameter-shared feature extraction module for dual-view residual connections; Simultaneously, from the labeled pure dataset X of the source domain dataset... sc Anomaly labeled dataset X s and target domain dataset X t Deep fault features are extracted from them. The feature extraction process for viewpoint 1 and viewpoint 2 is as follows: In the above formula, θ is the operation function of the feature extraction module. Feature This refers to the set of parameters to be optimized in the feature extraction module. Step 3: Use the Softmax function σ Softmax To predict the probability of each input feature associated with a health state, we obtain predicted labels from two perspectives: in and Represents the training parameters for viewpoint 1 and viewpoint 2; Calculate the objective function L of the dual-view transfer diagnostic model c And update the model parameter θ in step 2. Feature : In the formula, K represents the total number of categories in a sample set, and I(·) represents a binary indicator function, which is 1 if the condition is true and 0 otherwise. This represents the probability that a sample in the m-th labeled pure dataset is associated with a sample in the n-th labeled dataset. The optimization objective is: Step 4: When the number of iterations exceeds the set number ξ, calculate the distance between the m-th clean sample and the n-th labeled outlier sample in the source domain from both perspectives. and and gradient and In the above formula, These represent the m-th and n-th features of the anomaly-labeled dataset, respectively. This indicates that the calculation is performed in viewpoint 1. This indicates that the calculation is performed in viewpoint 2; L(·,·) is the cross-entropy loss function. and Let m represent the m-th predicted label and the m-th original label in the pure label dataset, respectively. and Let represent the nth predicted label in the pure label dataset and the nth original label in the anomaly label dataset, respectively. Combining the distance and gradient calculation formulas, we obtain the label purity factor of the nth labeled anomaly sample from both perspectives: Step 5: Fit the label purity factor from Step 4 using a binarized Gaussian mixture model. Based on the mean and variance in the model parameters, determine whether label correction is needed. In the formula, |μ v | and |σ v | These are the mean and standard deviation of the clusters with larger label purity factors; This indicates that the nth labeled sample is pure; otherwise, it indicates an anomaly. Step 6: Combining the method for identifying anomalous labeled datasets in Step 5, correct the source domain dataset. The label of the nth sample after correction is: In the formula, and These represent the nth predicted label under viewpoint 1 and the nth predicted label under viewpoint 2, respectively. Step 7: From the anomaly-labeled dataset X of the source domain dataset s and target domain dataset X t The minimum batch size N is randomly sampled. Step 8: The optimization goal is: In the formula, L s To obtain the loss function under source domain anomaly labeling by exchanging the judgment results from the two perspectives, λ is the penalty coefficient for feature distribution adaptation; L da The optimal transmission distance for the migration of target domain samples to source domain samples under dual-view conditions is determined; then, the Adam optimization algorithm is used to update the parameter set θ in the target domain. Feature To minimize the objective function; Step 9: Repeat steps 2-8 to iteratively optimize the parameter-sharing feature extraction module of the residual connection, and output the health label corresponding to the unlabeled sample in the target domain.

2. The method according to claim 1, characterized in that, The construction process of the parameter-sharing feature extraction module for residual connections in step 2 is as follows: 2.1) Construct a feature input layer, whose structure consists of a wide-kernel convolutional layer and a max-pooling layer in sequence; 2.2) Construct residual blocks and stack several of them; 2.3) After the residual unit, the extracted deep fault features are flattened into a one-dimensional vector through a tiling layer and then normalized to map the features onto the unit hypersphere. 2.4) After the tiling is completed, proceed to the diagnostic decision-making level.

3. The method according to claim 1, characterized in that, Step 8 describes the construction of the objective function L. da The specific process is as follows: 8.1) Based on the corrected labels obtained in step 6, the loss function under source domain anomaly labeling is obtained by exchanging the judgment results under the dual-view perspective: In the formula, v includes two perspectives, v1 and v2, μ is the differential penalty coefficient of the dual-view classifier, and KL(·) represents the calculation of KL divergence; 8.2) Initialize the relation matrix The specific format is as follows: For each a in the matrix i,j If the i-th clean label is associated with the j-th target domain label, then a i,j =1 otherwise 0; 8.3) For two perspectives v∈{v1,v2} and The adaptive trajectory between the i-th source domain sample and the j-th target domain sample is: In the formula, ε is a constant that controls the cost of cross-domain adaptation, and ε → +∞ when the i-th source domain sample is mislabeled. i s Represents the i-th feature of the source domain. This represents the j-th feature of the target domain; 8.4) Calculate the OT divergence of cross-domain features, in the following form: In the formula, T t→s M represents the transformation matrix from the target domain distribution to the source domain distribution. t→s Let a represent the sample distance matrix. i Let b represent the probability quality of the i-th source domain sample. k Representing the probability quality of the k-th target domain sample, the optimal transmission distance for the migration of the target domain sample to the source domain sample is obtained from the above formula: Therefore, the objective function in step 8 can be calculated.

Images