Power semiconductor module temperature distribution estimation method and system based on artificial intelligence

Abstract

The invention discloses a power semiconductor module temperature distribution estimation method and system based on artificial intelligence, and the method comprises the steps: firstly configuring training data according to a temperature curve graph of a temperature sensor, decoupling power loss and heat dissipation conditions through the temperature curve of the temperature sensor, and achieving the full coverage of an actual working condition; secondly, a model based on an artificial intelligence algorithm is constructed and debugged to achieve accurate reconstruction of a module thermal field, and temperature estimation of the power semiconductor module is achieved through the mapping relation between the temperature of the temperature sensor and module thermal field distribution. According to the method, decoupling of working condition parameters is achieved, accurate temperature field distribution of the whole module under any actual working condition can be obtained while a complete non-intrusive mode is guaranteed, the calculation speed is high, universality is high, and the method is suitable for online temperature monitoring of the power semiconductor module.

Description

Technical Field

[0001] The present invention belongs to the field of power semiconductor module status monitoring, and in particular relates to a method and system for estimating temperature distribution of a power semiconductor module based on artificial intelligence. Background Art

[0002] Power semiconductor modules are widely used in new energy conversion, smart grid, electric vehicles, industrial drives, aerospace and other fields. The influence of complex environmental stress and working conditions in different application scenarios may cause module performance degradation or failure. Timely and accurate status monitoring of power semiconductor modules can grasp their health status in real time, and take measures as soon as possible when failures may occur to reduce possible losses, which is crucial to improving the performance and reliability of the entire system. Among them, temperature is an important indicator for status monitoring. The semiconductor junction temperature directly affects the life and reliability of the chip, and the temperature of the packaging material also has a significant impact on the aging of materials such as the solder layer. Therefore, it is necessary to accurately monitor the temperature of the power semiconductor module. However, due to the complex structure of the power semiconductor module and its close coupling with working conditions and environmental conditions, it is difficult to obtain high-resolution and high-precision junction temperature results.

[0003] For online monitoring of junction temperature of power semiconductor modules, the industry generally adopts thermistor electrical parameter method or thermal model method. Thermistor electrical parameter method uses the semiconductor chip itself as a sensor and obtains 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 electrical parameters requires high-precision measurement equipment, which is bound to increase cost and complexity, and thus affect the reliability of the entire system. And because different degrees of nonlinearity require corresponding complex calibration procedures, there are certain difficulties in practical applications.

[0004] The thermal model method combines the module packaging structure and boundary conditions to build a thermal network model, uses simulation results to extract thermal network parameters, and calculates junction temperature through real-time loss information. In contrast, the use of thermal models to estimate junction temperature is relatively simple and non-invasive, without the need to introduce additional components, and can achieve junction temperature monitoring without affecting the reliability of the entire system. However, most existing power semiconductor modules use multiple chips in parallel to improve their current flow capacity. The close thermal cross-coupling between chips increases the difficulty of modeling and limits the improvement of thermal model resolution. At the same time, it is difficult to obtain accurate boundary condition information under actual working conditions, which greatly limits the accuracy of the thermal model. Boundary conditions include power loss and heat dissipation conditions. Among them, power loss is directly used as the input of the thermal model, while heat dissipation conditions will indirectly affect the model parameters, both of which affect the accuracy of the output results of the thermal model.

[0005] At present, many power semiconductor modules are equipped with temperature sensors, usually a negative thermal coefficient thermistor (NTC). The NTC and the chip are located on the same ceramic substrate. By measuring the NTC resistance online, it is very convenient to obtain the temperature change of that point. Using a temperature sensor as a temperature observation point can simplify the thermal model of the module and reduce the dependence on accurate heat dissipation conditions to a certain extent. However, the existing monitoring method based on temperature sensors is still restricted by the acquisition of power loss and can only reflect the temperature of a single point. There is a lack of a power semiconductor module temperature monitoring method that can obtain the temperature of all points in the module and decouple the operating conditions, which cannot meet the requirements of online monitoring of the temperature distribution of power semiconductor modules. Summary of the invention

[0006] The purpose of the present invention is to provide a method and system for estimating the temperature distribution of a power semiconductor module based on artificial intelligence to overcome the defects of the prior art. While ensuring that it is completely non-invasive, the present invention can obtain the accurate temperature field distribution of the entire module under any actual working conditions, and has fast calculation speed and strong versatility, and is suitable for online temperature monitoring of power semiconductor modules.

[0007] In order to achieve the above object, the present invention adopts the following technical scheme:

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

[0009] S1: Determine the layout geometric parameters and packaging material parameters of the module to be estimated, including: the size and position of the chip, the position of the temperature sensor, the size and thickness of each packaging layer, and the thermal physical property parameters of each layer material;

[0010] S2: Based on the parameters obtained in S1, a thermal simulation model of the module to be estimated is established in the simulation software, and a three-dimensional space grid is constructed by selecting a grid size based on the minimum unit side length of each part of the module;

[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, and the time step is determined according to the actual hardware computing capacity to perform thermal simulation to obtain multiple groups of thermal field data of the module to be estimated under typical operating conditions;

[0012] S4: from the results obtained in S3, the temperature data and its time differential of the temperature sensor position under different working conditions, the training data of the temperature field of the module to be estimated, and the time and space coordinates of the matching points are respectively derived, and the data is preprocessed according to the input and output normalization, and the derived data is divided into a training set and a validation set;

[0013] S5: Preliminarily determine the network structure and artificial intelligence algorithm, write the training objective function of the estimation model in combination with the packaging material parameters in S1, and establish an estimation model based on the artificial intelligence algorithm;

[0014] S6: training the estimation model, and adjusting its hyperparameters in turn according to the performance of the estimation model on the training set and the validation set, ensuring the output accuracy while making the estimation model converge quickly and stably, and obtaining a trained estimation model including the mapping relationship between the temperature curve of the temperature sensor and the thermal field within the working condition range of the module to be estimated;

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

[0016] Furthermore, 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, which are respectively located on the surface or inside the module package, and the temperature measured by the temperature sensor is used as the observation point, and the temperature is respectively recorded as {T obs1 (t),T obs2 (t),...,T obsM (t)}, respectively representing the temperatures measured by the 1st, 2nd, ..., Mth temperature sensors.

[0017] Furthermore, in step S2, in the thermal simulation software, firstly, a geometric model of the module to be estimated is established and the corresponding material thermal parameters are assigned; secondly, a three-dimensional grid is constructed by selecting a grid size according to the minimum unit side length of each part of the module, and 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.

[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 diagram of the temperature sensor, while ensuring reasonable spacing between the curves, the temperature curve cluster of the entire temperature sensor covers the largest range, that is, 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 multiple groups of thermal field data of the module to be estimated under typical operating conditions.

[0019] Furthermore, in step S4, three types of thermal simulation data are derived at the time point corresponding to the time step selected in S3, namely, the temperature data of the temperature sensor position and the numerical calculation result of its time differential, the temperature field of the module to be estimated, and the time-space coordinates of the matching points; wherein the matching points are selected only in the non-heat source part, that is, the area outside the chip layer;

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

[0021] The normalized data is divided into a training set and a validation set 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] The network structure includes fully connected neural network, residual neural network, convolutional neural network, recurrent neural network and other derived new networks;

[0024] The training objective function includes pure data loss function, physical information loss function or mixed loss function;

[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 the normalized spatial coordinates, Corresponding to the time coordinate, Corresponding to the normalized temperatures measured by the 1st, 2nd, ... Mth temperature sensors;

[0027] The plurality of intermediate layers can adopt various network structures;

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

[0029] Furthermore, the training objective function includes training data error and physical law residual;

[0030] loss = w D MSE D +w F MSE F

[0031] Among them, the training data error MSE D is the mean square error between the model output temperature and the thermal simulation result temperature, and the physical law residual MSE F is the residual of the heat conduction partial differential equation of the training collocation point, w D 、w F are the weights of the two items respectively;

[0032] The specific calculation method of the training data error term is:

[0033]

[0034] Among them, N D is the number of input training data, is the output temperature of the artificial intelligence network at the i-th point, is the actual temperature at the i-th point, all derived from S4;

[0035] The physical law residual term is the mean square residual of the heat conduction partial differential equation calculated at the matching points in the module. Its original form is:

[0036]

[0037] Among them, N F is the number of collocation points, and the space-time coordinates of the i-th collocation point are (x i ,y i ,z i ,t i ), the temperature is T i , k, ρ, c are the thermal conductivity, density, and constant-pressure heat capacity of the material in the region where the collocation point is located, respectively, which are determined in S1. The collocation point is the training point selected for applying physical constraints. The main difference between the collocation point and the ordinary training data is that the collocation point only needs the space-time coordinate input, and does not need the actual temperature value, which is derived from S4;

[0038] Considering the data normalization preprocessing in S4, the calculation formula is expressed by the normalized variables and the calculation formula of the residual term of the physical law is obtained as follows:

[0039]

[0040] in, They are all the differentials of the output of the artificial intelligence network to the input, and are calculated using automatic differentiation during the training process. is the derivative value of the temperature of the temperature sensor with respect to time, which is directly derived after being calculated using the numerical differentiation method in S4.

[0041] Furthermore, in step S6, the training set obtained in S4 is input into the estimation model in S5, the training objective function is calculated by automatic differentiation, and the optimization algorithm corresponding to the selected artificial intelligence network is selected to train the estimation model;

[0042] Based on the convergence of the estimation model, debug the learning rate in the optimization algorithm to make the estimation model converge stably and quickly; debug the structure of the estimation model based on the fitting results of the estimation model on the training set and the actual computing resources; comprehensively analyze the performance of the estimation model on the training set and the validation set, ensure the regularization effect of the physical constraint terms without affecting the output accuracy, and debug the weight of the training objective function described in S5 to make the training set error within an acceptable range and the test set error as small as possible, so as to obtain a trained estimation model.

[0043] Furthermore, in step S7, the spatial coordinates, time coordinate sequence and corresponding temperature sensor temperature sequence of any point in the module to be estimated are input into the trained estimation model, that is, the temperature change sequence of the 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 the specific time point is output.

[0044] The artificial intelligence-based power semiconductor module temperature distribution estimation system includes:

[0045] Parameter determination module: used to determine the layout geometric parameters and packaging material parameters of the module to be estimated, including: the size and position of the chip, the position of the temperature sensor, the size and thickness of each packaging layer, and the thermal physical property parameters of each layer of material;

[0046] Thermal simulation model building module: used to determine the parameters obtained in the module according to the parameters, establish the thermal simulation model of the module to be estimated in the simulation software, and select the grid size according to the minimum unit side length of each part of the module to construct a three-dimensional space grid;

[0047] Thermal simulation module: used in the thermal simulation model to select a series of typical operating conditions according to the actual operating range of the module to be estimated, determine the time step according to the actual hardware computing power, perform thermal simulation, and obtain multiple sets of thermal field data of the module to be estimated under typical operating conditions;

[0048] Data export module: It is used to export the temperature data and its time differential of the temperature sensor position under different working conditions, the training data of the temperature field of the module to be estimated, and the time and space coordinates of the matching points from the results obtained by the thermal simulation module, perform data preprocessing according to the input and output normalization, and divide the exported data into training set and verification set;

[0049] Estimation model building module: used to preliminarily determine the network structure and artificial intelligence algorithm, write the training objective function of the estimation model in combination with the packaging material parameters, and establish an estimation model based on the artificial intelligence algorithm;

[0050] Training module: used to train the estimation model. According to the performance of the estimation model on the training set and the validation set, its hyperparameters are adjusted in turn to ensure the output accuracy while making the estimation model converge quickly and stably, so as to obtain a trained estimation model containing the temperature curve of the temperature sensor and the thermal field mapping relationship within the working range of the module to be estimated;

[0051] Estimation module: It is used to input the temperature curve of the temperature sensor under any actual working conditions based on the trained estimation model, and complete the temperature distribution estimation of the module to be estimated under the corresponding conditions.

[0052] Compared with the prior art, the present invention has the following beneficial technical effects:

[0053] The present invention uses a method of decoupling power loss and heat dissipation conditions through the temperature curve of the temperature sensor, configures training data according to the temperature curve of the temperature sensor, and achieves full coverage of actual working conditions; the model based on the artificial intelligence algorithm realizes accurate reconstruction of the module thermal field. Different from other methods, this method realizes the decoupling of working condition parameters, and can obtain the accurate temperature field distribution of the entire module under any actual working condition while ensuring complete non-intrusion, and has fast calculation speed and strong versatility, which is suitable for online temperature monitoring of power semiconductor modules. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The drawings in the specification are used to provide further understanding of the present invention and constitute a part of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0055] Figure 1 This is the package structure diagram of the ROHM BSM250D17P2E004 SiC module, which shows the location of each chip and NTC;

[0056] Figure 2 The upper surface of the module DBC and T at t = 30s 7 Hybrid neural network results, thermal simulation results and relative error diagram of the chip surface temperature field. (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));

[0057] Figure 3 is the module junction temperature T j 7 Comparison of measured results 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));

[0058] Figure 4 The figure is a flow chart of a method for estimating junction temperature of a power semiconductor module based on a built-in NTC sensor. DETAILED DESCRIPTION

[0059] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0060] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0061] Embodiment 1

[0062] The present invention provides a method for estimating temperature distribution of a power semiconductor module based on artificial intelligence, which comprises the following steps:

[0063] S1: Determine the layout geometry parameters and packaging material parameters of the module to be estimated, including: the size and position of the chip, the position of the temperature sensor, the size and thickness of each packaging layer, and the thermal physical property parameters of each layer of material;

[0064] In step S1, the module to be estimated contains one or more temperature sensors, including but not limited to negative thermal coefficient thermistors, thermocouples, etc., which are respectively located on the surface or inside the module package. The temperatures measured by these sensors are used as observation points to carry out subsequent steps. The temperatures are respectively recorded as {T obs1 (t),T obs2 (t),...,T obsM (t)}, respectively representing the temperatures measured by the 1st, 2nd, ..., Mth temperature sensors.

[0065] S2: Based on the results obtained in S1, a thermal simulation model of the module is established in the simulation software, and a suitable density is selected to construct a three-dimensional space grid (specifically, a grid size is selected according to the minimum unit side length of each part of the module to construct the three-dimensional space grid);

[0066] 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 grid of appropriate size is constructed according to the geometric dimensions, and the grid density needs to be determined in combination with the actual hardware computing power.

[0067] S3: Based on the thermal simulation model, according to the actual operating range of the module, a series of typical operating conditions are selected, and 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 groups of module thermal field data under typical operating conditions;

[0068] In step S3, firstly, the actual operating condition range of the module to be estimated is determined. Then, a series of typical operating conditions (i.e., power loss P loss With the heat convection coefficient h), based on the temperature curve of the temperature sensor, under the premise of ensuring the reasonable spacing of each curve, the temperature curve cluster of the entire temperature sensor covers as large a range as possible (preferably the largest range), that is, to achieve full coverage of the actual possible working condition range. Then, according to the actual computing resources, a reasonable time step is selected to perform transient thermal simulation to obtain multiple groups of module thermal field data under typical working conditions.

[0069] S4: export the temperature data (and its time differential) of the temperature sensor position under different working conditions, the module temperature field training data and the time and space coordinates of the matching points respectively, perform data preprocessing according to input and output normalization, and divide the exported data into a training set and a validation set;

[0070] In step S4, three types of thermal simulation data are derived at the time point selected in S3, namely, the temperature data of the temperature sensor position and its time differential, the module temperature field, and the time-space coordinates of the matching points. In order to ensure that no new hyperparameters are introduced, only the matching points are selected in the non-heat source part;

[0071] The exported thermal simulation data contains three types of data: spatial coordinates, time coordinates, and temperature. The spatial coordinates are normalized to 0 using the same scaling scale, and then the time and temperature are scaled to ~[0,1] respectively. The specific value of the scaling scale needs to be determined in combination with the geometric dimensions and rated operating conditions of the module to be estimated.

[0072] The normalized data is divided into training set and validation set according to a certain ratio to complete the construction of the data set.

[0073] S5: Preliminarily determine the basic architecture and algorithm, write the training objective function of the model based on the parameters of each layer of packaging materials in S1, and establish a model based on artificial intelligence algorithm;

[0074] In step S5, the human intelligence algorithms that can be used to establish the model include but are not limited to various supervised learning methods, and the network structure may include fully connected neural networks, residual neural networks, convolutional neural networks, recurrent neural networks, and other derived new networks; the training objective function may include pure data loss function, physical information loss function, hybrid loss function and other similar forms;

[0075] The model consists of: an input layer, each input channel is They correspond to the spatial coordinates, time coordinates, and temperature of the temperature sensor (a total of M), Corresponding to the normalized spatial coordinates, Corresponding to the time coordinate, Corresponding to the normalized temperatures measured by the 1st, 2nd, ... Mth temperature sensors; several intermediate layers, which can adopt various network structures; a 1-channel output layer, corresponding to the module temperature field;

[0076] The training objective function of the model consists of two parts: training data error and physical law residual:

[0077] loss = w D MSE D +w F MSE F

[0078] Among them, the training data error MSE D is the mean square error between the model output temperature and the thermal simulation result temperature, and the physical law residual MSE F is the residual of the heat conduction partial differential equation of the training collocation point. D 、w F are the weights of the two items, which can be flexibly adjusted according to the actual data situation. If a pure data loss function is used, then w F = 0. Similarly, if the physical information loss function is used, then w D= 0. If a mixed loss function is used, it is necessary to determine a reasonable value in the subsequent debugging process to make the size of the two losses relatively balanced and ensure that the network converges stably and quickly.

[0079] The specific calculation method of the training data error term is:

[0080]

[0081] Among them, N D is the number of input training data, is the output temperature of the artificial intelligence network at the i-th point, is the actual temperature at the i-th point, which is derived from S4.

[0082] The physical law residual term is the mean square residual of the heat conduction partial differential equation calculated at the selected points (called collocation points) in the module. Its original form is:

[0083]

[0084] Among them, N F is the number of collocation points, and the space-time coordinates of the i-th collocation point are (x i ,y i ,z i ,t i ), the temperature is 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 collocation point is the training point selected for applying physical constraints. The main difference between the collocation point and the ordinary training data is that the collocation point only needs the space-time coordinate input, and does not require the actual temperature value, which is derived from S4;

[0085] Considering the data normalization preprocessing in S4, the calculation formula should be expressed in normalized variables. Substituting in, the calculation formula for the residual term of the physical law is:

[0086]

[0087] in, They are all the differentials of the output of the artificial intelligence network relative to the input, and are calculated using automatic differentiation during the training process. is the derivative value of the temperature of the temperature sensor with respect to time, which is directly derived after being calculated using the numerical differentiation method in S4;

[0088] In summary, both terms in the training objective function are in the form of mean squared error. When minimizing the loss function loss→0, there must be MSE D →0 and MSE F→0. The former indicates that the output temperature field of the model is accurate under the typical working conditions of the training data; the latter forces the output temperature field of the model to conform to the underlying thermal physics law (i.e., the partial differential equation of heat conduction), ensuring its strong generalization ability and full interpretability. Therefore, the model can fully guarantee the high accuracy and high resolution of thermal field reconstruction;

[0089] At this point, the model based on artificial intelligence algorithm has been established.

[0090] S6: Train the model and adjust its hyperparameters in turn according to the performance of the model on the training set and the validation set to ensure the output accuracy while making the model converge quickly and stably;

[0091] In step S6, the training set obtained in S4 is configured into a small batch of appropriate size, input into the model in S5, the training objective function is calculated by automatic differentiation, and a suitable optimization algorithm training model is selected (specifically, an optimization algorithm training estimation model corresponding to the selected artificial intelligence network is selected);

[0092] Based on the convergence of the model, debug the learning rate in the optimization algorithm to make the model converge stably and quickly; debug the structure of the model based on the fitting results of the model on the training set and the actual computing resources; comprehensively analyze the performance of the model on the training set and the validation set, ensure the regularization effect of the physical constraint terms without affecting the output accuracy, and debug the weight of the training objective function described in S5 to make the training set error within an acceptable range and the test set error as small as possible; complete the debugging of the model;

[0093] Finally, complete the training of the debugged model.

[0094] S7: Obtain a model including the temperature curve of the temperature sensor and the thermal field mapping relationship within the module working condition range. Input the temperature curve of the temperature sensor under any actual working condition to complete the module temperature distribution estimation under the corresponding conditions.

[0095] In step S7, the trained model realizes the mapping of the spatiotemporal coordinates and the temperature of the temperature sensor to the temperature field of the entire module, wherein the temperature of the temperature sensor indirectly reflects the operating conditions of the module, that is, the artificial intelligence thermal field reconstruction based on the temperature observation point is realized;

[0096] Input the temperature curve of the temperature sensor under any working condition, and the model outputs the module temperature field distribution under the corresponding working condition to complete the temperature estimation.

[0097] In practical applications, the spatial coordinates, time coordinate sequence and corresponding temperature sensor temperature sequence of any point in the module to be estimated can be input to output the temperature change sequence of the point under the current working condition; or the module spatial coordinate matrix and the corresponding temperature sensor temperature at a specific time point can be input to output the temperature field distribution of the entire module under the current working condition and at a specific time point;

[0098] At this point, the power semiconductor module temperature estimation of the decoupled operating condition parameters is completed.

[0099] Embodiment 2

[0100] by Figure 1 The present invention is further described in detail by taking the ROHM BSM250D17P2E004 multi-chip parallel SiC module as an example.

[0101] The method for estimating temperature distribution of a power semiconductor module based on artificial intelligence proposed in the present invention specifically comprises the following steps:

[0102] S1: Determine the layout geometry parameters and packaging material parameters of the module to be estimated, including: the size and position of the chip, the position of the temperature sensor, the size and thickness of each packaging layer, and the thermal physical property parameters of each layer of material;

[0103] S2: Based on the results obtained in S1, a thermal simulation model of the module is established in the simulation software, and a suitable density is selected to construct a three-dimensional space grid;

[0104] S3: According to the actual operating range of the module, a series of typical operating conditions are selected, a reasonable time step is determined, and thermal simulation is performed to obtain multiple groups of module thermal field data under typical operating conditions;

[0105] S4: export the temperature data (and its time differential) of the temperature sensor position under different working conditions, the module temperature field training data and the time and space coordinates of the matching points respectively, perform data preprocessing according to input and output normalization, and divide the exported data into a training set and a validation set;

[0106] S5: Preliminarily determine the basic architecture and algorithm, write the training objective function of the model based on the parameters of each layer of packaging materials in S1, and establish a model based on artificial intelligence algorithm;

[0107] S6: Train the model and adjust its hyperparameters in turn according to the performance of the model on the training set and the validation set to ensure the output accuracy while making the model converge quickly and stably;

[0108] S7: Obtain a model including the temperature curve of the temperature sensor and the thermal field mapping relationship within the module working condition range. Input the temperature curve of the temperature sensor under any actual working condition to complete the module temperature distribution estimation under the corresponding conditions.

[0109] In step S1, the module is the BSM250D17P2E004 SiC half-bridge module of ROHM, and the upper and lower bridge arms have 8 chips connected in parallel. The temperature sensor NTC is located near the boundary of the upper bridge arm. Figure 1 The module can be divided into 6 main layers from top to bottom, including chip layer, solder layer, DBC upper copper layer, DBC ceramic layer, DBC lower copper layer, and heat sink layer. The thicknesses are (0.38, 0.1, 0.3, 0.38, 0.3, 3) mm, corresponding to SiC, SAC305, Cu, and Al, respectively. 2 O 3 , Cu, Cu, a total of 4 packaging materials. SiC, SAC305, Cu, Al 2 O 3 The thermal conductivity is (450, 50, 400, 35) W / (m·K), the heat capacity at constant pressure is (1200, 237, 385, 790) J / (kg·K), and the density is (3200, 7345, 8960, 3965) kg / m 3 .

[0110] In step S2, according to the results obtained in S1, a geometric model of the module is established in COMSOL and the corresponding material thermal parameters are assigned to construct a three-dimensional grid of appropriate size, which contains a total of 44,909 spatial points.

[0111] In step S3, taking the upper bridge arm of the module as the heat source, considering the power cycle condition, the heating time t on =30s, cooling time t off =30s. Select the typical working condition with relatively low cooling water flow rate, corresponding to the typical heat convection coefficient h = 1500W / (m 2 ·K). For simplicity, only the simulation results under the single heat dissipation condition of h are used as training data. The proposed hybrid neural network model can cover the changes of heat dissipation conditions within a certain range. loss ={150, 250, 350, 450, 550}W five groups of typical working conditions are used as training data to ensure T NTC The (t) curves have a certain spacing, while making the curve cluster completely cover the possible power losses.

[0112] The ambient temperature is set to 20°C, and 10 heating-cooling cycles are simulated under each operating condition parameter until the module temperature field reaches a steady state. The temperature field results of 30 time points in t = [480s, 540s] are selected as training data, and the time step Δt = 2s. Finally, transient thermal simulation is performed to obtain module temperature field data under five typical operating conditions.

[0113] In step S4, the following three types of module thermal simulation data are derived at the time point selected in S3: 1) NTC temperature and its time differential, a total of 2×30×5=300. 2) Module temperature field data, in order to reduce the amount of calculation, is derived on the simplified grid (a total of 7724 points in the entire module), a total of 7724×30×5=1158600. 3) The spatiotemporal coordinates of the matching points, that is, the spatiotemporal coordinates of the grid points in the non-heat source area, a total of 4732×30×5=709800. Among them, in order to save the amount of calculation, only the training matching points are selected in the main layer (that is, the layer with relatively large thickness) in the module to be estimated. Considering the convenience of implementation and obtaining a better regularization effect, the spatiotemporal coordinates of the module temperature field data are selected as the training matching points. The simplified grid points (that is, the training matching points) are basically evenly distributed in the region, and the density is appropriately increased at the boundary. In addition, in order to achieve the purpose of decoupling the working condition parameters, only the grid points in the non-source area are selected to impose physical constraints, that is, the point coordinates of the chip layer are discarded. Otherwise, difficult-to-measure heat source terms will be introduced in the physical constraint residuals.

[0114] In the exported data, the spatial coordinates are normalized to 0 using the same scaling scale, and the time and temperature are scaled to ~[0,1] respectively:

[0115]

[0116] Among them, combined with the actual module parameters, the spatial scaling scale μ xyz =200, time scaling μ t =1 / 6, the highest and lowest reference temperatures T max =390K and T min =290K. At this point, the training set is complete.

[0117] Based on the training set working condition, the interpolation result P of power loss is selected loss = {200, 300, 400, 500} W, 20% extrapolation of the convection coefficient, i.e. (1 ± 0.2) h = {1200, 1800} W / (m 2 K) The corresponding thermal simulation data is used as the validation set.

[0118] In step S5, a hybrid neural network model is initially constructed in MATLAB's Deep Learning Toolbox and initialized using Xavier. The hybrid neural network model includes: an input layer, each input channel is They correspond to spatial coordinates, time coordinates and NTC temperature respectively; several intermediate layers, consisting of fully connected layers and tanh activation function layers, using the residual block form of jump connection; a 1-channel output layer, corresponding to the module temperature field The hybrid neural network model contains L residual blocks, so there are 2L+1 hidden layers, each of which contains N neurons. L and N are the network structure hyperparameters that need to be debugged to determine the optimal values.

[0119] The hybrid loss function of the hybrid neural network model consists of two parts: training data error and physical law residual:

[0120] loss = w D MSE D +w F MSE F

[0121] Among them, the training data error MSE D is the mean square error between the output temperature of the hybrid neural network and the actual temperature of the training point, and the physical law residual MSE F is the residual of the heat conduction partial differential equation of the training collocation point. D 、w F They are the weights of the two items, and reasonable values ​​need to be determined in the subsequent debugging process to make the size of the two losses relatively balanced and ensure the stable and rapid convergence of the network. Initially, they are both taken as 1.

[0122] The specific calculation method of the training data error term is:

[0123]

[0124] Among them, the number of input training data N D =1158600, is the output temperature of HNNM at the i-th training point, is the actual temperature at the i-th training point, which is derived from S4.

[0125] The physical law residual term is the mean square residual of the heat conduction partial differential equation calculated at the selected points (called collocation points) in the module. Its original form is:

[0126]

[0127] Among them, the number of collocation points N F =709800, the space-time coordinates of the i-th matching point are (x i ,y i ,z i ,t i ), the temperature is 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 collocation point is the training point selected for applying physical constraints. The main difference between the collocation point and the ordinary training data is that the collocation point only needs the space-time coordinate input, and does not require the actual temperature value, which is derived from S4.

[0128] Considering the data normalization preprocessing in S4, the calculation formula should be expressed in normalized variables. For the spatial differential term (taking the x coordinate as an example, the rest of the forms are the same), the chain rule is applied and the input and output normalization is:

[0129]

[0130] For the time derivative term, note that the input variable It is also a function of time. Using the chain rule of multivariate functions and normalizing the input and output, we have:

[0131]

[0132] Substituting into the original calculation formula, the calculation formula for the residual term of the physical law is obtained as follows:

[0133]

[0134] in, They are all differentials of the output of the hybrid neural network with respect to the input, and are calculated using automatic differentiation during the training process. is the derivative value of NTC temperature with respect to time, which is calculated and derived in S4 using the numerical differentiation method. It should also be noted that this numerical differentiation term is only used for training the network and is not required for actual estimation, so it has no effect on the convenience of application.

[0135] In summary, both terms in the total loss function are in the form of mean squared error. When minimizing the loss function loss→0, there must be MSE D →0 and MSE F →0. The former shows that the output temperature field of the hybrid neural network model is accurate under the typical working conditions of the training data; the latter forces the output temperature field of the hybrid neural network model to conform to the underlying thermal physics law (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 by the hybrid neural network model can be fully guaranteed.

[0136] In step S6, the training set obtained in S4 is configured into 10 small batches and input into the hybrid neural network model in S5. In the Deep Learning Toolbox of MATLAB, the hybrid loss function is calculated using automatic differentiation and the Adam optimization algorithm is used for training.

[0137] The learning rate α directly affects the convergence of the network, and the strategy of gradually decaying the learning rate with the number of iterations epoch is adopted:

[0138]

[0139] Among them, the initial learning rate α 0 =1e-2. The decayRate is initially set to a larger value of 0.1 to ensure stable convergence of the NN. To speed up convergence, the decayRate is gradually reduced, and convergence is accelerated as much as possible while ensuring basic stability. Finally, the decayRate is set to 0.005.

[0140] The number of hidden layers and the number of neurons in each layer characterize the network's fitting ability. The optimal network structure needs to be determined based on actual computing resources. The output performance of the network on the training set is mainly referenced, and the value of the loss function after the same training time is used to intuitively measure it, and the loss values ​​after the same training time under different structures are compared. Taking into account the training time and output accuracy, (L=4,N=128) is selected as the more optimized hybrid neural network structure.

[0141] The weights in the loss function need to be debugged to balance the relative sizes of different items. The performance of the hybrid neural network on the training set and the validation set is comprehensively analyzed to ensure the regularization effect of the physical constraint terms without affecting the output accuracy. The appropriate loss function weight is determined as w F =0.3, w D = 1. In addition, according to the characteristics of the specific problem, the weight of the training data of the focus part can be further fine-tuned to achieve a better fitting effect. Therefore, we choose w D_high =1.2, that is, a greater weight is given 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 the error in the high temperature zone during the training process, improving the accuracy of the estimation. In addition, it should be pointed out that since the loss function contains different weights and physical law residual terms, its size can only be used to qualitatively judge the degree of closeness between the current network output result and the true solution, and cannot be directly used as a measure of the result error.

[0142] Finally, after the debugged hybrid neural network was trained for epoch=8000, the loss of ~1.3867e-4 tended to be stable, and it was considered that the training of the hybrid neural network model was completed.

[0143] In step S7, the trained hybrid neural network model realizes the mapping of time-space coordinates and NTC temperature to the entire module temperature field, where NTC temperature indirectly reflects the operating conditions of the module, that is, the thermal field reconstruction based on NTC temperature is realized. Input the NTC temperature curve under any working condition, and the hybrid neural network model outputs the module temperature field distribution under the corresponding working condition to complete the temperature estimation.

[0144] To verify the estimated resolution, the training set and the validation set P loss Take the maximum as an example, Figure 2The upper surface of DBC and T under different working conditions at t = 30s 7 The hybrid neural network results, thermal simulation results and relative errors of the chip surface temperature field. Linear interpolation is performed on the discontinuities of the copper layer on the DBC when drawing. Figure 2 It can be seen that the temperature distribution of the two is basically the same, indicating that the hybrid neural network containing the underlying physical law constraints can well reproduce the module temperature field distribution, and the output result is fully reasonable. The relative error distribution shows that the relative error of most areas except the chip gap is ≤5%, and the relative error of the chip surface temperature field fitting is also ≤5%. It is verified that the thermal field reconstruction method based on the hybrid neural network proposed in the present invention has extremely high resolution.

[0145] To verify the accuracy of the estimation, power cycling verification was carried out under a series of actual operating conditions, and an infrared camera was used to accurately measure the chip junction temperature.

[0146] The chip coordinates and the experimentally measured NTC temperature curve are input into the trained hybrid neural network model to obtain the junction temperature curve under the corresponding working conditions. j 7 For example, Figure 3 The relative errors between the estimation results of the hybrid neural network model and the junction temperature measurement results under different working conditions are all less than 3%. The proposed method decouples a wide range of power loss and a certain range of heat dissipation conditions through the NTC temperature curve, and has a high estimation accuracy within the corresponding working condition range.

[0147] Embodiment 3

[0148] The present invention provides a power semiconductor module temperature distribution estimation system based on artificial intelligence, comprising:

[0149] Parameter determination module: used to determine the layout geometric parameters and packaging material parameters of the module to be estimated, including: the size and position of the chip, the position of the temperature sensor, the size and thickness of each packaging layer, and the thermal physical property parameters of each layer of material;

[0150] Thermal simulation model building module: used to determine the parameters obtained in the module according to the parameters, establish the thermal simulation model of the module to be estimated in the simulation software, and select the grid size according to the minimum unit side length of each part of the module to construct a three-dimensional space grid;

[0151] Thermal simulation module: used in the thermal simulation model to select a series of typical operating conditions according to the actual operating range of the module to be estimated, determine the time step according to the actual hardware computing power, perform thermal simulation, and obtain multiple sets of thermal field data of the module to be estimated under typical operating conditions;

[0152] Data export module: It is used to export the temperature data and its time differential of the temperature sensor position under different working conditions, the training data of the temperature field of the module to be estimated, and the time and space coordinates of the matching points from the results obtained by the thermal simulation module, perform data preprocessing according to the input and output normalization, and divide the exported data into training set and verification set;

[0153] Estimation model building module: used to preliminarily determine the network structure and artificial intelligence algorithm, write the training objective function of the estimation model in combination with the packaging material parameters, and establish an estimation model based on the artificial intelligence algorithm;

[0154] Training module: used to train the estimation model. According to the performance of the estimation model on the training set and the validation set, its hyperparameters are adjusted in turn to ensure the output accuracy while making the estimation model converge quickly and stably, so as to obtain a trained estimation model containing the temperature curve of the temperature sensor and the thermal field mapping relationship within the working range of the module to be estimated;

[0155] Estimation module: It is used to input the temperature curve of the temperature sensor under any actual working conditions based on the trained estimation model, and complete the temperature distribution estimation of the module to be estimated under the corresponding conditions.

[0156] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented 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.

[0157] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0158] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0159] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0160] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit its protection scope. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that after reading the present invention, those skilled in the art can still make various changes, modifications or equivalent substitutions to the specific implementation methods of the invention, but these changes, modifications or equivalent substitutions are all within the protection scope 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 in that: The steps include: S1: Determine the layout geometric parameters and packaging material parameters of the module to be estimated, including: the size and position of the chip, the position of the temperature sensor, the size and thickness of each packaging layer, and the thermal physical property parameters of each layer material; S2: Based on the parameters obtained in S1, a thermal simulation model of the module to be estimated is established in the simulation software, and a three-dimensional space grid is constructed by selecting a grid size based on the minimum unit side length of each part of the module; 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, and the time step is determined according to the actual hardware computing capacity to perform thermal simulation to obtain multiple groups of thermal field data of the module to be estimated under typical operating conditions; S4: from the results obtained in S3, the temperature data and its time differential of the temperature sensor position under different working conditions, the training data of the temperature field of the module to be estimated, and the time and space coordinates of the matching points are respectively derived, and the data is preprocessed according to the input and output normalization, and the derived data is divided into a training set and a validation set; S5: Preliminarily determine the network structure and artificial intelligence algorithm, write the training objective function of the estimation model in combination with the packaging material parameters in S1, and establish an estimation model based on the artificial intelligence algorithm; S6: training the estimation model, and adjusting its hyperparameters in turn according to the performance of the estimation model on the training set and the validation set, ensuring the output accuracy while making the estimation model converge quickly and stably, and obtaining a trained estimation model including the mapping relationship between the temperature curve of the temperature sensor and the thermal field within the working condition range of the module to be estimated; S7: Based on the estimation model trained in S6, the temperature curve of the temperature sensor under any actual working condition is input to complete the estimation of the temperature distribution of the module to be estimated under the corresponding conditions.

2. The method for estimating temperature distribution of a power semiconductor module based on artificial intelligence according to claim 1, characterized in that: 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, which are respectively located on the surface or inside the module package, and the temperature measured by the temperature sensor is used as the observation point, and the temperature is respectively recorded as {T obs1 (t),T obs2 (t),...,T obsM (t)}, respectively representing the temperatures measured by the 1st, 2nd, ..., Mth temperature sensors.

3. The method for estimating temperature distribution of a power semiconductor module based on artificial intelligence according to claim 1, characterized in that: In step S2, in the thermal simulation software, firstly, a geometric model of the module to be estimated is established and the corresponding material thermal parameters are assigned; secondly, a three-dimensional grid is constructed by selecting a grid size according to the minimum unit side length of each part of the module, and 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 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 diagram of the temperature sensor, the temperature curve cluster of the entire temperature sensor covers the largest range while ensuring reasonable spacing between the curves, that is, 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 multiple groups of thermal field data of the module to be estimated under typical operating conditions.

5. The method for estimating 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 derived at the time point corresponding to the time step selected in S3, namely, the temperature data of the temperature sensor position and the numerical calculation result of its time differential, the temperature field of the module to be estimated, and the time-space coordinates of the matching points; wherein the matching points are selected only in the non-heat source part, that is, the area outside the chip layer; The exported thermal simulation data contains three types of data: spatial coordinates, time coordinates, and temperature. The spatial coordinates are normalized to 0 as the center using the same scaling scale, and then the time and temperature are scaled to approximately [0,1] range. The specific value of the scaling scale is determined in combination with the geometric dimensions and rated operating conditions of the module to be estimated. The normalized data is divided into a training set and a validation set according to a preset ratio.

6. The method for estimating 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; The network structure includes fully connected neural network, residual neural network, convolutional neural network, recurrent neural network and other derived new networks; The training objective function includes pure data loss function, physical information loss function or mixed loss function; 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 the normalized spatial coordinates, Corresponding to the time coordinate, Corresponding to the normalized temperatures measured by the 1st, 2nd, ... Mth temperature sensors; The plurality of intermediate layers can adopt various network structures; The output layer corresponds to the module temperature field.

7. The method for estimating 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 residual; loss=w D MSE D +w F MSE F Among them, the training data error MSE D is the mean square error between the model output temperature and the thermal simulation result temperature, and the physical law residual MSE F is the residual of the heat conduction partial differential equation of the training collocation point, w D 、w F are the weights of the two items respectively; The specific calculation method of the training data error term is: Among them, N D is the number of input training data, is the output temperature of the artificial intelligence network at the i-th point, is the actual temperature at the i-th point, all derived from S4; The physical law residual term is the mean square residual of the heat conduction partial differential equation calculated at the matching points in the module. Its original form is: Among them, N F is the number of collocation points, and the space-time coordinates of the i-th collocation point are (x i ,y i ,z i ,t i ), the temperature is T i , k, ρ, c are the thermal conductivity, density, and constant-pressure heat capacity of the material in the region where the collocation point is located, respectively, which are determined in S1. The collocation point is the training point selected for applying physical constraints. The main difference between the collocation point and the ordinary training data is that the collocation point only needs the space-time coordinate input, and does not need the actual temperature value, which is derived from S4; Considering the data normalization preprocessing in S4, the calculation formula is expressed by the normalized variables and the calculation formula of the residual term of the physical law is obtained as follows: in, They are all the differentials of the output of the artificial intelligence network to the input, and are calculated using automatic differentiation during the training process. is the derivative value of the temperature of the temperature sensor with respect to time, which is directly derived after being calculated using the numerical differentiation method in S4.

8. The method for estimating 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 by 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, debug the learning rate in the optimization algorithm to make the estimation model converge stably and quickly; debug the structure of the estimation model based on the fitting results of the estimation model on the training set and the actual computing resources; comprehensively analyze the performance of the estimation model on the training set and the validation set, ensure the regularization effect of the physical constraint terms without affecting the output accuracy, and debug the weight of the training objective function described in S5 to make the training set error within an acceptable range and the test set error as small as possible, so as to obtain a trained estimation model.

9. The method for estimating temperature distribution of a power semiconductor module based on artificial intelligence according to claim 1, characterized in that: In step S7, the spatial coordinates, time coordinate sequence and corresponding temperature sensor temperature sequence of any point in the module to be estimated are input into the trained estimation model, that is, the temperature change sequence of the 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 the specific time point is output.

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 geometric parameters and packaging material parameters of the module to be estimated, including: the size and position of the chip, the position of the temperature sensor, the size and thickness of each packaging layer, and the thermal physical property parameters of each layer of material; Thermal simulation model building module: used to determine the parameters obtained in the module according to the parameters, establish the thermal simulation model of the module to be estimated in the simulation software, and select the grid size according to the minimum unit side length of each part of the module to construct a three-dimensional space grid; Thermal simulation module: used in the thermal simulation model to select a series of typical operating conditions according to the actual operating range of the module to be estimated, determine the time step according to the actual hardware computing power, perform thermal simulation, and obtain multiple sets of thermal field data of the module to be estimated under typical operating conditions; Data export module: It is used to export the temperature data and its time differential of the temperature sensor position under different working conditions, the training data of the temperature field of the module to be estimated, and the time and space coordinates of the matching points from the results obtained by the thermal simulation module, perform data preprocessing according to the input and output normalization, and divide the exported data into training set and verification set; Estimation model building module: used to preliminarily determine the network structure and artificial intelligence algorithm, write the training objective function of the estimation model in combination with the packaging material parameters, and establish an estimation model based on the artificial intelligence algorithm; Training module: used to train the estimation model. According to the performance of the estimation model on the training set and the validation set, its hyperparameters are adjusted in turn to ensure the output accuracy while making the estimation model converge quickly and stably, so as to obtain a trained estimation model containing the temperature curve of the temperature sensor and the thermal field mapping relationship within the working range of the module to be estimated; Estimation module: It is used to input the temperature curve of the temperature sensor under any actual working conditions based on the trained estimation model, and complete the temperature distribution estimation of the module to be estimated under the corresponding conditions.

Images