A multi-source domain equipment fault diagnosis method based on fine feature decoupling

Abstract

The application discloses a multi-source domain equipment fault diagnosis method based on fine feature decoupling, which divides field features into inter-domain invariant features, intra-domain invariant features and domain-specific features and extracts the features by using corresponding feature extractors; the inter-domain invariant features and the intra-domain invariant features respectively focus on the knowledge of cross-domain learning and the inherent fault class knowledge in each field; the domain-specific features reflect the unique knowledge of each field; the feature decoupling is beneficial to accurately obtaining the knowledge related to the fault diagnosis task; the CORAL distance is used to measure the differences of the inter-domain invariant features and the domain-specific features, the CENTER distance is used to measure the difference of the intra-domain invariant features, and a domain adaptation loss is obtained; the inter-domain invariant features and the intra-domain invariant features are combined as domain-invariant features, and a classification loss is calculated; the classification loss and the domain adaptation loss are added to obtain a total loss function value, and model parameter optimization is updated through back propagation iteration.

Description

Technical Field

[0001] This invention relates to a method for diagnosing equipment faults, and more particularly to a method for diagnosing multi-source domain equipment faults based on refined feature decoupling. Background Technology

[0002] Conducting condition monitoring and diagnosis of mechanical equipment is of great practical importance for ensuring reliable, continuous, and stable operation of machinery, reducing economic losses and operating costs, and preventing major accidents. In actual engineering projects, equipment monitoring data has the following characteristics: ① Large data distribution deviation: Due to equipment aging, frequent changes in operating conditions, and industrial noise interference, the distribution of equipment monitoring data often differs across different scenarios; ② Low availability: Only a very small amount of monitoring data corresponds to an known equipment health status, while most monitoring data requires labeling for health status. However, labeling data is costly, such as the inability to frequently shut down for self-inspection and the time-consuming and labor-intensive nature of manual data labeling, resulting in a lack of health labeling information in the monitoring data. Considering the characteristics of monitoring data, a deep transfer learning diagnostic model for equipment faults is established by combining deep neural networks and domain adaptation strategies. This model addresses the domain distribution deviation problem between the labeled source domain (training set) and the unlabeled target domain (test set), improving the model's generalization ability in different test scenarios and making it more suitable for practical fault diagnosis applications.

[0003] Extracting domain-invariant features from both the source and target domains is central to deep transfer learning diagnostic models. While several mature deep transfer learning methods exist, including Domain Adaption Networks (DANs), Joint Adaptation Networks (JANs), Domain Adversarial Neural Networks (DANNs), and Conditional Domain Adversarial Networks (CDANs), most only consider knowledge transfer from a single source domain. This fails to provide sufficient diagnostic knowledge for the target diagnostic task, leading to suboptimal diagnostic results. In practice, multiple source domains with different data distributions can be collected from various machine operating scenarios to compensate for the limitations of single-source-domain diagnostic knowledge. However, involving more source domains in model training implies greater distribution bias, as domain distribution bias exists not only between a single source and target domain but also between different source domains. Therefore, reducing domain distribution bias across all domains to extract domain-invariant features is challenging.

[0004] In conclusion, finding an effective method for diagnosing faults in multi-source domain devices, so as to make full use of diagnostic knowledge in multi-source domains, has become an urgent problem to be solved. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a multi-source domain equipment fault diagnosis method based on refined feature decoupling. This method effectively reduces the domain distribution bias across all domains when utilizing multi-source domain diagnostic knowledge, significantly improving the diagnostic accuracy of the equipment fault diagnosis model in the target domain and enabling intelligent diagnosis of equipment faults.

[0006] The technical solution of the present invention includes the following steps:

[0007] Step 1: Divide the data into N source domains and one target domain according to different operating conditions;

[0008] Step 2: Construct a neural network model for equipment fault diagnosis and initialize its parameters;

[0009] The described equipment fault diagnosis neural network consists of four parts: an inter-domain invariant feature extractor, an intra-domain invariant feature extractor, a domain-specific feature extractor, and a fault classifier. The three feature extractors have identical structures, each consisting of multiple convolutional layers. The fault classifier consists of a fully connected hidden layer and a Softmax activation function. Each of the three feature extractors' convolutional layers is followed by batch normalization and a ReLU activation function. The adaptive pooling layer uses max pooling to downsample features and is placed after the last convolutional layer. The outputs of the inter-domain invariant feature extractor and the intra-domain invariant feature extractor are combined and connected to the fully connected hidden layer of the fault classifier. The Softmax activation function is used to output the probability value of the data belonging to each fault category.

[0010] Step 3: Input N source domain and one target domain data into the neural network, calculate the inter-domain invariant feature difference loss, intra-domain invariant feature difference loss and domain-specific feature difference loss to obtain the domain adaptation loss;

[0011] This method designs a refined feature decoupling strategy and a domain adaptation loss. To effectively eliminate domain distribution bias, the refined feature decoupling strategy divides domain features into inter-domain invariant features, intra-domain invariant features, and domain-specific features. Different features are extracted by corresponding feature extractors in the neural network. Inter-domain invariant features refer to knowledge learned jointly by multiple domains; intra-domain invariant features are features related to fault classification, generated within each domain and unaffected by other domains; and domain-specific features are features unique to each domain. The domain adaptation loss consists of loss functions corresponding to the three types of features. The optimization process of the domain adaptation loss should minimize the differences between inter-domain invariant features and intra-domain invariant features, and maximize the differences in domain-specific features, thereby accurately separating domain-invariant features from the domain features and using these features for fault classification. The domain adaptation loss is calculated as follows:

[0012] L da =L inter +L intra +L spe (1)

[0013] In the formula, L inter L intra L spe These represent the inter-domain invariant feature difference loss, the intra-domain invariant feature difference loss, and the domain-specific feature difference loss, respectively. The three important components in the formula will be further explained below.

[0014] L inter The specific calculations are as follows:

[0015]

[0016] In the formula, f inter (·) represents the output of the inter-domain invariant feature extractor, X T and Let represent the original data of the target domain and the k-th source domain, respectively, where N represents the number of source domains, and d coral The CORAL distance represents the domain-invariant features between the source and target domains, and reflects the differences in different distribution features.

[0017] L intra The specific calculations are as follows:

[0018]

[0019] In the formula, f intra (·) represents the output of the domain-invariant feature extractor. and Let d represent the class centers of the target domain and the k-th source domain, respectively. center The CENTER distance represents the domain-invariant feature to the class center, and the CENTER distance reflects the difference between the feature and the class center.

[0020] L spe The specific calculations are as follows:

[0021]

[0022] In the formula, f spe (·) represents the output of the domain-specific feature extractor. and Let represent the original data of the i-th and j-th source domains, respectively; unlike formula (2), the domain adaptation loss optimization process should maximize the differences in domain-specific features between the source and target domains, as well as the differences in domain-specific features between different source domains, i.e., the CORAL distance d. coralThe larger the value, the better, to represent the domain-specific features extracted from different fields;

[0023] Step 4: Combine the inter-domain invariant features and intra-domain invariant features of the source domain as domain invariant features, and calculate the classification loss of the fault classifier on the domain invariant features;

[0024] Step 5: Add the classification loss and the domain adaptation loss to obtain the total loss function value, and then perform iterative training to update the model parameters to obtain the final model;

[0025] Step 7: During equipment fault diagnosis, input the target domain data into the final model to obtain the equipment fault diagnosis results.

[0026] Preferably, step one specifically involves: obtaining N source domains S = {S1, S2, ..., S...} based on different operating conditions of the equipment. N} and a target domain T, the sensor from the k-th source domain S k The collected sample data and corresponding category labels are represented as follows: in Represents sample data, This represents the category label corresponding to the sample data, where i represents the i-th sample in the source domain. Represents the source domain S k The number of samples included; the samples collected by the sensor from the target domain T are represented as in Let i represent the i-th sample in the target domain, and n represent the sample data. T This indicates the number of samples contained in the target domain T; the working conditions of the source and target domain data are different.

[0027] Preferably, in step two, the parameters of the equipment fault diagnosis neural network are initialized using a normal distribution random initialization method, and the parameters are updated using the Adam algorithm.

[0028] Preferably, in step four, the classification loss is calculated as follows:

[0029]

[0030] In the formula, C(·) represents the output of the fault classifier, F(·) represents the domain-invariant feature after combining the sample data, and J adopts the cross-entropy loss function.

[0031] Preferably, in step five, the classification loss and the domain adaptation loss are added together to obtain the total loss function value, specifically as follows:

[0032] L = L cls +λL da (6)

[0033] In the formula, λ represents the trade-off parameter used to control the domain adaptation loss L. da Weights in the overall loss function.

[0034] Preferably, the inter-domain invariant feature is a common fault occurrence mechanism for both the source and target domains.

[0035] Preferably, the domain-invariant feature is a feature that distinguishes fault categories.

[0036] As a preferred option, the domain-specific features are the equipment operating conditions and the equipment model.

[0037] The beneficial effects of this invention are as follows: This invention proposes a multi-source domain refined feature decoupling method for situations where diagnostic knowledge in a single operating condition source domain is limited and the source domain distribution deviations in different operating conditions are large. This method fully utilizes source domain data from multiple operating conditions to train the model, providing rich diagnostic knowledge. Domain features are divided into inter-domain invariant features, intra-domain invariant features, and domain-specific features, and extracted using corresponding feature extractors. Inter-domain invariant features and intra-domain invariant features focus on cross-domain learning knowledge and fault category knowledge inherent in each domain, respectively. Domain-specific features reflect knowledge unique to each domain. Refined feature decoupling facilitates the accurate acquisition of knowledge related to the fault diagnosis task, namely, inter-domain invariant features and intra-domain invariant features. The CORAL distance is used to measure the differences between inter-domain invariant features and domain-specific features, and the CENTER distance is used to measure the differences between intra-domain invariant features, to obtain the domain adaptation loss. Inter-domain invariant features and intra-domain invariant features are combined as domain invariant features to calculate the classification loss. The classification loss and domain adaptation loss are added to obtain the total loss function value, and the model parameters are iteratively updated and optimized through backpropagation. Based on the powerful learning capabilities of deep learning models, the feature representations of the target domain and different source domains are learned. Furthermore, the multi-source domain method with refined feature decoupling makes full use of the rich diagnostic knowledge of the multi-source domain and separates the inter-domain invariant features and intra-domain invariant features from the domain features. This effectively obtains knowledge related to the fault diagnosis task and solves the problem of limited diagnostic knowledge in the single-source domain and large distribution deviation in the multi-source domain, thereby realizing intelligent diagnosis of equipment faults. Attached Figure Description

[0038] Figure 1 This is a flowchart of the method of the present invention;

[0039] Figure 2 This is a diagram illustrating the overall structure of the neural network of the present invention.

[0040] Figure 3 This is a flowchart of the multi-source domain refined feature decoupling algorithm of the present invention;

[0041] Figure 4 This is a diagram of the diagnostic results. Detailed Implementation

[0042] The present invention will now be described in further detail with reference to the accompanying drawings.

[0043] like Figure 1 As shown, the present invention includes the following steps.

[0044] 1) Divide the data into N source domains and one target domain according to different operating conditions;

[0045] 2) Construct a neural network model for equipment fault diagnosis and initialize its parameters;

[0046] 3) Clear the gradients of the neural network model parameters to zero and perform one round of iterative training;

[0047] 4) Input N source domain and one target domain data into the neural network, calculate the inter-domain invariant feature difference loss, intra-domain invariant feature difference loss and domain-specific feature difference loss to obtain the domain adaptation loss;

[0048] 5) Combine the inter-domain invariant features and intra-domain invariant features of the source domain as domain invariant features, and calculate the classification loss of the fault classifier on the domain invariant features;

[0049] 6) Add the classification loss and the domain adaptation loss together to obtain the total loss function value;

[0050] 7) Backpropagation is performed based on the total loss value, and the Adam algorithm is used for optimization.

[0051] 8) Iterate through training, update model parameters, determine if the required number of iterations has been reached, and save the final model;

[0052] 9) When diagnosing equipment faults, the target domain data is input into the final model to obtain the equipment fault diagnosis results.

[0053] In step 1), based on the different operating conditions of the equipment, we obtain N source domains S = {S1, S2, ..., S...} N} and a target domain T, the sensor from the k-th source domain S k The collected sample data and corresponding category labels are represented as follows: in Represents sample data, This represents the category label corresponding to the sample data, where i represents the i-th sample in the source domain. Represents the source domain S k The number of samples included; the samples collected by the sensor from the target domain T are represented as in Let i represent the i-th sample in the target domain, and n represent the sample data. T This indicates the number of samples contained in the target domain T. The data in the source and target domains have different operating conditions.

[0054] In step 2), the overall structure diagram of the equipment fault diagnosis model is shown below. Figure 2 As shown, the specific construction steps are as follows:

[0055] The equipment fault diagnosis neural network consists of four parts: an inter-domain invariant feature extractor, an intra-domain invariant feature extractor, a domain-specific feature extractor, and a fault classifier. The three feature extractors have identical structures, each consisting of multiple convolutional layers. The fault classifier consists of a fully connected hidden layer and a Softmax activation function. Each of the three feature extractors' convolutional layers is followed by batch normalization and a ReLU activation function. The adaptive pooling layer uses max pooling to downsample features and is placed after the last convolutional layer. The outputs of the inter-domain invariant and intra-domain invariant feature extractors are combined and connected to the fully connected hidden layer of the fault classifier. The Softmax activation function then outputs the probability value of the data belonging to each fault category.

[0056] In step 4), the refined feature decoupling strategy divides domain features into inter-domain invariant features, intra-domain invariant features, and domain-specific features. Different features are extracted by the corresponding feature extractors in the neural network. Inter-domain invariant features refer to knowledge learned jointly by multiple domains, such as the common fault occurrence mechanism of the source and target domains; intra-domain invariant features refer to features related to fault classification, generated within each domain and unaffected by other domains, such as features that distinguish fault categories; domain-specific features refer to features unique to each domain, such as equipment operating conditions and equipment model. The CORAL (Correlation Alignment) distance is used to calculate the differences in inter-domain invariant features between the source and target domains, as well as the differences in domain-specific features between different domains. Taking the calculation of the feature difference between the target domain and the k-th source domain as an example, the CORAL distance calculation formula is as follows:

[0057]

[0058] In the formula, ||·|| F Represents the Frobenius norm, source domain data and target domain data X T The characteristic covariance matrix is ​​calculated as follows:

[0059]

[0060]

[0061] In the formula, 1 is a column vector in which all elements are equal to 1. This is used to measure differences in invariant features across domains. h T =f inter (X TWhen measuring differences in domain-specific features, h T =f spe (X T ).

[0062] CORAL is a statistical criterion that calculates the distance between different distributional characteristics by measuring the second-order statistic (covariance) of two domains. The domain adaptation loss optimization process should minimize the distance between inter-domain invariant features of the source and target domains and maximize the distance between domain-specific features of different domains, thereby representing the extraction of common inter-domain invariant features and unique domain-specific features.

[0063] The difference between the intra-domain invariant features of the source and target domains and the class centers is calculated using the CENTER distance, as follows:

[0064]

[0065]

[0066] In the formula, This represents the Euclidean distance. A smaller CENTER distance means that the features of the domain and the category center are closer, thus achieving the goal of reducing intra-class distance and maintaining intra-class compactness. A smaller intra-class distance means that the features inherent in the domain that distinguish fault categories have been extracted.

[0067] To illustrate the multi-source domain refined feature decoupling method more concretely and intuitively, the overall computational process is as follows: Figure 3 As shown, the calculation steps are as follows:

[0068] Step 1: Calculate the inter-domain invariant feature difference loss, intra-domain invariant feature difference loss, and domain-specific feature difference loss for each domain, and sum them to obtain L. inter L intra L spe The domain adaptation loss L is obtained by adding the loss function values ​​corresponding to the three types of features. da ;

[0069] Step 2: Combine the inter-domain invariant features and intra-domain invariant features of the source domain as domain invariant features, and calculate the classification loss L of the fault classifier on the domain invariant features. cls ;

[0070] Step 4: Add the classification loss and the domain adaptation loss to obtain the total loss function value L;

[0071] Step 5: Iteratively optimize the parameters of the equipment fault diagnosis model based on the total loss function value, and finally obtain the final model.

[0072] Step 4) of this invention proposes a multi-source domain refined feature decoupling method for situations where diagnostic knowledge of a single operating condition source domain is limited and the source domain distribution deviations of different operating conditions are large. This method fully utilizes source domain data from multiple operating conditions to train the model, providing rich diagnostic knowledge. Domain features are divided into inter-domain invariant features, intra-domain invariant features, and domain-specific features, and extracted using corresponding feature extractors. Inter-domain invariant features and intra-domain invariant features focus on cross-domain learning knowledge and fault category knowledge inherent in each domain, respectively. Domain-specific features reflect knowledge unique to each domain. Refined feature decoupling facilitates the accurate acquisition of knowledge related to the fault diagnosis task, namely inter-domain invariant features and intra-domain invariant features. The CORAL distance is used to measure the difference between inter-domain invariant features and the difference between domain-specific features, and the CENTER distance is used to measure the difference between intra-domain invariant features, to obtain the domain adaptation loss. The inter-domain invariant features and intra-domain invariant features are combined as domain invariant features to calculate the classification loss. The classification loss and the domain adaptation loss are added to obtain the total loss function value, and the model parameters are iteratively updated and optimized through backpropagation. Based on the powerful learning capabilities of deep learning models, the feature representations of the target domain and different source domains are learned. Furthermore, the multi-source domain method with refined feature decoupling makes full use of the rich diagnostic knowledge of the multi-source domain and separates the inter-domain invariant features and intra-domain invariant features from the domain features. This effectively obtains knowledge related to the fault diagnosis task and solves the problem of limited diagnostic knowledge in the single-source domain and large distribution deviation in the multi-source domain, thereby realizing intelligent diagnosis of equipment faults.

[0073] Specific embodiments of the present invention:

[0074] The original experimental dataset contained data collected under three different operating conditions. Each condition recorded samples from four fault categories, with 488 samples per fault category and 1952 samples per condition. The data was divided into two source domains and one target domain based on the operating conditions. The source domain contained 3904 samples, and the target domain contained 1952 samples. The source domain data was labeled, while the target domain data was unlabeled. Data from different domains were labeled from 0 to 2. Within the target domain data, the original training set and test set were randomly divided in a 4:1 ratio. The model was trained using the training sets from both the source and target domains using the method of this invention. The test set from the target domain data was then input into the final model to obtain the final fault diagnosis results.

[0075] The experimental environment for this invention is as follows: CPU is Core TM The system uses an i7-6700K processor at 4.00GHz, a GTX 1080 Ti GPU with 11GB of VRAM, a Windows 10 operating system, and PyTorch 1.8 as the deep learning framework. It utilizes Python 3.8 to implement a test method for multi-source domain device fault diagnosis based on refined feature decoupling.

[0076] To verify the effectiveness of the multi-source domain device fault diagnosis method based on refined feature decoupling, this method was compared with the benchmark method (a single-source domain deep neural network without domain adaptation strategy) and three deep transfer methods without refined feature decoupling strategies (ignoring inter-domain invariant features, intra-domain invariant features, and domain-specific features, respectively). Models were trained using different methods, and target domain data was input into the final model to obtain the final target domain diagnosis results, as shown below. Figure 4 As shown in the figure, the horizontal axis represents the transfer task, and the last column represents the average accuracy. The results show that the method of this invention has the highest average accuracy on the three transfer tasks. Compared with the benchmark method, the method of this invention significantly improves the diagnostic accuracy on the 1 and 2-0 transfer tasks. Compared with three deep transfer methods without refined feature decoupling strategies, the method of this invention improves the accuracy on all three transfer tasks. In other words, ignoring any type of feature will lead to a decrease in model performance, proving the importance of separating the three types of features. The improvement in accuracy is because the method of this invention fully utilizes the rich diagnostic knowledge of the multi-source domain. At the same time, the refined feature decoupling strategy meticulously divides the domain features into features with different knowledge focuses, which is conducive to accurately obtaining knowledge related to the fault diagnosis task, thereby improving the network performance. It can be seen that the multi-source domain equipment fault diagnosis method based on refined feature decoupling is an effective diagnostic method that can effectively utilize multi-source domain diagnostic knowledge to generalize the equipment fault diagnosis model to the target domain.

Claims

1. A method for fault diagnosis of multi-source domain devices based on refined feature decoupling, characterized in that, The method specifically includes the following steps: Step 1: Divide the data into N source domains and one target domain according to different operating conditions; Step 2: Construct a neural network model for equipment fault diagnosis and initialize its parameters; The described equipment fault diagnosis neural network consists of four parts: an inter-domain invariant feature extractor, an intra-domain invariant feature extractor, a domain-specific feature extractor, and a fault classifier. The three feature extractors have identical structures, each consisting of multiple convolutional layers. The fault classifier consists of a fully connected hidden layer and a Softmax activation function. Each of the three feature extractors' convolutional layers is followed by batch normalization and a ReLU activation function. The adaptive pooling layer uses max pooling to downsample features and is placed after the last convolutional layer. The outputs of the inter-domain invariant feature extractor and the intra-domain invariant feature extractor are combined and connected to the fully connected hidden layer of the fault classifier. The Softmax activation function is used to output the probability value of the data belonging to each fault category. Step 3: Input N source domain and one target domain data into the neural network, calculate the inter-domain invariant feature difference loss, intra-domain invariant feature difference loss and domain-specific feature difference loss to obtain the domain adaptation loss; Inter-domain invariant features refer to knowledge learned jointly by multiple domains. Intra-domain invariant features refer to features related to fault classification, generated within each domain and unaffected by other domains. Domain-specific features refer to features unique to each domain. The domain adaptation loss consists of loss functions corresponding to the three types of features. The optimization process of the domain adaptation loss should minimize the differences between inter-domain invariant features and intra-domain invariant features, and maximize the differences in domain-specific features, thereby accurately separating domain-invariant features from domain features and using these features for fault classification. The domain adaptation loss is calculated as follows: (1) In the formula, L inter L intra L spe These represent the inter-domain invariant feature difference loss, the intra-domain invariant feature difference loss, and the domain-specific feature difference loss, respectively. The three important components in the formula will be further explained below. L inter The specific calculations are as follows: (2) In the formula, f inter (·) represents the output of the inter-domain invariant feature extractor. and Let represent the original data of the target domain and the k-th source domain, respectively, where N represents the number of source domains, and d coral The CORAL distance represents the domain-invariant features between the source and target domains, and reflects the differences in different distribution features. L intra The specific calculations are as follows: (3) In the formula, f intra (·) represents the output of the domain-invariant feature extractor. and Let d represent the class centers of the target domain and the k-th source domain, respectively. center The CENTER distance represents the domain-invariant feature to the class center, and the CENTER distance reflects the difference between the feature and the class center. L spe The specific calculations are as follows: (4) In the formula, f spe (·) represents the output of the domain-specific feature extractor. and Let represent the original data of the i-th and j-th source domains, respectively; unlike formula (2), the domain adaptation loss optimization process should maximize the differences in domain-specific features between the source and target domains, as well as the differences in domain-specific features between different source domains, i.e., the CORAL distance d. coral The larger the value, the better, to represent the domain-specific features extracted from different fields; Step 4: Combine the inter-domain invariant features and intra-domain invariant features of the source domain as domain invariant features, and calculate the classification loss of the fault classifier on the domain invariant features; Step 5: Add the classification loss and the domain adaptation loss to obtain the total loss function value: (5) In the formula, λ represents the trade-off parameter used to control the domain adaptation loss L. da Weights in the entire loss function; Then, iterative training is performed to update the model parameters, resulting in the final model; Step 7: During equipment fault diagnosis, input the target domain data into the final model to obtain the equipment fault diagnosis results.

2. The multi-source domain device fault diagnosis method based on refined feature decoupling according to claim 1, characterized in that: Step one specifically involves: obtaining N source domains based on different operating conditions of the equipment. Given a target domain T, the sensor draws from the k-th source domain S. k The collected sample data and corresponding category labels are represented as follows: ,in Represents sample data, This represents the category label corresponding to the sample data, where i represents the i-th sample in the source domain. Represents the source domain S k Number of samples included; sensor from the target domain The collected samples are represented as ,in Let i represent the sample data, where i represents the i-th sample in the target domain. Represents the target domain The number of samples included; the working conditions of the source domain and the target domain data are different.

3. The multi-source domain device fault diagnosis method based on refined feature decoupling according to claim 1, characterized in that: In step two, the parameters of the equipment fault diagnosis neural network are initialized using a normal distribution random initialization method, and the parameters are updated using the Adam algorithm.

4. The multi-source domain device fault diagnosis method based on refined feature decoupling according to claim 1, characterized in that: In step four, the classification loss is calculated as follows: (6) In the formula, C(·) represents the output of the fault classifier, F(·) represents the domain-invariant feature after combining the sample data, and J adopts the cross-entropy loss function.

5. The multi-source domain device fault diagnosis method based on refined feature decoupling according to claim 1, characterized in that: The inter-domain invariant feature is a common fault occurrence mechanism for both the source and target domains.

6. The multi-source domain device fault diagnosis method based on refined feature decoupling according to claim 1, characterized in that: The domain-invariant features are those used to distinguish fault categories.

7. The multi-source domain device fault diagnosis method based on refined feature decoupling according to claim 1, characterized in that: Domain-specific characteristics include equipment operating conditions and equipment model.

Images