Power electronic device fault early warning method based on multi-physics field coupling

Abstract

The invention discloses a power electronic device fault early warning method based on multi-physics field coupling, and the method comprises the steps: (1) collecting aging data of a power electronic device under the condition of electric field and thermal field coupling, and dividing the aging data into a source domain data set and a target domain data set; (2) establishing a power electronic device aging parameter screening method, selecting a parameter which can best represent an aging rule as an aging precursor, and converting discrete voltage and current signals into time sequence signals which can be accepted by a neural network; (3) compressing the time sequence signal set into sliding window graph nodes, calculating a distance relationship between the graph nodes, and obtaining a graph matrix containing data structure information; (4) building a space-time fusion fault predictor, wherein the space-time fusion fault predictor comprises a feature extractor, a time sequence predictor and a domain discriminator; and (5) a weak supervision adversarial learning strategy is adopted, a small amount of target domain label data participates in the model training process, and the source domain prediction loss, the target domain prediction loss and the discrimination loss are combined for training.

Description

Technical Field

[0001] The present invention belongs to the field of fault prediction of power electronic devices, and particularly relates to a fault warning method for power electronic devices based on multi-physical field coupling. Background Art

[0002] Under the guidance of the new energy strategy and the support of the intelligent manufacturing industry, the development of new energy vehicles is gradually maturing. However, charging has become a shortcoming of the industrial development, significantly lagging behind the development of electric vehicles. The safety anxiety of users towards charging piles has become a constraint and obstruction to the development of new energy vehicles. The most critical component in a charging pile is the charging module, which is related to the overall performance and charging safety of the charging pile. The power electronic devices in the charging module include power devices, capacitors, resistors, etc. Among them, the power device can realize electric energy conversion and circuit control, and the power device is prone to failure in an environment with high temperature difference, high humidity, and high vibration. The existing super charging pile industry requires the power device to have high reliability. Therefore, it is necessary to explore the reliability research method of power electronic devices, so as to reduce the maintenance cost of the power electronic device system, reduce the probability of failure, and improve the reliability of the charging pile.

[0003] The electric field, thermal field, and stress field involved in the actual operation of power electronic devices will interact and couple with each other on the internal structure of the module. Under the action of these physical fields, different failure types will be caused due to different material parameters of the internal structure. Therefore, traditional models and data-driven methods are difficult to extract the coupling features with obvious degradation trends, resulting in low fault prediction accuracy. As a branch of machine learning, deep learning can learn the complex coupling relationships in multi-modal data and can realize the intelligent recognition and prediction of input data without complex model assumptions. Therefore, the fault prediction method of power electronic devices based on deep learning can better explore the internal relationships and non-linear laws between data and improve the fault prediction accuracy. However, deep learning models require a large amount of labeled data for training. However, in the actual application scenario of power electronic devices, the equipment is usually in a normal operating state and it is difficult to operate with faults. Therefore, collecting sufficient labeled training data is usually costly, time-consuming, and even unrealistic. In addition, the generalization ability of existing deep models is limited. When dealing with fault prediction under new working conditions, it is often necessary to retrain the model to obtain satisfactory prediction results, which greatly increases the computational cost.

[0004] In the Chinese patent “A state detection method for IGBT devices (CN117233566A)”, Hu Jie et al. performed trust transfer learning based on the low-dimensional feature vector of each sample in the high-level feature dimensionality reduction data, thereby obtaining a prediction model that can effectively predict different fault types and trust states, thereby improving the generalization ability and scalability of the model. However, the above invention only solves the domain generalization problem between different IGBT fault features, and is difficult to apply to the fault prediction problem of different IGBT devices, and is only suitable for power electronic device state detection under a single physical field. Therefore, it is necessary to study a fault prediction model that can extract multi-physical field coupling features and has strong generalization to adapt to the challenges brought by dynamic changes and uncertainty of data in industrial scenarios. Summary of the invention

[0005] In order to solve the problems existing in the prior art, the present invention provides a deep migration fault prediction method for multi-physical field coupling. The method establishes a method for screening aging parameters of power electronic devices, which can eliminate redundant information in the state monitoring data and retain the aging parameters that best reflect the degradation trend of the device; at the same time, a sliding window spatiotemporal fault predictor is constructed, which can extract class information, domain information and structural information in the time series data, and adaptively predict the failure time of power electronic devices; in order to further improve the generalization ability of the prediction model under different working conditions, a weakly supervised adversarial learning training strategy is designed to optimize the prediction model, so that the model can achieve satisfactory prediction accuracy in both the source domain and the target domain.

[0006] The power electronic device fault prediction method based on multi-physical field coupling provided by the present invention inputs the test data of the power electronic components into a pre-trained universal domain adaptive fault prediction model, and outputs the prediction result; the universal domain adaptive fault prediction model adopts a weakly supervised adversarial learning strategy, a small amount of target domain label data participates in the model training process, and the three losses of source domain prediction loss, target domain prediction loss and domain discrimination loss are combined for training.

[0007] Furthermore, the training of the general domain adaptive fault prediction model includes the following steps:

[0008] S1: Collect aging data of power electronic components under electric field and thermal field coupling conditions (including collector-emitter current, collector-emitter voltage and gate voltage), and divide them into source domain data set and target domain data set;

[0009] S2: Select the parameters that best characterize the aging laws of power electronic components as aging precursors; and convert discrete voltage and current signals into timing signals that can be accepted by the neural network;

[0010] S3: Compress the set of timing signals into sliding window graph nodes, calculate the distance relationship between each graph node, and obtain a graph matrix containing data structure information;

[0011] S4: Construct a spatio-temporal fusion fault prediction framework, including a feature extractor, a timing predictor, and a domain discriminator;

[0012] S5: Adopt a weakly supervised adversarial learning strategy, with a small amount of target domain label data participating in the model training process, and combine three losses: source domain prediction loss, target domain prediction loss, and domain discrimination loss for training.

[0013] Further, in step S1, the aging data is divided into a source domain dataset and a target domain dataset where represents the source domain data, represents the source domain data label, represents the target domain data, represents a small amount of target domain data labels, n s represents the number of source domain label data, n t represents the number of target domain data. Assume that the failure cycle numbers of power electronic components in the source domain and target domain are N s and N t , respectively, then N c represents the current cycle number.

[0014] Further, in step S2, first, use monotonicity and robustness indicators to screen parameters with significant aging characteristics. Then, convert discrete data points into a timing set, c i = Window(X n ), i = N - w, X n represents the set of screened aging data, Window represents the time window function, w represents the window length, and c i represents the timing set.

[0015] Further, in step S3, intercept the timing signal with the length of w and assign different weight parameters to the intercepted signal set to obtain compressed graph nodes; then, use the Radius distance to measure the distance between each graph node and establish an adjacency matrix of the graph nodes.

[0016] Further, in step S4, a feature extractor G f is constructed to extract spatial structure information; a source domain fault predictor G p and a target domain fault predictor are constructed to extract time scale information; a domain discriminator G d is constructed to identify whether the data belongs to the source domain or the target domain.

[0017] S41: The feature extractor G f includes a convolutional graph embedding network (CNN-GraphSage), which uses CNN to extract the fault information of a single node and defines a mean function AGGREGATE(·) and an aggregation function CONCAT(·) to aggregate the neighborhood information of nodes.

[0018] S42: The source domain fault predictor G p includes a regularized long short-term memory network (Dropout-BiLSTM), which consists of a forward LSTM network (F_LSTM) and a backward LSTM network (B_LSTM), and uses a Dropout layer to prevent overfitting. The target domain fault predictor has the same structure as the source domain fault predictor G p is the same.

[0019] S43: The domain discriminator G d contains two layers of fully connected networks (FC) and uses the Softmax function to determine whether the data belongs to the source domain or the target domain.

[0020] Furthermore, in step S5, the weakly supervised adversarial learning strategy includes a weakly supervised learning strategy and a conditional domain adversarial learning strategy, which can learn the non-linear mapping relationship between different domain features. The specific implementation steps are as follows:

[0021] S51: Construct the objective function of the source domain fault predictor G p : denotes the error of G p during the training process, L p (·) denotes the loss function of G p , θ f denotes the learnable parameters of G f , θ p denotes the learnable parameters of G p , denotes the source domain data, denotes the source domain data label, n s denotes the number of source domain label data;

[0022] S52: Construct the objective function of the target domain fault predictor : denotes the error during the training process, denotes the loss function of denotes the learnable parameters of, denotes the small amount of label information in the target domain, Denote the target domain data, n t Denote the number of target domain data;

[0023] S53: Based on the conditional domain adversarial idea, construct the domain discriminator G d 's objective function:

[0024] Denote the non-linear mapping function, Denote G d The error during the training process, θ d Denote G d 's learnable parameters, Denote the source domain data, Denote the source domain data labels, n s Denote the number of source domain label data, Denote the small amount of label information in the target domain, Denote the target domain data, n t Denote the number of target domain data;

[0025] S54: The optimization training process of the said objective function includes: θ f Denote G f 's learnable parameters, θ p Denote G p 's learnable parameters, θ d Denote G d 's learnable parameters, Denote 's learnable parameters, argmax means to take the maximum value, argmin means to take the minimum value.

[0026] Compared with the prior art, the beneficial effects of the present invention are at least as follows:

[0027] 1. In order to reduce the redundant information in the aging dataset and improve the data utilization rate. The present invention proposes a method for evaluating the state monitoring signals of power electronic components. It uses monotonicity and robustness indicators to screen the parameters with significant aging characteristics, which helps to reduce the complexity of the fault prediction model and improve the model training efficiency.

[0028] 2. In order to capture the structural information in the non-stationary data of power electronic components, the present invention assigns different weight parameters to the intercepted signal set, compresses the time series data into the nodes of the sliding window graph, and uses Radius to calculate the distance between each graph node, and finally converts it into a graph information matrix containing rich structural information.

[0029] 3. To overcome the problem of poor network prediction accuracy caused by the mismatch of spatio-temporal correlation information, the present invention constructs a spatio-temporal fusion fault prediction framework, which can automatically learn the unified representation of spatio-temporal correlation interaction information, provides a promising solution for the practical deployment of intelligent fault prediction algorithms, and will help improve industrial operation efficiency and equipment reliability.

[0030] 4. To solve the domain shift problem of power electronic components under different working conditions, the present invention proposes a weakly supervised adversarial learning strategy to achieve knowledge transfer under cross-working conditions, ensures that the model can achieve fast generalization in complex scenarios, and provides a feasible and effective solution for model self-regulation and optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 is the overall flowchart of a power electronic device fault warning method provided by an embodiment of the present invention.

[0032] Figure 2 is the flowchart of a weakly supervised adversarial learning strategy provided by an embodiment of the present invention.

[0033] Figure 3 is a schematic diagram of the RMSE index of the method proposed in an embodiment of the present invention in four types of transfer tasks.

[0034] Figure 4 is a schematic diagram of the SF index of the method proposed in an embodiment of the present invention in four types of transfer tasks. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0035] In order to make the technical solutions and objectives of the present invention clearer and more understandable, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific implementation steps described herein are only used to better illustrate the application of the present invention, but the technical features involved in the implementation mode of the present invention are not limited thereto.

[0036] Please refer to Figure 1 , a power electronic device fault warning method based on multi-physical field coupling in this embodiment specifically includes the following steps:

[0037] Step 1: Collect the aging data of IGBTs under different working conditions and divide them into a source domain dataset and a target domain dataset where represents the source domain data, represents the source domain data label, represents the target domain data, represents a small amount of target domain data labels, n s represents the number of source domain label data, n trepresents the number of target domain data. Assume that the failure cycle numbers of power electronic components in the source domain and the target domain are N s and N t ,but N c Indicates the current cycle number.

[0038] Step 2: Use the monotonicity and robustness indicators to screen the parameters with significant aging characteristics in the IGBT data. The determination principle of the monotonicity indicator is: The determination principle of the robustness index is: X represents the aging data set, x n Indicates t n The data point at the moment, N represents the data length, smoothed_x n represents the smoothed data points, and represents the number of numerical points in the data set whose derivative is greater than 0 and less than 0. Then, the discrete data points are converted into a time series set using the time window function. i =Window(X n ),i=Nw,X n represents the filtered aging data set, Window represents the time window function, w represents the window length, c i Represents a time series collection.

[0039] Step 3: Compress the time series signal set into a sliding window graph node. The time series signal is intercepted with a length of w, and different weight parameters are assigned to the intercepted signal set to obtain the compressed graph node. Specifically expressed as: ν m =W*[c0,c1,...,c i ], W represents the weight matrix, ν m represents the graph node, and m represents the length of the graph node. Then, the distance between each graph node is measured using the Radius distance to obtain the node ν m Neighboring nodes, specifically expressed as: Ne(ν m )=Radius(ν m ,V),ifRadius(ν m ,V)>0, Radius(·) represents the cosine distance measurement formula, Ne represents the set of adjacent nodes, and V represents the set of graph nodes. Finally, based on the distance measurement relationship between each node, the adjacency matrix of the graph nodes is established.

[0040] Step 4: In order to capture the spatiotemporal degradation information of IGBT, a spatiotemporal fusion fault prediction framework is constructed. It includes the feature extractor G f , source domain fault predictor G p , Target Domain Fault Predictor And domain discriminator G d . The specific implementation steps are as follows:

[0041] S41: The feature extractor G f includes a convolutional graph embedding network (CNN-GraphSage). The convolutional kernel of the CNN overlaps and slides on the feature map to extract detailed fault information; GraphSage defines a mean function AGGREGATE(·) and an aggregation function CONCAT(·) to aggregate the neighborhood information of nodes. For example, for a certain graph node ν m there is a neighborhood set After aggregating k times, the graph feature can be expressed as: where represents the set of node features in the neighborhood of node ν m and represents the single node feature of the adjacent node of node ν m , and the finally extracted graph feature, Θ k represents learnable parameters, and σ represents the mapping function Relu.

[0042] S42: The source domain fault predictor G p includes a regular long short-term memory network (Dropout-BiLSTM), which consists of a forward LSTM network (F_LSTM) and a backward LSTM network (B_LSTM), and uses a Dropout layer to prevent overfitting. F_LSTM and B_LSTM have the same network structure. F_LSTM receives the sequence in the original order, while B_LSTM receives the same input in the reverse order, and their outputs are combined into the final prediction result as follows: α represents the output weight of F_LSTM, β represents the output weight of B_LSTM, and o t represents the output of the BiLSTM network. The target domain fault predictor and the source domain fault predictor G p have the same structure.

[0043] S44: The domain discriminator G d includes two layers of fully connected networks (FC), and uses the Softmax function to determine whether the data belongs to the source domain or the target domain.

[0044] Step five: Optimize the training model using a weakly supervised adversarial learning strategy, mainly including a weakly supervised learning strategy and a conditional domain adversarial learning strategy. Please refer to Figure 2 . The specific implementation steps are as follows:

[0045] S51: Construct the objective function of the source domain fault predictor G p : Denote G p The error during training, L p (·) denotes G p 's loss function, θ f Denote G f 's learnable parameters, θ p Denote G p 's learnable parameters, Denote the source domain data, Denote the source domain data labels, n s Denote the number of source domain label data;

[0046] S52: Construct the target domain fault predictor 's objective function: Denote The error during training, Denote 's loss function, Denote 's learnable parameters, Denote a small amount of label information in the target domain, Denote the target domain data, n t Denote the number of target domain data;

[0047] S53: Based on the conditional domain adversarial idea, construct the domain discriminator G d 's objective function:

[0048] Denote the non-linear mapping function, Denote G d The error during training, θ d Denote G d 's learnable parameters, Denote the source domain data, Denote the source domain data labels, n s Denote the number of source domain label data, Denote a small amount of label information in the target domain, Denote the target domain data, n t Denote the number of target domain data;

[0049] S54: The optimization training process of the said objective function includes: θ f Denote G f 's learnable parameters, θ p Denote G p 's learnable parameters, θ d Denote G dLearnable parameters, denote Learnable parameters, argmax represents taking the maximum value, and argmin represents taking the minimum value.

[0050] In step S5, the source domain prediction loss is the objective function of the source domain fault predictor G p of The target domain prediction loss is the objective function of the target domain fault predictor of and the domain discrimination loss is the objective function of the domain discriminator G d of Define an overall loss where λ takes the value of 10.

[0051] The present invention will be further described below in conjunction with the accompanying drawings and experimental cases. To evaluate the performance of the proposed method, the publicly available IGBT aging dataset is used. Four sets of IGBT aging data (Device2, Device3, Device4, Device5) under different working conditions from the University of Padova in Italy are selected to test the proposed prediction network. In this experiment, the package temperature is controlled within a range outside the rated temperature of the device, aiming to accelerate the aging of the device. Several parameters are monitored, such as the collector current, collector voltage, gate voltage, and package temperature.

[0052] Visualize the voltage, current, and temperature data in the IGBT monitoring data, and apply the aging parameter screening method in step two. The evaluation results are shown in Table 1. The collector-emitter voltage (V CE ) has strong monotonicity throughout the entire life cycle, can correctly characterize the degradation process, and has stronger robustness. Select V CE parameter as the failure parameter for evaluating the IGBT failure time.

[0053] Table 1 Evaluation results of IGBT monitoring parameters

[0054]

[0055] The IGBT datasets are respectively labeled as datasets Device2, Device3, Device4, and Device5, and all four datasets can be used as the source domain or the target domain. It is assumed that there are the same proportion of collector-emitter voltage (V CE ) samples in the target domain dataset. For example, the migration task Case 0 (20% target domain labeled) means using datasets Device2, Device3, and Device4 as the source domain and Device5 as the target domain.

[0056] This invention compresses a set of time series signals into sliding window graph nodes. It intercepts time series signals with a length of w = 5 and assigns different weight parameters to the intercepted signal set to obtain compressed graph nodes. The spatio-temporal fusion fault predictor designed in this invention includes a feature extractor G f , a source domain fault predictor G p , a target domain fault predictor , and a domain discriminator G d . The feature extractor G f includes 2 Conv blocks, 2 GConv blocks, and a BN block. The filter size of the Conv block is 3×3. The filter size of the previous GConv block is 128, and the filter size of the latter GConv block is 32. The filter size of the BN block is the same as that of the GConv block. The source domain fault predictor G p and the target domain fault predictor have the same structure and parameters. The BiLSTM is connected by the same output weights in different directions of the LSTM, with a kernel size of 32 and a dropout rate of 0.2 for network parameters. The domain discriminator G d consists of two FC blocks, with 8 and 16 nodes respectively. Using the Softmax activation function, it assigns a binary label 1 to the source data and a binary label 0 to all target data.

[0057] The performance of the fault predictor designed in this invention in four types of transfer tasks is as shown in Figure 3 and 4 . Using the common model evaluation metrics SF and RMSE to evaluate the model performance, in the four types of transfer tasks, the SF and RMSE metrics of the model are the smallest in Case2, and the SF and RMSE metrics of the model in all tasks are relatively low, indicating strong generalization ability.

[0058] In summary, in view of the problems of low utilization rate of structural information and poor model generalization ability between different working conditions in the cross-domain fault prediction of power electronic devices, this invention designs a fault warning method for power electronic devices based on multi-physical field coupling, realizes the screening of multi-source information under the background of multi-physical field coupling, and enables the model to quickly adapt and accurately predict faults in multiple different transfer tasks, which has high application value in actual industrial production.

[0059] For the fault warning method for power electronic devices based on multi-physical field coupling disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the description in the method section.

[0060] Finally, it should be noted that although the implementation of the present invention has been described in detail with reference to the examples, it is easy for those skilled in the art to understand that any modifications, substitutions, improvements, etc. made within the spirit and principles of the present invention described in the appended claims should be included within the protection scope of the present invention.

Claims

1. A method for predicting power electronic device faults based on multi-physical field coupling, characterized in that: The test data of power electronic components is input into a pre-trained universal domain adaptive fault prediction model, and the prediction results are output; the universal domain adaptive fault prediction model adopts a weakly supervised adversarial learning strategy, a small amount of target domain label data participates in the model training process, and the three losses of source domain prediction loss, target domain prediction loss and domain discrimination loss are combined for training.

2. The power electronic device fault prediction method based on multi-physical field coupling according to claim 1 is characterized in that: The training of the general domain adaptive fault prediction model includes the following steps: S1: Collect aging data of power electronic components under the coupled electric and thermal field conditions; S2: Select the parameters that best characterize the aging laws of power electronic components as aging precursors; and convert discrete voltage and current signals into timing signals that can be accepted by the neural network; S3: Construct sliding window graph nodes and calculate the distances between graph nodes to obtain a graph matrix containing data structure information; S4: Build a spatiotemporal fusion fault predictor, including feature extractor, timing predictor and domain discriminator; S5: A weakly supervised adversarial learning strategy is adopted. A small amount of target domain data labels are involved in the model training process. The source domain prediction loss, target domain prediction loss, and domain discrimination loss are combined for training.

3. The power electronic device fault prediction method based on multi-physical field coupling according to claim 2 is characterized in that: In step S1, the aging data is divided into source domain data sets and target domain dataset in Represents source domain data, represents the source domain data label, represents the target domain data, represents a small amount of target domain data labels, n s Indicates the number of source domain label data, n t represents the number of target domain data; assuming that the failure cycle numbers of power electronic components in the source domain and the target domain are N s and N t ,but represents the current cycle number in source domain training, The value range is represents the current cycle number in the target domain training, The value range is 4. The power electronic device fault prediction method based on multi-physical field coupling according to claim 2 is characterized in that: Step S2 includes: using the monotonicity index and the robustness index to screen parameters with significant aging characteristics; the determination principle of the monotonicity index is: The determination principle of the robustness index is: X represents the aging data set, x n Indicates t n The data point at the moment, N represents the data length, smoothed_x n represents the smoothed data points, and represents the numerical points in the data set where the derivative is greater than 0 and less than 0, x n Indicates the aforementioned t n The data point at the time, Represents the partial derivative of the current point, n represents the position of the data point, and the value range of n is n∈[0,N].

5. The method for predicting power electronic device faults based on multi-physical field coupling according to claim 2, characterized in that: In step S3, the signal is intercepted by the time window function, and the time window length is w to intercept the time series signal, and multiple signal sets with length w are obtained, and different weight parameters are assigned to the intercepted signal sets to obtain the compressed graph nodes; then, the distance between each graph node is measured by the Radius distance to obtain all graph nodes ν m adjacent nodes; finally, according to the distance measurement relationship between each node, a graph matrix containing data structure information is established, where the data structure information refers to the distance measurement relationship between the graph node and its own adjacent nodes.

6. The method for predicting power electronic device faults based on multi-physical field coupling according to claim 2, characterized in that: In step S4, a feature extractor G is constructed f , used to extract spatial structure information; Construct a timing predictor, which includes a source domain fault predictor G p and target domain fault predictor Used to extract time scale information; construct a domain discriminator G d , used to identify whether the data belongs to the source domain or the target domain.

7. The method for predicting power electronic device faults based on multi-physical field coupling according to claim 6, characterized in that: In step S4, The feature extractor G f It includes a convolutional graph embedding network (CNN-GraphSage), which uses CNN to extract fault information of a single graph node, and defines the mean function AGGREGATE(·) and aggregation function CONCAT(·) to aggregate the neighborhood information of the node; The source domain fault predictor G p Including a regularized long short-term memory network (Dropout-BiLSTM), which consists of a forward LSTM network (F_LSTM) and a reverse LSTM network (B_LSTM), and uses a Dropout layer to prevent overfitting; the target domain fault predictor and the source domain fault predictor G p The structure is the same; The domain discriminator G d It contains two layers of fully connected networks (FC) and uses the Softmax function to determine whether the data belongs to the source domain or the target domain.

8. The method for predicting power electronic device faults based on multi-physical field coupling according to claim 2, characterized in that: In step S5, the weakly supervised adversarial learning strategy includes a weakly supervised learning strategy and a conditional domain adversarial learning strategy, which can learn the nonlinear mapping relationship between features in different fields; specifically, the steps include: S51: Construct source domain fault predictor G p The objective function is: Represents G p The error during training, L p (·) indicates G p The loss function, θ f Represents G f The learnable parameters, θ p Represents G p The learnable parameters of Represents source domain data, represents the source domain data label, n s Indicates the number of source domain label data; S52: Building a target domain fault predictor The objective function is: express The error during training, express The loss function is express The learnable parameters of Represents a small amount of label information in the target domain, represents the target domain data, n t Indicates the amount of target domain data; S53: Based on the idea of ​​conditional domain confrontation, construct a domain discriminator G d The objective function is: represents a nonlinear mapping function, Represents G d The error during training, θ d Represents G d The learnable parameters of Represents source domain data, represents the source domain data label, n s Indicates the number of source domain label data, Represents a small amount of label information in the target domain, represents the target domain data, n t Indicates the amount of target domain data; S54: The optimization training process of the objective function includes: θ f Represents G f The learnable parameters, θ p Represents G p The learnable parameters, θ d Represents G d The learnable parameters of express The learnable parameters are: argmax means taking the maximum value, and argmin means taking the minimum value.

9. The method for predicting power electronic device faults based on multi-physical field coupling according to claim 2, characterized in that: In step S5, the source domain prediction loss is the source domain fault predictor G p The objective function The target domain prediction loss is the target domain fault predictor The objective function And the domain discrimination loss is the domain discriminator G d The objective function 10. The method for predicting power electronic device faults based on multi-physical field coupling according to claim 2, characterized in that: In step S5, an overall loss is defined Participate in the overall training optimization of the model, λ represents the coefficient of balancing the numerical value of the loss.

Images