Method for multi-working condition industrial process fault diagnosis based on dynamic adaptive domain adversarial network

Abstract

The application discloses a multi-working-condition industrial process fault diagnosis method based on a dynamic adaptive domain adversarial network, and comprises the following steps: standardizing process data and taking a sliding window to obtain dynamic data; training a dynamic adaptive domain adversarial network by using labeled dynamic data of a source working condition and unlabeled dynamic data of a target working condition; the dynamic adaptive domain adversarial network comprises a feature extractor, a label classifier, a global domain discriminator and a local domain discriminator; the global domain discriminator aligns the marginal distribution of the source working condition and the target working condition; the local domain discriminator aligns the conditional distribution of the source working condition and the target working condition; a kind of learnable parameter is proposed to adaptively evaluate the relative importance of the marginal distribution and the conditional distribution, so as to better extract domain-invariant features; a kind of common center loss is proposed to improve the inter-class separability and intra-class compactness of data, and further improve the fault diagnosis precision; and the label classifier is used to predict the fault category of a sample. The application establishes an accurate fault diagnosis model for the target working condition.

Description

Technical Field

[0001] This invention relates to the field of data-driven industrial process fault diagnosis, and in particular to a multi-condition industrial process fault diagnosis method based on dynamic adaptive domain adversarial networks for fault diagnosis under the condition of scarce labeled samples in multi-condition industrial processes. Background Technology

[0002] With the increasing complexity of modern industrial processes, ensuring their safe and stable operation is becoming increasingly challenging. Fault diagnosis plays a crucial role in improving the reliability and safety of industrial processes. Fault diagnosis methods include model-based methods, knowledge-based methods, and data-driven methods. Due to advancements in computer and sensor technologies, large amounts of historical data from industrial processes can be collected and stored, leading to widespread attention being paid to data-driven fault diagnosis methods. These methods include multivariate statistical analysis, artificial neural networks, and deep learning. Deep learning methods, with their excellent feature extraction capabilities and ability to extract deep, nonlinear features from data, have received significant attention in industrial process fault diagnosis. Data-driven fault diagnosis methods typically involve supervised learning, requiring substantial labeled data. However, in multi-condition industrial processes, labeling samples is time-consuming and labor-intensive, and labeled samples for the target condition may be scarce when it is newly commissioned, making it difficult to establish a fault diagnosis model for the target condition. Deep transfer learning can utilize labeled data from other conditions and unlabeled data from the target condition to build a fault diagnosis model for the target condition.

[0003] Deep transfer learning methods are categorized into instance-based, model-based, and feature-based methods. Instance-based methods select instances in the source domain that are beneficial for training the target domain model and increase the target domain data by adjusting the instance weighting strategy. Model-based methods utilize pre-trained models in the source domain to improve the performance of the target domain model. Feature-based methods transform the original data from the source and target domains from different feature spaces into a common latent feature space, reducing the differences between domains and obtaining domain-invariant features. Feature-based transfer learning methods can achieve good results in feature selection and transformation, and are therefore adopted by most transfer learning methods. Adversarial domain adaptation is a feature-based transfer learning method, and some research results show that adversarial domain adaptation methods outperform traditional deep domain adaptation methods. Xu et al. proposed a deep adversarial neural network (DANN), which confuses the source and target domain datasets, maximizes the domain classification error to minimize the distribution distance between the source and target domains, and extracts domain-invariant features. However, this method only aligns the marginal distributions of the source and target domains. Liu et al. proposed a multi-adversarial domain adaptation (MADA) transfer learning method, which captures multimodal structures and achieves fine-grained alignment of different data distributions based on multiple domain discriminators. However, this method only aligns the conditional distributions of the source and target domains. Aligning only the conditional or marginal distributions of the source and target domains may be detrimental to extracting domain-invariant features. Kuang et al. proposed a class-imbalance adversarial transfer learning (CIATL) network. This method aligns the marginal and conditional distributions of the source and target domain data; however, the relative importance of conditional and marginal distributions differs during training, and CIATL does not measure the relative importance of conditional and marginal distributions.

[0004] The above-mentioned adversarial domain adaptation methods have achieved good results. However, in industrial processes, there is still a need for a parameter that can effectively evaluate the relative importance of marginal and conditional distributions. At the same time, how to improve the intra-class compactness and inter-class separability of source and target domain data and improve the accuracy of fault diagnosis is also a problem worth studying. Summary of the Invention

[0005] To address the problem of scarce target operating condition label samples making it difficult to establish fault diagnosis models, this invention proposes a multi-operating condition industrial process fault diagnosis method based on a dynamic adaptive domain adversarial network (DADAN). The main innovations of this invention are: (1) utilizing learnable parameters to evaluate the relative importance of marginal and conditional distributions, adaptively aligning the marginal and conditional distributions of source and target operating condition data; (2) proposing a common center loss function to improve the inter-class separability and intra-class compactness of source and target operating condition data.

[0006] A multi-condition industrial process fault diagnosis method based on dynamic adaptive domain adversarial networks is characterized by the following steps:

[0007] Step 1: Data Preprocessing

[0008] 1) Obtain historical data for both the source and target operating conditions, where the source operating condition data consists of labeled samples, denoted as {X}. S ,Y S}, Let R be the sample set of the source operating conditions, and let x be the set of real numbers. S,k n represents the source operating condition sample collected at the k-th sampling time. S Let m represent the number of source operating condition samples, and m represent the number of variables in the samples. Let C be the one-hot label of the source working condition sample, and C represent the number of classes in the sample. The target working condition is an unlabeled sample, denoted as . n T x represents the number of samples for the target operating condition. T,k This represents the target operating condition sample collected at the k-th sampling time.

[0009] 2) Standardize the training data by calculating the mean and standard deviation of the source and target operating condition samples. The mean of the source operating condition samples is expressed as... The standard deviation is S S The average value of the target working condition is The standard deviation is S T The standardization of the source operating condition and the target operating condition is achieved through equations (1) and (2) respectively, resulting in standardized source operating condition samples. and target working condition sample

[0010]

[0011]

[0012] 3) To address the dynamic nature of industrial processes, dynamic data is obtained by taking sliding windows for standardized source and target operating condition samples. The sliding windows for the source and target operating condition samples at the k-th sampling time are as follows:

[0013]

[0014]

[0015] Where w represents the width of the sliding window, the sliding window step size is 1, and n is obtained from the source operating data. S -w sliding windows, n target operating condition data are obtained. T -w sliding windows are used as training data;

[0016] Step 2: Model Building

[0017] A dynamic adaptive domain adversarial network model is established, comprising four parts: a feature extractor, a class label classifier, a global domain discriminator, and a local domain discriminator. The feature extractor is a convolutional neural network used to extract features. Its structure is: Convolutional Layer 1 - BatchNorm Layer 1 - Max Pooling Layer 1 - Convolutional Layer 2 - BatchNorm Layer 2 - Convolutional Layer 3 - BatchNorm Layer 3 - Dropout Layer. The kernel size of Convolutional Layer 1 is 3x3, the stride is 1, the number of channels is 64, and the activation function is ReLU. The feature dimension of BatchNorm Layer 1 is... The structure is 64. The stride of max-pooling layer 1 is 2. Convolutional layer 2 has a kernel size of 2*4, a stride of 1, 50 channels, and uses ReLU activation. BatchNorm layer 2 has a feature dimension of 50. Convolutional layer 3 has a kernel size of 3*4, a stride of 1, 50 channels, and uses ReLU activation. BatchNorm layer 3 also has a feature dimension of 50. The probability of neurons in the Dropout layer being set to zero is 0.4. A label classifier is used to predict the fault category of source and target condition samples, and its structure is a fully connected layer 1-BatchNorm. The sequence is: Layer 1 - Fully Connected Layer 2 - BatchNorm Layer 2 - Dropout Layer - Softmax Layer. Fully Connected Layer 1 has 500 neurons and uses the ReLU activation function. BatchNorm Layer 1 has a feature dimension of 500. Fully Connected Layer 2 has 100 neurons and uses the ReLU activation function. BatchNorm Layer 2 has a feature dimension of 100. The Dropout layer has a neuron zeroing probability of 0.4. The Softmax layer has C neurons. A global discriminator is used to determine whether a sample belongs to the source condition or the target condition. The target working condition is aligned with the edge distribution of the source and target working condition samples. The structure is: fully connected layer 1 - BatchNorm layer 1 - fully connected layer 2 - BatchNorm layer 2 - Dropout layer - Softmax layer. The fully connected layer 1 has 500 neurons and the activation function is ReLU. The feature dimension of BatchNorm layer 1 is 500. The fully connected layer 2 has 100 neurons and the activation function is ReLU. The feature dimension of BatchNorm layer 2 is 100. The neurons in the Dropout layer have a zero probability of being set to 0.4. The Softmax layer has 2 neurons, and there are C local discriminators. The c-th local discriminator is used to determine whether the c-th class of samples belongs to the source condition or the target condition, aligning the conditional distributions of the source and target condition samples. During training, the C local discriminators have no order, and each local discriminator has the same structure: Fully Connected Layer 1 - BatchNorm Layer 1 - Fully Connected Layer 2 - BatchNorm Layer 2 - Dropout Layer - Softmax Layer. Fully Connected Layer 1 has 500 neurons with the ReLU activation function, and BatchNorm Layer 1 has a feature dimension of 500. Fully Connected Layer 2 has 100 neurons with the ReLU activation function, and BatchNorm Layer 2 has a feature dimension of 100. The Dropout layer has a zero-probability neuron setting of 0.4, and the Softmax layer has 2 neurons.

[0018] Step 3: Model Training

[0019] 1) Train a classifier using source and target operating condition data. The loss function of the classifier is:

[0020]

[0021] Where θ f θ is the parameter of the feature extractor. y G is the parameter of the label classifier. f G represents the feature extractor. y Represents the label classifier, y i s The label representing the source operating condition sample, n S The number of source condition samples is represented by γ, which is a hyperparameter with a value of 0.0001. The loss function is... cc The common center loss function is expressed as:

[0022]

[0023] Where f i c represents the source or target operating condition features extracted by the last fully connected layer of the class label classifier. c The class center is a randomly initialized learnable parameter. The features of the source and target operating conditions share a set of class centers. If there are C classes of samples, then C class centers are set. By minimizing the distance between the features and their corresponding class centers, the intra-class compactness and inter-class separability of the source and target operating condition samples are improved.

[0024] 2) Train a global discriminator using samples from the source and target operating conditions, performing supervised learning. The loss function is the binary cross-entropy loss function:

[0025]

[0026] Where d i The field label can be either 0 or 1, where 0 indicates the sample belongs to the source condition and 1 indicates the sample belongs to the target condition. d Represents the global domain discriminator, θ d The parameter x represents the global domain discriminator. i Training data representing the source or target operating conditions;

[0027] 3) The local domain discriminator is also trained using samples from the source and target conditions, with the loss function being:

[0028]

[0029] in This represents the c-th local domain discriminator. This represents the label of a sample. If the sample is from the source working condition, the label is its true label; if the sample is from the target working condition, the label is the label predicted by the label classifier. C represents the number of classes in the sample. n represents the parameters of the c-th local domain discriminator. c This represents the number of samples for the c-th source and target operating conditions.

[0030] 4) The overall loss function consists of the loss functions of the label classifier, the global discriminator, and the local discriminator, and is expressed as:

[0031]

[0032] Where λ is a hyperparameter with a value of 0.4, w is a learnable parameter with an initial value of 0, used to adaptively evaluate the relative importance of marginal and conditional distributions, and e represents the natural constant with a value of 2.71828.

[0033] 5) The model is trained by minimizing the overall loss function. The optimization algorithm is Adam, the learning rate is 0.001, and the maximum batch size is 300. Training is stopped when the overall loss function no longer decreases after 50 consecutive batches or reaches the maximum batch size. The network parameters of the fault diagnosis model for the target working condition are obtained, and the values ​​of the learnable parameter w and the class center c are determined.

[0034] c

[0035] Step 4: Online Diagnosis

[0036] Input the online samples of the target working condition into the trained model, and the output of the class label classifier is the fault diagnosis result.

[0037] Beneficial effects

[0038] This invention designs a multi-condition industrial process fault diagnosis method based on a dynamic adaptive domain adversarial network, achieving migration fault diagnosis even with a limited number of target condition label samples in multi-condition industrial processes. The designed method uses a dynamic adaptive domain adversarial network to adaptively align the marginal and conditional distributions of the source and target conditions. The proposed learnable parameters adaptively evaluate the relative importance of the conditional and marginal distributions to better extract domain-invariant features. The proposed common center loss function improves inter-class separability and intra-class compactness, further enhancing the accuracy of fault diagnosis. This invention significantly improves fault diagnosis accuracy for multi-condition industrial processes with limited target condition label samples, and is of great importance for fault diagnosis in multi-condition industrial processes. Attached Figure Description

[0039] Figure 1 The flowchart shown is a flowchart of the present invention;

[0040] Figure 2 The diagram shown is a model structure diagram of DADAN proposed in this invention;

[0041] Figure 3 The diagram shown is a flowchart of the TE process;

[0042] Figure 4 The following is the confusion matrix for LMMD, DANN, MADA, and DADAN;

[0043] Figure 5 The TSNE charts for LMMD, DANN, MADA, and DADAN are shown below. Detailed Implementation

[0044] The Tennessee Eastman (TE) process is a simulation of a chemical process used to validate the performance of industrial process fault diagnosis methods. The TE process can generate normal data and 21 types of fault data, collecting 53 variables. By modifying the parameter settings of the variables, data for three operating conditions were generated, as shown in Table 1. The experimental data uses 31 of these variables, which are highlighted in bold in the table. The sampling interval was 1 minute, generating fault data for faults 1, 2, 4, 5, 7, 8, 12, 14, and 17. The simulation ran for 28 hours, with faults introduced after the 8th hour. Each type of fault data contained 1200 samples, where the source condition samples were labeled samples and the target condition samples were unlabeled samples.

[0045] Table 1. Parameter settings for three operating conditions.

[0046]

[0047]

[0048]

[0049] Based on the above description, and in accordance with the invention, the specific process is implemented as follows:

[0050] Step 1: Data Preprocessing

[0051] 1) Standardize the source operating condition data and the target operating condition data using equations (1) and (2) respectively;

[0052] 2) Take sliding windows with a window width of 20 and a step size of 1 for the standardized source and target operating condition data respectively to obtain dynamic data as training data. Then there are 1180 training samples for each type of fault in the source and target operating conditions.

[0053] Step 2: Model Building

[0054] A dynamic adaptive domain adversarial network model is established, comprising four parts: a feature extractor, a class label classifier, a global domain discriminator, and a local domain discriminator. The feature extractor is a convolutional neural network used to extract features. Its structure is: Convolutional Layer 1 - BatchNorm Layer 1 - Max Pooling Layer 1 - Convolutional Layer 2 - BatchNorm Layer 2 - Convolutional Layer 3 - BatchNorm Layer 3 - Dropout Layer. The kernel size of Convolutional Layer 1 is 3x3, the stride is 1, the number of channels is 64, and the activation function is ReLU. The feature dimension of BatchNorm Layer 1 is... The structure is 64. The stride of max-pooling layer 1 is 2. Convolutional layer 2 has a kernel size of 2*4, a stride of 1, 50 channels, and uses ReLU activation. BatchNorm layer 2 has a feature dimension of 50. Convolutional layer 3 has a kernel size of 3*4, a stride of 1, 50 channels, and uses ReLU activation. BatchNorm layer 3 also has a feature dimension of 50. The probability of neurons in the Dropout layer being set to zero is 0.4. A label classifier is used to predict the fault category of source and target condition samples, and its structure is a fully connected layer 1-BatchNorm. The sequence is: Layer 1 - Fully Connected Layer 2 - BatchNorm Layer 2 - Dropout Layer - Softmax Layer. Fully Connected Layer 1 has 500 neurons and uses the ReLU activation function. BatchNorm Layer 1 has a feature dimension of 500. Fully Connected Layer 2 has 100 neurons and uses the ReLU activation function. BatchNorm Layer 2 has a feature dimension of 100. The Dropout layer has a neuron zeroing probability of 0.4. The Softmax layer has 9 neurons. A global discriminator is used to determine whether a sample belongs to the source condition or the target condition. The target working condition is aligned with the edge distribution of the source and target working condition samples. The structure is: fully connected layer 1 - BatchNorm layer 1 - fully connected layer 2 - BatchNorm layer 2 - Dropout layer - Softmax layer. The fully connected layer 1 has 500 neurons and the activation function is ReLU. The feature dimension of BatchNorm layer 1 is 500. The fully connected layer 2 has 100 neurons and the activation function is ReLU. The feature dimension of BatchNorm layer 2 is 100. The neurons in the Dropout layer have a zero probability of being set to 0.4. The Softmax layer has 2 neurons, and there are 9 local discriminators. The c-th local discriminator is used to determine whether the c-th class of samples belongs to the source condition or the target condition (where c = 1, 2, ..., 9), aligning the conditional distributions of the source and target condition samples. During training, the 9 local discriminators have no order, and each local discriminator has the same structure: Fully Connected Layer 1 - BatchNorm Layer 1 - Fully Connected Layer 2 - BatchNorm Layer 2 - Dropout Layer - Softmax Layer. Fully Connected Layer 1 has 500 neurons with the ReLU activation function, and BatchNorm Layer 1 has a feature dimension of 500. Fully Connected Layer 2 has 100 neurons with the ReLU activation function, and BatchNorm Layer 2 has a feature dimension of 100. The Dropout layer has a neuron zeroing probability of 0.4, and the Softmax layer has 2 neurons.

[0055] Step 3: Model Training

[0056] 1) Train a classifier using source and target operating condition data. The loss function of the classifier is:

[0057]

[0058] Where θ f θ is the parameter of the feature extractor. y G is the parameter of the label classifier. f G represents the feature extractor. y Represents a label classifier. This represents the i-th source condition sample. The label represents the source operating condition sample, γ is a hyperparameter with a value of 0.0001, and the loss... cc The common center loss function is expressed as:

[0059]

[0060] Where f i c represents the source or target operating condition features extracted by the last fully connected layer of the class label classifier. c The class center is a randomly initialized learnable parameter. The features of the source and target operating conditions share a set of class centers. If there are 9 classes of samples, then 9 class centers are set. By minimizing the distance between the features and their corresponding class centers, the intra-class compactness and inter-class separability of the source and target operating condition samples are improved.

[0061] 2) Train a global discriminator using samples from the source and target operating conditions, performing supervised learning. The loss function is the binary cross-entropy loss function:

[0062]

[0063] Where d i The field label can be either 0 or 1, where 0 indicates the sample belongs to the source condition and 1 indicates the sample belongs to the target condition. d Represents the global domain discriminator, θ d The parameter x represents the global domain discriminator. i Training data representing the source or target operating conditions;

[0064] 3) The local domain discriminator is also trained using samples from the source and target conditions, with the loss function being:

[0065]

[0066] in This represents the c-th type of local domain discriminator. The label represents the sample. If the sample is from the source working condition, the label is its true label; if the sample is from the target working condition, the label is the label predicted by the label classifier. This represents the parameters of the c-th type of local domain discriminator;

[0067] 4) The overall loss function consists of the loss functions of the label classifier, the global discriminator, and the local discriminator, and is expressed as:

[0068]

[0069] Where λ is a hyperparameter, set to 0.4, w is a learnable parameter, initially set to 0, used to adaptively evaluate the relative importance of marginal and conditional distributions, and e is the natural constant, set to 2.71828. The model is trained by minimizing the overall loss function using the Adam optimization algorithm, with a learning rate of 0.001 and a maximum of 300 iterations. Training stops when the overall loss function no longer decreases after 50 consecutive batches or reaches the maximum batch size. This yields the network parameters of the fault diagnosis model for the target operating condition, and simultaneously determines the learnable parameter w and the class center c. c The possible values ​​of ;

[0070] Step 4: Online Diagnosis

[0071] Input the online samples of the target working condition into the trained model, and the output of the class label classifier is the fault diagnosis result.

[0072] To verify the fault diagnosis performance of this invention, a comparative experiment was designed, comparing this invention with three transfer learning methods: Local Maximum Mean Discrepancy (LMMD), DANN, and MADA. Table 2 shows the fault diagnosis accuracy of the four methods across three operating conditions. Figure 4The table shows the confusion matrices for the four methods transferring from condition 0 to condition 1. As can be seen from Table 2, our method achieves the highest fault diagnosis accuracy across the three condition transfers. LMMD is a transfer learning method based on difference metrics, while the other three methods are domain adversarial transfer learning methods. LMMD's diagnostic accuracy is lower than the other three methods; therefore, in industrial processes, domain adversarial transfer learning methods outperform difference-metric-based transfer learning methods in fault diagnosis. DANN aligns the marginal distributions of the source and target conditions, while MADA aligns the conditional distributions of the source and target conditions. DANN's diagnostic accuracy is higher than MADA's; therefore, marginal distributions are more important than conditional distributions in transfer learning for industrial processes. Our method's diagnostic accuracy is higher than the other three methods; therefore, aligning the marginal and conditional distributions of the source and target conditions is beneficial for extracting domain-invariant features, and learnable parameters can better evaluate the importance of conditional and marginal distributions.

[0073] Table 2. Fault diagnosis accuracy of the four methods

[0074]

[0075] To illustrate the ability of this invention to extract domain-invariant features, the TSNE graphs of this invention are compared with those of three other methods, such as... Figure 5 The diagram shows the TSNE diagrams for the four methods of transitioning from operating condition 0 to operating condition 1.

[0076] Figure 5 In the TSNE graph of LMMD, the feature overlap between different categories of the source and target load cases is quite significant, indicating that LMMD has limited ability to extract domain-invariant features. This is consistent with the lower accuracy of LMMD compared to the other three methods shown in Table 3. Compared to DANN and MADA, the proposed method DADAN shows less feature overlap between different categories in its TSNE graph, indicating better inter-class separability. The above analysis demonstrates that DADAN is superior to the other three methods in extracting domain-invariant features.

Claims

1. A method for industrial process fault diagnosis based on dynamic adversarial domain adaptive networks, characterized in that... Includes the following steps: Step 1: Data Preprocessing 1) Obtain historical data for both the source and target operating conditions, where the source operating condition data consists of labeled samples, represented as follows: , For the sample set of source operating conditions, R Represents the set of real numbers. Indicates the first k Source operating condition samples collected at each sampling time. The number of source operating condition samples. m The number of variables representing the sample. One-hot labels for source operating condition samples. C This indicates the number of categories in the sample. The target condition is unlabeled samples, represented as... , Indicates the number of samples for the target operating condition. Indicates the first k Target operating condition samples collected at each sampling time; Specific variables are as follows: A feed, D feed, E feed, A+C feed, recirculation flow rate, reactor feed rate, reactor pressure, reactor class, reactor temperature, discharge rate, product separator temperature, product separator level, product separator pressure, product separator bottom flow rate, stripper class, stripper pressure, stripper bottom flow rate, stripper temperature, stripper flow rate, compressor power, reactor cooling water outlet temperature, separator cooling water outlet temperature, composition A to F of stream 6, composition A to H of stream 9, composition D to H of stream 11, D feed, E feed, A feed, A+C feed, compressor recirculation valve, discharge valve, separator tank liquid flow rate, stripper liquid product flow rate, stripper water flow valve, reactor cooling water flow rate, condenser cooling water flow rate, stirring speed; 2) Standardize the training data by calculating the mean and standard deviation of the source and target operating condition samples. The mean of the source operating condition samples is expressed as... The standard deviation is The average value of the target working condition is The standard deviation is The standardization of the source operating condition and the target operating condition is achieved through equations (1) and (2) respectively, resulting in standardized source operating condition samples. and target working condition sample : ； (2)； 3) To address the dynamic nature of industrial processes, dynamic data is obtained by taking sliding windows from standardized source and target operating condition samples. k The sliding window for the source and target operating condition samples at each sampling time is: (3) (4)； in w This represents the width of the sliding window, with a sliding window step size of 1. The source operating data is obtained. A sliding window is used to obtain the target operating condition data. A sliding window is used as training data; Step 2: Model Building A dynamic adaptive domain adversarial network model is established, consisting of four parts: a feature extractor, a class label classifier, a global domain discriminator, and a local domain discriminator. The feature extractor is a convolutional neural network (CNN) used to extract features. Its structure is: Convolutional Layer 1 - BatchNorm Layer 1 - Max Pooling Layer 1 - Convolutional Layer 2 - BatchNorm Layer 2 - Convolutional Layer 3 - BatchNorm Layer 3 - Dropout Layer. Convolutional Layer 1 has a 3x3 kernel size, a stride of 1, 64 channels, and uses ReLU activation. BatchNorm Layer 1 has a feature dimension of 64, and Max Pooling Layer 1 has a stride of 2. Convolutional Layer 2 has a 2x4 kernel size, a stride of 1, 50 channels, and uses ReLU activation. BatchNorm Layer 2 has a feature dimension of 50. Convolutional Layer 3 has a 3x4 kernel size, a stride of 1, 50 channels, and uses ReLU activation. The BatchNorm layer 3 has a feature dimension of 50, and the Dropout layer has a neuron zeroing probability of 0.

4. A label classifier is used to predict the fault category of source and target condition samples. The structure is: Fully Connected Layer 1 - BatchNorm Layer 1 - Fully Connected Layer 2 - BatchNorm Layer 2 - Dropout Layer - Softmax Layer. Fully Connected Layer 1 has 500 neurons with ReLU activation. BatchNorm Layer 1 has a feature dimension of 500. Fully Connected Layer 2 has 100 neurons with ReLU activation. BatchNorm Layer 2 has a feature dimension of 100. The Dropout layer has a neuron zeroing probability of 0.4, and the Softmax layer has... C The global discriminator is used to determine whether a sample belongs to the source condition or the target condition, aligning the edge distribution of the source and target condition samples. Its structure is: Fully connected layer 1 - BatchNorm layer 1 - Fully connected layer 2 - BatchNorm layer 2 - Dropout layer - Softmax layer. Fully connected layer 1 has 500 neurons with the ReLU activation function, BatchNorm layer 1 has a feature dimension of 500, fully connected layer 2 has 100 neurons with the ReLU activation function, BatchNorm layer 2 has a feature dimension of 100, the Dropout layer has a neuron zeroing probability of 0.4, and the Softmax layer has 2 neurons. The local discriminator has... C The, the c The local discriminant is used to determine the first... c Whether a sample belongs to the source or target operating condition, align the conditional distributions of the source and target operating condition samples during training. C The local discriminators have no specific order, and each local discriminator has the same structure: fully connected layer 1 - BatchNorm layer 1 - fully connected layer 2 - BatchNorm layer 2 - Dropout layer - Softmax layer. The fully connected layer 1 has 500 neurons and uses ReLU activation function. The BatchNorm layer 1 has a feature dimension of 500. The fully connected layer 2 has 100 neurons and uses ReLU activation function. The BatchNorm layer 2 has a feature dimension of 100. The Dropout layer has a neuron zeroing probability of 0.

4. The Softmax layer has 2 neurons. Step 3: Model Training 1) Train a classifier using source and target operating condition data. The loss function of the classifier is: ； in θ f These are the parameters of the feature extractor. θ y These are the parameters of the label classifier. G f Indicates feature extractor, G y Represents a label classifier. Labels indicating source operating condition samples, Indicates the number of source operating condition samples. This is a hyperparameter with a value of 0.0001. loss cc The common center loss function is expressed as: （6） in f i This represents the source or target operating condition features extracted by the last fully connected layer of the class label classifier. c c The class centers are randomly initialized learnable parameters. The features of the source and target operating conditions share a common set of class centers. The samples have... C Class, then set C Each class center improves the intra-class compactness and inter-class separability of source and target condition samples by minimizing the distance between features and their corresponding class centers. 2) Train a global discriminator using samples from the source and target operating conditions, performing supervised learning. The loss function is the binary cross-entropy loss function: (7) in d i This represents the field label, with a value of 0 or 1. 0 indicates that the sample belongs to the source operating condition, and 1 indicates that the sample belongs to the target operating condition. G d Represents a global domain discriminator. θ d This represents the parameters of the global domain discriminator. x i Training data representing the source or target operating conditions; 3) The local domain discriminator is also trained using samples from the source and target operating conditions, with the loss function being: ;in Indicates the first c A local domain discriminator, The label represents the sample. If the sample is from the source working condition, the label is its true label; if the sample is from the target working condition, the label is the label predicted by the label classifier. C Indicates the number of categories in the sample. Indicates the first c Parameters of the class locality discriminator n c Indicates the first c The number of samples for the source and target operating conditions; 4) The overall loss function consists of the loss functions of the label classifier, the global discriminator, and the local discriminator, and is expressed as: (9) in λ For hyperparameters, λ The value is 0.

4. w This is a learnable parameter, initially set to 0, used to adaptively evaluate the relative importance of marginal and conditional distributions. e This represents the natural constant, with a value of 2.71828. 5) The model is trained by minimizing the overall loss function. The optimization algorithm is Adam, the learning rate is 0.001, and the maximum batch size is 300. Training stops when the overall loss function no longer decreases after 50 consecutive batches or reaches the maximum batch size. This yields the network parameters of the fault diagnosis model for the target operating condition, and simultaneously determines the learnable parameters. w and class center c c The possible values ​​of ; Step 4: Online Diagnosis Input the online samples of the target working condition into the trained model, and the output of the class label classifier is the fault diagnosis result.

Images