IGBT (Insulated Gate Bipolar Translator) temperature field calculation method, system, medium and equipment

Abstract

The invention discloses an IGBT temperature field calculation method, system, medium and equipment based on data driving, and the method comprises the steps: building a finite element thermal model based on the geometric structure and material thermal parameters of an IGBT module; simulating steady-state temperature field distribution under each working condition by using a finite element thermal model, and constructing a multi-working-condition training data set comprising input variables and corresponding temperature field output; constructing a main network-branch network neural network model, and training the main network-branch network neural network model by using the multi-working-condition training data set; inputting the power loss value and heat dissipation parameters measured in real time into an IGBT temperature field prediction model to obtain an initial temperature field prediction result; updating the power loss value according to the initial temperature field prediction result, and inputting the updated power loss into the IGBT temperature field prediction model again for a new round of temperature field prediction; and repeatedly executing until the power loss difference value between two adjacent iterations is smaller than a preset convergence threshold value, stopping iteration and outputting a final steady-state temperature field.

Description

Technical Field

[0001] This invention relates to the field of IGBT module testing technology, and in particular to a data-driven method, system, medium, and device for calculating the temperature field of IGBTs. Background Technology

[0002] IGBTs have been widely used in electric vehicles, rail transportation, renewable energy systems, and high-voltage direct current transmission. However, the reliability of IGBTs is often limited by thermal failure, which accounts for more than half of device failures. Efficiently and accurately measuring the critical temperature distribution using a limited number of sensing points is an urgent problem to be solved, as it is crucial for health management and lifetime prediction.

[0003] The information disclosed in the background section is only for enhancing the understanding of the background of this invention, and therefore may contain information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0004] This invention provides a data-driven IGBT temperature field calculation method, system, medium, and device that can efficiently and accurately calculate the temperature field distribution under various operating conditions, resolving the contradiction between efficiency and accuracy in finite element-based temperature field calculation methods.

[0005] A data-driven method for calculating the temperature field of IGBTs includes:

[0006] Step 1: Establish a finite element thermal model based on the geometry and material thermal parameters of the IGBT module;

[0007] Step 2: Using the power loss distribution of the IGBT chip and the FWD chip as the first input variable, and the cooling medium flow rate and temperature as the second input variables, multiple operating condition combinations are generated within the preset parameter range through Latin hypercube sampling.

[0008] Step 3: Use the finite element thermal model to simulate the steady-state temperature field distribution under various working conditions, and construct a multi-working-condition training dataset containing input variables and corresponding temperature field outputs.

[0009] Step 4: Construct a main network-branch network neural network model, in which the main network uses a Fourier neural operator network to process the spatial distribution characteristics of power loss, and the branch network uses a fully connected neural network to process the heat dissipation boundary conditions; after fusing the output features of the main network and the branch network, predict the spatial temperature field distribution of the IGBT module.

[0010] Step 5: Train the main network-branch network neural network model using the multi-condition training dataset, and optimize the network weights using the Sobolev H¹ loss function to obtain the trained IGBT temperature field prediction model.

[0011] Step 6: Under actual operating conditions, obtain the load current, switching frequency and bus voltage of the IGBT module, and calculate the initial power loss value of the IGBT and FWD chip in combination with the pre-calibrated power loss model; input the power loss value and the real-time measured heat dissipation parameters into the IGBT temperature field prediction model to obtain the initial temperature field prediction result.

[0012] Step 7: Update the power loss value based on the initial temperature field prediction result, and re-input the updated power loss into the IGBT temperature field prediction model for a new round of temperature field prediction; repeat until the power loss difference between two adjacent iterations is less than the preset convergence threshold, then stop the iteration and output the final steady-state temperature field.

[0013] In the data-driven IGBT temperature field calculation method, the main network maps the positions of the IGBT and FWD chip and their corresponding power losses into a two-dimensional power loss map; the scalar power of each grid point is converted into a high-dimensional feature vector through the dimension-upgrading layer; convolution operation is performed in the frequency domain through multi-layer Fourier transform and inverse transform operations, and spatial-frequency features are extracted by combining residual connections and nonlinear activation functions; finally, a fixed-dimensional feature matrix is ​​output through the projection layer.

[0014] In the data-driven IGBT temperature field calculation method, the branch network consists of a three-layer fully connected neural network, which encodes the cooling medium flow rate and temperature into a set of high-dimensional feature vectors as modulation weights for subsequent feature fusion.

[0015] In the data-driven IGBT temperature field calculation method, in step 4, the feature fusion involves multiplying the feature matrix output by the main network and the feature vector output by the branch network element by element, and then processing them through a convolutional layer and a nonlinear activation function to finally output a single-channel temperature field image.

[0016] In the data-driven IGBT temperature field calculation method, in step 1, the surface temperature distribution of the IGBT module is measured by an infrared thermal imager and compared with the finite element simulation results. If the maximum temperature difference does not exceed 1℃, the finite element model is deemed valid.

[0017] In the data-driven IGBT temperature field calculation method, the power loss model is an analytical or lookup table model obtained by fitting the parameters of the IGBT device manual with measured loss data under various operating conditions. It calculates the conduction loss and switching loss based on the load current, bus voltage and switching frequency.

[0018] In the data-driven IGBT temperature field calculation method, the IGBT temperature field prediction model has a single inference time of less than 0.2 seconds and a maximum temperature field prediction error of less than 2℃.

[0019] A system for implementing the method includes:

[0020] The data acquisition unit is used to obtain the operating parameters and heat dissipation conditions of the IGBT module.

[0021] The power loss calculation unit is used to calculate chip loss based on the power loss model invoked by the operating parameters.

[0022] The neural network inference unit integrates a pre-trained main network-branch network neural network model to receive power loss and heat dissipation parameters and output the temperature field.

[0023] The iterative control unit is used to determine whether the electrothermal iteration has converged and output the final temperature field result;

[0024] The display and storage unit is used to visualize the temperature field distribution and save historical data.

[0025] A computer storage medium including computer instructions that, when run on a computer, cause the computer to perform the method.

[0026] An electronic device, the electronic device comprising:

[0027] Memory, processor, and computer programs stored in memory and executable on the processor, wherein,

[0028] The processor implements the method when executing the program.

[0029] Compared with existing technologies, the present invention has the following advantages: The present invention constructs a heat conduction equation in the frequency domain and a periodic loss input, adopts a two-step iterative strategy of first steady-state average temperature and then frequency domain harmonic disturbance, and combines frequency-time domain transformation technology to achieve efficient and high-precision simulation calculation of the periodic steady-state temperature fluctuation distribution of IGBT modules, overcoming the problems of long time consumption and difficulty in engineering application of traditional transient thermal simulation. Attached Figure Description

[0030] Various other advantages and benefits of the present invention will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiments below. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. Furthermore, the same reference numerals denote the same parts throughout the drawings.

[0031] In the attached diagram:

[0032] Figure 1 This is a schematic diagram of the IGBT of the present invention;

[0033] Figure 2 This is a schematic diagram of the IGBT explosion of the present invention;

[0034] Figure 3 This is a schematic diagram of the meshing results of the IGBT finite element model of the present invention;

[0035] Figure 4 This is a schematic diagram comparing the simulated and measured surface temperature distribution of the IGBT module according to the present invention. Figure 4 (a) Temperature distribution of the IGBT in the simulation. Figure 4 (b) IGBT temperature distribution measured in mid-infrared spectroscopy;

[0036] Figure 5 This is a schematic diagram of the neural network framework of the present invention;

[0037] Figure 6 This is a schematic diagram of the electrothermal iteration of the present invention;

[0038] Figure 7 This is a schematic diagram comparing the temperature fields of the IGBTs of this invention. Figure 7 In the middle (a), the temperature field of the neural network is shown. Figure 7 In diagram (b), the temperature field is represented by a finite element method. Figure 7 (c) represents the temperature field captured by infrared imaging.

[0039] The present invention will be further explained below with reference to the accompanying drawings and embodiments. Detailed Implementation

[0040] Specific embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While specific embodiments of the invention are shown in the drawings, it should be understood that the invention may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0041] It should be noted that certain terms are used in the specification and claims to refer to specific components. Those skilled in the art will understand that different terms may be used to refer to the same component. This specification and claims do not distinguish components based on differences in terminology, but rather on differences in function. The terms "comprising" or "including" used throughout the specification and claims are open-ended and should be interpreted as "comprising but not limited to." The following descriptions are preferred embodiments for carrying out the invention; however, these descriptions are for the purpose of understanding the general principles of the specification and are not intended to limit the scope of the invention. The scope of protection of this invention is determined by the appended claims.

[0042] To facilitate understanding of the embodiments of the present invention, further explanations and descriptions will be provided below with reference to the accompanying drawings and specific embodiments. The accompanying drawings do not constitute a limitation on the embodiments of the present invention.

[0043] like Figures 1 to 7 As shown, the data-driven IGBT temperature field calculation method includes the following steps:

[0044] Step 1: Establish a finite element thermal model based on the geometric structure and material thermal parameters of the IGBT module; Geometric structure: Obtain the thickness, length, and width dimensions of each packaging layer, such as the chip, copper layer, ceramic layer, solder layer, and substrate, by dissecting and measuring the IGBT module.

[0045] Material thermal parameters: By consulting relevant literature and data handbooks, we can obtain the thermal conductivity, specific heat capacity, and other material thermal parameters of silicon-based chips, copper, ceramics, and other materials.

[0046] A specific finite element model refers to a model that has undergone geometric construction, material parameter assignment, mesh generation, and physics field construction. Mesh generation needs to be performed based on the specific geometric structure, and specific operations include sweeping, free tetrahedrons, etc. In this model, the physics field construction mainly uses solid heat transfer, solving the heat conduction equation. The meshed model is shown below. Figure 3 The solution is as follows Figure 4 (a)

[0047] Step 2: Using the power loss distribution of the IGBT chip and the FWD chip as the first input variable, and the cooling medium flow rate and temperature as the second input variables, multiple operating condition combinations are generated within a preset parameter range using Latin hypercube sampling. The preset parameter range is set according to the actual operating condition range of the power module, where the IGBT loss is set to [0W, 400W], the FWD loss is set to [0W, 200W], the cooling medium flow rate is [0L / min, 10L / min], and the cooling medium temperature is [20℃, 40℃].

[0048] Step 3: Use the finite element thermal model to simulate the steady-state temperature field distribution under various working conditions, and construct a multi-working-condition training dataset containing input variables and corresponding temperature field outputs.

[0049] Step 4: Construct a main network-branch network neural network model. The main network uses a Fourier neural operator network to process the spatial distribution characteristics of power loss, while the branch network uses a fully connected neural network to process the heat dissipation boundary conditions. After fusing the output features of the main network and the branch network, the spatial temperature field distribution of the IGBT module is predicted. The heat dissipation boundary conditions refer to the flow rate and temperature of the cooling medium, collectively referred to as heat dissipation boundary conditions.

[0050] Step 5: Train the main network-branch network neural network model using the multi-condition training dataset. Optimize the network weights using the Sobolev H¹ loss function to obtain the trained IGBT temperature field prediction model. Training steps: First, read sample data containing condition parameters, power loss field, and temperature field. Standardize the condition vector using Z-score and divide the training and validation sets in an 8:2 ratio. The network structure consists of two parts: the branch network uses three fully connected layers (2→64→128→128) with Tanh activation function to extract condition features; the backbone network is a two-dimensional Fourier neural operator (FNO), inputting a single-channel power loss distribution (1×128×256) and outputting a 128-channel feature field. The two are multiplied and fused element-wise along the channel dimension, and further feature transformation is performed through residual connections and two 1×1 convolutions (containing Tanh activation). Finally, a single-channel temperature field is output via a 1×1 convolution. Training uses the Adam optimizer (learning rate 1×10⁻⁶). -3 The model employs StepLR, halving the learning rate every 20 epochs for a total of 200 epochs. The loss function is H¹ loss, which comprehensively considers the L² error between the predicted and actual temperature fields and the difference in their spatial gradients. The model is evaluated on a validation set every 5 epochs, and the optimal weights are automatically saved. To ensure training convergence, the model combines learning rate decay, gradient smoothing (H¹ regularization), and data normalization strategies. Gradient clipping, weight decay, or early stopping mechanisms can be used as needed to prevent non-convergence and overfitting, thereby ensuring stable network convergence and achieving the expected accuracy.

[0051] Step 6: Under actual operating conditions, obtain the load current, switching frequency, and bus voltage of the IGBT module. Calculate the initial power loss values ​​of the IGBT and FWD chips using a pre-calibrated power loss model. Input the power loss values ​​and real-time measured heat dissipation parameters into the IGBT temperature field prediction model to obtain the initial temperature field prediction results. The heat dissipation parameters include the cooling medium flow rate and cooling medium temperature. The power loss model includes average conduction loss and average switching loss. The power loss value is the sum of the average conduction loss and average switching loss. The average conduction loss P of the IGBT and FWD in the power module within one power frequency cycle is...Tcon_ave and P Dcon_ave They are respectively:

[0052]

[0053]

[0054] In the formula, V CE and V F These represent the on-state voltage drop of the IGBT and the forward voltage drop of the FWD, respectively; i T and i D These represent the current flowing through the IGBT and FWD respectively; f0 is the operating frequency; M T and M D These represent the duty cycles of the current flowing through the IGBT and FWD, respectively. By utilizing the relationship between the IGBT and FWD's current-voltage curves and temperature, the average conduction loss P of the IGBT and FWD over one power frequency cycle can be calculated. Tcon_ave and P Dcon_ave .

[0055] The average switching loss P of the IGBT and FWD in the power module over one power frequency cycle Tsw_ave and P Drec_ave They are respectively:

[0056]

[0057]

[0058] In the formula, E Tsw and E rec E represents the switching energy loss and recovery energy loss of the IGBT and FWD, respectively. Tsw To enable energy loss E on and the energy loss during shutdown E off The sum of; f sw This refers to the IGBT switching frequency. All of the above parameters can be obtained from the datasheet.

[0059] Step 7: Update the power loss value based on the initial temperature field prediction result, and re-input the updated power loss into the IGBT temperature field prediction model for a new round of temperature field prediction; repeat until the power loss difference between two adjacent iterations is less than the preset convergence threshold (the loss difference between two iterations is less than 1e-3), then stop the iteration and output the final steady-state temperature field.

[0060] In a preferred embodiment of the data-driven IGBT temperature field calculation method, the main network maps the positions of the IGBT and FWD chip and their corresponding power losses into a two-dimensional power loss map; the scalar power of each grid point is converted into a high-dimensional feature vector through an upscaling layer; convolution operations are performed in the frequency domain through multi-layer Fourier transform and inverse transform operations, and spatial-frequency features are extracted by combining residual connections and nonlinear activation functions; finally, a fixed-dimensional feature matrix is ​​output through a projection layer.

[0061] In a preferred embodiment of the data-driven IGBT temperature field calculation method, the branch network consists of a three-layer fully connected neural network, which encodes the cooling medium flow rate and temperature into a set of high-dimensional feature vectors as modulation weights for subsequent feature fusion.

[0062] In a preferred embodiment of the data-driven IGBT temperature field calculation method, in step 4, the feature fusion involves multiplying the feature matrix output by the main network and the feature vector output by the branch network element by element, then processing them through a convolutional layer and a nonlinear activation function to finally output a single-channel temperature field image.

[0063] In a preferred embodiment of the data-driven IGBT temperature field calculation method, in step 1, the surface temperature distribution of the IGBT module is measured by an infrared thermal imager and compared with the finite element simulation results. If the maximum temperature difference does not exceed 1℃, the finite element model is deemed valid.

[0064] In a preferred embodiment of the data-driven IGBT temperature field calculation method, the power loss model is an analytical or lookup table model obtained by fitting the parameters in the IGBT device manual with measured loss data under various operating conditions. It calculates conduction loss and switching loss based on load current, bus voltage and switching frequency.

[0065] In a preferred embodiment of the data-driven IGBT temperature field calculation method, the IGBT temperature field prediction model has a single inference time of less than 0.2 seconds and a maximum temperature field prediction error of less than 2℃.

[0066] A system for implementing the method includes:

[0067] The data acquisition unit is used to obtain the operating parameters and heat dissipation conditions of the IGBT module.

[0068] The power loss calculation unit is used to calculate chip loss based on the power loss model invoked by the operating parameters.

[0069] The neural network inference unit integrates a pre-trained main network-branch network neural network model to receive power loss and heat dissipation parameters and output the temperature field.

[0070] The iterative control unit is used to determine whether the electrothermal iteration has converged and output the final temperature field result;

[0071] The display and storage unit is used to visualize the temperature field distribution and save historical data.

[0072] A computer storage medium including computer instructions that, when run on a computer, cause the computer to perform the method.

[0073] An electronic device, the electronic device comprising:

[0074] Memory, processor, and computer programs stored in memory and executable on the processor, wherein,

[0075] The processor implements the method when executing the program.

[0076] In one embodiment, a main-branch neural network is used as the main network, with power loss and heat dissipation parameters as inputs and temperature field distribution as the output. The main network uses a Fourier neural operator network to facilitate the extraction of spatial-frequency domain features from power loss to temperature distribution. The branch networks use fully connected neural networks to encode heat dissipation parameters into a high-dimensional feature space. Finally, the outputs of the main network and branch networks are fused to output the IGBT temperature field distribution. Although the input parameters of the neural network framework include power loss and heat dissipation parameters, in practical IGBT applications, power loss cannot be directly measured. It is usually necessary to calculate the power loss value using a power loss model based on external operating parameters (such as load current, switching frequency, bus voltage, etc.). Subsequently, the calculated power loss and the actually measurable heat dissipation parameters are used as inputs to the neural network model. However, there is a non-linear coupling relationship between power loss and temperature. The power loss is iteratively updated according to the calculated temperature field. When the difference in power loss between two adjacent iterations is lower than a preset threshold, the iteration is considered to have converged. The final converged temperature field is shown in the figure.

[0077] In one embodiment, a commercially available solderable IGBT module (NI450B12E6K4) is used as an example.

[0078] IGBT Finite Element Method Establishment

[0079] Taking the soldered IGBT module (FF150R12ME3G) as an example, Figure 1 and Figure 2 The images show both the physical diagram and the exploded view of the IGBT module structure. From top to bottom, they are: the chip layer consisting of the IGBT chip and diode chip, the chip solder layer, the upper copper layer of DBC (Direct Bonded Copper, DBC), the substrate alumina ceramic layer, the lower copper layer of DBC, the substrate solder layer, the copper heat dissipation substrate, and the thermal grease layer.

[0080] The materials and their specific applications are as follows:

[0081] Copper (Cu): Used for the top and bottom copper layers of the DBC and as substrate material, exhibiting good thermal conductivity; Silicon (Si): Used for the IGBT and FWD chip layers, possessing high thermal conductivity; Aluminum oxide (Al2O3): Used as the DBC insulating layer, providing electrical insulation and a heat conduction path; Tin-lead alloy (Sn-Pb): Used for the chip solder layer and substrate solder layer, possessing moderate thermal conductivity; Thermal grease (TIM): Primarily used to fill the gap between the chip and the heat sink, improving heat conduction. Based on the IGBT's geometric dimensions and material parameters, a finite element model of the IGBT was established, and the IGBT was meshed. The meshing result is shown in the figure below. Figure 3 As shown. To verify the accuracy of the finite element model, the IGBT was set to a normally on state and a DC current was applied. An infrared camera was used to measure the surface temperature distribution of the IGBT, and the results of the finite element model were compared with the actual measurement results. The comparison results are shown below. Figure 4 As shown, the results indicate that the maximum error in surface temperature is only 0.66℃, verifying the correctness of the finite element model.

[0082] Neural networks are used to achieve efficient and accurate calculation of the temperature field of IGBTs. Power loss and heat dissipation parameters are used as inputs, and the temperature field distribution is used as the output.

[0083] (1) Mainnet

[0084] The main network is used to extract the spatial-frequency features of power loss distribution. Its inputs are chip layout information and corresponding power loss values, and it mainly includes the following steps:

[0085] 1) Power loss map construction: Generate a two-dimensional power loss distribution map based on the chip location and input power;

[0086] 2) Dimensionality Upgrading Layer: A multilayer perceptron (MLP) is used to map the scalar input into a high-dimensional feature representation to enhance the feature representation capability;

[0087] 3) Fourier layer: Performs multi-layer Fourier transform and inverse transform operations on the feature map, completes convolution operations in the frequency domain, and combines residual connections and nonlinear activation functions to achieve efficient extraction of spatial-frequency features;

[0088] 4) Projection layer: The output of the Fourier layer is mapped to a fixed-dimensional feature matrix through one-dimensional convolution, which is used for subsequent feature fusion.

[0089] (2) Branch network

[0090] The branch network is used to encode the cooling conditions. Its inputs are water flow rate and water temperature, which are processed by a three-layer multilayer perceptron (MLP), and the output is a set of feature vectors as modulation weights.

[0091] (3) Feature fusion module

[0092] The outputs of the main network and the branch networks are first fused by element-wise multiplication, and then stability is enhanced by a residual connection structure. Subsequently, the fused features are processed by multi-layer convolution and nonlinear activation functions, and finally, a single-channel spatial temperature field prediction result is output.

[0093] Construction of multi-condition training dataset

[0094] To ensure that the neural network can accurately learn the multi-condition thermal field distribution characteristics of the power module, a training dataset covering a wide range of parameters was constructed. The specific method is as follows:

[0095] (1) Parameter range setting

[0096] Considering the power loss characteristics of IGBT and FWD chips, power loss ranges are set for IGBT and FWD respectively. In the heat dissipation boundary conditions, water flow velocity range and water temperature range are set.

[0097] (2) Operating condition sampling method

[0098] Because the thermal field distribution of the power module has a complex nonlinear mapping relationship with the aforementioned input parameters, insufficient sampling will lead to poor model generalization. Therefore, this invention employs the Latin hypercube sampling (LHS) method to uniformly select m operating points in a defined parameter space, ensuring uniform coverage of training data within the input space. The value of m is set according to the parameter range and the degree of nonlinearity.

[0099] (3) Generation of thermal field data

[0100] After obtaining the sampling conditions, the corresponding temperature field is calculated using the finite element method, and the results are used as training labels to form a mapping dataset between input parameters (power loss, heat dissipation boundary conditions) and output temperature field.

[0101] Neural network training

[0102] (1) Loss function design

[0103] To improve the accuracy of model predictions and its ability to characterize spatial distributions, this invention introduces the Sobolev H1 loss function during training. This function considers gradient error in addition to point error, which can effectively enhance the model's sensitivity to changes in spatial gradient.

[0104] (2) During training, the number of modes of the Fourier layer is set. The optimizer uses the Adam algorithm, sets an initial learning rate, and halves the learning rate after multiple iterations through a step-by-step learning rate scheduling strategy to improve the robustness of training.

[0105] Electrothermal Iteration Method

[0106] (1) External parameter measurement and power loss calculation

[0107] In actual operation, only parameters such as load current, switching frequency, and DC bus voltage can be directly measured. Therefore, this invention first calculates the power loss values ​​of the IGBT and FWD chips based on the aforementioned measurable operating parameters using a pre-established power loss model. This power loss model is constructed by combining experimental test data and datasheet parameters from multiple devices, and can accurately estimate device power loss under different operating conditions.

[0108] (2) Neural network computation and electrothermal coupling iteration

[0109] Since power loss is inherently temperature-dependent, there is a coupling relationship between the two. To address this issue, this invention employs an electrothermal iterative calculation method: the calculated power loss is input into a pre-trained neural network model to obtain an initial temperature field calculation result; the power loss is updated based on the calculated temperature field, and then input into the neural network again for temperature field calculation; through multiple iterations, the deviation between power loss and temperature field is gradually corrected. When the difference in power loss calculated in two adjacent iterations is less than a preset threshold, the electrothermal iterative process is considered to have converged, and the final steady-state temperature field result is output. Figure 6 As shown. In this example, under the three-phase inverter operating condition, its operating parameters and heat dissipation parameters are as follows: current 255A, switching frequency 2kHz, bus voltage, water flow rate 3.11L / min, and water temperature 25℃. The IGBT temperature field is obtained by solving. To verify the accuracy of the calculation results, the results of the neural network calculation are compared with the results of the finite element calculation and the results of the infrared camera, as shown. Figure 7 As shown in the figure. The results show that the single calculation time is 0.1s, and the maximum error in temperature field calculation is less than 2℃. Considering that the infrared camera itself has an error of about 1℃, the feasibility of this method is considered to be high.

[0110] Furthermore, this invention employs a dual-channel neural network architecture consisting of a main network and branch networks, achieving decoupled modeling and feature fusion of spatial heat source distribution and boundary heat dissipation conditions. The main network introduces a Fourier neural operator (FNO), which can directly capture the nonlocal mapping relationship between power loss and temperature field in the frequency domain, effectively extracting the spatial periodicity and multi-scale features of the chip-layer heat source distribution, overcoming the limitations of traditional convolutional neural networks (CNNs) in terms of limited receptive field and difficulty in modeling long-range dependencies. Simultaneously, the branch networks use fully connected networks to encode scalar boundary parameters such as water flow velocity and water temperature in high dimensions, transforming them into modulated feature vectors, enabling the model to flexibly adapt to different cooling conditions. The two networks are multiplied element-wise and connected via a feature fusion module, which not only enhances the model's responsiveness to changes in boundary conditions but also improves the overall prediction stability and generalization performance.

[0111] Secondly, by constructing a multi-condition training dataset based on finite element simulation and Latin hypercube sampling, high coverage and representativeness of the input space were ensured, avoiding prediction distortion caused by blind spots in the operating conditions. Combined with the design of the Sobolev H¹ loss function, the optimization process not only minimizes the error of the temperature value itself but also simultaneously constrains the deviation of the temperature gradient. This allows the model to accurately reproduce the temperature gradient changes in hotspot areas while maintaining the accuracy of the overall temperature level, significantly improving the predictive ability for localized thermal stress concentration areas. This is of great significance for thermal fatigue life assessment.

[0112] Finally, an electrothermal iterative mechanism is introduced to solve the core problem of power loss being unmeasurable and strongly coupled with temperature in practical applications. This mechanism embeds a data-driven model into a physical closed loop: the initial power loss is estimated from measurable operating parameters, and after the temperature field is quickly predicted by a neural network, it is fed back to the loss model for correction, forming an iterative closed loop of "electricity → heat → electricity". This method achieves dynamic approximation of the internal thermal behavior of IGBTs under real operating conditions without the need for additional sensors. It retains the efficiency of the data-driven model (single inference time <0.1s) while incorporating the constraints of physical laws, effectively suppressing model extrapolation errors and improving prediction robustness under long-term operation.

[0113] Although embodiments of the present invention have been described above in conjunction with the accompanying drawings, the present invention is not limited to the specific embodiments and application fields described above. The specific embodiments described above are merely illustrative and instructive, and not restrictive. Those skilled in the art can make many other forms based on the guidance of this specification and without departing from the scope of protection of the claims of the present invention, and all of these are within the scope of protection of the present invention.

Claims

1. A data-driven method for calculating the temperature field of an IGBT, characterized in that, Includes the following steps: Step 1: Establish a finite element thermal model based on the geometry and material thermal parameters of the IGBT module; Step 2: Using the power loss distribution of the IGBT chip and the FWD chip as the first input variable, and the cooling medium flow rate and temperature as the second input variables, multiple operating condition combinations are generated within the preset parameter range through Latin hypercube sampling. Step 3: Use the finite element thermal model to simulate the steady-state temperature field distribution under various working conditions, and construct a multi-working-condition training dataset containing input variables and corresponding temperature field outputs. Step 4: Construct a main network-branch network neural network model, in which the main network uses a Fourier neural operator network to process the spatial distribution characteristics of power loss, and the branch network uses a fully connected neural network to process the heat dissipation boundary conditions; after fusing the output features of the main network and the branch network, predict the spatial temperature field distribution of the IGBT module. Step 5: Train the main network-branch network neural network model using the multi-condition training dataset, and optimize the network weights using the Sobolev H¹ loss function to obtain the trained IGBT temperature field prediction model. Step 6: Under actual operating conditions, obtain the load current, switching frequency and bus voltage of the IGBT module, and calculate the initial power loss value of the IGBT and FWD chip in combination with the pre-calibrated power loss model; input the power loss value and the real-time measured heat dissipation parameters into the IGBT temperature field prediction model to obtain the initial temperature field prediction result. Step 7: Update the power loss value based on the initial temperature field prediction result, and re-input the updated power loss into the IGBT temperature field prediction model for a new round of temperature field prediction. Repeat the process until the power loss difference between two adjacent iterations is less than the preset convergence threshold, then stop the iteration and output the final steady-state temperature field.

2. The data-driven IGBT temperature field calculation method according to claim 1, characterized in that, Preferably, the main network maps the positions of the IGBT and FWD chips and their corresponding power losses into a two-dimensional power loss map; the scalar power of each grid point is converted into a high-dimensional feature vector through the dimension-upgrading layer; convolution operations are performed in the frequency domain through multi-layer Fourier transform and inverse transform operations, and spatial-frequency features are extracted by combining residual connections and nonlinear activation functions; finally, a fixed-dimensional feature matrix is ​​output through the projection layer.

3. The data-driven IGBT temperature field calculation method according to claim 1, characterized in that, The subnet consists of a three-layer fully connected neural network, which encodes the cooling medium flow rate and temperature into a set of high-dimensional feature vectors, which serve as modulation weights for subsequent feature fusion.

4. The data-driven IGBT temperature field calculation method according to claim 1, characterized in that, In step 4, the feature fusion involves multiplying the feature matrix output by the main network and the feature vector output by the branch network element by element, then processing them through a convolutional layer and a nonlinear activation function to finally output a single-channel temperature field image.

5. The data-driven IGBT temperature field calculation method according to claim 1, characterized in that, In step 1, the surface temperature distribution of the IGBT module is measured by an infrared thermal imager and compared with the finite element simulation results. If the maximum temperature difference does not exceed 1℃, the finite element model is deemed valid.

6. The data-driven IGBT temperature field calculation method according to claim 1, characterized in that, The power loss model is an analytical or lookup table model obtained by fitting the parameters of the IGBT device manual with measured loss data under various operating conditions. It calculates conduction loss and switching loss based on load current, bus voltage and switching frequency.

7. The data-driven IGBT temperature field calculation method according to claim 1, characterized in that, The IGBT temperature field prediction model has a single inference time of less than 0.2 seconds and a maximum temperature field prediction error of less than 2℃.

8. A system for implementing the method according to any one of claims 1-7, characterized in that, It includes: The data acquisition unit is used to obtain the operating parameters and heat dissipation conditions of the IGBT module. The power loss calculation unit is used to calculate chip loss based on the power loss model invoked by the operating parameters. The neural network inference unit integrates a pre-trained main network-branch network neural network model to receive power loss and heat dissipation parameters and output the temperature field. The iterative control unit is used to determine whether the electrothermal iteration has converged and output the final temperature field result; The display and storage unit is used to visualize the temperature field distribution and save historical data.

9. A computer storage medium, characterized in that, The storage medium includes computer instructions that, when executed on a computer, cause the computer to perform the method as described in any one of claims 1-7.

10. An electronic device, characterized in that, The electronic device includes: Memory, processor, and computer programs stored in memory and executable on the processor, wherein, When the processor executes the program, it implements the method as described in any one of claims 1-7.

Images