IGBT junction temperature prediction method and system based on improved model of transfer learning

Abstract

The application discloses an IGBT junction temperature prediction method and system based on a transfer learning improved model, relates to the field of IGBT application, and comprises the following steps: constructing a discretized heat transfer ODE equation; generating a finite element simulation data set and an experimental data set; constructing an LSTM-ODE model embedded with a physical layer, extracting a thermal resistance and thermal capacity mapping relationship through a sub neural network, and pre-training by using simulation data; bridging simulation and experimental deviations through transfer learning fine-tuning; and predicting junction temperature by using the fine-tuned model; the IGBT junction temperature prediction method based on the transfer learning improved model fuses a physical heat transfer process and an LSTM network, introduces transfer learning and a UKF algorithm, overcomes the limitations of existing methods, realizes accurate and rapid prediction of the junction temperature of an IGBT module, and effectively improves the reliability and economy of IGBT module thermal management.

Description

Technical Field

[0001] This invention relates to IGBT application technology, specifically to an IGBT junction temperature prediction method and system based on an improved model of transfer learning. Background Technology

[0002] As core switching devices in power electronic systems, IGBT modules are widely used in new energy vehicles, rail transportation, wind power generation, and high-voltage direct current transmission. Accurate acquisition of the chip junction temperature is directly related to the module's thermal management design, over-temperature protection strategy formulation, and operational reliability assessment. However, IGBT modules employ a multi-layered packaging structure, with the chip encased in silicone; therefore, real-time and precise measurement of the junction temperature remains a critical technical bottleneck that urgently needs to be overcome.

[0003] Currently, methods for obtaining IGBT junction temperature mainly include contact sensor methods, infrared thermal imaging methods, thermosensitive electrical parameter methods, and thermal network methods. Contact sensor methods require the temperature sensing element to contact the chip, which is an invasive measurement and can damage the integrity of the module package. Infrared thermal imaging methods require opening the module, which is complex and cannot achieve online monitoring. The thermosensitive electrical parameter method is greatly affected by stray inductance and electromagnetic interference, and the calibration process is cumbersome and consistency is difficult to guarantee. The thermal network method estimates the junction temperature by establishing a thermal resistance-thermal capacity network model, which does not damage the device structure and is easy to calculate offline, thus gaining widespread popularity. However, traditional thermal network methods usually treat thermal resistance and thermal capacity as constant parameters, making it difficult to reflect the nonlinear heat transfer characteristics of the module under different operating conditions, resulting in limited prediction accuracy.

[0004] To address the shortcomings of traditional thermal networks, data-driven methods, such as Long Short-Term Memory (LSTM) networks, have been introduced into the field of junction temperature prediction. Leveraging their gating mechanisms and cell state design, they exhibit powerful capabilities in modeling nonlinear dynamic systems. However, purely data-driven models lack the constraints of heat transfer physics and are prone to overfitting and weak generalization ability when training data coverage is insufficient.

[0005] To address the dual problems of static physical model parameters and lack of constraints in data-driven models, the applicant previously filed an invention patent with publication number CN121306313B, entitled "A Method and System for IGBT Module Junction Temperature Prediction Based on LSTM Thermal Neural Network." This method analyzes the unsteady-state heat transfer characteristics of multilayer materials inside the IGBT module, constructs a discretized heat transfer equation, and embeds it as a physical layer into an LSTM network to form an LSTM-ODE neural unit. Through the sub-neural network, the mapping relationship between thermal resistance and heat capacity is dynamically extracted, enabling the model to possess both the nonlinear fitting capability of deep learning and strict adherence to the physical laws of heat transfer, thus achieving an effective integration of physical modeling and data-driven approaches.

[0006] Despite the breakthroughs achieved in the theoretical aspects of the aforementioned preliminary research, a core technical bottleneck remains in the process of translating these findings into practical engineering applications: the inherent discrepancy between the simulation domain and the experimental domain means that models trained solely on simulation data cannot accurately reflect the heat transfer behavior of real devices.

[0007] Specifically, the pre-training of the LSTM-ODE model constructed in the aforementioned comparative documents primarily relies on datasets generated by finite element method (FEM) simulation. While FEM simulation can provide complete temperature response data covering multiple operating conditions with high time resolution and at a controllable cost, its accuracy is highly dependent on the precision of the geometric model, the reliability of material property parameters, and the rationality of boundary condition settings. In real-world production, IGBT module manufacturing involves process tolerances—small deviations in solder layer thickness and unavoidable differences in void ratio can lead to discrepancies between actual thermal characteristics and the simulation model; material thermal properties can also fluctuate due to different supplier batches; furthermore, in actual testing environments, factors such as sensor measurement noise, uncertainties in contact thermal resistance introduced during assembly, and differences between cooling conditions and ideal boundaries combine to create a systematic domain bias between the simulation model and the actual physical device. This domain bias has a hierarchical impact: at the input feature level, there is a shift in the statistical distribution between simulation data and actual operating condition data; at the physical parameter level, there are quantifiable differences between the thermal resistance and thermal capacity parameters assumed by the finite element simulation and the measured values ​​of the device. If the model is trained solely on simulation data and then deployed directly in an experimental environment, the model's prediction accuracy for real devices will significantly deteriorate. On the other hand, if training is entirely based on experimental data, it faces the dilemma of high data acquisition costs and limited coverage of test conditions—real-world data under extreme conditions is often difficult to collect, making it impossible to verify the model's reliability under boundary conditions.

[0008] Therefore, how to effectively bridge the gap between the simulation domain and the experimental domain while maintaining the advantages of the physical information neural network architecture, so that the model can quickly adapt to the heat transfer characteristics of real devices with only a small amount of experimental data, has become a key technical problem that urgently needs to be solved. Summary of the Invention

[0009] The purpose of this invention is to provide an IGBT junction temperature prediction method and system based on an improved model of transfer learning, in order to solve the problems in the prior art where static parameters of the thermal network method cannot reflect nonlinear heat transfer characteristics, the lack of physical constraints in the pure data-driven model leads to weak generalization ability, and the deviation between the simulation model and the real device leads to insufficient prediction accuracy.

[0010] To achieve the above objectives, the present invention provides the following technical solution: an IGBT junction temperature prediction method based on an improved model of transfer learning, comprising the following steps:

[0011] S1: Analyze the unsteady-state heat transfer characteristics of the multilayer materials inside the IGBT module, and construct the ODE equation based on the structure and material properties;

[0012] S2: Generate finite element datasets and experimental datasets containing temperature response data under different power loss conditions through FEM simulation and experiments;

[0013] S3: Construct an LSTM-TL-ODE junction temperature prediction model. In the pre-training stage, embed the physical layer to form an LSTM-ODE neural unit. The ODE equation is embedded into the network as a forward module in the training process. The model is pre-trained using a finite element dataset. The pre-trained LSTM-TL-ODE junction temperature prediction model is used to predict the junction temperature of the IGBT module under different operating conditions.

[0014] S4: Use experimental data to fine-tune the pre-trained model through transfer learning. The fine-tuning stage includes domain bias quantification analysis, fine-tuning strategy optimization design, loss function design, fine-tuning process control, and prediction of junction temperature of IGBT module under different operating conditions.

[0015] S5: Use the UKF algorithm to identify the thermal parameters of the IGBT module under deterioration, feed the identified thermal parameters back to the LSTM-TL-ODE model, and predict the junction temperature of the IGBT module under different operating conditions and different degrees of deterioration.

[0016] Furthermore, in step S1, the ODE equation is:

[0017] In step S1, the ODE equation is:

[0018]

[0019]

[0020]

[0021]

[0022] Mode to middle, , , and These are the chip center temperature, chip solder layer center temperature, substrate solder layer center temperature, and copper substrate center temperature, respectively. This refers to ambient temperature or coolant temperature. and The thermal resistance and thermal capacitance of the connection between the chip center and the chip solder layer center. and The thermal resistance and thermal capacitance between the center of the chip solder layer and the center of the substrate solder layer are considered. and The thermal resistance and thermal capacity of the connection between the center of the solder layer and the center of the copper substrate are calculated. and The thermal resistance and thermal capacity are the connection between the center of the copper substrate and the environment. To mitigate the power loss of the chip, a matrix is ​​used. To describe the connectivity between the centers of each material layer, the chip center temperature is regarded as the chip junction temperature;

[0023] when At that time, the chip junction temperature and the center temperature of each material layer are equal to the ambient temperature. In subsequent time steps, the formula can be calculated sequentially. to To obtain the junction temperature .

[0024] Furthermore, the process of generating the finite element dataset in step S2 includes:

[0025] Construct an FEM simulation model that includes the geometry, material properties, and boundary conditions of the IGBT module;

[0026] Temperature monitoring points are set at the center of each material layer of the IGBT module;

[0027] Apply different power losses to simulate different operating conditions in actual operation;

[0028] The temperature response curves of each monitoring point from the initial moment to the steady state were recorded with high temporal resolution and integrated to form a training dataset.

[0029] The process of generating the experimental dataset in step S2 includes:

[0030] Construct an experimental setup consisting of a power supply circuit and a cooling circuit;

[0031] The power loss of the IGBT chip can be controlled by adjusting the amplitude of the AC voltage in the power supply circuit.

[0032] The cooling circuit maintains constant current and constant temperature heat dissipation conditions.

[0033] The junction temperature of the IGBT module is measured by an optical fiber temperature sensor. Data is recorded when the junction temperature fluctuation stabilizes and the maximum value no longer increases, and the data is integrated to form an experimental dataset.

[0034] Furthermore, in step S3, the sub-neural network and Output dynamic thermal resistance and dynamic heat capacity Dynamic thermal resistance and thermal capacity are concatenated with input features to form an enhanced feature vector. This will enhance the feature vector. The data is input into an LSTM-ODE neural unit for time series modeling.

[0035] Represents the centers of two material layers, namely: , , or ;

[0036] The final hidden layer output of the sub-neural network during the pre-training phase Mapped to thermal resistance matrices respectively With heat capacity vector :

[0037]

[0038]

[0039] in, , The output layer weights are the thermal resistance and thermal capacity. , The output layer bias corresponds to thermal resistance and thermal capacitance. This represents the hidden state of the output layer.

[0040] Furthermore, in step S3, the output of the LSTM layer is mapped to the target temperature through a fully connected layer, and the prediction result of the sub-neural network is substituted into the ODE equation to calculate the node temperature at the next moment, thereby realizing real-time prediction of the IGBT module temperature.

[0041] Further, in step S3, the LSTM-TL-ODE network junction temperature prediction model pre-trained using the finite element dataset is specifically as follows:

[0042] S301: Perform feature extraction and normalization on the dataset generated by finite element simulation;

[0043] S302: Define the loss function for the pre-training phase, including mean squared error loss and first derivative difference loss, to minimize the difference between the actual temperature value and the predicted value.

[0044] S303: Hyperparameter tuning is performed using time series cross-validation and Bayesian optimization methods.

[0045] Furthermore, in step S4, to quantify the degree of offset in this domain, the maximum mean difference (MMD) is used to quantify the input feature distribution deviation, and the relative error is used to quantify the thermal resistance. Heat capacity Parameter deviation:

[0046]

[0047]

[0048]

[0049] in, , Input feature sets are the source domain and the target domain, respectively. , These represent the sample sizes of the two datasets, respectively. For feature mapping function, For the regenerating nucleus Hilbert space, , These are the thermal resistance and thermal capacity parameters predicted by the source domain pre-trained model, respectively. , These are the thermal resistance and thermal capacity parameters measured in the experiment.

[0050] Furthermore, in step S4, the LSTM feature extraction layer is divided into shallow, medium, and deep layers, and the thermal parameters are fine-tuned accordingly. An attention mechanism is introduced into the physical parameter sub-network to enhance the understanding of the contact thermal resistance of the solder layer. Chip thermal capacity The predicted weights;

[0051]

[0052]

[0053] in For learnable scaling factors, The quantities of thermal resistance and thermal capacity parameters, and These are the deviation values ​​of thermal resistance and thermal capacity, respectively. The attention weight is positively correlated with the deviation value, ensuring that parameters with high deviations receive more fine-tuning weight.

[0054] Furthermore, in step S4, the loss function for fine-tuning transfer learning is optimized, and dynamic weight coefficients are introduced to achieve a dynamic balance between data fitting, physical consistency, and domain adaptation.

[0055]

[0056]

[0057] in, The loss function for transfer learning, For mean square error loss, For first-order difference loss, For domain adaptation loss, To adapt weights for the dynamic domain, For the first-order difference loss weights, This is the current number of fine-tuning iterations. This represents the total number of fine-tuning iterations.

[0058] Furthermore, in step S4, the learning rate adopts a cosine annealing decay strategy to avoid parameter oscillations during fine-tuning and improve the convergence speed.

[0059]

[0060] In the above formula This represents the current iteration number. For learning rate, , .

[0061] Furthermore, in step S4, the LSTM-TL-ODE model is fine-tuned using transfer learning during the pre-training stage using experimental data, and the junction temperature of the IGBT module under different operating conditions is predicted. Specifically, this includes the following steps:

[0062] S401: Perform feature extraction and normalization on the dataset obtained from the experiment;

[0063] S402: Analyze the differences between source domain simulation data and target domain experimental data to provide a theoretical basis for fine-tuning strategy design;

[0064] S403: Optimize the design fine-tuning strategy. Based on freezing the feature layer and fine-tuning the physical layer, introduce a hierarchical fine-tuning and learning rate decay strategy to achieve accurate transfer of source domain knowledge and correction of domain bias.

[0065] S404: Optimize the loss function for fine-tuning transfer learning by introducing dynamic weight coefficients to achieve a dynamic balance between data fitting, physical consistency, and domain adaptation.

[0066] Furthermore, in step S5, based on State vector at time step and its covariance matrix Thirteen Sigma points were sampled; during the time update phase, all Sigma points were substituted into the discretized thermal network state equation for nonlinear propagation.

[0067]

[0068]

[0069]

[0070] in Let be the state dimension. This is the scale parameter.

[0071] Furthermore, in step S5, the Kalman gain... for:

[0072]

[0073]

[0074]

[0075] in For covariance weights, For the observation equation, This is the observed mean.

[0076] Furthermore, in step S5, the optimal state estimate and the state covariance matrix are updated:

[0077]

[0078]

[0079] in To observe and predict the covariance matrix, For Kalman gain.

[0080] Further, in step S5, the UKF algorithm is used to identify the thermal parameters of the IGBT module under deterioration conditions, and the identified thermal parameters are fed back to the LSTM-TL-ODE model. The junction temperature of the IGBT module under different operating conditions and different degrees of deterioration is then predicted, specifically including:

[0081] S501: Use the finite element method to simulate different degradation conditions of IGBT modules and obtain the corresponding datasets;

[0082] S502: Preprocess the dataset and perform temporal alignment;

[0083] S503: UKF algorithm initialization, constructing augmented state vector, performing Sigma point sampling and time update;

[0084] S504: The algorithm performs measurement updates and feeds back the thermal parameter identification results to the LSTM-TL-ODE model to predict the junction temperature under deterioration conditions.

[0085] An IGBT junction temperature prediction system based on an improved model using transfer learning includes:

[0086] The equation module is used to analyze the unsteady heat transfer characteristics of the multilayer materials inside the IGBT module and to construct the ODE equation based on the structural and material properties.

[0087] The data module generates a finite element dataset containing temperature response data under different power loss conditions through FEM simulation; and obtains an experimental dataset of temperature response data under different power loss conditions through experiments.

[0088] The building block is used to construct the LSTM-TL-ODE junction temperature prediction model and embeds physical layers during the LSTM-TL-ODE model pre-training stage to form LSTM-ODE neural units. Each LSTM-ODE neural unit contains two sub-neural networks. and In addition, the ODE layer, two sub-neural networks are used to extract the mapping relationship between input features and thermal resistance and thermal capacity, respectively; transfer learning is introduced to fine-tune the physical layer; the UKF algorithm is introduced to identify thermal parameters under deterioration conditions and feed them back to the LSTM-TL-ODE junction temperature prediction model for correction.

[0089] The prediction module uses an LSTM-TL-ODE junction temperature prediction model pre-trained with a finite element dataset. The ODE equation is embedded into the network as a forward module in the training process and transformed into an LSTM-ODE neural unit. The pre-trained model, the LSTM-TL-ODE model, and the LSTM-TL-ODE model corrected by UKF thermal parameter identification are used to predict the junction temperature of the IGBT module under different operating conditions.

[0090] Furthermore, the AdamW optimizer is used for fine-tuning, with the batch size consistent with that in the pre-training stage, and a cosine annealing decay strategy is employed for the learning rate. This avoids parameter oscillations during fine-tuning and improves the convergence speed.

[0091] Furthermore, a process for identifying thermal parameters under different degradation levels using the UKF algorithm was designed, including input data acquisition and preprocessing, UKF algorithm initialization settings, Sigma point sampling and time updating, algorithm measurement updates, and model correction. The junction temperature of the IGBT module under different operating conditions and different degradation levels was also predicted.

[0092] Furthermore, using directly measurable power losses during IGBT operation as input, the output is a dynamic identification value of the thermal parameters related to the chip solder layer.

[0093] Furthermore, the Sigma points are mapped to the observation domain, the optimal state estimate and the state covariance matrix are updated, and the thermal resistance and thermal capacity parameters identified by the UKF algorithm are substituted into the LSTM-TL-ODE model. The junction temperature time series curve is corrected by model prediction.

[0094] Compared with existing technologies, the IGBT junction temperature prediction method and system based on transfer learning improved model provided by this invention, at the physical modeling level, embeds the discretized heat transfer ODE equation into a neural network as the physical layer, so that the model training process follows the physical laws of heat transfer. This overcomes the problems of static parameters and difficulty in dynamically reflecting nonlinear heat transfer characteristics in traditional thermal network methods, and avoids the overfitting risk caused by the lack of physical constraints in purely data-driven models. At the data utilization level, a transfer learning strategy of "simulation pre-training + experimental fine-tuning" is adopted. First, the basic training of the model is completed using finite element simulation data with wide coverage of operating conditions and controllable cost, and then a small number of experimental data are used for the final training. By fine-tuning the physical layer parameters, the domain deviation caused by manufacturing tolerances, uncertainties in contact thermal resistance, and differences in boundary conditions between the simulation model and the real device is effectively bridged. This makes the prediction results closer to the actual heat transfer behavior of the device, solving the accuracy degradation problem caused by relying solely on simulation data for training and the high cost and limited operating condition coverage problem faced by relying entirely on experimental data for training. In terms of prediction performance, the model fine-tuned by transfer learning achieves high-precision and rapid prediction of junction temperature under different operating conditions, combining physical consistency and practical adaptability, meeting the needs of online evaluation, and effectively improving the reliability and economy of IGBT module thermal management. Attached Figure Description

[0095] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0096] Figure 1 A schematic diagram of the material layers in the FEM simulation of an IGBT module;

[0097] Figure 2 This is a diagram showing the temperature monitoring points and thermal resistance / capacity network in an IGBT module.

[0098] Figure 3 This is a diagram of the IGBT module junction temperature prediction model based on the LSTM-TL-ODE model of this invention.

[0099] Figure 4 This is a flowchart of the thermal parameter identification process based on the UKF algorithm of this invention;

[0100] Figure 5 Comparison of IGBT module junction temperatures under different operating conditions using FEM simulation, experiments, LSTM-ODE model, and LSTM-TL-ODE model.

[0101] Figure 6A schematic diagram showing the results of identifying thermal parameters using the UKF algorithm under different degrees of degradation and different operating conditions;

[0102] Figure 7 This is a comparison chart of IGBT module junction temperatures under different degradation levels in this invention, using FEM simulation, LSTM-TL-ODE model, and LSTM-TL-ODE model with thermal parameters corrected by UKF algorithm. Detailed Implementation

[0103] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0104] Please see the appendix Figure 3 Appendix Figure 4 :

[0105] This invention provides a method for predicting IGBT junction temperature based on an improved model using transfer learning, comprising the following steps:

[0106] S1: Analyze the unsteady-state heat transfer characteristics within the IGBT module and construct the ODE equation. Use the ODE equation to represent the heat transfer relationship between nodes, describe the node connections in matrix form, and derive the formula for the node temperature at the next moment. The thermal resistance and heat capacity in the ODE equation can vary with the material properties and boundary conditions of the IGBT module. Therefore, the constructed ODE equation can dynamically reflect the nonlinear characteristics of the heat transfer process in the IGBT module.

[0107] Please see the appendix Figure 1 Based on the structure and material properties of the IGBT module, an ODE equation is constructed and represented as a thermal resistance-thermal capacity network, which includes heat source power loss. Thermal resistance and heat capacity The system's state space is discretized, and the connection relationships between nodes are described in matrix form. Formulas for the node temperatures at the next time step are derived. This includes temperature monitoring points for each material layer, with each pair of thermal resistance and thermal capacity representing several material layers along the heat transfer path.

[0108] The heat transfer relationships between nodes in an IGBT module can be represented in matrix form, where the temperature change of each node can be described by the ODE equation:

[0109]

[0110]

[0111]

[0112]

[0113] Mode to middle, , , and These are the chip center temperature, chip solder layer center temperature, substrate solder layer center temperature, and copper substrate center temperature, respectively. This refers to ambient temperature or coolant temperature. and The thermal resistance and thermal capacitance of the connection between the chip center and the chip solder layer center. and The thermal resistance and thermal capacitance between the center of the chip solder layer and the center of the substrate solder layer are considered. and The thermal resistance and thermal capacity of the connection between the center of the solder layer and the center of the copper substrate are calculated. and The thermal resistance and thermal capacity are the connection between the center of the copper substrate and the environment. To mitigate the power loss of the chip, a matrix is ​​used. To describe the connectivity between the centers of each material layer, the chip center temperature is regarded as the chip junction temperature;

[0114] when At that time, the chip junction temperature and the center temperature of each material layer are equal to the ambient temperature. In subsequent time steps, the formula can be calculated sequentially. to To obtain the junction temperature .

[0115] S2: Generate a complete dataset based on FEM simulation for pre-training the LSTM-TL-ODE junction temperature prediction model; obtain an accurate dataset based on experiments for fine-tuning through transfer learning.

[0116] The pre-training stage of constructing the LSTM-TL-ODE junction temperature prediction model includes an input layer, an LSTM layer, and a fully connected layer. The input is the power loss of the IGBT module. The output is the junction temperature. The model extracts local features step by step through convolutional layers, obtains global high-level features through pooling layers, and concatenates features and outputs data through fully connected layers. It also extracts time series features step by step through LSTM layers, obtains global high-level features using cell states and hidden states, and concatenates features and outputs data through fully connected layers. Considering that the output feature information of the thermal resistance curve is relatively small, a three-layer LSTM and one-layer fully connected LSTM neural network thermal resistance prediction model is used.

[0117] Please see the appendix Figure 2In ANSYS Fluent, an IGBT module FEM simulation was established, temperature monitoring points were set, and temperature response data under asynchronous power loss conditions were extracted.

[0118] In ANSYS Fluent, the material properties of the IGBT module were set, and the inlet temperature of the water-cooled heatsink was configured to 20°C. Temperature monitoring points were set at the center of the chip and the critical material layer to record temperature changes under different power loss conditions. The data sampling frequency was 1000Hz, and the recording duration was 10 seconds, generating a training dataset containing 10,000 samples.

[0119] Please see the appendix Figure 3 The LSTM-TL-ODE junction temperature prediction model uses power loss and node connectivity as input signals, and IGBT junction temperature as the output. Data normalization is performed, and the weighted average and standard deviation of each input variable at each time point are calculated using Exponentially Weighted Moving Average (EWMA) and Exponentially Weighted Moving Standard Deviation (EWMS) methods. These values ​​are then used as input features of the model. Thermal resistance and thermal capacity from the ODE equation are embedded into the physical layer of the LSTM thermal neural network junction temperature prediction model, forming an LSTM-ODE unit. This unit is trained in parallel by sub-neural networks, extracting thermal parameters from the input power loss and combining them with the connectivity of the thermal network to predict the temperature at the next time step. The forget gate, input gate, and output gate mechanisms of LSTM are used to accurately capture the temporal characteristics of the heat transfer process.

[0120] The process of generating a complete dataset based on FEM simulation for pre-training the LSTM-TL-ODE junction temperature prediction model is as follows:

[0121] S201: Construct an FEM simulation that includes the geometry, material properties, and boundary conditions of the IGBT module to ensure that the trained thermal neural network can accurately reflect the heat transfer characteristics of the IGBT module under different operating conditions.

[0122] S202: Set up multiple temperature monitoring points in the FEM simulation. These monitoring points are distributed in key locations of the IGBT module, including the chip center, chip solder layer center, substrate solder layer center, and copper substrate center, etc., to obtain temperature response data at each key location.

[0123] S203: Apply different power losses to simulate different operating conditions that IGBT modules may encounter in practical applications;

[0124] S204: Acquire the temperature response curves of each temperature monitoring point. These curves are recorded with high time resolution (e.g., 1000Hz) and cover the complete temperature change process from the initial moment to the steady state.

[0125] S205: Integrate the temperature response data obtained under all operating conditions into a complete dataset. This dataset contains rich temperature information and can provide comprehensive and diverse data support for the pre-training stage of the LSTM-TL-ODE junction temperature prediction model.

[0126] S3: A physical layer is constructed within the neural network framework, dynamically extracting thermal resistance and heat capacity through sub-neural networks. Specifically, a physical layer containing two sub-neural networks is embedded in the pre-training stage of the LSTM-TL-ODE junction temperature prediction model, mapping the relationship between input power loss and thermal resistance and heat capacity, respectively. After training, the extracted thermal resistance and heat capacity are substituted into the ODE equation to calculate the node temperature at the next time step. This combination leverages prior physical knowledge with the nonlinear fitting capability of deep learning to improve the accuracy and generalization ability of temperature prediction.

[0127] The PyTorch deep learning framework was chosen, combining momentum optimization and the Adam optimizer with an adaptive learning rate. A loss function was defined for the pre-training phase, including mean squared error loss and first-order derivative difference loss, to minimize the difference between actual and predicted temperature values. The dataset was divided into training (70%), validation (15%), and test (15%) sets, and time-series cross-validation was employed to prevent time leakage bias. Bayesian optimization was used to optimize the hyperparameters (learning rate, batch size, number of hidden layer nodes, span) of the LSTM-TL-ODE junction temperature prediction model, achieving fast and accurate optimization of the objective function.

[0128] During the pre-training phase of the LSTM-TL-ODE model, an LSTM-ODE neural unit is constructed, containing two sub-neural networks. and These are used to extract the mapping relationship between input features and thermal resistance and thermal capacity, respectively. and The outputs are dynamic thermal resistance. and dynamic heat capacity Its mathematical expression is:

[0129]

[0130]

[0131] In the above formula, , The output layer weights are the thermal resistance and thermal capacity. , The output layer bias corresponds to thermal resistance and thermal capacitance. This represents the hidden state of the output layer.

[0132] A pre-trained LSTM-TL-ODE junction temperature prediction model is used to predict the junction temperature of IGBT modules under different operating conditions. The thermal parameters predicted by the LSTM thermal neural network junction temperature prediction model are combined with the ODE equation to calculate the junction temperature at the next moment based on the current junction temperature, thus achieving rapid prediction of the IGBT module junction temperature. The pre-training process includes:

[0133] S301: Perform feature extraction and normalization on the dataset generated by FEM simulation;

[0134] The specific process is as follows:

[0135] S301: Extract features from the temperature response data generated by FEM simulation, extract key features related to temperature changes, and normalize these features to facilitate the pre-training of the LSTM-TL-ODE junction temperature prediction model.

[0136] S302: Define the loss function in the pre-training phase, including mean squared error loss and first derivative difference loss, to minimize the difference between the actual temperature value and the predicted value; by optimizing the loss function, the LSTM thermal neural network junction temperature prediction model can more accurately predict the junction temperature change of the IGBT module.

[0137] S303: Hyperparameter tuning employs time-series cross-validation and Bayesian optimization. Time-series cross-validation effectively avoids time leakage in time-series data, while Bayesian optimization efficiently searches for the optimal hyperparameter combination, thereby improving the predictive performance of the LSTM thermal neural network. A pre-trained LSTM-TL-ODE junction temperature prediction model is used, based on the input power loss... Predicting the junction temperature of IGBT modules under different operating conditions The accuracy of the prediction results was verified by comparing them with FEM data and experimental data.

[0138] S4: In the transfer learning fine-tuning stage of the LSTM-TL-ODE model, feature extraction and normalization are performed on the dataset obtained from the experiment; the differences between the source domain finite element simulation data and the target domain experimental data are analyzed to provide a theoretical basis for the fine-tuning strategy design; the fine-tuning strategy is optimized, and based on freezing the feature layer and fine-tuning the physical layer, a layered fine-tuning and learning rate decay strategy is introduced to achieve accurate transfer of source domain knowledge and correction of domain bias; the loss function of transfer learning fine-tuning is optimized, and dynamic weight coefficients are introduced to achieve a dynamic balance between data fitting, physical consistency and domain adaptation.

[0139] During the fine-tuning phase of transfer learning, the domain deviation between the finite element method and the experimental dataset is quantitatively analyzed. Maximum mean difference (MMD) is used to quantify the input feature distribution deviation, and relative error is used to quantify the thermal resistance. Heat capacity Parameter deviation:

[0140]

[0141]

[0142]

[0143] in, , Input feature sets are the source domain and the target domain, respectively. , These represent the sample sizes of the two datasets, respectively. For feature mapping function, For the regenerating nucleus Hilbert space, , These are the thermal resistance and thermal capacity parameters predicted by the source domain pre-trained model, respectively. , These are the thermal resistance and thermal capacity parameters measured in the experiment.

[0144] To optimize the transfer learning fine-tuning strategy, the first step is to perform layered freezing and fine-tuning. The LSTM feature extraction layer is divided into shallow, medium, and deep layers. Secondly, targeted fine-tuning of thermal parameters is performed by introducing an attention mechanism into the physical parameter sub-network to enhance the understanding of solder layer contact thermal resistance. Chip thermal capacity The predicted weights;

[0145]

[0146]

[0147] in For learnable scaling factors, The quantities of thermal resistance and thermal capacity parameters, and These are the deviation values ​​of thermal resistance and thermal capacity, respectively. The attention weight is positively correlated with the deviation value, ensuring that parameters with high deviations receive more fine-tuning weight.

[0148] The loss function for fine-tuning transfer learning is optimized by introducing dynamic weight coefficients to achieve a dynamic balance between data fitting, physical consistency, and domain adaptation. Initially, larger weights are used to minimize domain bias; later, the weights are gradually reduced to optimize the accuracy of data fitting to the target domain. The mathematical expression of the loss function is:

[0149]

[0150]

[0151] in, The loss function for transfer learning, For mean square error loss, For first-order difference loss, For domain adaptation loss, To adapt weights for the dynamic domain, For the first-order difference loss weights, This is the current number of fine-tuning iterations. This represents the total number of fine-tuning iterations;

[0152] The AdamW optimizer is used for fine-tuning, with the batch size consistent with the pre-training stage, and a cosine annealing decay strategy is employed for the learning rate. This avoids parameter oscillations during fine-tuning and improves convergence speed. The specific mathematical expression is as follows:

[0153]

[0154] The design of the UKF algorithm for identifying thermal parameters under different degradation levels includes input data acquisition and preprocessing, UKF algorithm initialization settings, Sigma point sampling and time updating, algorithm measurement updates, and model correction. It also predicts the junction temperature of IGBT modules under different operating conditions and degradation levels.

[0155] Using directly measurable power losses during IGBT operation as input, the output consists of dynamically identified values ​​of thermal parameters related to the chip solder layer, specifically including thermal resistance. , heat capacity , .use The criteria are used to remove outliers, and the mean values ​​of power loss and temperature time series for each material layer are calculated. and standard deviation .

[0156] The augmented state vector is initialized, and the augmented state vector includes junction temperature, ambient temperature, and thermal resistance. , heat capacity , Initialize the state covariance matrix, process noise covariance matrix, and observation noise covariance matrix.

[0157] based on State vector at time step and its covariance matrix Thirteen Sigma points were sampled; during the time update phase, all Sigma points were substituted into the discretized thermal network state equation for nonlinear propagation.

[0158]

[0159]

[0160]

[0161] In the above formula Let be the state dimension. This is the scale parameter.

[0162] The Sigma points are mapped to the observation domain, the optimal state estimate and the state covariance matrix are updated, the thermal resistance and thermal capacity parameters identified by the UKF algorithm are substituted into the LSTM-TL-ODE model, and the junction temperature time series curve is corrected by model prediction.

[0163] S5: Use the UKF algorithm to identify the thermal parameters of the IGBT module under deterioration, feed the identified thermal parameters back to the LSTM-TL-ODE model, and predict the junction temperature of the IGBT module under different operating conditions and different degrees of deterioration.

[0164] In step S5, the Kalman gain for:

[0165]

[0166]

[0167]

[0168] in For covariance weights, For the observation equation, This is the observed mean.

[0169] In step S5, update the optimal state estimate and the state covariance matrix:

[0170]

[0171]

[0172] In the above formula To observe and predict the covariance matrix, For Kalman gain.

[0173] In step S5, the UKF algorithm is used to identify the thermal parameters of the IGBT module under deterioration conditions, and the identified thermal parameters are fed back into the LSTM-TL-ODE model. The junction temperature of the IGBT module under different operating conditions and different degrees of deterioration is then predicted, specifically including:

[0174] S501: Simulate different degradation scenarios of IGBT modules using the finite element method and obtain the corresponding datasets.

[0175] S502: Preprocess the dataset and perform temporal alignment;

[0176] S503: UKF algorithm initialization, constructing augmented state vector, performing Sigma point sampling and time update;

[0177] S504: The algorithm performs measurement updates and feeds back the thermal parameter identification results to the LSTM-TL-ODE model to predict the junction temperature under deterioration conditions.

[0178] In one embodiment of the present invention, an IGBT junction temperature prediction system based on an improved transfer learning model is also provided. This system is used to implement the aforementioned IGBT junction temperature prediction method based on an improved transfer learning model. Specifically, the IGBT junction temperature prediction system based on an improved transfer learning model includes an equation module, a data module, a construction module, and a prediction module, wherein:

[0179] The equation module analyzes the unsteady heat transfer characteristics of the multilayer materials inside the IGBT module and constructs the ODE equation based on the structural and material properties.

[0180] The data module generates a finite element dataset containing temperature response data under different power loss conditions through FEM simulation; and obtains an experimental dataset of temperature response data under different power loss conditions through experiments.

[0181] The module constructs an LSTM-TL-ODE junction temperature prediction model. During the pre-training phase of the LSTM-TL-ODE model, physical layers are embedded to form LSTM-ODE neural units, each containing two sub-neural networks. and In addition, there is an ODE layer, with two sub-neural networks used to extract the mapping relationship between input features and thermal resistance and thermal capacity, respectively; transfer learning is introduced to fine-tune the physical layer; the UKF algorithm is introduced to identify thermal parameters under deterioration conditions and feed them back to the LSTM-TL-ODE junction temperature prediction model for correction.

[0182] The prediction module uses an LSTM-TL-ODE junction temperature prediction model pre-trained with a finite element dataset. The ODE equation is embedded into the network as a forward module in the training process and transformed into an LSTM-ODE neural unit. The pre-trained model, the LSTM-TL-ODE model, and the LSTM-TL-ODE model corrected by UKF thermal parameter identification are used to predict the junction temperature of the IGBT module under different operating conditions.

[0183] This invention also provides a terminal device, which includes a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, graphics processing units (GPUs), tensor processing units (TPUs), digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions to achieve a corresponding method flow or function. The processor described in this embodiment can be used for the operation of an IGBT module junction temperature prediction method based on a UKF-improved LSTM-TL-ODE model.

[0184] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0185] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0186] This invention presents an IGBT junction temperature prediction method based on an improved transfer learning model, which underwent comprehensive experimental evaluation under different operating conditions (converter power 200kVA, 300kVA, 400kVA, and 500kVA) and different degradation levels (degradation level 0.2, 0.4, 0.6, and 0.8). Experimental results show that this method outperforms traditional finite element method (FEM) simulation in both prediction accuracy and computational efficiency.

[0187] Please see the appendix Figure 5 , attached Figure 5 (a) Converter power 200 kVA, (b) Converter power 300 kVA, (c) Converter power 400 kVA, (d) Converter power 500 kVA. We compared the junction temperature prediction results of the finite element simulation, experimental data, the LSTM-TL-ODE pre-training stage, and the LSTM-TL-ODE model under different load conditions. It can be seen that the LSTM-TL-ODE junction temperature prediction model proposed in this paper highly matches the experimental dynamic performance under all operating conditions (200–500 kVA) and outperforms the finite element method (FEM) and the LSTM-TL-ODE pre-training stage in key error indicators. To further demonstrate the accuracy of the LSTM-TL-ODE model in predicting actual temperatures, mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²) were also used. 2 As an evaluation metric, MAE directly reflects the average deviation and is insensitive to outliers. RMSE, on the other hand, amplifies the impact of extreme errors. 2 The dispersion of prediction error is measured to reflect the stability of the model prediction. The MAE of LSTM-TL-ODE ranges from 0.003 to 0.058, and the RMSE ranges from 0.014 to 0.06. Compared with the Finite Element Method (FEM) and LSTM-ODE, both MAE and RMSE are reduced by one to two orders of magnitude, indicating that the prediction results of LSTM-TL-ODE are more stable and closer to the experimental results. The R² of LSTM-TL-ODE is 99.97% to 99.98%. Furthermore, the R² of LSTM-TL-ODE is higher than that of the Finite Element Method (78.5% to 98.52%) and the R² of LSTM-ODE (92.36% to 99.71%), indicating that the prediction results of LSTM-TL-ODE are highly consistent with the experimental data.

[0188] For local data analysis, it can be observed that under different operating conditions, the peak-to-peak relative errors of FEM, pre-trained LSTM-ODE, and LSTM-TL-ODE are similar, indicating that the LSTM-TL-ODE thermal neural network can predict stable fluctuations in junction temperature. The maximum relative error of LSTM-TL-ODE is 0.13%–0.21%, lower than that of FEM (1.08%–2.13%) and LSTM-ODE (0.36%–1.15%), indicating that the LSTM-TL-ODE thermal neural network provides more accurate junction temperature predictions and has better prediction performance than the LSTM-ODE model and FEM simulation. The results show that the LSTM-TL-ODE model can more accurately capture the dynamic changes in IGBT module junction temperature and has higher prediction reliability.

[0189] Please see the appendix Figure 6 Under different degrees of degradation, the changes in thermal resistance and heat capacity of the solder layer exhibit significant asymmetric characteristics, and this characteristic becomes more pronounced with changes in operating conditions. Under the same operating conditions, as the void ratio increases from 20% to 80%, the effective heat conduction area of ​​the solder layer continuously decreases, the thermal resistance shows a non-linear, rapid upward trend, while the heat capacity shows a linear downward trend. Taking the 500kVA operating condition as an example, when the void ratio is 0, The measured value is 0.0215 K / W. The initial value was 0.2602 K / W, and when the void ratio increased to 80%, the two values ​​increased to 0.14083 K / W and 0.3707 K / W, respectively; while Then it increases from 0.02 J / K to 0.03326 J / K. The thermal resistance decreased from 0.075 J / K to 0.035 J / K. The expansion of voids directly reduces the effective heat transfer path, leading to an exponential increase in thermal resistance, while the heat capacity decreases synchronously due to the linear reduction in the effective solder mass. Furthermore, at the same level of degradation, improving operating conditions further exacerbates the increase in thermal resistance; at an 80% void ratio, the thermal resistance at a 200 kVA operating condition... The measured value is 0.12890 K / W, while under 500kVA conditions, the value rises to 0.14210 K / W. This is because the high junction temperature under high operating conditions will accelerate the aging of the solder layer and further increase the thermal resistance.

[0190] Please see the appendix Figure 7 , attached Figure 7 The table shows the junction temperature predictions for degradation levels d=0.2, d=0.4, d=0.6, and d=0.8 for a 500kVA VA system. When d=0, UKF does not perform any correction, so the LSTM-TL-ODE model result is... Figure 6At d=0.2, the predicted curve of the LSTM-TL-ODE model begins to gradually deviate from the actual junction temperature, and the error starts to rise slightly. The MAE is approximately 0.8℃, mainly manifested in a lower predicted peak junction temperature and a more similar trough value. After UKF correction, the error remains around 0.2℃. At d=0.4 and 0.6, the error between the LSTM-TL-ODE model and the actual value increases nonlinearly, with MAEs of 3.8℃ and 8.2℃, respectively. After UKF correction, the error does not increase significantly and remains below 0.35℃. As the degradation continues to worsen to d=0.8, the MAE error between the LSTM-TL-ODE model and the actual value increases to 18.5℃, the prediction bias intensifies significantly, and the waveform trend shows a clear misalignment. However, the UKF-corrected model, by updating the solder layer thermal resistance and thermal capacity parameters identified by UKF in real time, can still accurately predict the actual junction temperature under degradation conditions.

[0191] In summary, this invention presents an IGBT junction temperature prediction method and system based on an improved model using transfer learning. Combining physical constraints with the temporal modeling capabilities of deep learning, it dynamically extracts thermal resistance and heat capacity. Multi-condition training datasets are generated through FEM simulation and experiments, allowing the model to adapt to different power losses and overcome the limitations of traditional methods in handling nonlinear heat transfer. The model is pre-trained using a finite element dataset, and then fine-tuned through transfer learning based on experimental data, making the prediction results closer to the experimental data. The LSTM-TL-ODE model offers fast prediction speed, meeting real-time requirements, while the embedded physical layer ensures that the prediction conforms to heat transfer laws. The UKF algorithm is introduced to identify thermal parameters at different degradation levels and feed them back to the LSTM-TL-ODE model for correction, enabling the model to predict junction temperatures under different degradation levels, thus improving the reliability and economy of IGBT module thermal management.

[0192] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A method for predicting IGBT junction temperature based on an improved model using transfer learning, characterized in that, Includes the following steps: S1: Analyze the unsteady-state heat transfer characteristics of the multilayer materials inside the IGBT module, and construct the ODE equation based on the structure and material properties; S2: Generate finite element datasets and experimental datasets containing temperature response data under different power loss conditions through FEM simulation and experiments; S3: Construct an LSTM-TL-ODE junction temperature prediction model. In the pre-training stage, embed the physical layer to form an LSTM-ODE neural unit. The ODE equation is embedded into the network as a forward module in the training process. The model is pre-trained using a finite element dataset. The pre-trained LSTM-TL-ODE junction temperature prediction model is used to predict the junction temperature of the IGBT module under different operating conditions. S4: Use experimental data to fine-tune the pre-trained model through transfer learning. The fine-tuning stage includes domain bias quantification analysis, fine-tuning strategy optimization design, loss function design, fine-tuning process control, and prediction of junction temperature of IGBT module under different operating conditions. S5: Use the UKF algorithm to identify the thermal parameters of the IGBT module under deterioration, feed the identified thermal parameters back to the LSTM-TL-ODE model, and predict the junction temperature of the IGBT module under different operating conditions and different degrees of deterioration.

2. The IGBT junction temperature prediction method based on the improved model of transfer learning according to claim 1, characterized in that, In step S1, the ODE equation is: Mode to middle, , , and These are the chip center temperature, chip solder layer center temperature, substrate solder layer center temperature, and copper substrate center temperature, respectively. This refers to ambient temperature or coolant temperature. and The thermal resistance and thermal capacitance of the connection between the chip center and the chip solder layer center. and The thermal resistance and thermal capacitance between the center of the chip solder layer and the center of the substrate solder layer are considered. and The thermal resistance and thermal capacity of the connection between the center of the solder layer and the center of the copper substrate are calculated. and The thermal resistance and thermal capacity are the connection between the center of the copper substrate and the environment. To mitigate the power loss of the chip, a matrix is ​​used. This describes the connectivity between the centers of each material layer.

3. The IGBT junction temperature prediction method based on the improved model of transfer learning according to claim 1, characterized in that, The process of generating the finite element dataset in step S2 includes: Construct an FEM simulation model that includes the geometry, material properties, and boundary conditions of the IGBT module; Temperature monitoring points are set at the center of each material layer of the IGBT module; Apply different power losses to simulate different operating conditions in actual operation; The temperature response curves of each monitoring point from the initial moment to the steady state were recorded with high temporal resolution and integrated to form a training dataset. The process of generating the experimental dataset in step S2 includes: Construct an experimental setup consisting of a power supply circuit and a cooling circuit; The power loss of the IGBT chip can be controlled by adjusting the amplitude of the AC voltage in the power supply circuit. The cooling circuit maintains constant current and constant temperature heat dissipation conditions. The junction temperature of the IGBT module is measured by an optical fiber temperature sensor. Data is recorded when the junction temperature fluctuation stabilizes and the maximum value no longer increases, and the data is integrated to form an experimental dataset.

4. The IGBT junction temperature prediction method based on the improved model of transfer learning according to claim 1, characterized in that, In step S3, the sub-neural network and Output dynamic thermal resistance and dynamic heat capacity Dynamic thermal resistance and thermal capacity are concatenated with input features to form an enhanced feature vector. This will enhance the feature vector. The data is input into an LSTM-ODE neural unit for time series modeling. Represents the centers of two material layers, namely: , , or ; The final hidden layer output of the sub-neural network during the pre-training phase Mapped to thermal resistance matrices respectively With heat capacity vector : in, , The output layer weights are the thermal resistance and thermal capacity. , The output layer bias corresponds to thermal resistance and thermal capacitance. This refers to the hidden state of the output layer. In step S3, the output of the LSTM layer is mapped to the target temperature through a fully connected layer, and the prediction result of the sub-neural network is substituted into the ODE equation to calculate the node temperature at the next moment, thereby realizing real-time prediction of the IGBT module temperature. In step S3, the LSTM-TL-ODE network junction temperature prediction model pre-trained using the finite element dataset is specifically as follows: S301: Perform feature extraction and normalization on the dataset generated by finite element simulation; S302: Define the loss function for the pre-training phase, including mean squared error loss and first derivative difference loss, to minimize the difference between the actual temperature value and the predicted value. S303: Hyperparameter tuning is performed using time series cross-validation and Bayesian optimization methods.

5. The IGBT junction temperature prediction method based on the improved model of transfer learning according to claim 1, characterized in that, In step S4, to quantify the degree of offset in this domain, the maximum mean difference (MMD) is used to quantify the input feature distribution deviation, and the relative error is used to quantify the thermal resistance. Heat capacity Parameter deviation: in, , Input feature sets are the source domain and the target domain, respectively. , These represent the sample sizes of the two datasets, respectively. For feature mapping function, For the regenerating nucleus Hilbert space, , These are the thermal resistance and thermal capacity parameters predicted by the source domain pre-trained model, respectively. , These are the thermal resistance and thermal capacity parameters measured experimentally. In step S4, the LSTM feature extraction layer is divided into shallow, medium, and deep layers, and the thermal parameters are fine-tuned accordingly. An attention mechanism is introduced into the physical parameter sub-network to enhance the understanding of the contact thermal resistance of the solder layer. Chip thermal capacity The predicted weights; in For learnable scaling factors, The quantities of thermal resistance and thermal capacity parameters, and These are the deviation values ​​of thermal resistance and thermal capacity, respectively. The attention weight is positively correlated with the deviation value, ensuring that parameters with high deviation receive more fine-tuning weight. In step S4, the loss function for fine-tuning transfer learning is further optimized by introducing dynamic weight coefficients to achieve a dynamic balance between data fitting, physical consistency, and domain adaptation. in, The loss function for transfer learning, For mean square error loss, For first-order difference loss, For domain adaptation loss, To adapt weights for the dynamic domain, For the first-order difference loss weights, This is the current number of fine-tuning iterations. This represents the total number of fine-tuning iterations; In step S4, the learning rate adopts a cosine annealing decay strategy to avoid parameter oscillations during fine-tuning and improve the convergence speed. Mode middle This represents the current iteration number. For learning rate, , .

6. The IGBT junction temperature prediction method based on the improved model of transfer learning according to claim 1, characterized in that, In step S4, the LSTM-TL-ODE model is fine-tuned using transfer learning during the pre-training stage using experimental data, and the junction temperature of the IGBT module under different operating conditions is predicted. Specifically, this includes the following steps: S401: Perform feature extraction and normalization on the dataset obtained from the experiment; S402: Analyze the differences between source domain simulation data and target domain experimental data to provide a theoretical basis for fine-tuning strategy design; S403: Optimize the design fine-tuning strategy. Based on freezing the feature layer and fine-tuning the physical layer, introduce a hierarchical fine-tuning and learning rate decay strategy to achieve accurate transfer of source domain knowledge and correction of domain bias. S404: Optimize the loss function for fine-tuning transfer learning by introducing dynamic weight coefficients to achieve a dynamic balance between data fitting, physical consistency, and domain adaptation.

7. The IGBT junction temperature prediction method based on the improved model of transfer learning according to claim 1, characterized in that, In step S5, based on State vector at time step and its covariance matrix Thirteen Sigma points were sampled; during the time update phase, all Sigma points were substituted into the discretized thermal network state equation for nonlinear propagation. in Let be the state dimension. For scale parameters; In step S5, the Kalman gain for: in For covariance weights, For the observation equation, This is the observed mean.

8. The IGBT junction temperature prediction method based on the improved model of transfer learning according to claim 1, characterized in that, In step S5, update the optimal state estimate and the state covariance matrix: in To observe and predict the covariance matrix, For Kalman gain.

9. The IGBT junction temperature prediction method based on the improved model of transfer learning according to claim 1, characterized in that, In step S5, the UKF algorithm is used to identify the thermal parameters of the IGBT module under deterioration conditions, and the identified thermal parameters are fed back into the LSTM-TL-ODE model. The junction temperature of the IGBT module under different operating conditions and different degrees of deterioration is then predicted, specifically including: S501: Use the finite element method to simulate different degradation conditions of IGBT modules and obtain the corresponding datasets; S502: Preprocess the dataset and perform temporal alignment; S503: UKF algorithm initialization, constructing augmented state vector, performing Sigma point sampling and time update; S504: The algorithm performs measurement updates and feeds back the thermal parameter identification results to the LSTM-TL-ODE model to predict the junction temperature under deterioration conditions.

10. An IGBT junction temperature prediction system based on an improved model of transfer learning, used to implement the prediction method according to any one of claims 1-9, characterized in that, include: The equation module is used to analyze the unsteady heat transfer characteristics of the multilayer materials inside the IGBT module and to construct the ODE equation based on the structural and material properties. The data module generates a finite element dataset containing temperature response data under different power loss conditions through FEM simulation; and obtains an experimental dataset of temperature response data under different power loss conditions through experiments. The building block is used to construct the LSTM-TL-ODE junction temperature prediction model and embeds physical layers during the LSTM-TL-ODE model pre-training stage to form LSTM-ODE neural units. Each LSTM-ODE neural unit contains two sub-neural networks. and In addition, the ODE layer, two sub-neural networks are used to extract the mapping relationship between input features and thermal resistance and thermal capacity, respectively; transfer learning is introduced to fine-tune the physical layer; the UKF algorithm is introduced to identify thermal parameters under deterioration conditions and feed them back to the LSTM-TL-ODE junction temperature prediction model for correction. The prediction module uses an LSTM-TL-ODE junction temperature prediction model pre-trained with a finite element dataset. The ODE equation is embedded into the network as a forward module in the training process and transformed into an LSTM-ODE neural unit. The pre-trained model, the LSTM-TL-ODE model, and the LSTM-TL-ODE model corrected by UKF thermal parameter identification are used to predict the junction temperature of the IGBT module under different operating conditions.

Images