Artificial intelligence-based method and system for estimating temperature distribution of power semiconductor module

Abstract

Provided are an artificial intelligence-based method and system for estimating a temperature distribution of a power semiconductor module. First, training data is configured on the basis of a temperature curve graph of a temperature sensor, and a power loss and a heat dissipation condition are decoupled by means of temperature curves of the temperature sensor, so as to realize full coverage of actual operating conditions; and second, an artificial intelligence algorithm-based model is constructed and debugged, so as to realize the accurate reconstruction of a thermal field of a module, and the temperature estimation of the power semiconductor module is realized by means of a mapping relationship between the temperature of the temperature sensor and a thermal field distribution of the module. The present application realizes the decoupling of operating condition parameters, can obtain an accurate temperature field distribution of the entire module under any actual operating conditions while ensuring complete non-invasiveness, has a fast computing speed and strong universality, and is suitable for online temperature monitoring of the power semiconductor module.

Description

A method and system for estimating temperature distribution in power semiconductor modules based on artificial intelligence Technical Field

[0001] This invention belongs to the field of power semiconductor module condition monitoring, specifically relating to a method and system for estimating the temperature distribution of power semiconductor modules based on artificial intelligence. Background Technology

[0002] Power semiconductor modules are widely used in new energy conversion, smart grids, electric vehicles, industrial drives, aerospace, and other fields. The complex environmental stresses and operating conditions in different application scenarios can lead to performance degradation or failure of these modules. Timely and accurate condition monitoring of power semiconductor modules is crucial for understanding their health status in real time and taking early action to minimize potential losses, thus improving the performance and reliability of the entire system. Temperature is a key indicator for condition monitoring; the semiconductor junction temperature directly affects the chip's lifespan and reliability, and the temperature of the packaging materials also significantly impacts the aging of materials such as solder layers. Therefore, accurate temperature monitoring of power semiconductor modules is necessary. However, due to the complex structure of power semiconductor modules and their close coupling with operating and environmental conditions, obtaining high-resolution, high-precision junction temperature results is challenging.

[0003] For online monitoring of junction temperature in power semiconductor modules, the industry generally employs the thermistor method or the thermal model method. The thermistor method uses the semiconductor chip itself as a sensor, obtaining the chip junction temperature by measuring external macroscopic electrical signals related to temperature. This method has a fast dynamic response and is suitable for online monitoring. However, the measurement of thermistor parameters requires high-precision measuring equipment, inevitably increasing cost and complexity, which in turn affects the reliability of the entire system. Furthermore, the varying degrees of nonlinearity require corresponding complex calibration procedures, posing certain difficulties in practical applications.

[0004] The thermal model method combines the module packaging structure and boundary conditions to construct a thermal network model, extracts thermal network parameters using simulation results, and calculates the junction temperature using real-time loss information. In comparison, using a thermal model for junction temperature estimation is simpler and non-invasive, requiring no additional components and enabling junction temperature monitoring without affecting the overall system reliability. However, most existing power semiconductor modules use multiple chips connected in parallel to improve their current carrying capacity. The tight thermal cross-coupling between chips increases the modeling difficulty and limits the improvement of thermal model resolution. Furthermore, the difficulty in obtaining accurate boundary condition information under actual operating conditions also greatly limits the accuracy of the thermal model. Boundary conditions include power loss and heat dissipation conditions. Power loss directly serves as the input to the thermal model, while heat dissipation conditions indirectly affect model parameters, both impacting the accuracy of the thermal model's output results.

[0005] Many power semiconductor modules are equipped with temperature sensors, typically negative thermal coefficient thermistors (NTCs). The NTC and the chip reside on the same ceramic substrate, and online measurement of the NTC resistance provides a convenient way to determine the temperature change at that point. Using temperature sensors as observation points simplifies the module's thermal model and reduces reliance on accurate heat dissipation conditions. However, existing temperature sensor-based monitoring methods are still limited by power loss acquisition and can only reflect the temperature at a single point. There is a lack of power semiconductor module temperature monitoring methods that can obtain the temperature of all points within the module while decoupling operating conditions, thus failing to meet the requirements for online monitoring of temperature distribution in power semiconductor modules. Summary of the Invention

[0006] The purpose of this invention is to provide an artificial intelligence-based method and system for estimating the temperature distribution of power semiconductor modules, in order to overcome the shortcomings of existing technologies. This invention can obtain the accurate temperature field distribution of the entire module under any actual operating conditions while ensuring that it is completely non-invasive. It also has fast calculation speed, strong versatility, and is suitable for online temperature monitoring of power semiconductor modules.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] The method for estimating the temperature distribution of power semiconductor modules based on artificial intelligence includes the following steps:

[0009] S1: Determine the layout geometry parameters and packaging material parameters of the module to be estimated, specifically including: the size and position of the chip, the position of the temperature sensor, the size and thickness of each packaging layer, and the thermophysical properties of each layer material;

[0010] S2: Based on the parameters obtained in S1, establish a thermal simulation model of the module to be estimated in the simulation software, and select the mesh size according to the minimum unit side length of each part of the module to construct a three-dimensional spatial mesh.

[0011] S3: In the thermal simulation model established in S2, a series of typical operating conditions are selected according to the actual operating range of the module to be estimated. The time step is determined according to the actual hardware computing power, and thermal simulation is performed to obtain thermal field data of the module to be estimated under multiple typical operating conditions.

[0012] S4: In the results obtained in S3, the temperature data and time derivative of the temperature sensor location under different working conditions, the temperature field training data of the module to be estimated, and the spatiotemporal coordinates of the matching point are exported respectively. Data preprocessing is performed according to input and output normalization, and the exported data is divided into training set and validation set.

[0013] S5: Initially determine the network structure and artificial intelligence algorithm, combine the encapsulated material parameters in S1 to write the training objective function of the estimation model, and establish an estimation model based on the artificial intelligence algorithm;

[0014] S6: Train the estimation model. Adjust its hyperparameters according to the performance of the estimation model on the training set and validation set to ensure the output accuracy while making the estimation model converge quickly and stably. This will result in a trained estimation model that includes the temperature curve of the temperature sensor and the thermal field mapping relationship within the operating range of the module to be estimated.

[0015] S7: Based on the estimation model trained in S6, input the temperature curve of the temperature sensor under any actual working condition to complete the estimation of the temperature distribution of the module to be estimated under the corresponding conditions.

[0016] Further, in step S1, the module to be estimated includes one or more temperature sensors, including but not limited to negative thermal coefficient thermistors and thermocouples, located on the module's encapsulation surface or inside, respectively. The temperature measured by the temperature sensors is used as the observation point, and the temperature is denoted as {T}. obs1 (t),T obs2 (t),...,T obsM ...

[0017] Furthermore, in step S2, in the thermal simulation software, firstly, a geometric model of the module to be estimated is established and corresponding material thermal parameters are assigned; secondly, the grid size is selected according to the minimum unit side length of each part of the module to construct a three-dimensional grid, and the grid density needs to be determined in conjunction with the actual hardware computing power to complete the establishment of the module thermal simulation model.

[0018] Furthermore, in step S3, the actual operating range of the module to be estimated is first determined, and then a series of typical operating conditions are selected. Based on the temperature curve of the temperature sensor, the temperature curve cluster of the entire temperature sensor covers the largest possible range while ensuring reasonable spacing between the curves, that is, to achieve full coverage of the actual possible operating range. Then, a reasonable time step is selected according to the actual hardware computing power to perform transient thermal simulation and obtain thermal field data of the module to be estimated under multiple typical operating conditions.

[0019] Furthermore, in step S4, three types of thermal simulation data are exported at the time point corresponding to the time step selected in S3: temperature data of the temperature sensor location and the numerical calculation result of its time derivative, temperature field of the module to be estimated, and spatiotemporal coordinates of the mating point. Among them, mating points are selected only in the non-heat source part, that is, in the area outside the chip layer.

[0020] The exported thermal simulation data contains three types of data: spatial coordinates, time coordinates, and temperature data. The spatial coordinates are normalized to 0 using the same scaling scale. Then, the time and temperature are scaled to approximately [0,1]. The specific values ​​of the scaling scale are determined in conjunction with the geometric dimensions of the module to be estimated and the rated operating conditions.

[0021] The normalized data is divided into training and validation sets according to a preset ratio.

[0022] Furthermore, in step S5, the human intelligence algorithm includes, but is not limited to, various supervised learning methods;

[0023] Network architectures include fully connected neural networks, residual neural networks, convolutional neural networks, recurrent neural networks, and other novel networks derived from them;

[0024] Training objective functions include pure data loss functions, physical information loss functions, or hybrid loss functions;

[0025] The estimation model includes an input layer, several intermediate layers, and a 1-channel output layer;

[0026] The input channels of the input layer are in, Corresponding to normalized spatial coordinates, Corresponding time coordinates The temperatures measured by the corresponding normalized 1st, 2nd...Mth temperature sensors;

[0027] The aforementioned intermediate layers can adopt various different network structures;

[0028] The output layer corresponds to the module temperature field.

[0029] Furthermore, the training objective function includes the training data error and the physical law residual; loss = w D MSE D +w F MSE F

[0030] Among them, the training data error MSE D The mean square error (MSE) between the model output temperature and the thermal simulation result temperature is the physical law residual. F To train the residuals of the partial differential equation for heat conduction at the matching points, w D w F These are the weights of the two items, respectively.

[0031] The specific calculation method for the training data error term is as follows:

[0032] Where, ND The number of input training data, Let i be the output temperature of the artificial intelligence network at the i-th point. The actual temperature at the i-th point is derived from S4;

[0033] The physical law residual term is the mean square residual calculated at the combination points in the module for the partial differential equation of heat conduction. Its original form is:

[0034] Where, N F To represent the number of syntagmatic points, the spatiotemporal coordinates of the i-th syntagmatic point are (x... i ,y i ,z i ,t i ), temperature T i k, ρ, and c are the thermal conductivity, density, and constant-pressure heat capacity of the material in the region where the pairing point is located, respectively, and are determined in S1. The pairing point is the selected training point for applying physical constraints. Its main difference from ordinary training data is that the pairing point only requires spatiotemporal coordinate input and does not require actual temperature values, which are all derived from S4.

[0035] Considering the data normalization preprocessing in S4, the calculation formula is expressed using normalized variables. Substituting these variables into the formula yields the following formula for calculating the physical law residual term:

[0036] in, These are all the derivatives of the output of the artificial intelligence network with respect to the input, calculated automatically during training. The temperature is the derivative of the temperature sensor with respect to time, which is calculated in S4 using the numerical differentiation method and then directly exported.

[0037] Further, in step S6, the training set obtained in S4 is input into the estimation model in S5, the training objective function is calculated using automatic differentiation, and the optimization algorithm corresponding to the selected artificial intelligence network is selected to train the estimation model.

[0038] Based on the convergence of the estimation model, the learning rate in the optimization algorithm is adjusted to ensure stable and rapid convergence of the estimation model. The structure of the estimation model is adjusted according to the fitting results of the estimation model on the training set, taking into account actual computing resources. The performance of the estimation model on both the training and validation sets is comprehensively analyzed. While ensuring the regularization effect of the physical constraint terms without affecting the output accuracy, the weights of the training objective function described in S5 are adjusted to ensure that the training set error is within an acceptable range and the test set error is as small as possible, thus obtaining a well-trained estimation model.

[0039] Furthermore, in step S7, the spatial coordinates and time coordinate sequence of any point in the module to be estimated, as well as the corresponding temperature sensor temperature sequence, are input into the trained estimation model, that is, the temperature change sequence of that point under the current working condition is output; or the module spatial coordinate matrix and the corresponding temperature sensor temperature at a specific time point are input, that is, the temperature field distribution of the entire module under the current working condition and at a specific time point is output.

[0040] An AI-based power semiconductor module temperature distribution estimation system includes:

[0041] Parameter determination module: used to determine the layout geometry parameters and packaging material parameters of the module to be estimated, specifically including: chip size and position, temperature sensor position, size and thickness of each packaging layer, and thermophysical property parameters of each layer material;

[0042] Thermal simulation model building module: It is used to determine the parameters obtained in the module based on the parameters, build a thermal simulation model of the module to be estimated in the simulation software, and select the mesh size to construct a three-dimensional spatial mesh according to the minimum unit side length of each part of the module.

[0043] Thermal simulation module: In the thermal simulation model, a series of typical operating conditions are selected according to the actual operating range of the module to be estimated. The time step is determined according to the actual hardware computing power to perform thermal simulation and obtain thermal field data of the module to be estimated under multiple typical operating conditions.

[0044] Data export module: It is used to export the temperature data and time derivative of the temperature sensor location under different working conditions, the temperature field training data of the module to be estimated, and the spatiotemporal coordinates of the matching points from the results obtained by the thermal simulation module. It performs data preprocessing according to the input and output normalization and divides the exported data into training set and validation set.

[0045] The estimation model building module is used to initially determine the network structure and artificial intelligence algorithm, write the training objective function of the estimation model in combination with the encapsulation material parameters, and establish an estimation model based on the artificial intelligence algorithm.

[0046] Training module: Used to train the estimation model. Based on the performance of the estimation model on the training set and validation set, its hyperparameters are adjusted sequentially to ensure output accuracy while enabling the estimation model to converge quickly and stably, resulting in a trained estimation model that includes the temperature curve of the temperature sensor and the thermal field mapping relationship within the operating range of the module to be estimated.

[0047] Estimation module: Based on the trained estimation model, it takes the temperature curve of the temperature sensor under any actual working condition as input and completes the estimation of the temperature distribution of the module to be estimated under the corresponding conditions.

[0048] The beneficial effects of this application are as follows: This invention uses a method to decouple power loss and heat dissipation conditions through temperature curves from temperature sensors, and configures training data based on these temperature curves to achieve full coverage of actual operating conditions; a model based on artificial intelligence algorithms achieves accurate reconstruction of the module's thermal field. Unlike other methods, this method decouples operating parameters, ensuring complete non-intrusiveness while obtaining accurate temperature field distribution of the entire module under any actual operating condition. Furthermore, it boasts fast calculation speed, strong versatility, and is suitable for online temperature monitoring of power semiconductor modules. Attached Figure Description

[0049] The accompanying drawings are provided to further understand the invention and constitute a part of this invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0050] Figure 1 shows the package structure of the ROHM BSM250D17P2E004 SiC module, with the locations of each chip and NTC marked.

[0051] Figure 2 shows the hybrid neural network results, thermal simulation results, and relative error diagram of the temperature field on the upper surface of module DBC and the upper surface of T7 chip at t=30s. (a) Training set (P loss =550W,h=1500W / (m 2 ·K)), (b) Validation set (P) loss =500W,h=1200W / (m 2 ·K)), (c) Validation set (P) loss =500W,h=1800W / (m 2 ·K));

[0052] Figure 3 shows the module junction temperature T. j 7 Comparison of measured and estimated results, (a) Test set (P) loss =181.7W,h=1220W / (m 2 ·K)), (b) test set (P) loss =305.6W,h=1510W / (m 2 ·K)), (c) test set (P) loss =543.3W,h=1740W / (m 2 ·K));

[0053] Figure 4 is a flowchart of the junction temperature estimation method for power semiconductor modules based on a built-in NTC sensor. Detailed Implementation

[0054] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0055] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention 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 the invention 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 a 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.

[0056] Example 1

[0057] This invention provides a method for estimating the temperature distribution of power semiconductor modules based on artificial intelligence, which includes the following steps:

[0058] S1: Determine the layout geometry parameters and packaging material parameters of the module to be estimated. Specifically, this includes: the size and location of the chip, the location of the temperature sensor, the size and thickness of each packaging layer, and the thermophysical properties of each layer's material.

[0059] In step S1, the module to be estimated contains one or more temperature sensors, including but not limited to negative thermal coefficient thermistors and thermocouples, located on the module's encapsulation surface or inside. The temperatures measured by these sensors are used as observation points for subsequent steps, and their temperatures are denoted as {T}. obs1 (t),T obs2 (t),...,T obsM ...

[0060] S2: Based on the results obtained in S1, establish a thermal simulation model of the module in the simulation software, and select an appropriate density to construct a three-dimensional spatial mesh (specifically, select the mesh size to construct a three-dimensional spatial mesh based on the minimum unit side length of each part of the module).

[0061] In step S2, in commercial thermal simulation software (such as COMSOL, ANSYS, etc.), firstly, a geometric model of the module to be estimated is established and the corresponding material thermal parameters are assigned; secondly, a three-dimensional mesh of appropriate size is constructed according to the geometric dimensions, and the mesh density needs to be determined in combination with the actual hardware computing power.

[0062] S3: Based on the thermal simulation model, according to the actual operating conditions of the module, a series of typical operating conditions are selected, a reasonable time step is determined (the time step is determined according to the actual hardware computing power), and thermal simulation is performed to obtain multiple sets of module thermal field data under typical operating conditions.

[0063] In step S3, the actual operating conditions of the module to be estimated are first determined. Then, a series of typical operating conditions (i.e., power loss P) are selected. loss Based on the temperature curves of the temperature sensor (and the thermal convection coefficient h), and ensuring reasonable spacing between curves, the temperature curve cluster of the entire temperature sensor is designed to cover as large a range as possible (preferably the largest range), thus achieving full coverage of the actual possible operating conditions. Then, a reasonable time step is selected based on available computing resources to perform transient thermal simulations, obtaining module thermal field data under multiple sets of typical operating conditions.

[0064] S4: Export the temperature data (and its time derivative) of the temperature sensor location under different working conditions, the module temperature field training data, and the spatiotemporal coordinates of the matching points. Perform data preprocessing according to the input and output normalization, and divide the exported data into training set and validation set.

[0065] In step S4, temperature data at the temperature sensor location and its time derivative, module temperature field, and spatiotemporal coordinates of the mating point are exported at the time points selected in S3, forming three types of thermal simulation data. To avoid introducing new hyperparameters, mating points are selected only in the non-heat source region.

[0066] The exported thermal simulation data contains three types of data: spatial coordinates, time coordinates, and temperature data. The spatial coordinates are normalized to 0 using the same scaling scale, and the time and temperature are scaled to ~[0,1]. The specific values ​​of the scaling scale need to be determined in combination with the geometric dimensions and rated operating conditions of the module to be estimated.

[0067] The normalized data is divided into training and validation sets according to a certain ratio to complete the construction of the dataset.

[0068] S5: Initially determine the basic architecture and algorithm, combine the encapsulation material parameters of each layer in S1 to write the training objective function of the model, and establish a model based on artificial intelligence algorithm;

[0069] In step S5, the human intelligence algorithms that can be used to build the model include, but are not limited to, various supervised learning methods. In terms of network structure, they can include fully connected neural networks, residual neural networks, convolutional neural networks, recurrent neural networks, and other derived novel networks. In terms of training objective function, they can include pure data loss function, physical information loss function, and hybrid loss function, as well as other similar forms.

[0070] The model consists of: an input layer, with each input channel being... These correspond to spatial coordinates, time coordinates, and the temperatures of temperature sensors (a total of M). Corresponding to normalized spatial coordinates, Corresponding time coordinates The temperature corresponds to the normalized temperature measured by the 1st, 2nd...Mth temperature sensor; several intermediate layers, which can adopt various different network structures; a 1-channel output layer, corresponding to the module temperature field;

[0071] The model's training objective function consists of two parts: the training data error and the physical law residual: loss = w D MSE D +w F MSE F

[0072] Among them, the training data error MSE D The mean square error (MSE) between the model output temperature and the thermal simulation result temperature is the physical law residual. F To train the residuals of the partial differential equation for heat conduction at the matching points. D w F These are the weights for the two items, which can be adjusted flexibly according to the actual data. If a pure data loss function is used, then w is taken as... F =0 is sufficient. Similarly, if the physical information loss function is used, then w is taken as 0. D =0 is sufficient. If a hybrid loss function is used, a reasonable value needs to be determined during subsequent debugging to ensure that the magnitudes of the two losses are relatively balanced, thus ensuring stable and rapid convergence of the network.

[0073] The specific calculation method for the training data error term is as follows:

[0074] Where, N D The number of input training data, Let i be the output temperature of the artificial intelligence network at the i-th point. The actual temperature at the i-th point is derived from S4.

[0075] The physical law residual term is the mean square residual of the heat conduction partial differential equation calculated at selected points (called compositing points) in the module. Its original form is:

[0076] Where, N F To represent the number of syntagmatic points, the spatiotemporal coordinates of the i-th syntagmatic point are (x... i ,y i ,z i ,t i ), temperature T i k, ρ, and c are the thermal conductivity, density, and constant-pressure heat capacity of the material in the region where the point is located, respectively, and are determined in S1. The lattice point is the training point selected to apply physical constraints. Its main difference from ordinary training data is that the lattice point only requires spatiotemporal coordinate input and does not require actual temperature values, which are all derived from S4.

[0077] Considering the data normalization preprocessing in S4, the calculation formula should be expressed using normalized variables. Substituting these variables, the calculation formula for the physical law residual term can be obtained as follows:

[0078] in, These are all the derivatives of the output of the artificial intelligence network with respect to the input, and automatic differentiation is used during training. The temperature of the temperature sensor is the derivative of temperature with respect to time. It is calculated in S4 using the numerical differentiation method and then directly exported.

[0079] In summary, since both terms in the training objective function are in the form of mean square error, minimizing the loss function loss→0 will always result in MSE. D →0 and MSE F →0. The former indicates that the model's output temperature field is accurate under typical conditions of the training data; the latter forces the model's output temperature field to conform to the underlying thermophysical laws (i.e., the partial differential equation of heat conduction), ensuring its strong generalization ability and sufficient interpretability. Therefore, it can fully guarantee the high accuracy and high resolution of the model in reconstructing the thermal field;

[0080] At this point, the model based on artificial intelligence algorithms has been established.

[0081] S6: Train the model and adjust its hyperparameters according to the model’s performance on the training and validation sets to ensure output accuracy while enabling the model to converge quickly and stably.

[0082] In step S6, the training set obtained in S4 is configured into a small batch of appropriate size and input into the model in S5. The training objective function is calculated using automatic differentiation, and a suitable optimization algorithm is selected to train the model (specifically, the optimization algorithm corresponding to the selected artificial intelligence network is selected to train the estimation model).

[0083] Based on the model's convergence performance, adjust and optimize the learning rate in the algorithm to ensure stable and rapid model convergence; based on the model's fitting results to the training set, adjust the model's structure according to actual computing resources; comprehensively analyze the model's performance on the training and validation sets, ensuring the regularization effect of the physical constraint terms without affecting output accuracy, and adjust the weights of the training objective function described in S5 to keep the training set error within an acceptable range and minimize the test set error; complete the model debugging.

[0084] Finally, the training of the debugged model is completed.

[0085] S7: Obtain a model that includes the mapping relationship between the temperature curves of the temperature sensors and the thermal field within the module's operating range. Input the temperature curves of the temperature sensors under any actual operating conditions to complete the estimation of the module's temperature distribution under the corresponding conditions.

[0086] In step S7, the trained model realizes the mapping of spatiotemporal coordinates and temperature sensor temperature to the temperature field of the entire module. The temperature sensor temperature indirectly reflects the operating conditions of the module, that is, it realizes artificial intelligence thermal field reconstruction based on temperature observation points.

[0087] Input the temperature curve of the temperature sensor under any operating condition, and the model outputs the module temperature field distribution under the corresponding operating condition to complete the temperature estimation.

[0088] In practical applications, by inputting the spatial coordinates and time coordinate sequence of any point in the module to be estimated, as well as the corresponding temperature sensor temperature sequence, the temperature change sequence of that point under the current operating conditions can be output; or by inputting the module spatial coordinate matrix and the corresponding temperature sensor temperature at a specific time point, the temperature field distribution of the entire module under the current operating conditions and at that specific time point can be output.

[0089] At this point, the temperature estimation of the power semiconductor module with decoupled operating parameters is complete.

[0090] Example 2

[0091] The present invention will be further described in detail using the ROHM BSM250D17P2E004 multi-chip parallel SiC module shown in Figure 1 as an example.

[0092] The proposed method for estimating the temperature distribution of power semiconductor modules based on artificial intelligence includes the following steps:

[0093] S1: Determine the layout geometry parameters and packaging material parameters of the module to be estimated. Specifically, this includes: the size and location of the chip, the location of the temperature sensor, the size and thickness of each packaging layer, and the thermophysical properties of each layer's material.

[0094] S2: Based on the results obtained in S1, establish a thermal simulation model of the module in the simulation software, and select an appropriate density to construct a three-dimensional spatial mesh;

[0095] S3: Based on the actual operating conditions of the module, select a series of typical operating conditions, determine a reasonable time step, perform thermal simulation, and obtain module thermal field data under multiple typical operating conditions.

[0096] S4: Export the temperature data (and its time derivative) of the temperature sensor location under different working conditions, the module temperature field training data, and the spatiotemporal coordinates of the matching points. Perform data preprocessing according to the input and output normalization, and divide the exported data into training set and validation set.

[0097] S5: Initially determine the basic architecture and algorithm, combine the encapsulation material parameters of each layer in S1 to write the training objective function of the model, and establish a model based on artificial intelligence algorithm;

[0098] S6: Train the model and adjust its hyperparameters according to the model’s performance on the training and validation sets to ensure output accuracy while enabling the model to converge quickly and stably.

[0099] S7: Obtain a model that includes the mapping relationship between the temperature curves of the temperature sensors and the thermal field within the module's operating range. Input the temperature curves of the temperature sensors under any actual operating conditions to complete the estimation of the module's temperature distribution under the corresponding conditions.

[0100] In step S1, the module is a ROHM BSM250D17P2E004 SiC half-bridge module, with eight chips connected in parallel on both the upper and lower bridge arms. The temperature sensor NTC is located near the boundary on one side of the upper bridge arm, as shown in Figure 1. The module can be divided into six main layers from top to bottom: chip layer, solder layer, upper copper layer of DBC, ceramic layer of DBC, lower copper layer of DBC, and heat sink layer. Their thicknesses are (0.38, 0.1, 0.3, 0.38, 0.3, 3) mm, corresponding to four packaging materials: SiC, SAC305, Cu, Al2O3, Cu, and Cu. The thermal conductivity of SiC, SAC305, Cu, and Al2O3 are (450, 50, 400, 35) W / (m·K), their constant-pressure heat capacities are (1200, 237, 385, 790) J / (kg·K), and their densities are (3200, 7345, 8960, 3965) kg / m³. 3 .

[0101] In step S2, based on the results obtained in S1, the geometric model of the module is established in COMSOL and the corresponding material thermal parameters are assigned. A three-dimensional mesh of appropriate size is constructed, containing a total of 44,909 spatial points.

[0102] In step S3, taking the upper bridge arm of the module as a heat source as an example, considering the power cycle condition, the heating time t on =30s, cooldown time t off =30s. A typical operating condition with a relatively low cooling water flow rate is selected, corresponding to a typical thermal convection coefficient h = 1500 W / (m²). 2 For simplicity, only simulation results under a single heat dissipation condition (h) are used as training data. The proposed hybrid neural network model can cover variations in heat dissipation conditions within a certain range. Referring to the NTC temperature curve of the reference module, P is selected. loss = Five typical working conditions of {150, 250, 350, 450, 550}W were used as training data to ensure T NTC The (t) curves have a certain spacing, which allows the curve cluster to completely cover the possible power loss.

[0103] The ambient temperature was set to 20℃. Ten heating-cooling cycles were simulated under each operating condition until the module temperature field reached a steady state. The temperature field results at 30 time points in t = [480s, 540s] were selected as training data, with a time step Δt = 2s. Finally, transient thermal simulation was performed to obtain module temperature field data under five typical operating conditions.

[0104] In step S4, the following three types of module thermal simulation data are exported at the selected time points in S3: 1) NTC temperature and its time derivative, totaling 2 × 30 × 5 = 300 data points. 2) Module temperature field data, exported from the simplified grid (7724 points in the entire module) to reduce computational load, totaling 7724 × 30 × 5 = 1158600 data points. 3) Spatiotemporal coordinates of matching points, i.e., the spatiotemporal coordinates of grid points in non-heat source areas, totaling 4732 × 30 × 5 = 709800 data points. To save computational load, training matching points are selected only in the main layers (i.e., relatively thicker layers) of the module to be estimated. Considering the convenience of implementation and to obtain better regularization effects, the spatiotemporal coordinates of the module temperature field data are selected as training matching points. The simplified grid points (i.e., training matching points) are basically evenly distributed in the region, with an appropriate increase in density at the boundaries. Furthermore, to decouple operating parameters, physical constraints are applied only to grid points in non-source regions, i.e., the point coordinates of the chip layer are discarded. Otherwise, difficult-to-measure heat source terms would be introduced into the physical constraint residuals.

[0105] In the exported data, the spatial coordinates are normalized to 0-centered using the same scaling scale, and then the time and temperature are scaled to ~[0,1] respectively:

[0106] Among them, the spatial scaling scale μ is determined by combining the actual module parameters. xyz=200, time scaling scale μ t = 1 / 6, the highest and lowest reference temperatures T max =390K and T min =290K. The training set is now complete.

[0107] Based on the training set operating conditions, the interpolation result P of the power loss is selected. loss ={200,300,400,500}W, the 20% extrapolation result of the convection coefficient, i.e., (1±0.2)h={1200,1800}W / (m 2 The corresponding thermal simulation data of K) is used as the validation set.

[0108] In step S5, a hybrid neural network model is initially constructed in MATLAB's Deep Learning Toolbox, using Xavier initialization. The hybrid neural network model consists of: an input layer, with each input channel... These correspond to spatial coordinates, temporal coordinates, and NTC temperature, respectively; several intermediate layers, consisting of fully connected layers and tanh activation function layers, are implemented using residual blocks with jump connections; and a single-channel output layer corresponds to the module's temperature field. A hybrid neural network model containing L residual blocks has a total of 2L+1 hidden layers, each containing N neurons. L and N are the network structure hyperparameters that need to be tuned to determine their optimal values.

[0109] The hybrid loss function of a hybrid neural network model consists of two parts: training data error and physical law residuals: loss = w D MSE D +w F MSE F

[0110] Among them, the training data error MSE D The mean square error (MSE) between the output temperature of the hybrid neural network and the actual temperature at the training point is represented by the physical law residual. F To train the residuals of the partial differential equation for heat conduction at the matching points. D w F These are the weights for the two items, and their reasonable values ​​need to be determined during subsequent debugging to ensure that the magnitudes of the two losses are relatively balanced, thus ensuring stable and rapid convergence of the network. Initially, both are set to 1.

[0111] The specific calculation method for the training data error term is as follows:

[0112] Wherein, the number of input training data N D =1158600, Let HNNM be the output temperature at the i-th training point. The actual temperature at the i-th training point is derived from S4.

[0113] The physical law residual term is the mean square residual of the heat conduction partial differential equation calculated at selected points (called compositing points) in the module. Its original form is:

[0114] Among them, the number of matching points N F =709800, the spatiotemporal coordinates of the i-th pairing point are (x i ,y i ,z i ,t i ), temperature T i k, ρ, and c represent the thermal conductivity, density, and constant-pressure heat capacity of the material in the region where the point is located, respectively, and are determined in S1. The lattice point is the training point selected to apply physical constraints. Its main difference from ordinary training data is that the lattice point only requires spatiotemporal coordinate input and does not require actual temperature values, which are all derived from S4.

[0115] Considering the data normalization preprocessing in S4, the calculation formulas should be expressed using normalized variables. For the spatial differential term (taking the x-coordinate as an example, the rest are in the same form), applying the chain rule and substituting the input and output normalization, we have:

[0116] For the time derivative, note the input variables. It is also a function of time. Applying the chain rule for multivariable functions and substituting the input and output normalization, we have:

[0117] Substituting into the original calculation formula, we obtain the formula for calculating the physical law residual term:

[0118] in, These are all the derivatives of the output of the hybrid neural network with respect to the input, and automatic differentiation is used during training. The derivative of NTC temperature with respect to time is calculated and derived in S4 using numerical differentiation. It's important to note that this numerical differentiation term is only used for network training and is not needed for actual estimation; therefore, it does not affect the ease of application.

[0119] In summary, since both terms in the total loss function are in the form of mean square error, minimizing the loss function as loss→0 will always result in MSE. D →0 and MSE F→0. The former indicates that the output temperature field of the hybrid neural network model is accurate under typical operating conditions of the training data; the latter forces the output temperature field of the hybrid neural network model to conform to the underlying thermophysical laws (i.e., the partial differential equation of heat conduction), ensuring its strong generalization ability and sufficient interpretability. Therefore, the high accuracy and high resolution of thermal field reconstruction achieved through the hybrid neural network model can be fully guaranteed.

[0120] In step S6, the training set obtained in S4 is configured into 10 mini-batches and input into the hybrid neural network model in S5. In MATLAB's Deep Learning Toolbox, the hybrid loss function is calculated using automatic differentiation, and the Adam optimization algorithm is used for training.

[0121] The learning rate α directly affects the convergence of the network. A strategy is adopted to gradually decrease the learning rate with the number of epochs:

[0122] The initial learning rate α0 = 1e-2. A relatively large decayRate of 0.1 is initially set to ensure stable convergence of the neural network. To accelerate convergence, decayRate is gradually decreased, aiming to speed up convergence as much as possible while maintaining basic stability. The final decayRate is set to 0.005.

[0123] The number of hidden layers and the number of neurons per layer characterize the network's fitting ability. The optimal network structure needs to be determined based on available computing resources. The main reference is the network's output performance on the training set, using the loss function value after the same training time as a direct measure, comparing the loss values ​​of different structures after the same training time. Considering both training time and output accuracy, (L=4, N=128) is chosen as a relatively optimized hybrid neural network structure.

[0124] The weights in the loss function need to be adjusted to balance the relative magnitudes of different terms. A comprehensive analysis of the hybrid neural network's performance on both the training and validation sets is conducted to ensure the regularization effect of the physical constraint terms without affecting output accuracy. Therefore, a suitable weight for the loss function, w, is determined. F =0.3, w D =1. Furthermore, considering the specific characteristics of the problem, the weights of the training data for the key areas of focus can be fine-tuned to achieve a better fit. Therefore, w is chosen. D_high=1.2, meaning that greater weight is assigned to the training data of the module chip layer and the copper layer on the DBC. Therefore, the proposed hybrid neural network focuses more on reducing errors in the high-temperature region during training, improving estimation accuracy. It is also important to note that because the loss function contains different weights and physical law residual terms, its magnitude can only be used to qualitatively judge how close the current network output is to the true solution, and cannot be directly used as a measure of the result error.

[0125] Finally, after training for 8000 epochs, the loss of the calibrated hybrid neural network stabilized at ~1.3867e-4, at which point the training of the hybrid neural network model was considered complete.

[0126] In step S7, the trained hybrid neural network model realizes the mapping between spatiotemporal coordinates and NTC temperature to the entire module temperature field. The NTC temperature indirectly reflects the module's operating conditions, thus achieving thermal field reconstruction based on NTC temperature. Inputting the NTC temperature curve under any operating condition, the hybrid neural network model outputs the module temperature field distribution under the corresponding condition, completing the temperature estimation.

[0127] To verify the estimated resolution, P was used in both the training and validation sets. loss Taking the maximum value as an example, Figure 2 shows the hybrid neural network results, thermal simulation results, and relative errors of the temperature fields on the upper surfaces of the DBC and T7 chips under different operating conditions at t=30s. Linear interpolation was used for the discontinuities in the copper layer on the DBC during plotting. As can be seen from Figure 2, the temperature distributions of the two are basically the same, indicating that the hybrid neural network, which includes underlying physical constraints, can effectively reproduce the module's temperature field distribution, and the output results are sufficiently reasonable. The relative error distribution shows that the relative error in most areas, except for the chip gaps, is ≤5%, and the relative error of the chip surface temperature field fitting is also ≤5%. This verifies that the thermal field reconstruction method based on hybrid neural networks proposed in this invention has extremely high resolution.

[0128] To verify the accuracy of the estimation, power cycling verification was carried out under a series of actual operating conditions, and the chip junction temperature was accurately measured using an infrared camera.

[0129] By inputting the chip's coordinates and the experimentally measured NTC temperature curve into the trained hybrid neural network model, the junction temperature curve under the corresponding operating conditions can be obtained, using the chip with a higher junction temperature T... j 7 For example, as shown in Figure 3, the relative error between the estimation results of the hybrid neural network model and the junction temperature measurement results under different operating conditions is less than 3%. The proposed method decouples a wide range of power losses from a certain range of heat dissipation conditions through the NTC temperature curve, and has high estimation accuracy within the corresponding operating condition range.

[0130] Example 3

[0131] This invention provides an artificial intelligence-based system for estimating the temperature distribution of power semiconductor modules, comprising:

[0132] Parameter determination module: used to determine the layout geometry parameters and packaging material parameters of the module to be estimated, specifically including: chip size and position, temperature sensor position, size and thickness of each packaging layer, and thermophysical property parameters of each layer material;

[0133] Thermal simulation model building module: It is used to determine the parameters obtained in the module based on the parameters, build a thermal simulation model of the module to be estimated in the simulation software, and select the mesh size to construct a three-dimensional spatial mesh according to the minimum unit side length of each part of the module.

[0134] Thermal simulation module: In the thermal simulation model, a series of typical operating conditions are selected according to the actual operating range of the module to be estimated. The time step is determined according to the actual hardware computing power to perform thermal simulation and obtain thermal field data of the module to be estimated under multiple typical operating conditions.

[0135] Data export module: It is used to export the temperature data and time derivative of the temperature sensor location under different working conditions, the temperature field training data of the module to be estimated, and the spatiotemporal coordinates of the matching points from the results obtained by the thermal simulation module. It performs data preprocessing according to the input and output normalization and divides the exported data into training set and validation set.

[0136] The estimation model building module is used to initially determine the network structure and artificial intelligence algorithm, write the training objective function of the estimation model in combination with the encapsulation material parameters, and establish an estimation model based on the artificial intelligence algorithm.

[0137] Training module: Used to train the estimation model. Based on the performance of the estimation model on the training set and validation set, its hyperparameters are adjusted sequentially to ensure output accuracy while enabling the estimation model to converge quickly and stably, resulting in a trained estimation model that includes the temperature curve of the temperature sensor and the thermal field mapping relationship within the operating range of the module to be estimated.

[0138] Estimation module: Based on the trained estimation model, it takes the temperature curve of the temperature sensor under any actual working condition as input and completes the estimation of the temperature distribution of the module to be estimated under the corresponding conditions.

[0139] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0140] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in one or more blocks of the flowchart illustrations and / or one or more blocks of the block diagrams.

[0141] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means that implement the functions specified in one or more flowcharts and / or one or more block diagrams.

[0142] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, such that the instructions, which execute on the computer or other programmable apparatus, provide steps for implementing the functions specified in one or more flowcharts and / or one or more block diagrams.

[0143] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit its scope of protection. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading the present invention, they can still make various changes, modifications or equivalent substitutions to the specific implementation of the invention, but these changes, modifications or equivalent substitutions are all within the scope of protection of the pending claims of the invention.

Claims

1. A method for estimating temperature distribution of a power semiconductor module based on artificial intelligence, characterized by, Includes the following steps: S1: Determine the layout geometry parameters and packaging material parameters of the module to be estimated, specifically including: the size and position of the chip, the position of the temperature sensor, the size and thickness of each packaging layer, and the thermophysical properties of each layer material; S2: Based on the parameters obtained in S1, establish a thermal simulation model of the module to be estimated in the simulation software, and select the mesh size according to the minimum unit side length of each part of the module to construct a three-dimensional spatial mesh. S3: In the thermal simulation model established in S2, a series of typical operating conditions are selected according to the actual operating range of the module to be estimated. The time step is determined according to the actual hardware computing power, and thermal simulation is performed to obtain thermal field data of the module to be estimated under multiple typical operating conditions. S4: In the results obtained in S3, the temperature data and time derivative of the temperature sensor location under different working conditions, the temperature field training data of the module to be estimated, and the spatiotemporal coordinates of the matching point are exported respectively. Data preprocessing is performed according to the input and output normalization, and the exported data is divided into training set and validation set. S5: Initially determine the network structure and artificial intelligence algorithm, combine the encapsulated material parameters in S1 to write the training objective function of the estimation model, and establish an estimation model based on the artificial intelligence algorithm; S6: Train the estimation model. Adjust its hyperparameters according to the performance of the estimation model on the training set and validation set to ensure the output accuracy while making the estimation model converge quickly and stably. This will result in a trained estimation model that includes the temperature curve of the temperature sensor and the thermal field mapping relationship within the operating range of the module to be estimated. S7: Based on the estimation model trained in S6, input the temperature curve of the temperature sensor under any actual working condition to complete the estimation of the temperature distribution of the module to be estimated under the corresponding conditions.

2. The power semiconductor module temperature distribution estimation method based on artificial intelligence according to claim 1, characterized by, In step S1, the module to be estimated contains one or more temperature sensors, including but not limited to negative thermal coefficient thermistors and thermocouples, which are located on the surface or inside the module package, respectively. The temperature measured by the temperature sensor is used as the observation point, and the temperature is denoted as {T}. obs1 (t),T obs2 (t),...,T obsM ...

3. The power semiconductor module temperature distribution estimation method based on artificial intelligence according to claim 1, characterized by, In step S2, the geometric model of the module to be estimated is first established in the thermal simulation software and the corresponding material thermal parameters are assigned. Then, the grid size is selected according to the minimum unit side length of each part of the module to construct a three-dimensional grid. The grid density needs to be determined in combination with the actual hardware computing power to complete the establishment of the module thermal simulation model.

4. The method for estimating the temperature distribution of a power semiconductor module based on artificial intelligence according to claim 1, characterized in that, In step S3, the actual operating range of the module to be estimated is first determined, and then a series of typical operating conditions are selected. Based on the temperature curve of the temperature sensor, the temperature curve cluster of the entire temperature sensor covers the largest possible range while ensuring reasonable spacing between the curves. That is, the full coverage of the actual possible operating range is achieved. Then, a reasonable time step is selected according to the actual hardware computing power, and transient thermal simulation is performed to obtain thermal field data of the module to be estimated under multiple typical operating conditions.

5. The method for estimating the temperature distribution of a power semiconductor module based on artificial intelligence according to claim 1, characterized in that, In step S4, three types of thermal simulation data are exported at the time point corresponding to the time step selected in S3: temperature data of the temperature sensor location and the numerical calculation result of its time derivative, temperature field of the module to be estimated, and spatiotemporal coordinates of the mating point. Among them, mating points are selected only in the non-heat source part, that is, in the area outside the chip layer. The exported thermal simulation data contains three types of data: spatial coordinates, time coordinates, and temperature data. The spatial coordinates are normalized to 0 using the same scaling scale. Then, the time and temperature are scaled to approximately [0,1]. The specific values ​​of the scaling scale are determined in conjunction with the geometric dimensions of the module to be estimated and the rated operating conditions. The normalized data is divided into training and validation sets according to a preset ratio.

6. The method for estimating the temperature distribution of a power semiconductor module based on artificial intelligence according to claim 5, characterized in that, In step S5, the human intelligence algorithm includes, but is not limited to, various supervised learning methods; Network architectures include fully connected neural networks, residual neural networks, convolutional neural networks, recurrent neural networks, and other novel networks derived from them; Training objective functions include pure data loss functions, physical information loss functions, or hybrid loss functions; The estimation model includes an input layer, several intermediate layers, and a 1-channel output layer; The input channels of the input layer are in, Corresponding to normalized spatial coordinates, Corresponding time coordinates The temperatures measured by the corresponding normalized 1st, 2nd...Mth temperature sensors; The aforementioned intermediate layers can adopt various different network structures; The output layer corresponds to the module temperature field.

7. The method for estimating the temperature distribution of a power semiconductor module based on artificial intelligence according to claim 6, characterized in that, The training objective function includes training data error and physical law residuals; loss = w D MSE D +w F MSE F Wherein, the training data error MSE D is the mean square error of the model output temperature and the thermal simulation result temperature, the physical law residual error MSE F is the heat conduction partial differential equation residual error of the training collocation point, w D , w F are the weights of the two terms respectively. The specific calculation method for the training data error term is as follows: Where, N D The number of input training data, Let i be the output temperature of the artificial intelligence network at the i-th point. The actual temperature at the i-th point is derived from S4; The physical law residual term is the mean square residual calculated at the combination points in the module for the partial differential equation of heat conduction. Its original form is: Where, N F To represent the number of syntagmatic points, the spatiotemporal coordinates of the i-th syntagmatic point are (x... i ,y i ,z i ,t i ), temperature T i k, ρ, and c are the thermal conductivity, density, and constant-pressure heat capacity of the material in the region where the pairing point is located, respectively, and are determined in S1. The pairing point is the selected training point for applying physical constraints. Its main difference from ordinary training data is that the pairing point only requires spatiotemporal coordinate input and does not require actual temperature values, which are all derived from S4. Considering the data normalization preprocessing in S4, the calculation formula is expressed using normalized variables. Substituting these variables into the formula yields the following formula for calculating the physical law residual term: in, These are all the derivatives of the output of the artificial intelligence network with respect to the input, calculated automatically during training. The temperature is the derivative of the temperature sensor with respect to time, which is calculated in S4 using the numerical differentiation method and then directly exported.

8. The method for estimating the temperature distribution of a power semiconductor module based on artificial intelligence according to claim 1, characterized in that, In step S6, the training set obtained in S4 is input into the estimation model in S5, the training objective function is calculated using automatic differentiation, and the optimization algorithm corresponding to the selected artificial intelligence network is selected to train the estimation model. Based on the convergence of the estimation model, the learning rate in the optimization algorithm is adjusted to ensure stable and rapid convergence of the estimation model. The structure of the estimation model is adjusted according to the fitting results of the estimation model on the training set, taking into account actual computing resources. The performance of the estimation model on both the training and validation sets is comprehensively analyzed. While ensuring the regularization effect of the physical constraint terms without affecting the output accuracy, the weights of the training objective function described in S5 are adjusted to ensure that the training set error is within an acceptable range and the test set error is as small as possible, thus obtaining a well-trained estimation model.

9. The method for estimating the temperature distribution of a power semiconductor module based on artificial intelligence according to claim 1, characterized in that, In step S7, the spatial coordinates and time coordinate sequence of any point in the module to be estimated, as well as the corresponding temperature sensor temperature sequence, are input into the trained estimation model, thereby outputting the temperature change sequence of that point under the current operating condition; or the module spatial coordinate matrix and the corresponding temperature sensor temperature at a specific time point are input, thereby outputting the temperature field distribution of the entire module under the current operating condition and at a specific time point.

10. A power semiconductor module temperature distribution estimation system based on artificial intelligence, characterized in that, include: Parameter determination module: used to determine the layout geometry parameters and packaging material parameters of the module to be estimated, specifically including: chip size and position, temperature sensor position, size and thickness of each packaging layer, and thermophysical property parameters of each layer material; Thermal simulation model building module: It is used to determine the parameters obtained in the module based on the parameters, build a thermal simulation model of the module to be estimated in the simulation software, and select the mesh size to construct a three-dimensional spatial mesh according to the minimum unit side length of each part of the module. Thermal simulation module: In the thermal simulation model, a series of typical operating conditions are selected according to the actual operating range of the module to be estimated. The time step is determined according to the actual hardware computing power to perform thermal simulation and obtain thermal field data of the module to be estimated under multiple typical operating conditions. Data export module: It is used to export the temperature data and time derivative of the temperature sensor location under different working conditions, the temperature field training data of the module to be estimated, and the spatiotemporal coordinates of the matching points from the results obtained by the thermal simulation module. It performs data preprocessing according to the input and output normalization and divides the exported data into training set and validation set. The estimation model building module is used to initially determine the network structure and artificial intelligence algorithm, write the training objective function of the estimation model in combination with the encapsulation material parameters, and establish an estimation model based on the artificial intelligence algorithm. Training module: Used to train the estimation model. Based on the performance of the estimation model on the training set and validation set, its hyperparameters are adjusted sequentially to ensure output accuracy while enabling the estimation model to converge quickly and stably, resulting in a trained estimation model that includes the temperature curve of the temperature sensor and the thermal field mapping relationship within the operating range of the module to be estimated. Estimation module: Based on the trained estimation model, it takes the temperature curve of the temperature sensor under any actual working condition as input and completes the estimation of the temperature distribution of the module to be estimated under the corresponding conditions.

Images