Method and system for evaluating service life of power electronic device based on electrical and thermal stress

Abstract

The invention discloses a power electronic device life evaluation method and system based on electrical and thermal stress. The method comprises the following steps: collecting operation data of the power electronic component under electric and thermal stress, selecting aging parameters from the operation data, and dividing the aging parameters into a source domain data set and a target domain data set; establishing a two-way feature extraction unit of the power electronic device, and extracting a global feature including a long-term dependency relationship between time steps and a local feature including local time dynamic information at the same time; establishing a power electronic device multi-mode feature fusion unit, and effectively fusing the degradation features of each specific mode through modeling of interaction between the features; constructing a prediction network based on multi-modal feature fusion, wherein the prediction network comprises a feature extractor, a time sequence predictor and a domain discriminator; an unsupervised adversarial learning strategy is adopted, unlabeled data of a target domain participates in a model training process, and life prediction loss and domain discrimination loss are combined for training.

Description

Technical Field

[0001] The present invention belongs to the field of life assessment of power electronic devices, and in particular relates to a method and system for life assessment of power electronic devices based on electrical and thermal stress. Background Art

[0002] Power electronic devices are core components of electronic devices such as converters and inverters, and play a vital role in energy conversion and transmission. Currently, power electronic devices have been widely used in various cutting-edge technology fields, including new energy vehicles, power systems, modern electronics, and aerospace. For example, in photovoltaic (PV) power generation systems, low reliability will increase maintenance costs and reduce system availability, thereby increasing the levelized cost of electricity, which will ultimately affect its market penetration. Offshore wind farms also have similar problems, and their maintenance is inconvenient or not cheap. In automotive, locomotive and avionics applications, safety requirements require almost zero fault tolerance. Therefore, it is necessary to explore reliability research methods for power electronic devices to reduce the maintenance cost of power electronic device systems, reduce the probability of failure, and improve the reliability of electronic equipment.

[0003] To ensure the reliable and stable operation of power electronic devices under adverse conditions, researchers typically employ online monitoring and data analysis techniques. This enables comprehensive analysis and modeling of the degradation process of power electronic devices, thereby predicting the remaining useful life (RUL) and ensuring the safety and reliability of the entire system. Deep learning, a key branch of machine learning, is also increasingly being applied to RUL prediction for power electronic devices, leveraging its ability to automatically learn features. However, power electronic devices operate under coupled electrical and thermal stresses, and interactions between their internal components lead to a degradation process that involves the combined influence of internal components and external environmental factors. Existing deep learning methods only consider single modal features and struggle to capture the complex nonlinear relationships and time-varying characteristics of power electronic device failure processes. Furthermore, aging tests for power electronic devices typically require a long time to simulate device degradation under real-world operating conditions, and test conditions often exceed normal operating parameters, resulting in prohibitive costs. Consequently, collecting sufficient labeled training data is often costly, time-consuming, or even impractical.

[0004] In the Chinese patent “Life Assessment Method, Apparatus and Computer Equipment for Power Electronic Devices (CN111079253A)”, Pan Guangze et al. obtained the optimal performance life distribution model corresponding to each performance degradation characteristic quantity based on the failure time corresponding to each performance degradation characteristic quantity; established a competitive failure model for each optimal performance life distribution model, and processed the competitive failure model based on the Monte Carlo method to obtain the life distribution model of the power electronic device, and output the life assessment value of the power electronic device based on the life distribution model. However, the above invention relies on manual experience to set the failure time of each degradation characteristic quantity, and it is difficult to capture the complex nonlinear relationship in the data, and the generalization ability for new data is also poor. Therefore, it is necessary to study a method and system for life assessment of power electronic devices to adapt to the challenges brought by dynamic changes and uncertainty of data in industrial scenarios. Summary of the Invention

[0005] To address the challenges of the existing technology, this paper provides a method and system for assessing the lifespan of power electronic devices based on electrical and thermal stresses. This method proposes a dual-path feature extraction method that simultaneously extracts global and local features from power electronic device aging data. Furthermore, a feature fusion unit is constructed to effectively integrate the degradation characteristics of each specific mode. To further improve the generalization capability of the prediction model in the absence of labeled data, an unsupervised adversarial learning training strategy is designed, enabling the model to achieve satisfactory prediction accuracy in both the source and target domains. The specific steps are as follows:

[0006] S1: Collect the operating data of power electronic components under electrical and thermal stress (including collector-emitter current, collector-emitter voltage, and gate voltage) and select aging parameters from them, dividing them into source domain data set and target domain data set;

[0007] S2: Establish a dual-path feature extraction unit for power electronic devices to simultaneously extract global features containing long-term dependencies between time steps and local features containing local temporal dynamic information;

[0008] S3: Establish a multimodal feature fusion unit for power electronic devices, which effectively fuses the degradation features of each specific mode by modeling the interaction between global and local features;

[0009] S4: Construct an adaptive lifespan prediction model based on multimodal feature fusion, including feature extractor, time series predictor and domain discriminator;

[0010] S5: The adaptive life prediction model adopts an unsupervised adversarial learning training strategy. The unlabeled data of the target domain participates in the adaptive life prediction model training process, and the life prediction loss is reduced. and domain discrimination loss Combined training is performed to obtain a trained adaptive lifespan prediction model;

[0011] S6: Input the data to be tested into the trained adaptive life prediction model and output the prediction results.

[0012] Furthermore, in step S1, based on the failure characteristics of power electronic devices, aging parameters that can characterize their degradation characteristics are determined. Afterwards, the aging data is divided into source domain data sets and target domain dataset in represents the i-th source domain data, represents the i-th source domain data label, represents the i-th target domain data, n s Indicates the number of source domain label data, n t Indicates 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 ,but N c Indicates the current cycle number.

[0013] Furthermore, in step S2, the dual-path feature extraction unit includes a CNN branch and a Transformer branch; the source domain data and the target domain data are input into the CNN branch and the Transformer branch of the dual-path feature extraction unit in the training and testing phases, respectively; wherein, the CNN branch relies on the convolution kernel to overlap and slide on the input source domain and target domain data to obtain a feature map rich in local details; the Transformer branch can utilize its own unique multi-head self-attention mechanism to capture sequence information from different angles such as the data's temporal dependency and the correlation between features, thereby enhancing the perception of global features.

[0014] Furthermore, in step S3, the multimodal feature fusion unit implements tensor fusion by modeling global and local modal features. It redefines the degenerate features using binary Cartesian products and expands the features with a dimension of 1, calculating the feature correlation between the two modalities while preserving the information of the native data modality.

[0015] Furthermore, in step S4, a feature extractor G is constructed. f , used to extract spatial structure information; a lifespan predictor G was constructed p , used to extract time scale information; construct the domain discriminator G d , used to identify whether the data belongs to the source domain or the target domain.

[0016] The feature extractor G fIt includes a dual-path feature extraction unit and a multimodal feature fusion unit; the two are connected in sequence and finally output the fusion degradation features of the power electronic device.

[0017] The lifespan predictor G p It includes two layers of fully connected networks (FC) and uses the ReLU function to achieve the mapping between degradation features and remaining lifespan.

[0018] The domain discriminator G d It includes a three-layer fully connected network (FC) and uses the Softmax function to determine whether the data belongs to the source domain or the target domain.

[0019] Furthermore, in step S5, the unsupervised adversarial learning strategy includes a conditional domain adversarial learning strategy, which can learn the nonlinear mapping relationship between features in different domains. The specific implementation steps are as follows:

[0020] S51: Constructing lifespan predictor G p Predicted loss of life

[0021] E Gp (θ f ,θ p ) represents G p The predicted loss of life, 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, represents the extracted fusion degradation features, Indicates the prediction result.

[0022] S52: Based on the idea of ​​conditional domain confrontation, construct a domain discriminator G d Domain discrimination loss:

[0023] represents a nonlinear mapping function, Represents G d Domain discrimination loss, 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, represents the extracted fusion degradation features, Represents the prediction result, Indicates the judgment result.

[0024] S53: The optimization training of the objective function utilizes features from the source domain and the target domain, and then these features are input into the domain discriminant G d , calculate the domain discrimination loss. In addition, the lifespan predictor G p Receive from feature extractor G f features to perform the prediction process: Therefore, using the formula and To optimize the training of the entire adaptive life prediction model, the parameter θ f ,θ p and θ d Represents the feature extractor G f , lifespan predictor G p and domain discrimination G d The trainable parameters of , argmin and argmax represent the minimum and maximum values ​​respectively.

[0025] The present invention also provides a power electronic device life assessment system based on electrical and thermal stress, which is used to implement the above-mentioned method, including the following modules: a data set construction module, which is used to collect status monitoring data of power electronic devices under different aging conditions and divide it into a source domain data set and a target domain data set; a feature extraction module, which is used to realize the extraction and fusion of global and local aging features of power electronic devices; a time series prediction module, which is used to establish a mapping relationship between multimodal degradation features and life of power electronic devices; a model training module, which is used to combine the two loss functions of life prediction loss and domain discrimination loss for training; and a prediction module, which is used to input the data to be tested into a trained multimodal feature-driven prediction network and output the prediction results.

[0026] Compared with the prior art, the present invention has at least the following beneficial effects:

[0027] To obtain multimodal feature information from power electronic devices, this paper proposes a dual-path feature extraction unit, consisting of a CNN branch and a Transformer branch. This unit can simultaneously process input aging data, forming a dual-path structure that extracts both global and local features, making feature extraction more efficient.

[0028] 2. To obtain combined features that contain both global and local information, this paper proposes a multimodal feature fusion unit. This unit redefines degenerate features using a binary Cartesian product. Furthermore, modal features are dimensionally expanded with a value of 1, calculating the feature correlation between the two modalities while preserving modality-specific information.

[0029] 3. In order to solve the domain shift problem of power electronic components under different working conditions, the present invention proposes an unsupervised adversarial learning strategy to realize the degradation knowledge transfer of unknown domains, reduce the model's dependence on labeled data, and ensure that the model can achieve rapid generalization in complex scenarios, providing a feasible solution for the application of the model in new industrial scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is an overall flow chart of a method for evaluating the life of a power electronic device provided by an embodiment of the present invention.

[0031] Figure 2 This is a dual-path feature extraction and fusion flow chart provided by an embodiment of the present invention.

[0032] Figure 3 This is a flow chart of an unsupervised adversarial learning strategy provided by an embodiment of the present invention.

[0033] Figure 4 Schematic diagram of the SF indicator of the method proposed in the embodiment of the present invention in four types of migration tasks.

[0034] Figure 5 3 is a schematic diagram of the RMSE indicator of the method proposed in the embodiment of the present invention in four types of migration tasks. DETAILED DESCRIPTION

[0035] In order to make the technical solutions and purposes of the present invention more clearly understood, the present invention is 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 methods of the present invention are not limited thereto.

[0036] See also Figure 1 The present invention provides a method and system for evaluating the life of power electronic devices based on electrical and thermal stress, which specifically includes the following steps:

[0037] Step 1: Collect the operating data of power electronic components under electrical and thermal stress (including collector-emitter current, collector-emitter voltage, and gate voltage). The collector-emitter voltage stands out due to its feasibility of online measurement, wide applicability, reliable calibration, high accuracy, linearity, and sensitivity. Divide into source domain datasets and target domain dataset in Represents source domain data, represents the source domain data label, represents the target domain data, n s Indicates the number of source domain label data, n t Indicates 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 ,but N c Indicates the current cycle number.

[0038] Step 2: Establish a dual-path feature extraction unit for power electronic devices, and simultaneously extract global features containing long-term dependencies between time steps and local features containing local time dynamic information (when the life assessment of power electronic devices needs to refer to the time series data before the current moment, the relationship between the data and the features beyond the current moment is expressed as a long-term dependency); the dual-path feature extraction unit includes a CNN branch and a Transformer branch; the CNN branch adopts a feature pyramid structure, and the resolution of the feature map extracted by the CNN branch decreases with the increase of the network depth, and the number of channels also increases. The entire branch is divided into five blocks, including four convolution blocks and one fully connected block. The fully connected block contains a linear layer to realize local feature mapping. The Transformer branch contains two sub-layers: a multi-head self-attention layer and a feedforward layer. The multi-head self-attention layer is designed to capture the dependencies between features. For the l-1 layer, its attention can be defined as: Among them H l-1 is the parallel attention function, and is the projection weight; after that, define Q, K and V to represent query, key and value, and the proportional dot product attention mechanism is defined as follows: Multi-head self-attention can be defined as: MultiHead(H l-1 )=[head1,head2,...,head h ]W O , where W O are trainable weights. The feed-forward layer consists of linear and nonlinear mapping channels, applied separately at each time step.

[0039] Step 3: Establish a multi-modal feature fusion unit for power electronic devices. Figure 2 By modeling the interaction between features, the degenerate features of each specific mode are effectively fused. The degenerate features are redefined using binary Cartesian products, and the dimension of the modal features is expanded with a value of 1 to produce a single modal feature: By embedding the dimension [z l 1] T 、[zv 1] T After that, each coordinate (z l ,z v ) can be regarded as tensors in Cartesian space. Then, the obtained features are outer-producted to obtain the bimodal features in tensor fusion: in represents the outer product of vectors, z l and z v It is a unimodal feature embedding, which constitutes the unimodal interaction in tensor fusion.

[0040] Step 4: Construct a prediction network based on multimodal feature fusion, including feature extractor G f , lifespan predictor G p and domain discriminator G d The specific implementation steps are as follows:

[0041] The feature extractor G f It includes a dual-path feature extraction unit and a multimodal feature fusion unit; the two are connected in sequence and ultimately output the fused degradation features of the power electronic device;

[0042] The lifespan predictor G p It includes two layers of fully connected networks (FC) and uses the ReLU function to achieve the mapping between degradation features and remaining lifespan.

[0043] The domain discriminator G d It includes a three-layer fully connected network (FC) and uses the Softmax function to determine whether the data belongs to the source domain or the target domain.

[0044] Step 5: Use unsupervised adversarial learning strategy to train the model, see Figure 3 The specific implementation steps are as follows:

[0045] S51: Constructing lifespan predictor G p Predicted loss of life Represents G p The predicted loss of life, 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, represents the extracted fusion degradation features, Indicates the prediction result.

[0046] S52: Based on the idea of ​​conditional domain confrontation, construct a domain discriminator G d Domain discrimination loss:

[0047] Represents a nonlinear mapping function. Represents G d Domain discrimination loss, 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, represents the extracted fusion degradation features, Represents the prediction result, In order to reduce the impact of uncertain prediction, the entropy criterion ω(H(p))=1+e is introduced. -H(p) Define the uncertainty of the prediction, the discriminator G d The objective function can be expressed as:

[0048] S54: Optimization training process of the objective function: (θ f ,θ p ,θ d ) participates in the adversarial training process and uses the Adam algorithm for iterative training. The network loss optimization process can be expressed as: λ represents the trainable weight parameter, and argmin represents the minimum value.

[0049] Step 6: Input the data to be tested into the trained adaptive life prediction model and output the prediction results.

[0050] The present invention also provides a power electronic device life assessment system based on electrical and thermal stress, which is used to implement the above-mentioned method, including the following modules: a data set construction module, which is used to collect status monitoring data of power electronic devices under different aging conditions and divide it into a source domain data set and a target domain data set; a feature extraction module, which is used to realize the extraction and fusion of global and local aging features of power electronic devices; a time series prediction module, which is used to establish a mapping relationship between multimodal degradation features and life of power electronic devices; a model training module, which is used to combine the two loss functions of life prediction loss and domain discrimination loss for training; and a prediction module, which is used to input the data to be tested into a trained multimodal feature-driven prediction network and output the prediction results.

[0051] The system of this embodiment can execute a power electronic device life assessment method based on electrical and thermal stress provided by the method embodiment of the present invention, can execute any combination of implementation steps of the method embodiment, and has the corresponding functions and beneficial effects of the method.

[0052] The present invention is further described below with reference to the accompanying figures and experimental cases. To evaluate the performance of the proposed method, a publicly available IGBT aging dataset was used. Four sets of IGBT aging data (Device 2, Device 3, Device 4, and Device 5) under different operating conditions from the University of Padova, Italy, were selected to test the proposed prediction network. The experiment controlled the package temperature outside the device's rated temperature in order to accelerate device aging. Several parameters were monitored, such as collector current, collector voltage, gate voltage, and package temperature.

[0053] The IGBT datasets are labeled as datasets Device2, Device3, Device4, and Device5. All four datasets can be used as source or target domains. Assume that the target domain dataset has the same ratio of collector-emitter voltage (V CE ) samples. For example, the migration task Case 0 (20% target domain labels) means that the datasets Device2, Device3, and Device4 are used as source domains, and Device5 is used as the target domain.

[0054] See also Figure 1 The feature extractor designed by the present invention includes a dual-path feature extraction unit and a feature fusion unit, where 1 represents the CNN branch, 2 represents the Transformer branch, and 3 represents the multimodal feature fusion unit. The CNN branch is divided into five blocks, including four convolution blocks and one fully connected block. The convolution blocks are based on 16, 32, 64, 128, and 256 2×2 filters of the same size. The Transformer branch includes a multi-head self-attention layer, a regularization layer, a feedforward layer, and a linear layer. The life expectancy predictor G designed by the present invention p It includes two layers of fully connected layers and Dropout layers. The domain discriminator G designed by the present invention d It consists of two FC blocks with 8 and 16 nodes respectively. Softmax activation function is used to assign binary label 1 to source data and binary label 0 to all target data.

[0055] The prediction performance of the multimodal feature fusion prediction network designed by the present invention after applying the unsupervised learning strategy is as follows: Figure 4 and Figure 5As shown. The common deep learning model evaluation indicators SF and RMSE are used to evaluate the results of the proposed life prediction model. RMSE is a statistic used to measure the error of the prediction model, while SF can evaluate the accuracy of the life prediction. These two indicators are calculated based on the final prediction results of each power electronic device. The smaller the value of the two indicators, the stronger the prediction ability of the model. In the four types of migration tasks, the indicators of SF and RMSE are relatively low, indicating that the method proposed in this invention not only captures the high-level degradation trend knowledge of power electronic devices, but also learns the unique information of power electronic devices. It also verifies the importance of unsupervised adversarial transfer learning strategy for achieving accurate life assessment of power electronic devices.

[0056] In summary, this paper addresses the issues of redundant degradation features, low utilization of multimodal features, and poor model generalization due to reliance on labeled data in power electronic device life assessment. By designing a power electronic device life assessment method and system based on electrical and thermal stresses, this method efficiently integrates global and local multimodal features of power electronic devices. Furthermore, an unsupervised adversarial transfer learning strategy is employed during network training, achieving high-precision life prediction across domains, demonstrating its high application value in practical industrial production.

[0057] As for the method and system for life assessment of power electronic devices based on electrical and thermal stress 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 of the method part.

[0058] Finally, it should be noted that although the implementation of the present invention has been described in detail with reference to examples, it is easy for those skilled in the art to understand that any modifications, substitutions and improvements made without departing from the spirit and principles of the present invention as described in the appended claims should be included in the scope of protection of the present invention.

Claims

1. A method for evaluating the life of power electronic devices based on electrical and thermal stress, characterized in that: The following steps are involved: S1: Collect operating data of power electronic components under electrical and thermal stress, select aging parameters from them, and divide them into source domain dataset and target domain dataset; S2: Establish a dual-path feature extraction unit for power electronic devices to simultaneously extract global features containing long-term dependencies between time steps and local features containing local temporal dynamic information; S3: Establish a multimodal feature fusion unit for power electronic devices, which effectively fuses the degradation features of each specific mode by modeling the interaction between global and local features; S4: Construct an adaptive lifespan prediction model based on multimodal feature fusion, including feature extractor, time series predictor and domain discriminator; S5: The adaptive life prediction model adopts an unsupervised adversarial learning training strategy. The unlabeled data of the target domain participates in the adaptive life prediction model training process, and the life prediction loss is reduced. and domain discrimination loss Combined training is performed to obtain a trained adaptive lifespan prediction model; S6: Input the data to be tested into the trained adaptive life prediction model and output the prediction results.

2. The method for evaluating the life of power electronic devices based on electrical and thermal stress according to claim 1, characterized in that: In step S1, based on the failure characteristics of power electronic devices, aging parameters that can characterize their degradation characteristics are determined; the aging data is divided into source domain data sets and target domain dataset in represents the i-th source domain data, represents the i-th source domain data label, represents the i-th target domain data, n s Indicates the number of source domain label data, n t Indicates the amount of target domain data.

3. The method for life assessment of power electronic devices based on electrical and thermal stress according to claim 1, characterized in that: In step S2, the dual-path feature extraction unit includes a CNN branch and a Transformer branch; the source domain data and the target domain data are input into the CNN branch and the Transformer branch of the dual-path feature extraction unit in the training and testing phases, respectively; wherein, the CNN branch relies on the convolution kernel to overlap and slide on the input source domain and target domain data to obtain a feature map rich in local details; the Transformer branch can use its own unique multi-head self-attention mechanism to capture sequence information from different angles such as the temporal dependency of data and the correlation between features, thereby enhancing the perception of global features.

4. The method for life assessment of power electronic devices based on electrical and thermal stress according to claim 1, characterized in that: In step S3, the multimodal feature fusion unit implements tensor fusion by modeling global and local modal features. It redefines the degenerate features using binary Cartesian products, and expands the features with a dimension of 1 to calculate the feature correlation between the two modalities while retaining the information of its own data modality.

5. The method for life assessment of power electronic devices based on electrical and thermal stress according to claim 1, characterized in that: In step S4, construct the feature extractor G f , used to extract spatial structure information; construct lifespan predictor G p , used to extract time scale information; construct domain discriminator G d , used to identify whether the data belongs to the source domain or the target domain.

6. The method for life assessment of power electronic devices based on electrical and thermal stress according to claim 5, characterized in that: The feature extractor G f It includes a dual-path feature extraction unit and a multimodal feature fusion unit; The two are connected in sequence, and the final output power electronic device integrates the degradation characteristics.

7. The method for life assessment of power electronic devices based on electrical and thermal stress according to claim 5, characterized in that: The lifespan predictor G p It includes two layers of fully connected networks (FC) and uses the ReLU function to achieve the mapping between degradation features and remaining lifespan.

8. The method for life assessment of power electronic devices based on electrical and thermal stress according to claim 5, characterized in that: The domain discriminator G d It includes a three-layer fully connected network (FC) and uses the Softmax function to determine whether the data belongs to the source domain or the target domain.

9. The method for life assessment of power electronic devices based on electrical and thermal stress according to claim 1, characterized in that: In step S5, the unsupervised adversarial learning strategy includes a conditional domain adversarial learning strategy, which can learn the nonlinear mapping relationship between features in different domains. The specific implementation steps are as follows: S51: Constructing lifespan predictor G p Predicted loss of life Represents G p The predicted loss of life, 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, represents the extracted fusion degradation features, Indicates the prediction result; S52: Based on the idea of ​​conditional domain confrontation, construct a domain discriminator G d Domain discrimination loss: Represents a nonlinear mapping function; Represents G d Domain discrimination loss, 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, represents the extracted fusion degradation features, Represents the prediction result, Indicates the judgment result; S53: The optimization training of the objective function utilizes features from the source domain and the target domain, and then these features are input into the domain discriminant G d , calculate the domain discrimination loss. In addition, the lifespan predictor G p Receive from feature extractor G f features to perform the prediction process: Therefore, using the formula and To optimize the training of the entire adaptive life prediction model, the parameter θ f ,θ p and θ d Represents the feature extractor G f , lifespan predictor G p and domain discrimination G d The trainable parameters of , argmin and argmax represent the minimum and maximum values ​​respectively.

10. A power electronic device life assessment system based on electrical and thermal stress, characterized in that: Used to implement the method according to any one of claims 1 to 7, comprising the following modules: a data set construction module for collecting condition monitoring data of power electronic devices under different aging conditions and dividing the data into a source domain data set and a target domain data set; The feature extraction module is used to extract and fuse the global and local aging features of power electronic devices. The timing prediction module is used to establish the mapping relationship between the multimodal degradation features and lifespan of power electronic devices. Model training module, which is used to combine the lifespan prediction loss and domain discrimination loss for training; The prediction module is used to input the data to be tested into the trained life prediction network and output the prediction results.

Images