A method and related equipment for predicting the thermal resistance performance parameters of IGCT based on the IWOA-BP model
By improving the hybrid prediction model that combines the whale optimization algorithm with the BP neural network, the deviation problem caused by changes in physical parameters in the prediction of IGCT thermal resistance performance parameters is solved. This achieves accurate prediction of the thermal resistance performance parameters of IGCT devices, adapts to device aging and changes in operating conditions, and improves the model's generalization ability and robustness.
Patent Information
- Application Number
- CN202611263074.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-19
- Publication Date
- 2026-11-10
AI Technical Summary
Existing methods for predicting the thermal resistance performance parameters of IGCT rely on precise physical parameters, but these parameters are easily affected by material batches, aging levels, and process differences, leading to prediction bias. Furthermore, traditional data-driven models have weak generalization ability when devices age or operating conditions change.
A hybrid prediction model combining the improved whale optimization algorithm (IWOA) and the backpropagation neural network is adopted. By acquiring the time-series monitoring data of the IGCT device, the data is cleaned and normalized to construct a sliding window sample set. The improved whale optimization algorithm is used to globally optimize the weights and thresholds of the backpropagation neural network. Combined with gradient descent training, the accurate prediction of the thermal resistance performance parameters of the IGCT is achieved.
It achieves high-precision prediction of steady-state and transient thermal resistance performance parameters of IGCT devices, improves the generalization ability and robustness of the model, adapts to device aging and operating condition changes, and meets the needs of rapid design and real-time monitoring.
Smart Images

Figure CN123068952A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of semiconductor device performance prediction technology, and in particular to a method and related equipment for predicting the thermal resistance performance parameters of IGCT based on the IWOA-BP model. Background Technology
[0002] In related technologies, the accuracy of physical models (especially finite element models) heavily depends on the precise input of a large number of physical parameters such as chip / package material properties and interface contact thermal resistance. However, these parameters often vary due to factors such as material batches, aging degree, and process differences, leading to prediction deviations.
[0003] In summary, the technical problems existing in the relevant technologies need to be improved. Summary of the Invention
[0004] The main objective of this application is to propose a method and related equipment for predicting the thermal resistance performance parameters of IGCT based on the IWOA-BP model, which can achieve accurate prediction of the thermal resistance performance parameters of IGCT devices.
[0005] To achieve the above objectives, one aspect of this application proposes a method for predicting the thermal resistance performance parameters of IGCT based on the IWOA-BP model, the method comprising the following steps: Acquire the monitoring data of the IGCT device under test in power cycling test according to the time sequence to obtain the original time-series monitoring data sequence; The original time-series monitoring data sequence is cleaned and normalized to obtain a standard time-series data sequence; The standard time series data sequence is divided into sliding window segments according to a preset window length and step size to generate several time series sample blocks. A mapping relationship is established between each time series sample block and the thermal resistance parameter at the corresponding prediction step size to obtain a total sample set. The total sample set is then divided into a training set and a test set according to a preset ratio. A hybrid prediction model combining an improved whale optimization algorithm and a BP neural network is constructed to obtain an initial hybrid prediction model with undetermined weights and thresholds. The improved whale optimization algorithm introduces Logistic chaotic mapping for population initialization, adopts a cosine nonlinear decreasing convergence factor, and introduces adaptive weights based on individual fitness in position updates. The improved whale optimization algorithm is used to perform global iterative optimization of the weights and thresholds of the BP neural network in the initial hybrid prediction model. The mean square error on the training set is used as the fitness evaluation criterion to obtain the optimized hybrid prediction model of the initial weights and thresholds. After assigning the optimized initial weights and thresholds, the hybrid prediction model is trained by gradient descent on the training set to obtain the trained IWOA-BP prediction model. The training IWOA-BP prediction model was validated using the test set, and the mean absolute percentage error, root mean square error, and coefficient of determination were used as performance evaluation indicators to obtain the final prediction results of the IGCT thermal resistance performance parameters.
[0006] In some embodiments, acquiring the monitoring data collected sequentially during power cycling testing of the IGCT device under test to obtain the original timing monitoring data sequence includes: Monitoring data of the IGCT device during power cycling test is continuously collected according to a preset sampling period; the monitoring data includes collector-emitter voltage, gate voltage, collector current, case temperature, and junction temperature; The monitoring data are arranged in chronological order to form the original time-series monitoring data sequence.
[0007] In some embodiments, the step of cleaning and normalizing the original time-series monitoring data sequence to obtain a standard time-series data sequence includes: The mean and standard deviation of each dimension of the feature variables in the original time-series monitoring data sequence are calculated, and data points that deviate from the mean by more than three times the standard deviation are removed as outliers. For gaps caused by outlier removal, linear interpolation of the nearest valid data points before and after the gap is used to fill them; when the beginning or end of the sequence is missing, nearest neighbor filling is used. The cleaned feature data of each dimension are transformed to the [0,1] interval using the Min-Max method to obtain the standard time series data sequence.
[0008] In some embodiments, the standard time-series data sequence is divided into several time-series sample blocks by a sliding window according to a preset window length and step size, and a mapping relationship is established between each time-series sample block and the thermal resistance parameter at the corresponding prediction step size to obtain a total sample set. The total sample set is then divided into a training set and a test set according to a preset ratio, including: The standard time-series data sequence is divided into sliding window segments according to a preset window length and step size. Each window contains feature fields of a preset dimension, and the data matrix within each window is used as an input sample. The feature fields include voltage, current, case temperature, junction temperature, timestamp, and power loss. Each input sample is mapped to the thermal resistance parameter at the preset prediction step size after the corresponding window ends. All windows are stacked to form an input tensor. The thermal resistance parameter corresponding to each window constitutes the output target vector. The input tensor and the output target vector are paired to form the total sample set. The total sample set is randomly divided into a training set and a test set according to a preset ratio; the training set is used for model training; and the test set is used for prediction accuracy verification.
[0009] In some embodiments, constructing a hybrid prediction model combining the improved whale optimization algorithm and the BP neural network to obtain an initial hybrid prediction model with undetermined weights and thresholds includes: Set the BP neural network as a three-layer feedforward structure consisting of an input layer, a hidden layer, and an output layer. Set the number of nodes in the input layer, the hidden layer, and the output layer. Define the parameters to be optimized as all connection weights and thresholds of the BP neural network to obtain the BP neural network structure settings. Set the population size and maximum number of iterations for the improved whale optimization algorithm, generate the initial population using Logistic chaotic mapping, map each chaotic value to the search range of each parameter to be optimized, and obtain the settings for the improved whale optimization algorithm. Set the convergence factor to decrease with the number of iterations in a cosine form; When setting the position update, the adaptive weight is calculated based on the difference between the fitness of each individual whale and the average fitness and minimum fitness of the population. The adaptive weight is used to weight the position update in the encirclement and random search phases. In the bubble net attack phase, the position is updated by selecting the shrinking encirclement or spiral ascent mode with a preset probability. An initial hybrid prediction model with undetermined weights and thresholds is constructed by setting the BP neural network structure and the improved whale optimization algorithm.
[0010] In some embodiments, the step of using the improved whale optimization algorithm to perform global iterative optimization of the weights and thresholds of the BP neural network in the initial hybrid prediction model, and using the mean squared error on the training set as the fitness evaluation criterion, to obtain the optimized hybrid prediction model with the initial weights and thresholds, includes: All the weights and thresholds to be optimized in the BP neural network of the initial hybrid prediction model are encoded as position vectors of each individual whale. The position vector of each individual whale is decoded into a set of weights and thresholds for a BP neural network. A forward propagation is performed on the training set, and the mean square error between the predicted value and the target output of the network is calculated. The mean square error is used as the fitness value of each individual whale. In the current iteration, calculate the current value of the convergence factor and the coefficient vector; For each individual whale, the surrounding predation update strategy, bubble net attack update strategy, or random search update strategy are selected based on the relationship between the modulus of the coefficient vector and the preset probability. In the surrounding predation and random search strategies, adaptive weights are used to perform position-weighted updates. Boundary detection is performed on the updated position vectors of each individual whale, and components that exceed the search range are truncated to the boundary values of the search range; Recalculate the fitness value of each individual whale, update and record the best individual in the current population and its position vector; Repeatedly execute the iterative process of calculating the convergence factor and the coefficient vector, selecting an update strategy for position update, boundary detection, recalculating fitness and updating the optimal individual, until the maximum number of iterations is reached. The position vector corresponding to the final optimal individual is decoded into a set of weights and thresholds of a BP neural network, which are then used as the optimized initial weights and thresholds and assigned to the hybrid prediction model.
[0011] In some embodiments, the step of assigning the optimized initial weights and thresholds, and then performing gradient descent training on the training set to obtain the trained IWOA-BP prediction model includes: The optimized initial weights and thresholds are used as the initial values for the connection weights and thresholds of each layer in the BP neural network of the hybrid prediction model; On the training set, the mean squared error is used as the loss function, and the gradient descent method is used to iteratively update the connection weights and thresholds of each layer of the BP neural network in the hybrid prediction model. The gradients of the output layer weights and thresholds are calculated by multiplying the partial derivative of the loss with respect to the net input of each output layer by the corresponding hidden layer output. The gradients of the hidden layer weights and thresholds are calculated by multiplying the partial derivative of the loss with respect to the net input of each hidden layer by the corresponding input features. The partial derivative of the net input of the hidden layer is determined by multiplying the derivative of the hidden layer activation function by the backpropagation error from the output layer. The connection weights and thresholds are updated by multiplying each gradient by a preset learning rate; Repeat the following steps on the training set: calculate the loss using mean squared error as the loss function, calculate the gradient of each layer's connection weights and thresholds using gradient descent, and update the connection weights and thresholds according to a preset learning rate, until the loss function converges or the preset maximum number of fine-tuning iterations is reached, to obtain the trained IWOA-BP prediction model.
[0012] In some embodiments, the step of verifying the prediction accuracy of the trained IWOA-BP prediction model using the test set, and obtaining the final prediction results of the IGCT thermal resistance performance parameters using the mean absolute percentage error, root mean square error, and coefficient of determination as model performance evaluation indicators, includes: Each input sample in the test set is input into the trained IWOA-BP prediction model, and the predicted value of the thermal resistance parameter of each test sample is obtained through forward propagation calculation. The predicted thermal resistance parameters are inversely normalized to their original dimensions to obtain the final predicted thermal resistance parameters for each test sample. For the steady-state thermal resistance output dimension, the absolute value of the difference between the predicted value and the true value of each test sample is calculated, divided by the true value, summed, averaged, and then multiplied by the percentage to obtain the average absolute percentage error of the steady-state thermal resistance. For the transient thermal impedance output dimension, the absolute value of the difference between the predicted value and the true value of each test sample is calculated, divided by the true value, summed, averaged, and then multiplied by the percentage to obtain the average absolute percentage error of the transient thermal impedance. The average absolute percentage error of the steady-state thermal resistance and the average absolute percentage error of the transient thermal resistance are taken as the overall average absolute percentage error evaluation index. For the steady-state thermal resistance output dimension, the root mean square error of the steady-state thermal resistance is obtained by taking the square root of the sum of the squares of the differences between the predicted and true values of each test sample and the total number of test samples. For the transient thermal impedance output dimension, the root mean square error of the transient thermal impedance is obtained by taking the square root of the sum of the squares of the differences between the predicted and true values of each test sample and the total number of test samples. The average value of the root mean square error of the steady-state thermal resistance and the root mean square error of the transient thermal resistance is taken as the overall root mean square error evaluation index. For the steady-state thermal resistance output dimension, calculate the ratio of the residual sum of squares to the total sum of squares to obtain the determination coefficient of the steady-state thermal resistance. The residual sum of squares is the sum of squares of the differences between the predicted value and the true value of each test sample, and the total sum of squares is the sum of squares of the differences between the true value and the mean of the true value of each test sample. For the transient thermal impedance output dimension, calculate the ratio of the residual sum of squares to the total sum of squares to obtain the determination coefficient of the transient thermal impedance; The average value of the determination coefficient of the steady-state thermal resistance and the determination coefficient of the transient thermal resistance is taken as the overall determination coefficient evaluation index. Output the final predicted results of steady-state thermal resistance and transient thermal impedance for each test sample, and output the mean absolute percentage error, the root mean square error and the coefficient of determination accordingly.
[0013] To achieve the above objectives, another aspect of this application proposes a device for predicting IGCT thermal resistance performance parameters based on the IWOA-BP model, the device comprising: The first module is used to acquire the monitoring data collected in sequence during the power cycling test of the IGCT device under test, and obtain the original timing monitoring data sequence. The second module is used to clean and normalize the original time-series monitoring data sequence to obtain a standard time-series data sequence. The third module is used to perform sliding window segmentation on the standard time series data sequence according to a preset window length and step size to generate a number of time series sample blocks, and to establish a mapping relationship between each time series sample block and the thermal resistance parameter at the corresponding prediction step size to obtain a total sample set. The total sample set is then divided into a training set and a test set according to a preset ratio. The fourth module is used to construct a hybrid prediction model that combines the improved whale optimization algorithm with the BP neural network, and to obtain an initial hybrid prediction model with the weights to be optimized and the thresholds to be determined. The improved whale optimization algorithm introduces Logistic chaotic mapping for population initialization, adopts a cosine nonlinear decreasing convergence factor, and introduces adaptive weights based on individual fitness in position updates. The fifth module is used to perform global iterative optimization of the weights and thresholds of the BP neural network in the initial hybrid prediction model using the improved whale optimization algorithm, and to obtain the optimized hybrid prediction model with the initial weights and thresholds using the mean squared error on the training set as the fitness evaluation criterion. The sixth module is used to assign the optimized initial weights and thresholds, and then perform gradient descent training on the hybrid prediction model on the training set to obtain the trained IWOA-BP prediction model. The seventh module is used to verify the prediction accuracy of the trained IWOA-BP prediction model using the test set, and to obtain the final prediction results of the IGCT thermal resistance performance parameters using the mean absolute percentage error, root mean square error and coefficient of determination as model performance evaluation indicators.
[0014] To achieve the above objectives, another aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described above.
[0015] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the methods described above.
[0016] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer program product, including a computer program that, when executed by a processor, implements the aforementioned method.
[0017] The embodiments of this application include at least the following beneficial effects: This application provides a method, apparatus, electronic device, storage medium, and program product for predicting the thermal resistance performance parameters of IGCT based on the IWOA-BP model. The solution includes: acquiring monitoring data collected sequentially during power cycling testing of the IGCT device under test to obtain an original time-series monitoring data sequence; obtaining a standard time-series data sequence; dividing the total sample set into a training set and a test set according to a preset ratio; constructing a hybrid prediction model combining an improved whale optimization algorithm and a BP neural network to obtain an initial hybrid prediction model with undetermined weights and thresholds; obtaining an optimized hybrid prediction model with initial weights and thresholds; obtaining a trained IWOA-BP prediction model; and verifying the prediction accuracy of the trained IWOA-BP prediction model using the test set to obtain the final prediction result of the IGCT thermal resistance performance parameters. This application can accurately predict the thermal resistance performance parameters of IGCT devices. Attached Figure Description
[0018] Figure 1 This is a flowchart of the IGCT thermal resistance performance parameter prediction method based on the IWOA-BP model provided in the embodiments of this application; Figure 2 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation
[0019] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.
[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.
[0021] Before providing a detailed description of the embodiments of this application, some of the nouns and terms involved in the embodiments of this application will be explained first. The nouns and terms involved in the embodiments of this application are subject to the following interpretations.
[0022] 1) IWOA-BP: An improved model combining the whale optimization algorithm and the BP neural network. It uses the improved whale optimization algorithm to globally optimize the initial weights and thresholds of the BP neural network, and then assigns the optimal parameters obtained by optimization to the BP neural network for fine training, so as to improve the prediction accuracy and convergence speed of the network.
[0023] 2) IGCT: Integrated Gate Commutated Thyristor. A high-power semiconductor switching device widely used in high-voltage direct current transmission, rail transit traction, industrial frequency conversion, and other fields. This scheme uses the IGCT as the prediction object, and predicts its internal thermal resistance performance parameters through its external electrical and thermal monitoring data.
[0024] 3) BP: Backpropagation Neural Network. A multi-layer feedforward neural network trained using the backpropagation algorithm, consisting of an input layer, hidden layers, and an output layer. It continuously adjusts the connection weights and thresholds between layers using gradient descent, making the network output gradually approach the target value. It is one of the fundamental models of deep learning.
[0025] 4) IWOA: Improved Whale Optimization Algorithm. A swarm intelligence optimization algorithm that simulates the bubble-web hunting behavior of humpback whales, finding the global optimum by iteratively updating individual positions in the search space. This scheme introduces three improvements to the standard whale optimization algorithm: Logistic chaotic mapping for population initialization, a cosine-type nonlinear decreasing convergence factor, and adaptive weights based on individual fitness to enhance global search capability and convergence speed.
[0026] 5) FEM: Finite Element Method. A numerical engineering calculation method for solving partial differential equations. It approximates the solution of physical field problems such as heat conduction and structural mechanics by discretizing the continuum into a finite number of small elements. In traditional methods for predicting thermal resistance in IGCT, FEM is often used for thermal simulation calculations, but it is computationally intensive and time-consuming. This solution uses a data-driven approach to replace traditional FEM simulation, achieving rapid prediction.
[0027] 6) TSEP: Temperature-Sensitive Electrical Parameter. Utilizing the linear relationship between certain electrical parameters of semiconductor devices (such as saturation voltage drop at low current, threshold voltage, etc.) and junction temperature, the junction temperature of the device can be indirectly calculated by measuring these electrical parameters. TSEP is a commonly used method for measuring junction temperature in IGCT, and the junction temperature monitoring data in this scheme comes from TSEP measurement technology.
[0028] 7) Logistic Mapping. This scheme uses Logistic mapping to generate chaotic sequences to initialize the whale population, ensuring that the initial individuals are evenly distributed in the search space, avoiding population aggregation, and enhancing global search capabilities.
[0029] This invention takes into account that, in related technologies, the prediction techniques for the thermal resistance parameters of IGCTs (Integrated Gate Commutated Thyristors) mainly rely on two major categories of methods: "physical modeling" and "data-driven" approaches. 1. Numerical simulation based on the finite element method (FEM): This is a mainstream physical modeling method. A 3D physical model of the IGCT is constructed using software such as ANSYS Icepak and Flotherm to simulate its internal temperature distribution and heat flow path, thereby calculating the thermal resistance. For example, some studies have established accurate 3D thermal models of IGCTs and fitted measured data by adjusting parameters such as contact thermal resistance to obtain internal thermal characteristics.
[0030] 2. Lumped-parameter thermal network model method: This method simplifies the complex thermal system into an equivalent circuit network composed of thermal resistance (R) and heat capacity (C), enabling rapid calculation of junction temperature. In high-power converter systems, thermal resistance network models are often used to calculate the die temperature rise of the IGCT and to design the cooling system accordingly. Furthermore, water-cooling system modeling methods extract equivalent thermal resistance parameters to achieve real-time calculation of the IGCT junction temperature.
[0031] 3. Structure Function Method: This method is a non-destructive analysis technique based on the principle of heat conduction. It can deduce the thermal resistance and thermal capacity of each layer of material inside the device, from the junction to the shell, from the transient thermal response. Studies have shown that the structure function method can obtain the thermal resistance and thermal capacity information of each layer inside the IGCT with high accuracy, and it is currently recognized as one of the most accurate measurement methods.
[0032] 4. Experimental Testing Methods (Electrical / Optical Methods): This method directly measures the junction temperature using the IGCT's own temperature-sensitive electrical parameter (TSEP) or infrared / fiber optic temperature measurement devices, then calculates the thermal resistance based on power loss. For example, some studies have used fiber optic temperature sensors to measure junction temperature and combined this with multiphysics simulations to determine the thermal resistance of SiC (silicon carbide) power modules. Furthermore, electrical methods are also used to measure temperature rise and thermal resistance in device thermal characteristic analysis.
[0033] Although the above methods have been widely used, they have gradually revealed some limitations when faced with the increasing complexity of applications: 1. Complex physical modeling and high computational cost: The process of building a 3D simulation model based on the finite element method is cumbersome. A high-precision thermal simulation usually takes several hours or even longer, which is difficult to meet the needs of rapid design iteration or real-time status monitoring.
[0034] 2. There is a contradiction between the accuracy and adaptability to complex scenarios in thermal network models: Traditional parallel thermal circuit models / series thermal circuit models have fixed orders, making it difficult to accurately capture the thermal behavior of IGCT under multiple heat source coupling and complex dynamic conditions.
[0035] 3. High dependence on precise physical parameters, accuracy is easily affected: The accuracy of physical models (especially finite element models) heavily depends on the precise input of a large number of physical parameters such as chip / package material properties and interface contact thermal resistance. However, these parameters often vary due to factors such as material batches, aging levels, and process differences, leading to prediction deviations. While model correction (such as correcting the junction environment thermal resistance matrix) can improve accuracy, the process is time-consuming.
[0036] 4. Limitations of experimental measurements, including destructive or difficult-to-monitor in-situ methods: High-precision experimental measurements, such as structure function methods and optical methods, involve expensive equipment and complex operations. Some methods (such as optical methods) require opening the module cover, which is destructive. Although electrical methods can be used for online measurement, TSEP extraction is often affected by noise and operating conditions, calibration is complex, and long-term application may result in decreased accuracy due to chip aging.
[0037] 5. Pure data-driven methods have weak generalization ability and poor long-term adaptability: Traditional pure data-driven models such as backpropagation (BP) neural networks, although fast in prediction, directly establish a black box model of the "loss-junction temperature" relationship. When the device ages or the operating conditions change drastically, the original model relationship may become invalid and need to be retrained, making it unsuitable for long-term thermal estimation tasks.
[0038] Predicting the thermal resistance performance parameters of IGCT requires a method that balances high accuracy, fast response, and strong generalization. The powerful global optimization capability of the IWOA algorithm finds the optimal initial weights and thresholds for the BP neural network, fundamentally improving the model's prediction performance. The powerful nonlinear mapping capability of the BP neural network can also replace complex physical models. The model combining the IWOA algorithm and the BP neural network can avoid the problem of the algorithm easily getting trapped in local optima, and can also better adapt to dynamic changes in data (such as device aging, operating condition drift, etc.), constructing a prediction model with stronger generalization ability and higher robustness than traditional offline physical models and standard neural networks, and realizing effective tracking of long-term operating thermal characteristics.
[0039] In view of this, this application provides a method and related equipment for predicting the thermal resistance performance parameters of IGCT based on the IWOA-BP model. This scheme obtains the timing monitoring data of IGCT devices in power cycling tests, and constructs a training sample set after data cleaning, normalization and sliding window feature construction. Then, a hybrid prediction model combining the improved whale optimization algorithm and the BP neural network is constructed. The improved whale optimization algorithm is used to globally optimize the initial weights and thresholds of the BP neural network, and then the optimized parameters are assigned to the BP neural network for gradient descent fine-tuning training. Finally, the accurate prediction of the two thermal resistance performance parameters of IGCT devices, namely steady-state thermal resistance and transient thermal impedance, is achieved.
[0040] Figure 1This is an optional flowchart of a method for predicting IGCT thermal resistance performance parameters based on the IWOA-BP model provided in an embodiment of this application. Figure 1 The method may include, but is not limited to, steps S101 to S107.
[0041] Step S101: Obtain the monitoring data collected in sequence during the power cycling test of the IGCT device under test to obtain the original timing monitoring data sequence; Step S102: Clean and normalize the original time-series monitoring data sequence to obtain a standard time-series data sequence; Step S103: The standard time series data sequence is divided into sliding window segments according to the preset window length and step size to generate several time series sample blocks. The mapping relationship between each time series sample block and the thermal resistance parameter at the corresponding prediction step size is established to obtain the total sample set. The total sample set is divided into training set and test set according to the preset ratio. Step S104: Construct a hybrid prediction model combining the improved whale optimization algorithm and the BP neural network to obtain an initial hybrid prediction model with undetermined weights and thresholds; The improved whale optimization algorithm introduces Logistic chaotic mapping for population initialization, adopts a cosine nonlinear decreasing convergence factor, and introduces adaptive weights based on individual fitness in position updates. Step S105: Use the improved whale optimization algorithm to perform global iterative optimization of the weights and thresholds of the BP neural network in the initial hybrid prediction model. Use the mean square error on the training set as the fitness evaluation criterion to obtain the optimized hybrid prediction model with the initial weights and thresholds. Step S106: After assigning the optimized initial weights and thresholds, perform gradient descent training on the hybrid prediction model on the training set to obtain the trained IWOA-BP prediction model. Step S107: The prediction accuracy of the trained IWOA-BP prediction model is verified using the test set, and the mean absolute percentage error, root mean square error and coefficient of determination are used as the model performance evaluation indicators to obtain the final prediction results of the IGCT thermal resistance performance parameters.
[0042] Steps S101 to S107 as illustrated in the embodiments of this application optimize the initial parameters of the BP network through an improved whale algorithm that constructs time-series samples using a sliding window, performs chaotic initialization and adaptive weight enhancement, and fine-tunes the training using gradient descent. This avoids the BP network from getting stuck in local optima and improves the convergence speed and generalization ability. Thus, by utilizing easily measurable external electrical and thermal monitoring data from IGCT, high-precision prediction of steady-state thermal resistance and transient thermal impedance can be achieved, providing a basis for the performance evaluation and reliability design of power semiconductor devices.
[0043] In some embodiments, step S101 may include, but is not limited to, steps S111 to S112: Step S111: Continuously collect monitoring data of the IGCT device during the power cycling test according to the preset sampling period; the monitoring data includes collector-emitter voltage, gate voltage, collector current, case temperature, and junction temperature; Step S112: Arrange the monitoring data in chronological order to form the original time-series monitoring data sequence.
[0044] In steps S111 to S112 of some embodiments, by acquiring timing monitoring data of the IGCT device during power cycling testing, and ensuring that the monitoring data includes at least collector-emitter voltage, gate voltage, collector current, case temperature, and junction temperature, raw data reflecting the electrical operating state and thermal response characteristics of the IGCT device can be comprehensively collected. This provides a sufficient information basis for subsequent prediction of thermal resistance parameters and avoids inaccurate predictions due to missing or insufficient data. Simultaneously, the timing-acquired data retains the dynamic changes of the IGCT during power cycling, enabling the prediction model to capture the timing characteristics of thermal resistance parameters evolving with operating conditions.
[0045] In some embodiments, step S102 may include, but is not limited to, steps S201 to S203: Step S201: Calculate the mean and standard deviation of each dimension of the feature variables in the original time series monitoring data sequence, and remove data points that deviate from the mean by more than three times the standard deviation as outliers. Step S202: For the missing positions generated after outlier removal, linear interpolation of the nearest valid data points before and after the missing position is used to fill the gaps. When the beginning or end of the sequence is missing, nearest neighbor filling is used. Step S203: The cleaned feature data of each dimension is transformed to the [0,1] interval using the Min-Max method to obtain the standard time series data sequence.
[0046] In steps S201 to S203 of some embodiments, the original time-series monitoring data sequence is cleaned and normalized—specifically including removing outliers using the 3σ principle, filling missing values with linear interpolation, and normalizing to the [0,1] interval using Min-Max—which can remove abnormal sampling points caused by sensor noise, electromagnetic interference, or occasional faults during the IGCT test, and restore the continuity and integrity of the data through linear interpolation, avoiding interference from outliers and missing values on model training; the normalization process unifies features such as voltage, current, and temperature of different dimensions to the same scale, eliminating the influence of dimensions, and avoiding gradient update imbalance caused by excessive differences in the numerical range of features in each dimension during the training process of the BP neural network, thereby improving the convergence speed and prediction stability of the model.
[0047] In some embodiments, step S103 may include, but is not limited to, steps S301 to S303: Step S301: Perform sliding window segmentation on the standard time series data sequence according to the preset window length and step size. Each window contains feature fields of preset dimensions. The data matrix in each window is used as an input sample. The feature fields include voltage, current, case temperature, junction temperature, timestamp, and power loss. Step S302: Establish a mapping between each input sample and the thermal resistance parameter at the preset prediction step size after the corresponding window ends. All windows are stacked to form an input tensor. The thermal resistance parameter corresponding to each window constitutes the output target vector. The input tensor and the output target vector are paired to form the total sample set. Step S303: The total sample set is randomly divided into a training set and a test set according to a preset ratio; the training set is used for model training; the test set is used for prediction accuracy verification.
[0048] In steps S301 to S303 of some embodiments, the standard time-series data sequence is divided into several time-series sample blocks by sliding window segmentation according to a preset window length and step size. A mapping relationship is established between each time-series sample block and the thermal resistance parameter at the corresponding prediction step size, resulting in a total sample set divided into a training set and a test set. This converts continuous long-term data into several fixed-length time-series sample blocks. Each sample block contains six-dimensional features: voltage, current, case temperature, junction temperature, timestamp, and power loss, completely preserving historical operating information during the formation of IGCT thermal resistance parameters. The window sliding step size is smaller than the window length, resulting in data overlap between adjacent windows. This increases the number of samples to meet the data requirements for neural network training while ensuring the continuity of time-series information, solving the problem that traditional static samples cannot reflect the thermal accumulation effect. Furthermore, dividing the total sample set into a training set and a test set according to a preset ratio, with the training set used for model training and the test set used for prediction accuracy verification, effectively evaluates the model's generalization ability, avoids overfitting, and ensures that the prediction model has high prediction accuracy for unknown data.
[0049] In some embodiments, step S104 may include, but is not limited to, steps S401 to S405: Step S401: Set the BP neural network as a three-layer feedforward structure consisting of an input layer, a hidden layer, and an output layer. Set the number of nodes in the input layer, the number of nodes in the hidden layer, and the number of nodes in the output layer. Define the parameters to be optimized as all connection weights and thresholds of the BP neural network to obtain the BP neural network structure settings. Step S402: Set the population size and maximum number of iterations for the improved whale optimization algorithm. Generate an initial population using Logistic chaotic mapping and map each chaotic value to the search range of each parameter to be optimized to obtain the settings for the improved whale optimization algorithm. Step S403: Set the convergence factor to decrease with the number of iterations in a cosine form; Step S404: When setting the position update, the adaptive weight is calculated based on the difference between the fitness of each individual whale and the average fitness and minimum fitness of the population. The adaptive weight is used to weight the position update in the encirclement and predation and random search phases. In the bubble net attack phase, the position is updated by selecting the shrinking encirclement or spiral ascent mode with a preset probability. Step S405: An initial hybrid prediction model with undetermined weights and thresholds is constructed by setting the BP neural network structure and the improved whale optimization algorithm.
[0050] In steps S401 to S405 of some embodiments, a hybrid prediction model combining the improved whale optimization algorithm and the BP neural network is constructed. The improved whale optimization algorithm incorporates Logistic chaotic mapping for population initialization, adopts a cosine-type nonlinear decreasing convergence factor, and introduces adaptive weights based on individual fitness in position updates. This ensures that the initial whale population is evenly distributed in the search space without clustering, avoiding the problems of insufficient population diversity and premature convergence to local optima caused by random initialization in the standard whale algorithm. The cosine-type nonlinear decreasing convergence factor allows the algorithm to maintain global search capability for a longer period in the early stages of iteration to fully explore the solution space, and then quickly switches to local exploration in the later stages to accelerate convergence. The adaptive weights dynamically adjust the step size of position updates according to individual performance. Individuals with excellent performance search more precisely near the optimal solution to improve accuracy, while individuals with poor performance explore with larger step sizes to escape local optima. Through these three improvements, IWOA can provide the BP neural network with a set of initial weights and thresholds close to the global optimum, overcoming the problems of the BP neural network's sensitivity to initial parameters and its tendency to get trapped in local minima.
[0051] In some embodiments, step S105 may include, but is not limited to, steps S501 to S507: Step S501: Encode all the weights and thresholds to be optimized in the BP neural network of the initial hybrid prediction model into position vectors for each individual whale. Step S502: Decode the position vector of each individual whale into a set of weights and thresholds for a BP neural network, perform a forward propagation on the training set, calculate the mean square error between the predicted value of the network output and the target output, and use the mean square error as the fitness value of each individual whale. Step S503: In the current iteration, calculate the current value of the convergence factor and the coefficient vector; Step S504: For each individual whale, select an encirclement predation update strategy, a bubble net attack update strategy, or a random search update strategy based on the relationship between the modulus of the coefficient vector and the preset probability, and use adaptive weights to perform position weighted updates in the encirclement predation and random search strategies. Step S505: Perform boundary detection on the updated position vector of each individual whale, and truncate the components that exceed the search range to the boundary value of the search range. Step S506: Recalculate the fitness value of each individual whale, update and record the best individual in the current population and its position vector; Step S507: Repeat the iterative process of calculating the convergence factor and coefficient vector, selecting an update strategy for position update, boundary detection, recalculating fitness and updating the optimal individual until the maximum number of iterations is reached. Then, decode the position vector corresponding to the final optimal individual into a set of weights and thresholds for the BP neural network, and assign them to the hybrid prediction model as the optimized initial weights and thresholds.
[0052] In steps S501 to S507 of some embodiments, the improved whale optimization algorithm is used to perform global iterative optimization of the weights and thresholds to be optimized in the BP neural network of the initial hybrid prediction model, and the mean squared error on the training set is used as the fitness evaluation criterion. Specifically, this includes encoding the parameters to be optimized as whale individual position vectors, decoding and calculating the mean squared error as fitness, selecting one of three update strategies based on |A| and p, boundary detection truncation, and iterating until the maximum number of iterations. This can utilize the global search capability of swarm intelligence algorithms to find a set of initial parameters that minimize the prediction error of the BP network in a parameter space of up to 110 dimensions, avoiding network performance instability caused by manual random initialization or empirical initialization. By switching between three strategies—surrounding predation, bubble net attack, and random search—the algorithm's global exploration and local development capabilities are taken into account. Boundary detection ensures that each dimension of the parameters of all whale individuals is always within the preset search range, avoiding invalid searches. The final output set of optimized initial weights and thresholds enables the BP neural network to be in a better parameter state at the beginning of training, laying the foundation for subsequent fine-tuning training.
[0053] In some embodiments, step S106 may include, but is not limited to, steps S601 to S604: Step S601: Use the optimized initial weights and thresholds as the initial values for the connection weights and thresholds of each layer in the BP neural network of the hybrid prediction model; Step S602: Using mean squared error as the loss function on the training set, the gradient descent method is used to iteratively update the connection weights and thresholds of each layer of the BP neural network in the hybrid prediction model. The gradients of the output layer weights and thresholds are calculated by multiplying the partial derivative of the loss with respect to the net input of each output layer by the corresponding hidden layer output. The gradients of the hidden layer weights and thresholds are calculated by multiplying the partial derivative of the loss with respect to the net input of each hidden layer by the corresponding input features. The partial derivative of the net input of the hidden layer is determined by multiplying the derivative of the hidden layer activation function by the backpropagation error from the output layer. Step S603: Update the connection weights and thresholds by multiplying each gradient by the preset learning rate; Step S604: Repeat the following operations on the training set: calculate the loss using mean squared error as the loss function, calculate the gradient of each layer's connection weights and thresholds using gradient descent, and update the connection weights and thresholds according to the preset learning rate, until the loss function converges or the preset maximum number of fine-tuning iterations is reached, to obtain the trained IWOA-BP prediction model.
[0054] In steps S601 to S604 of some embodiments, after assigning optimized initial weights and thresholds, the hybrid prediction model is fine-tuned on the training set using gradient descent. Specifically, this includes using the optimized parameters as the initial values of the weights and thresholds of each layer of the BP network, using the mean squared error as the loss function, and iteratively updating the parameters of each layer using the gradient descent method. This utilizes the high-quality initial parameters provided by IWOA, enabling the BP neural network to converge to a smaller loss value with fewer iterations during the fine-tuning phase. This overcomes the problem of the standard BP network being highly sensitive to initial parameters and fully leverages the efficiency and accuracy of the gradient descent method in local search. The chain rule calculation of the gradients of the output layer and hidden layers ensures that the error signal can be effectively backpropagated from the output layer to the input layer, allowing the weights and thresholds of each layer to be reasonably adjusted. Parameter updates are performed by multiplying the gradient by the preset learning rate, achieving fine-tuning of the BP network and ultimately obtaining a prediction model that combines global optimality and local accuracy.
[0055] In some embodiments, step S107 may include, but is not limited to, steps S701 to S706: Step S701: Input each input sample in the test set into the trained IWOA-BP prediction model, and calculate the predicted value of the thermal resistance parameter of each test sample through forward propagation. Step S702: Inverse normalize the predicted thermal resistance parameter values to the original dimensions to obtain the final predicted thermal resistance parameter results for each test sample. Step S703: For the steady-state thermal resistance output dimension, calculate the absolute value of the difference between the predicted value and the true value of each test sample, divide it by the true value, sum the results, take the average, and then multiply by the percentage to obtain the average absolute percentage error of the steady-state thermal resistance; for the transient thermal resistance output dimension, calculate the absolute value of the difference between the predicted value and the true value of each test sample, divide it by the true value, sum the results, take the average, and then multiply by the percentage to obtain the average absolute percentage error of the transient thermal resistance; take the average of the average absolute percentage error of the steady-state thermal resistance and the average absolute percentage error of the transient thermal resistance as the overall average absolute percentage error evaluation index. Step S704: For the steady-state thermal resistance output dimension, calculate the square root of the sum of the squares of the differences between the predicted and true values of each test sample, divided by the total number of test samples, to obtain the root mean square error of the steady-state thermal resistance; for the transient thermal impedance output dimension, calculate the square root of the sum of the squares of the differences between the predicted and true values of each test sample, divided by the total number of test samples, to obtain the root mean square error of the transient thermal impedance; take the average of the root mean square error of the steady-state thermal resistance and the root mean square error of the transient thermal impedance as the overall root mean square error evaluation index. Step S705: For the steady-state thermal resistance output dimension, calculate the ratio of the residual sum of squares to the total sum of squares to obtain the determination coefficient of the steady-state thermal resistance. The residual sum of squares is the sum of squares of the differences between the predicted value and the true value of each test sample, and the total sum of squares is the sum of squares of the differences between the true value and the mean of the true values of each test sample. For the transient thermal impedance output dimension, calculate the ratio of the residual sum of squares to the total sum of squares to obtain the determination coefficient of the transient thermal impedance. Take the average of the determination coefficients of the steady-state thermal resistance and the transient thermal impedance as the overall determination coefficient evaluation index. Step S706: Output the final predicted results of steady-state thermal resistance and transient thermal impedance for each test sample, and output the mean absolute percentage error, root mean square error and coefficient of determination accordingly.
[0056] In steps S701 to S706 of some embodiments, the prediction accuracy of the trained IWOA-BP prediction model is verified using a test set, and the mean absolute percentage error, root mean square error, and coefficient of determination are used as model performance evaluation indicators. Specifically, this includes inputting the test set into the trained model to obtain the predicted value of each test sample, inverse normalizing the predicted value to the original dimensions, and calculating the MAPE, RMSE, and R of the two output dimensions of steady-state thermal resistance and transient thermal resistance, respectively. 2 By averaging the results, the predictive performance of the model can be comprehensively evaluated from three different dimensions: relative error, absolute error, and goodness of fit. MAPE measures the percentage of prediction error relative to the true value, intuitively reflecting the prediction accuracy; RMSE is sensitive to larger errors, reflecting the dispersion of the predicted values; R... 2The three metrics measure the model's ability to explain the variation in real data, reflecting the model's fit. They evaluate the model from different perspectives, avoiding the limitations of a single metric and ensuring the comprehensiveness and reliability of the model performance assessment. Inverse normalization restores the predicted values to their original dimensions (°C / W), giving the error metrics practical physical meaning and facilitating direct comparison with traditional finite element simulation results or measured values.
[0057] As an optional implementation, the IGCT thermal resistance performance parameter prediction method based on the IWOA-BP model provided by the present invention includes the following steps: A. Data collection; B. Data processing; C. Model building; D. Error analysis.
[0058] Step A specifically involves acquiring timing monitoring data of the IGCT device during power cycling testing. The data includes collector-emitter voltage, gate voltage, collector current, case temperature, and junction temperature.
[0059] Step B specifically includes: B1. Data cleaning of time series monitoring data: outliers are removed by the 3σ principle and missing values are filled by linear interpolation; the normalization method is Min-Max normalization to the [0,1] interval. For the acquired IGCT time-series monitoring data (such as collector-emitter voltage) collector current Shell temperature (etc.), assuming a certain feature variable is Calculate its mean and standard deviation:
[0060]
[0061] in, The mean (average value) of a certain feature variable; The standard deviation of this feature variable reflects the degree of dispersion of the data relative to the mean. This represents the total number of data points for this feature variable. For the feature variable of this dimension, the first The values of each data point.
[0062] According to the "3σ rule" of the normal distribution, if a data point satisfies:
[0063] Then it is believed These are outliers and will be removed; after outlier removal, gaps will appear in the original time series. Let time be... The data at point is missing (or has been removed), and the nearest valid data points before and after it are respectively and Then, the missing values can be calculated and filled using the linear interpolation formula:
[0064] After removing outliers, gaps will appear in the original time series. Let time be... The data at point is missing (or has been removed), and the nearest valid data points before and after it are respectively ( ), ( Then, the fill value is calculated using the linear interpolation formula:
[0065] For cases where the beginning or end of the sequence is missing, nearest neighbor filling is used.
[0066] in, The moment when data is missing (or when outliers are removed and no data is available); In order to be in Data values that need to be filled in at any time; The time corresponding to the nearest valid data point before the missing time; The value is the nearest valid data point before the missing time. This is the time corresponding to the nearest valid data point after the missing time. This represents the value of the nearest valid data point after the missing time.
[0067] Min-Max normalization is used to transform all features to the [0,1] interval. For each feature dimension... (Cleaning complete), let its minimum value be... The maximum value is Then the normalized for:
[0068] in, If the value is greater than 0, the transformation preserves the original data distribution, eliminates the influence of dimensions, and facilitates subsequent BP neural network training.
[0069] in, The original data values of a certain feature variable; This is the result of normalizing the data value, ranging from [0,1]. It is the minimum value among the feature variables of this dimension; It is the maximum value among the feature variables of this dimension.
[0070] Within a certain sliding window For example, a sequence: Original data: [2.1, 2.3, 2.8, 25.0, 2.4] (where 25.0 is an outlier); After removing 25.0, the sequence is [2.1,2.3,2.8,2.4][2.1,2.3,2.8,2.4], with position indices 1,2,3,5, and a gap at index 4. Linear interpolation is then performed using the values at indices 3 and 5.
[0071] The padded sequence is [2.1,2.3,2.8,2.6,2.4].
[0072] Normalization: ,but:
[0073] The result is [0, 0.286, 1, 0.714, 0.429].
[0074] B2. Sliding window feature construction: The window length is 50 sampling points and the step size is 10 sampling points. Each window contains 6-dimensional feature fields: voltage, current, case temperature, junction temperature, timestamp, and power loss.
[0075] Assume the preprocessed IGCT time series dataset contains M sampling times, with each time time recording a 6-dimensional feature vector: ,
[0076] in, For a moment 6-dimensional feature vectors (column vectors); This refers to the collector-emitter voltage (electrical characteristic). This refers to the collector current (electrical characteristic). Case temperature — the temperature of the device's outer casing (thermal characteristics); Junction temperature—the temperature of the PN junction inside the device (thermal characteristic); The timestamp is a normalizable unit based on the sampling period. For power loss, usually ; This is a transpose, converting a row vector into a column vector.
[0077] Given: window length L = 50 (sampling points), stride S = 10 sampling points, feature dimension D = 6. The sliding window starts from the beginning of the sequence at position t=1, and moves S points to the right each time, generating a total of K windows:
[0078] in, This represents the total number of sliding windows (i.e., the number of samples generated). This represents the total number of sampling points in the preprocessed time-series dataset. This is the window length, i.e., the number of sampling points contained in each window; The step size is the number of sampling points moved to the right each time.
[0079] For the windows ( Its starting index is:
[0080] in, For the first The starting index of the window; The step size.
[0081] The range of sampling point indices contained in this window is:
[0082] in, For the first The set of sampling point indices contained in each window; This is the window length.
[0083] The data matrix within the window is:
[0084] Each window As an input sample, its corresponding target output is the thermal resistance parameter (such as steady-state thermal resistance) within a certain prediction step after the end of this window. ).
[0085] in, For the first The data matrix of each window (shape: ); This is the 6-dimensional feature vector at the starting index of the window; The 6-dimensional feature vector at the end index of the window; It is a real matrix with dimension . Line × List.
[0086] Stack all windows to form a three-dimensional input tensor:
[0087] Example: Assuming the original sequence length M = 200, then:
[0088] The starting indices of the windows are 1, 11, 21, …, 151. Each window takes 50 consecutive sampling points, covering 6 feature fields, and finally generates 16 input samples with a shape of 50×6.
[0089] Step C specifically includes: C1, the parameters to be optimized in the BP neural network are all connection weights and thresholds. Let the network structure be... The total number of parameters is:
[0090] in, The total number of parameters to be optimized in the BP neural network (all connection weights and thresholds); This represents the number of nodes in the input layer. This represents the number of hidden layer nodes. This represents the number of nodes in the output layer.
[0091] Each individual whale represents a set of parameter vectors. The fitness function is the mean squared error (MSE) of the BP network on the training set:
[0092] in, This is the fitness function value, i.e., the mean squared error; The position vector of an individual whale (encoding a complete set of parameters of the BP network); The total number of samples in the training set; For the first The model's predicted values for each training sample; For the first The true target value of each training sample.
[0093] The goal of IWOA (Improved Whale Optimization Algorithm) is to find a way to make... smallest .
[0094] The standard whale algorithm uses random initialization of the population, which can easily lead to insufficient population diversity. We introduce a Logistic chaotic mapping to generate the initial population, enhancing global search capabilities. The Logistic mapping formula is as follows:
[0095] in, For the first The chaos value of the step; For the first The chaos value of the step (current value).
[0096] Usually taken (Complete chaos) and .
[0097] For the The first whale The generation steps for the dimension parameters are as follows: 1. Randomly generate a chaotic initial value. ; 2. Iteration This yields a chaotic sequence:
[0098] 3. Map chaotic values to parameter ranges :
[0099] Where lb and ub are the lower and upper bounds of the weight and threshold, respectively (usually taken as...). or ); For the first The first individual whale Dimensional chaotic value (between [0,1]).
[0100] The standard WOA convergence factor α decreases linearly from 2 to 0, resulting in abrupt switching between early global search and later local development. An improvement is made using a cosine-based non-linear decreasing convergence factor α.
[0101] Where t is the current iteration number. The maximum number of iterations is denoted by α. This function decreases slowly in the early stages, maintaining a large α to sustain the global search; it decreases rapidly in the later stages, accelerating local convergence.
[0102] The calculation of coefficients A and C remains unchanged:
[0103] Where r is A random vector within.
[0104] Introducing Adaptive Weights in Whale Position Updates This weight is dynamically adjusted based on the current fitness of the individual, balancing global and local search. Define the first... The fitness of the whales in the current iteration is: The average fitness in the population is The optimal fitness is The formula for calculating the weight is as follows:
[0105] in To prevent small positive numbers with a denominator of zero. For individuals who are better than average. Individuals with larger weights are encouraged to perform finer searches near the current best; those with worse weights are given smaller weights to encourage them to escape local optima. For the first The whale in the Adaptive weights for the next iteration.
[0106] Weighted predator encirclement position update: Encirclement and predation phase :
[0107] Bubble web attack (shrinking encirclement and spiral ascent), choosing between shrinking encirclement or spiral update with probability p=0.5:
[0108] in, This is the updated whale position vector; The current position vector; This is the current optimal whale position vector (prey position).
[0109] Where b is the helical shape constant (usually taken as 1). for[ A random number within the range [1, 1].
[0110] Random search phase : in Individuals randomly selected from the current population.
[0111] A summary of the specific steps for optimizing BP in IWOA: 1. Parameter initialization: Set the whale population size Maximum number of iterations In the BP network structure, lb= 1,ub=1.
[0112] 2. Chaotic Population Initialization: Generating Population Using Logistic Mapping Each whale is an individual, and each individual encodes all the weights and thresholds of the backpropagation (BP).
[0113] 3. Calculate fitness: Decode each individual into the initial parameters of the BP network, train once on the training set (calculate the error during forward propagation), and fitness = MSE.
[0114] 4. Update the convergence factor a (the cosine form mentioned above) and the coefficients A and C.
[0115] 5. For each whale, according to and random probability Choose an update strategy (surround predation, spiral update, or random search) and apply adaptive weights. .
[0116] 6. Boundary handling: Truncate out-of-bounds parameters to [lb, ub].
[0117] 7. Recalculate fitness and retain the current best individual. .
[0118] 8. If Return to step 4; otherwise, directly output the globally optimal parameters. .
[0119] 9. The decoder serves as the initial weights and thresholds for the BP network, which is then finely trained on the full training set using gradient descent (or with added momentum) until convergence.
[0120] C2. The structure of the invented IWOA-BP model is as follows: the number of nodes in the input layer is 6, the number of nodes in the hidden layer is 12, the number of nodes in the output layer is 2, the activation function is Sigmoid, and the loss function is mean squared error.
[0121] Input layer: Number of nodes The corresponding 6-dimensional feature vector (voltage, current, case temperature, junction temperature, timestamp, power loss) is constructed using a sliding window.
[0122] Hidden layer: Number of nodes .
[0123] Output layer: Number of nodes The steady-state thermal resistance of the IGCT is output respectively. and transient thermal impedance The predicted value. The network is a fully connected feedforward structure, with layers connected by weights and biases.
[0124] Let the input sample be Hidden layer weight matrix Bias vector .
[0125] Net input to hidden layer:
[0126] Hidden layer activation output (using the Sigmoid function):
[0127] in Each component is calculated independently.
[0128] Output layer weight matrix Bias vector .
[0129] Output layer net input:
[0130] The output layer does not have an activation function (for regression tasks, directly outputting linear values), and the predicted value is:
[0131] Right now:
[0132] in, This is the predicted steady-state thermal resistance value; This is the predicted value of transient thermal impedance.
[0133] For a single sample, the true thermal resistance vector is The predicted value is The loss is:
[0134] in, For the first The predicted value of each output node ( correspond , correspond ); For the first The true target value of each output node; For batch For each sample, the total loss is taken as the average:
[0135] The training objective is to minimize . For the first The first sample The predicted value of each output node; For the first The first sample The actual value of each output node.
[0136] IWOA is used for initial weight and threshold optimization in BP networks, not to replace backpropagation. Specific process: 1. IWOA Optimization Phase: All parameters to be trained Encoded as an individual whale. Total number of parameters:
[0137] Each individual whale is represented by a 110-dimensional vector.
[0138] Fitness function: Perform a full forward propagation on the training set (without backpropagation updates) and calculate the MSE loss.
[0139] IWOA outputs the parameter combination that minimizes the initial MSE through chaotic initialization, nonlinear convergence factor, and adaptive weight iterative search. .
[0140] 2. BP fine-tuning phase: Will As initial parameters for the BP network, they are fine-tuned using standard gradient descent (or with added momentum) through backpropagation.
[0141] Gradient calculation (taking output layer weights as an example): For the output layer one neuron loss to The partial derivative is:
[0142] Output layer weights gradient:
[0143] Output layer bias gradient:
[0144] Hidden layer gradient (using the chain rule, via the Sigmoid derivative) :
[0145] Parameter update (learning rate) ): Iterate until the loss converges or the preset maximum number of iterations is reached (e.g., 500 times).
[0146]
[0147] Step D specifically includes: Suppose there are N samples in the test set, and the i-th The true thermal resistance value of each sample is (This can be steady-state thermal resistance) or transient thermal resistance The model's predicted value is The mean absolute percentage error, root mean square error, and coefficient of determination are used to measure the mean absolute percentage error, root mean square error, and coefficient of determination. To measure error.
[0148] 1. MAPE (Mean Absolute Percentage Error) measures the percentage of prediction error relative to the true value, directly reflecting prediction accuracy. In this patent, MAPE is calculated for each of the two nodes in the output layer, and the average value is taken as the comprehensive index. The calculation formula is as follows:
[0149] MAPE outputs units in percentage (%), when The formula has a division-by-zero issue. In IGCT thermal resistance prediction, the thermal resistance value is usually greater than zero, so it can be applied directly. If the value is close to zero, a small positive number can be added to the denominator. To avoid singularities, the smaller the MAPE, the more accurate the model predictions.
[0150] 2. RMSE (Root Mean Square Error) is sensitive to large errors and reflects the dispersion of predicted values. It has the same dimensions as the original data, and its dimensions are similar to... For the same parameters (e.g., °C / W), a smaller RMSE indicates a stronger model fit. Since the error is squared, RMSE is more sensitive to outliers than MAPE. These parameters are calculated separately in this patent. and The RMSE is calculated by taking the average of the RMSE values and then using it as the overall evaluation index. The formula for calculating RMSE is:
[0151] 3. The coefficient of determination measures how well a model explains the variation in real data, and its value ranges from ( ). [∞,1], the closer to 1, the better the fit. This patent calculates the two output dimensions separately. Then, take the average value and report it. The calculation formula is:
[0152] in The numerator is the mean of the true values, and the numerator is the sum of squared residuals. The denominator is the sum of squares ( When the model predictions are completely consistent with the actual values, When the model's predictions are not as good as those made by directly using the mean, In this patent, It should typically be higher than 0.95.
[0153] If the model output is a normalized value, it needs to be restored to its original dimensions before calculating the error index; if the model directly outputs predictions in the original dimensions, then inverse normalization is not necessary. This patent suggests that no activation function be added to the output layer, and that normalization is completed during the data preprocessing stage, so the model output value... and Inverse normalization is required:
[0154] MAPE and RMSE are calculated independently for each output dimension. List the three metrics for the two output dimensions in tabular form, or take their average as the overall model performance evaluation, as shown in Table 1.
[0155] Table 1. List of Indicators
[0156] This application also provides an IGCT thermal resistance performance parameter prediction device based on the IWOA-BP model, which can implement the above method. The device includes: The first module is used to acquire the monitoring data collected in sequence during the power cycling test of the IGCT device under test, and obtain the original timing monitoring data sequence. The second module is used to clean and normalize the original time-series monitoring data sequence to obtain a standard time-series data sequence. The third module is used to perform sliding window segmentation on the standard time series data sequence according to the preset window length and step size to generate several time series sample blocks, and to establish a mapping relationship between each time series sample block and the thermal resistance parameter at the corresponding prediction step size to obtain the total sample set. The total sample set is then divided into training set and test set according to a preset ratio. The fourth module is used to construct a hybrid prediction model that combines the improved whale optimization algorithm with the BP neural network, and obtains an initial hybrid prediction model with the weights to be optimized and the thresholds to be determined. The improved whale optimization algorithm introduces Logistic chaotic mapping for population initialization, adopts a cosine nonlinear decreasing convergence factor, and introduces adaptive weights based on individual fitness in position updates. The fifth module is used to perform global iterative optimization of the weights and thresholds of the BP neural network in the initial hybrid prediction model using the improved whale optimization algorithm. The mean square error on the training set is used as the fitness evaluation criterion to obtain the optimized hybrid prediction model with the initial weights and thresholds. The sixth module is used to assign optimized initial weights and thresholds, and then perform gradient descent training on the hybrid prediction model on the training set to obtain the trained IWOA-BP prediction model. The seventh module is used to verify the prediction accuracy of the trained IWOA-BP prediction model using the test set, and uses the mean absolute percentage error, root mean square error and coefficient of determination as the model performance evaluation indicators to obtain the final prediction results of the IGCT thermal resistance performance parameters.
[0157] It is understood that the content of the above method embodiments is applicable to the present device embodiments. The specific functions implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.
[0158] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method.
[0159] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.
[0160] Please see Figure 2 , Figure 2 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes: The processor 201 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application. The memory 202 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 202 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 202 and is called and executed by the processor 201 using the methods described in the embodiments of this application. Input / output interface 203 is used to implement information input and output; The communication interface 204 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 205 transmits information between various components of the device (e.g., processor 201, memory 202, input / output interface 203, and communication interface 204); The processor 201, memory 202, input / output interface 203 and communication interface 204 are connected to each other within the device via bus 205.
[0161] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.
[0162] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.
[0163] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.
[0164] It is understood that the content of the above method embodiments is applicable to the embodiments of this program product. The specific functions implemented by the embodiments of this program product are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.
[0165] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.
[0166] The method, apparatus, electronic device, storage medium, and program product for predicting IGCT thermal resistance performance parameters based on the IWOA-BP model provided in this application embodiment acquire, clean, normalize, and construct sliding window features of time-series monitoring data through a data acquisition and processing module to generate a training sample set; perform the fusion of an improved whale optimization algorithm and a BP neural network through a model building module, and use the improved whale optimization algorithm to globally optimize the initial weights and thresholds of the BP network; and perform optimized parameter assignment and gradient descent fine-tuning training through a training module to ultimately achieve accurate prediction of the steady-state thermal resistance and transient thermal impedance of IGCT devices.
[0167] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.
[0168] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.
[0169] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.
[0170] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.
[0171] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.
[0172] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.
[0173] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.
[0174] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.
[0175] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.
[0176] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0177] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.
Claims
1. A method for predicting the thermal resistance performance parameters of IGCT based on the IWOA-BP model, characterized in that, The method includes the following steps: Acquire the monitoring data of the IGCT device under test in power cycling test according to the time sequence to obtain the original time-series monitoring data sequence; The original time-series monitoring data sequence is cleaned and normalized to obtain a standard time-series data sequence; The standard time series data sequence is divided into sliding window segments according to a preset window length and step size to generate several time series sample blocks. A mapping relationship is established between each time series sample block and the thermal resistance parameter at the corresponding prediction step size to obtain a total sample set. The total sample set is then divided into a training set and a test set according to a preset ratio. A hybrid prediction model combining an improved whale optimization algorithm and a BP neural network is constructed to obtain an initial hybrid prediction model with undetermined weights and thresholds. The improved whale optimization algorithm introduces Logistic chaotic mapping for population initialization, adopts a cosine nonlinear decreasing convergence factor, and introduces adaptive weights based on individual fitness in position updates. The improved whale optimization algorithm is used to perform global iterative optimization of the weights and thresholds of the BP neural network in the initial hybrid prediction model. The mean square error on the training set is used as the fitness evaluation criterion to obtain the optimized hybrid prediction model of the initial weights and thresholds. After assigning the optimized initial weights and thresholds, the hybrid prediction model is trained by gradient descent on the training set to obtain the trained IWOA-BP prediction model. The training IWOA-BP prediction model was validated using the test set, and the mean absolute percentage error, root mean square error, and coefficient of determination were used as performance evaluation indicators to obtain the final prediction results of the IGCT thermal resistance performance parameters.
2. The method according to claim 1, characterized in that, The acquisition of monitoring data collected sequentially during power cycling testing of the IGCT device under test yields the original timing monitoring data sequence, including: Monitoring data of the IGCT device during power cycling test is continuously collected according to a preset sampling period; the monitoring data includes collector-emitter voltage, gate voltage, collector current, case temperature, and junction temperature; The monitoring data are arranged in chronological order to form the original time-series monitoring data sequence.
3. The method according to claim 1, characterized in that, The step of cleaning and normalizing the original time-series monitoring data sequence to obtain a standard time-series data sequence includes: The mean and standard deviation of each dimension of the feature variables in the original time-series monitoring data sequence are calculated, and data points that deviate from the mean by more than three times the standard deviation are removed as outliers. For gaps caused by outlier removal, linear interpolation of the nearest valid data points before and after the gap is used to fill them; when the beginning or end of the sequence is missing, nearest neighbor filling is used. The cleaned feature data of each dimension are transformed to the [0,1] interval using the Min-Max method to obtain the standard time series data sequence.
4. The method according to claim 1, characterized in that, The standard time-series data sequence is segmented into several time-series sample blocks by a sliding window with a preset window length and step size. A mapping relationship is established between each time-series sample block and the thermal resistance parameter at the corresponding prediction step size to obtain a total sample set. The total sample set is then divided into a training set and a test set according to a preset ratio, including: The standard time-series data sequence is divided into sliding window segments according to a preset window length and step size. Each window contains feature fields of a preset dimension, and the data matrix within each window is used as an input sample. The feature fields include voltage, current, case temperature, junction temperature, timestamp, and power loss. Each input sample is mapped to the thermal resistance parameter at the preset prediction step size after the corresponding window ends. All windows are stacked to form an input tensor. The thermal resistance parameter corresponding to each window constitutes the output target vector. The input tensor and the output target vector are paired to form the total sample set. The total sample set is randomly divided into a training set and a test set according to a preset ratio; the training set is used for model training; and the test set is used for prediction accuracy verification.
5. The method according to claim 1, characterized in that, The construction of the hybrid prediction model combining the improved whale optimization algorithm and the BP neural network yields an initial hybrid prediction model with undetermined weights and thresholds, including: Set the BP neural network as a three-layer feedforward structure consisting of an input layer, a hidden layer, and an output layer. Set the number of nodes in the input layer, the hidden layer, and the output layer. Define the parameters to be optimized as all connection weights and thresholds of the BP neural network to obtain the BP neural network structure settings. Set the population size and maximum number of iterations for the improved whale optimization algorithm, generate the initial population using Logistic chaotic mapping, map each chaotic value to the search range of each parameter to be optimized, and obtain the settings for the improved whale optimization algorithm. Set the convergence factor to decrease with the number of iterations in a cosine form; When setting the position update, the adaptive weight is calculated based on the difference between the fitness of each individual whale and the average fitness and minimum fitness of the population. The adaptive weight is used to weight the position update in the encirclement and random search phases. In the bubble net attack phase, the position is updated by selecting the shrinking encirclement or spiral ascent mode with a preset probability. An initial hybrid prediction model with undetermined weights and thresholds is constructed by setting the BP neural network structure and the improved whale optimization algorithm.
6. The method according to claim 1, characterized in that, The improved whale optimization algorithm is used to perform global iterative optimization of the weights and thresholds of the BP neural network in the initial hybrid prediction model, using the mean squared error on the training set as the fitness evaluation criterion, to obtain the optimized hybrid prediction model with the initial weights and thresholds, including: All the weights and thresholds to be optimized in the BP neural network of the initial hybrid prediction model are encoded as position vectors of each individual whale. The position vector of each individual whale is decoded into a set of weights and thresholds for a BP neural network. A forward propagation is performed on the training set, and the mean square error between the predicted value and the target output of the network is calculated. The mean square error is used as the fitness value of each individual whale. In the current iteration, calculate the current value of the convergence factor and the coefficient vector; For each individual whale, the surrounding predation update strategy, bubble net attack update strategy, or random search update strategy are selected based on the relationship between the modulus of the coefficient vector and the preset probability. In the surrounding predation and random search strategies, adaptive weights are used to perform position-weighted updates. Boundary detection is performed on the updated position vectors of each individual whale, and components that exceed the search range are truncated to the boundary values of the search range; Recalculate the fitness value of each individual whale, update and record the best individual in the current population and its position vector; Repeatedly execute the iterative process of calculating the convergence factor and the coefficient vector, selecting an update strategy for position update, boundary detection, recalculating fitness and updating the optimal individual, until the maximum number of iterations is reached. The position vector corresponding to the final optimal individual is decoded into a set of weights and thresholds of a BP neural network, which are then used as the optimized initial weights and thresholds and assigned to the hybrid prediction model.
7. The method according to claim 1, characterized in that, After assigning the optimized initial weights and thresholds, the hybrid prediction model is trained using gradient descent on the training set to obtain the trained IWOA-BP prediction model, including: The optimized initial weights and thresholds are used as the initial values for the connection weights and thresholds of each layer in the BP neural network of the hybrid prediction model; On the training set, the mean squared error is used as the loss function, and the gradient descent method is used to iteratively update the connection weights and thresholds of each layer of the BP neural network in the hybrid prediction model. The gradients of the output layer weights and thresholds are calculated by multiplying the partial derivative of the loss with respect to the net input of each output layer by the corresponding hidden layer output. The gradients of the hidden layer weights and thresholds are calculated by multiplying the partial derivative of the loss with respect to the net input of each hidden layer by the corresponding input features. The partial derivative of the net input of the hidden layer is determined by multiplying the derivative of the hidden layer activation function by the backpropagation error from the output layer. The connection weights and thresholds are updated by multiplying each gradient by a preset learning rate; Repeat the following steps on the training set: calculate the loss using mean squared error as the loss function, calculate the gradient of each layer's connection weights and thresholds using gradient descent, and update the connection weights and thresholds according to a preset learning rate, until the loss function converges or the preset maximum number of fine-tuning iterations is reached, to obtain the trained IWOA-BP prediction model.
8. The method according to claim 1, characterized in that, The training IWOA-BP prediction model is validated using the test set, and the mean absolute percentage error, root mean square error, and coefficient of determination are used as model performance evaluation indicators to obtain the final prediction results of IGCT thermal resistance performance parameters, including: Each input sample in the test set is input into the trained IWOA-BP prediction model, and the predicted value of the thermal resistance parameter of each test sample is obtained through forward propagation calculation. The predicted thermal resistance parameters are inversely normalized to their original dimensions to obtain the final predicted thermal resistance parameters for each test sample. For the steady-state thermal resistance output dimension, the absolute value of the difference between the predicted value and the true value of each test sample is calculated, divided by the true value, summed, averaged, and then multiplied by the percentage to obtain the average absolute percentage error of the steady-state thermal resistance. For the transient thermal impedance output dimension, the absolute value of the difference between the predicted value and the true value of each test sample is calculated, divided by the true value, summed, averaged, and then multiplied by the percentage to obtain the average absolute percentage error of the transient thermal impedance. The average absolute percentage error of the steady-state thermal resistance and the average absolute percentage error of the transient thermal resistance are taken as the overall average absolute percentage error evaluation index. For the steady-state thermal resistance output dimension, the root mean square error of the steady-state thermal resistance is obtained by taking the square root of the sum of the squares of the differences between the predicted and true values of each test sample and the total number of test samples. For the transient thermal impedance output dimension, the root mean square error of the transient thermal impedance is obtained by taking the square root of the sum of the squares of the differences between the predicted and true values of each test sample and the total number of test samples. The average value of the root mean square error of the steady-state thermal resistance and the root mean square error of the transient thermal resistance is taken as the overall root mean square error evaluation index. For the steady-state thermal resistance output dimension, calculate the ratio of the residual sum of squares to the total sum of squares to obtain the determination coefficient of the steady-state thermal resistance. The residual sum of squares is the sum of squares of the differences between the predicted value and the true value of each test sample, and the total sum of squares is the sum of squares of the differences between the true value and the mean of the true value of each test sample. For the transient thermal impedance output dimension, calculate the ratio of the residual sum of squares to the total sum of squares to obtain the determination coefficient of the transient thermal impedance; The average value of the determination coefficient of the steady-state thermal resistance and the determination coefficient of the transient thermal resistance is taken as the overall determination coefficient evaluation index. Output the final predicted results of steady-state thermal resistance and transient thermal impedance for each test sample, and output the mean absolute percentage error, the root mean square error and the coefficient of determination accordingly.
9. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method according to any one of claims 1 to 8.
10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 8.
