IGBT module state evaluation method based on substrate shell temperature distribution and neural network model
By selecting key shell temperature distribution feature parameters and using a neural network model trained with Bayesian optimization, the problem of feature parameter redundancy in IGBT module state assessment was solved, achieving efficient and accurate health status assessment.
Patent Information
- Application Number
- CN202511050965.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-29
- Publication Date
- 2025-11-11
AI Technical Summary
Existing IGBT module state assessment methods suffer from redundant feature parameters, resulting in high data acquisition and model training costs, high model complexity, and impacting online integration and real-time assessment.
By collecting substrate shell temperature distribution data, performing normality tests and correlation analyses, key feature parameters are selected, and a neural network model is trained using Bayesian optimization methods to achieve non-invasive IGBT module state assessment.
This improves the accuracy and precision of IGBT module health status assessment, reduces data acquisition costs and model complexity, and enhances the reliability of assessment results.
Smart Images
Figure CN120928142A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power device aging condition assessment technology, specifically to an IGBT module condition assessment method based on substrate case temperature distribution and neural network model. Background Technology
[0002] Insulated-gate bipolar transistors (IGBTs) experience package fatigue damage during long-term operation, leading to increased thermal resistance and junction temperature, ultimately affecting the safe operation of the power converter and even the entire system. Evaluating the health parameters of IGBT devices, such as package thermal resistance aging, is crucial for ensuring the safe and reliable operation of the system.
[0003] Data-driven modeling methods based on artificial neural networks offer numerous advantages, including simple modeling, high model accuracy, and stability. They can fit complex nonlinear mappings between external aging characteristics and internal health information of devices without requiring additional calibration of the coupling relationships between parameters. However, these methods suffer from feature parameter redundancy. Furthermore, in existing data-driven modeling methods, the non-intrusiveness of feature parameters is a critical requirement to ensure device packaging integrity and operational safety. However, to achieve a comprehensive characterization of system behavior, current research often uses a large number of feature parameters as model inputs without considering feature parameter redundancy. This not only increases the cost of data acquisition and model training but also raises model complexity and computational requirements, impacting online integration and real-time evaluation. Summary of the Invention
[0004] To address the aforementioned shortcomings in existing technologies, this invention provides an IGBT module state assessment method based on substrate shell temperature distribution and a neural network model. By using correlation analysis, the redundancy of input feature parameters is determined, thus overcoming the parameter redundancy defects of existing methods. This enables the assessment of IGBT module solder aging, chip loss, and junction temperature levels under thermal steady-state conditions using non-invasive sampling.
[0005] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: The IGBT module state assessment method based on substrate shell temperature distribution and neural network model includes the following steps: S1. Collect chip loss and junction temperature, device solder aging level, case temperature distribution and ambient temperature data of insulated gate bipolar transistor under different operating conditions. Through data preprocessing, generate preprocessed chip loss and junction temperature, device solder aging level, case temperature distribution and ambient temperature data. S2. Perform normality tests and correlation analyses on the preprocessed shell temperature distribution data to screen characteristic parameters of shell temperature distribution; S3. The shell temperature distribution characteristic parameters and the pre-processed ambient temperature are used as raw data. The pre-processed chip loss, chip junction temperature and solder aging level data are used as labeled data. The labeled raw data are input into the artificial neural network model. The Bayesian optimization method is used to optimize the hyperparameters of the artificial neural network model to generate a trained artificial neural network model. S4. Obtain the case temperature distribution and ambient temperature of the insulated gate bipolar transistor under different operating conditions, and input them into the trained artificial neural network model for state prediction to obtain the health status assessment results of the chip loss, junction temperature, and solder aging level of the insulated gate bipolar transistor under test.
[0006] The present invention has the following beneficial effects: The IGBT module status assessment method proposed in this invention, based on substrate case temperature distribution and neural network model, achieves health status assessment of insulated gate bipolar transistors (IGBTs) based on artificial neural networks through non-invasive sampling design and feature parameter correlation analysis. This not only improves the accuracy of the artificial neural network in assessing the health status of IGBTs, but also improves the accuracy of assessing solder aging level, chip loss, and chip junction temperature level of IGBTs. Attached Figure Description
[0007] Figure 1 This is a flowchart illustrating the IGBT module state evaluation method based on substrate shell temperature distribution and neural network model proposed in this invention. Figure 2 This is a schematic diagram showing the different layer materials and thermal expansion coefficients of the IGBT module in the embodiment; Figure 3 This is a schematic diagram of heat transfer in the IGBT module after solder aging, as shown in the example. Figure 4 This is a schematic diagram of the artificial neural network model in the embodiment; Figure 5 This is a structural diagram of the finite element thermal simulation model of the IGBT module in the embodiment; Figure 6 This is a schematic diagram of the shell temperature data acquisition location in the finite element thermal simulation model of the IGBT module in the embodiment; Figure 7 This is a schematic diagram of the half-bridge inverter experimental circuit structure in the embodiment; Figure 8 This is a schematic diagram of the half-bridge inverter experimental platform in the embodiment; Figure 9 This is a graph showing the solder aging level evaluation results in the embodiment; Figure 10 This is a schematic diagram of the IGBT module chip loss assessment results in the embodiment; Figure 11This is a schematic diagram of the diode chip loss evaluation results in the embodiment; Figure 12 This is a schematic diagram showing the junction temperature evaluation results of the IGBT module chip in the embodiment; Figure 13 This is a schematic diagram showing the diode chip junction temperature evaluation results in the embodiment; Detailed Implementation The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.
[0008] like Figure 1 As shown, the IGBT module state assessment method based on substrate shell temperature distribution and neural network model includes the following steps S1-S4: S1. Collect chip loss and junction temperature, device solder aging level, case temperature distribution and ambient temperature data of insulated gate bipolar transistor under different operating conditions. Through data preprocessing, generate preprocessed chip loss and junction temperature, device solder aging level, case temperature distribution and ambient temperature data.
[0009] In this embodiment, the purpose of using chip loss, junction temperature, and solder aging level of the insulated-gate bipolar transistor (IGBT) under different operating conditions is to: use these data as label data for the artificial neural network model, while using case temperature distribution and ambient temperature data as raw data. The labeled raw data is then input into the artificial neural network model for training, so that any subsequent input of case temperature distribution and ambient temperature data of the IGBT under test can predict the health status of the IGBT, i.e., obtain the chip loss, junction temperature, and solder aging level (grade) status of the IGBT. A detailed analysis follows: First, solder aging failure in Insulated Gate Bipolar Transistors (IGBTs) primarily manifests as solder porosity and cracks caused by thermal stress, leading to increased device thermal resistance. Unlike sudden chip-level failures, solder aging exhibits a gradual characteristic, with its aging level showing a clear positive correlation with the number of power cycles. During module operation, the conduction and switching losses generated by the semiconductor chip are converted into heat energy, subjecting the solder and other packaging structures to periodic temperature fluctuations and thermomechanical stresses, resulting in gradual structural aging failure. The typical materials and coefficient of thermal expansion (CTE) of each layer in an IGBT module are as follows: Figure 2As shown, the different coefficients of thermal expansion lead to varying degrees of thermal expansion in different layers of the IGBT module under thermal stress. Materials with higher coefficients of thermal expansion experience more significant volume expansion when the temperature changes, resulting in shear stress at the interfaces between different materials. This periodic stress can trigger various failure mechanisms, including solder failure, such as voids or cracks in the solder layer, breakage or detachment of bond wires, and chip warping and cracking. Among these, solder interface degradation is one of the main causes of thermomechanical failure. Gradual degradation of the solder structure leads to a gradual increase in the thermal resistance of the module's packaging structure, a continuous decline in the module's heat dissipation performance, and a gradual increase in the chip junction temperature, ultimately causing functional failure of the device.
[0010] Secondly, under certain heat dissipation conditions for IGBT modules, the case temperature distribution on the module's baseplate is affected by the solder aging level and the chip loss distribution within the module. Solder aging leads to the contraction of the effective heat transfer path inside the module, causing a change in the case temperature gradient. Specifically, the two forms of solder aging in IGBT modules are as follows: Figure 3 As shown, This refers to the casing temperature located directly below the chip (at the center of the chip). , These represent the two case temperatures at the edge of the IGBT module. Solder degradation in IGBT modules includes chip solder aging and substrate solder aging. Both degradation modes lead to heat flow channel concentration and exhibit the effects described above. Therefore, the case temperature distribution on the module substrate can reflect information about solder aging.
[0011] Furthermore, under constant solder aging levels and heat dissipation conditions, changes in chip losses can also alter the case temperature distribution. However, since the internal heat transfer path of the module remains constant, the overall case temperature variation pattern is consistent, and the mechanism by which this factor affects the case temperature distribution differs from that of aging. Therefore, the case temperature distribution information also includes the loss and junction temperature information of the IGBT and diode chips. It is worth noting that, because the neural network model is a data-driven model, the coupling relationship between junction temperature and loss is already reflected in the data, requiring no additional calibration.
[0012] Therefore, the case temperature distribution data contains information on the solder aging level, chip loss, and chip junction temperature of the IGBT module, which can be used as feature parameter inputs for the artificial neural network model (ANN). Furthermore, to calibrate the heat dissipation level information, ambient temperature also needs to be used as a feature parameter input.
[0013] Specifically, the case temperature distribution data includes case temperature data at different locations, including the chip center and the edge of the insulated gate bipolar transistor.
[0014] Specifically, data preprocessing includes outlier removal and data normalization.
[0015] In this embodiment, the purpose of removing outliers is to ensure data quality, reduce the risk of gradient explosion, oscillation, and overfitting during the training process of artificial neural network models, and improve training efficiency and quality. The purpose of normalization is to keep the data input to the artificial neural network within the range of [0, 1], unify the feature scale, and enable the gradient descent optimization method to find the optimal solution quickly and stably, thereby improving training efficiency.
[0016] S2. Perform normality tests and correlation analyses on the preprocessed shell temperature distribution data to screen characteristic parameters of shell temperature distribution.
[0017] In this embodiment, to reduce the dimensionality of the shell temperature data input and thus decrease sampling costs and the complexity of the artificial neural network model, a correlation analysis of the shell temperature data is performed using statistical methods. First, a normality test is conducted on each data point to select a suitable correlation analysis method. The Kolmokolov-Smilov test (KS test) is used to test the normality of the data. The statistical parameters of the KS test result reflect the maximum difference between the sample distribution and the normal distribution. The significance directly reflects whether the data is normally distributed. When the normality significance is less than 0.05, the data does not meet the normal distribution; otherwise, it does. However, the operating condition data of power devices often do not meet the normal distribution; therefore, the Spearman correlation coefficient is selected for correlation calculation. The specific operation process is as follows: Specifically, step S2 includes S21-S24: S21. Use the Kolmokolov-Smilov test to perform a normal distribution test on the preprocessed shell temperature distribution data to determine whether the preprocessed shell temperature distribution data meets the normal distribution. If yes, proceed to step S22; otherwise, proceed to step S23.
[0018] In this embodiment, the purpose of determining whether the preprocessed shell temperature distribution data meets the normal distribution is to select appropriate coefficients for subsequent correlation coefficient calculations, thereby improving the accuracy and reliability of the correlation calculation results.
[0019] S22. Based on the preprocessed shell temperature distribution data, calculate the Pearson correlation coefficient of the shell temperature data under different operating conditions at any two locations, i.e.:
[0020] in, This represents the Pearson correlation coefficient of shell temperature data under different operating conditions at any two locations. , Representing positions respectively ,Location In working condition Below the shell temperature data, Indicates position The average shell temperature data under all operating conditions Indicates position The average shell temperature data under all operating conditions.
[0021] S23. Based on the preprocessed shell temperature distribution data, sort the shell temperature data at each location under different operating conditions from smallest to largest to obtain the rank of the shell temperature data at each location under different operating conditions. Calculate the Spearman correlation coefficient between the shell temperature data at any two locations under different operating conditions, i.e.:
[0022]
[0023] in, This represents the Spearman correlation coefficient of shell temperature data under different operating conditions at any two locations. This indicates that any two positions are in the working condition. The difference in rank of the lower shell temperature data. Indicates the total number of operating conditions. Indicates rank or order.
[0024] S24. Sort the Pearson correlation coefficient or Spearman correlation coefficient of the case temperature data under different operating conditions at any two locations from smallest to largest. From the case temperature data under different operating conditions at any two locations corresponding to the smallest Pearson correlation coefficient or the smallest Spearman correlation coefficient, select the case temperature data that meets the condition of including the chip center position and the edge position of at least one insulated gate bipolar transistor, and use it as the case temperature distribution characteristic parameter.
[0025] In this embodiment, based on the physical location characteristics of each shell temperature point, at least two shell temperature points must be included to encompass shell temperature distribution information. Shell temperature distribution can be represented by shell temperature gradients; that is, the shell temperature at the center point is necessary, and the number of shell temperatures at the edge positions must be greater than or equal to one to reflect the lateral and longitudinal shell temperature gradients. Therefore, the above-mentioned constraints must be met when selecting shell temperature distribution characteristic parameters.
[0026] S3. The case temperature distribution characteristic parameters and the pre-processed ambient temperature are used as raw data. The pre-processed chip loss, chip junction temperature and solder aging level data are used as labeled data. The labeled raw data are input into the artificial neural network model. The Bayesian optimization method is used to optimize the hyperparameters of the artificial neural network model to generate a trained artificial neural network model.
[0027] In this embodiment, since the artificial neural network model has a strong nonlinear mapping fitting capability, the present invention selects the artificial neural network model for IGBT module state recognition. Its typical operation structure is as follows: Figure 4As shown, it includes an input layer, a hidden layer, and an output layer. The input layer is responsible for receiving feature parameter data, the hidden layer is responsible for performing weight matrix operations and nonlinear activation operations, and the output layer is responsible for outputting the operation results. Figure 4 Each arrow represents a data transmission relationship, and each arrow contains a weight. Each circle is equivalent to a neuron, capable of storing the results of previous stage data processing and passing data to subsequent stages. In the latter two stages, there exists an activation function that enables the neural network to perform non-linear fitting. Additionally, each hidden layer neuron has a bias. For a well-trained artificial neural network model, its calculation formula is:
[0028]
[0029] in, This represents the output of the input layer of an artificial neural network model, and its length is... , This represents the output of the hidden layer of an artificial neural network model, and its length is... , This represents the output of the output layer of the artificial neural network model, and its length is... , This represents the activation function of the hidden layer. This represents the activation function of the output layer. This represents the weight matrix of the hidden layer, with dimension 1. , This represents the bias matrix of the hidden layer, with a length of . , This represents the weight matrix of the output layer, with dimension 1. , This represents the bias matrix of the output layer, with a length of . , This indicates transpose.
[0030] In this embodiment, the activation function from the input layer to the hidden layer is the hyperbolic tangent function, which provides a strong nonlinear fitting capability for model training. The activation function from the hidden layer to the output layer is a linear activation function, which facilitates the output of aging levels.
[0031] Furthermore, in the training settings of artificial neural network models, hyperparameters such as the initial learning rate, batch size, number of hidden layers, and number of neurons in the hidden layers often need to be determined empirically, which involves considerable randomness and chance, leading to significant differences in model performance under different parameter combinations. Therefore, to ensure the effectiveness of network training, this invention proposes to use a Bayesian optimization method to assist in the training process to achieve global optimization in the hyperparameter space, thereby obtaining the optimal hyperparameter combination and thus obtaining a well-trained artificial neural network model with the optimal hyperparameter combination; the specific operation process is as follows: Specifically, step S3 includes S31-S38: S31. Use the case temperature distribution characteristic parameters and ambient temperature as raw data, and use chip loss, chip junction temperature and solder aging level data as labeled data. Divide the labeled raw data into training set and validation set according to the proportion.
[0032] S32. Set the hyperparameters to be optimized and the hyperparameter space, and set the objective function to evaluate the performance of the artificial neural network model. The input of the objective function is the combination of hyperparameters, and the output is the mean squared error under the validation set.
[0033] Specifically, the hyperparameters to be optimized include learning rate, number of neurons, and batch size.
[0034] In this embodiment, the hyperparameter space refers to the range of values for the hyperparameters to be optimized.
[0035] S33. Randomly sample the hyperparameters to be optimized to generate several sets of hyperparameter combinations. For each set of hyperparameters, input the training set into the artificial neural network model for training. After training, input the validation set into the artificial neural network model for validation, generate the mean squared error of the validation set under all hyperparameters, and use it as the current dataset.
[0036] S34. Fit a Gaussian process regression model using the current dataset, and obtain the predicted mean and variance under each hyperparameter combination by establishing the probability distribution of the objective function.
[0037] S35. Based on the predicted mean and variance of each hyperparameter combination, a uniform grid is generated in the hyperparameter space. Each grid point corresponds to a set of hyperparameter combinations. The acquisition function value of each grid point is calculated. The grid point corresponding to the maximum acquisition function value is selected as the updated hyperparameter combination.
[0038] Specifically, the formula for calculating the acquisition function value of each grid point is as follows:
[0039]
[0040] in, Represents grid points The value of the acquisition function, This represents the optimal value of the current objective function. Represents grid points The predicted mean, Indicates the exploration coefficient. Indicates intermediate variables. The cumulative distribution function represents the standard normal distribution. Represents grid points The predicted variance is below. This represents the probability density function of the standard normal distribution.
[0041] S36. Train the artificial neural network model under the updated hyperparameter combination. After training, input the validation set into the artificial neural network model for validation, generate the mean squared error of the validation set under the updated hyperparameter combination, and add it to the current dataset to generate the updated dataset.
[0042] S37. Determine whether the current iteration number is greater than or equal to the maximum iteration number or whether the mean square error of the validation set for N consecutive iterations has not decreased. If so, proceed to step S38. Otherwise, use the updated dataset to fit and update the Gaussian process regression model. By establishing the probability distribution of the objective function, obtain the predicted mean and variance under each hyperparameter combination, and proceed to step S35.
[0043] In this embodiment, the stopping condition for the optimization process is the maximum number of iterations, and the maximum number of iterations is 1000; N is 100.
[0044] S38. The combination of hyperparameters corresponding to the minimum mean square error in the updated dataset is taken as the optimal combination of hyperparameters, and the artificial neural network model trained using the optimal combination of hyperparameters is taken as the optimal trained artificial neural network model.
[0045] S4. Obtain the case temperature distribution and ambient temperature of the insulated gate bipolar transistor under different operating conditions, and input them into the trained artificial neural network model for state prediction to obtain the health status assessment results of the chip loss, junction temperature, and solder aging level of the insulated gate bipolar transistor under test.
[0046] In this embodiment, to verify the effectiveness of the IGBT module state evaluation method based on substrate shell temperature distribution and neural network model proposed in this invention, the following experiments were conducted: First, build such Figure 5 The finite element thermal simulation model of the IGBT module shown includes a single FF50R12RT4 type IGBT half-bridge module and an aluminum heat sink, and the case temperature data sampling locations are as follows. Figure 6 As shown, this includes a chip center location case temperature. and the case temperature at the edges of the four devices , , , The correlation coefficient is calculated using the Spearman correlation coefficient calculation formula proposed in this invention; Table 1 shows the correlation calculation results of the sampled shell temperature, namely: Table 1. Correlation calculation results of shell temperature after sampling
[0047] As can be seen from Table 1, for the shell temperature at several locations, the center location... Because it can reflect solder degradation as a whole, its correlation with other shell temperature points is relatively low; however, for shell temperature sampling points at the edges, and The correlation was 0.999. and The correlation of 1.000 indicates a high degree of overlap between the corresponding parameters. This is because the response of edge locations to solder degradation is relatively similar under the current aging settings. And because... Located below the diode chip, it can better reflect the chip's loss distribution, therefore its correlation with other data is relatively low. Taking all factors into consideration... , and The correlation with other shell temperatures was selected based on the degree of correlation. This reflects the information in this set of data. In summary, based on... , and The case temperature data at the location is used as the input parameter for case temperature distribution characteristics, which reduces the dimensionality of the model input and lowers the cost of actual temperature sampling. In addition, since the simulation verification model used in this embodiment is a 4-chip IGBT half-bridge module, the thermal behavior complexity of the model is limited. For IGBT modules with more chips, the solder degradation mode is more complex, so it is necessary to increase the number of case temperature sampling points to effectively reflect solder aging.
[0048] Secondly, adopting such Figure 7The half-bridge inverter circuit shown was used to collect experimental data under different operating conditions. By controlling the DC-side voltage and AC load, different loss and junction temperature distributions appeared on the IGBT and diode chips, resulting in a corresponding case temperature distribution on the IGBT module casing. Junction temperature was acquired using a high-resolution SC7000 FLIR infrared thermal imager. Loss was obtained using a lookup table method. A device loss lookup table was constructed using the device's output characteristic curves and switching energy curves from the datasheet. The device loss level was determined based on parameters such as voltage, current, and junction temperature under operating conditions. Case temperature and ambient temperature were acquired using K-type thermocouples. The experimental platform is shown below. Figure 8 As shown, for the fixed components in the circuit, the capacitance values of capacitors C1 and C2 are both set to 3000 μF to provide sufficient voltage regulation capability. The FF50R12RT4 half-bridge module, consistent with the simulation, is selected as the device under test, with a maximum withstand voltage of 1200 V and a maximum current of 50 A. The switching frequency is set to 10 kHz, and the dead time is set to 1 μs to prevent current tailing from causing a short circuit. Considering that the voltage and current across the two arms of the half-bridge module are basically the same, only a single arm (the lower arm) is used as the research object for experimental verification.
[0049] Then, the experimental data was shuffled, with 70% of the scenarios used for training the artificial neural network model and the remaining 30% used for testing. Simultaneously, Bayesian optimization was used to adjust the hyperparameters during training, resulting in a well-trained artificial neural network model. The test data was then input into the trained model, and the output was as follows. Figures 9-13 As shown, the evaluation results for solder aging level, IGBT module chip loss, diode chip loss, IGBT module chip junction temperature, and diode chip junction temperature are presented respectively. To visually demonstrate the relationship between error assessment level and power level, the load size is sorted here, with the horizontal axis representing the order of operating conditions and arranged according to power level from smallest to largest. Power level gradually increases from left to right, and error level and power level show a positive correlation trend; from... Figure 9 As can be seen, the accuracy rate of solder aging assessment reaches 96.3%; the model's assessment error shows an approximately positive correlation with power level. Figure 10 As can be seen, the maximum estimated error of IGBT module chip loss is approximately 0.4 W, with a percentage error of about 2.4%. Figure 11 As can be seen, the maximum estimated error of the diode chip loss is approximately 0.45 W, with a percentage error of about 5.6%. Figure 12 As can be seen, the maximum estimated error of the IGBT module chip junction temperature is approximately 2 ℃, with a percentage error of about 2.5%; from Figure 13It can be seen that the maximum estimation error of the diode junction temperature is about 3 ℃, and the percentage error is about 3.8%. Therefore, the state assessment model (i.e., the artificial neural network model) trained with experimental data can have excellent state assessment performance for the solder aging state, chip loss and junction temperature level of the IGBT module.
[0050] In summary, the IGBT module status assessment method based on substrate case temperature distribution and a neural network model proposed in this invention achieves IGBT module health status assessment based on artificial neural networks through non-invasive sampling design and feature parameter correlation analysis, improving the accuracy of solder aging level, chip loss, and junction temperature level assessment. Specifically, it involves: first, analyzing the external heat transfer performance of the IGBT module to determine the non-invasive case temperature as a feature parameter input for the internal health status; then, through correlation analysis, calculating the correlation degree of case temperatures at each location to eliminate case temperatures with high correlation, and using the retained case temperatures as input feature parameters. This reduces sampling costs and the complexity of the artificial neural network model by lowering the dimensionality of the case temperature data input. Secondly, state evaluation is performed based on the artificial neural network, and hyperparameter optimization is carried out using Bayesian optimization methods, which improves the recognition accuracy of the artificial neural network. In addition, data from the half-bridge inverter experimental circuit under different operating conditions are collected through experiments and used as input for training and testing of the artificial neural network. The results show that the method proposed in this invention achieves an accuracy of 96.3% in evaluating the aging level of module solder and an accuracy of over 94.4% in evaluating chip loss and junction temperature, and the artificial neural network has good generalization performance.
[0051] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.
[0052] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.
Claims
1. A method for evaluating the state of an IGBT module based on substrate shell temperature distribution and a neural network model, characterized in that, Includes the following steps: S1. Collect chip loss and junction temperature, device solder aging level, case temperature distribution and ambient temperature data of insulated gate bipolar transistor under different operating conditions. Through data preprocessing, generate preprocessed chip loss and junction temperature, device solder aging level, case temperature distribution and ambient temperature data. S2. Perform normality tests and correlation analyses on the preprocessed shell temperature distribution data to screen characteristic parameters of shell temperature distribution; S3. The shell temperature distribution characteristic parameters and the pre-processed ambient temperature are used as raw data. The pre-processed chip loss, chip junction temperature and solder aging level data are used as labeled data. The labeled raw data are input into the artificial neural network model. The Bayesian optimization method is used to optimize the hyperparameters of the artificial neural network model to generate a trained artificial neural network model. S4. Obtain the case temperature distribution and ambient temperature of the insulated gate bipolar transistor under different operating conditions, and input them into the trained artificial neural network model for state prediction to obtain the health status assessment results of the chip loss, junction temperature, and solder aging level of the insulated gate bipolar transistor under test.
2. The IGBT module state assessment method based on substrate shell temperature distribution and neural network model according to claim 1, characterized in that, The case temperature distribution data includes case temperature data at different locations, including the chip center and the edge of the insulated gate bipolar transistor.
3. The IGBT module state assessment method based on substrate shell temperature distribution and neural network model according to claim 2, characterized in that, Data preprocessing includes outlier removal and data normalization.
4. The IGBT module state assessment method based on substrate shell temperature distribution and neural network model according to claim 3, characterized in that, Step S3 specifically includes: S31. Use the Kolmokolov-Smilov test method to perform a normal distribution test on the preprocessed shell temperature distribution data to determine whether the preprocessed shell temperature distribution data meets the normal distribution. If yes, proceed to step S32; otherwise, proceed to step S33. S32. Based on the preprocessed shell temperature distribution data, calculate the Pearson correlation coefficient of the shell temperature data under different working conditions at any two locations; S33. Based on the preprocessed shell temperature distribution data, sort the shell temperature data under different working conditions at each location in ascending order of data size to obtain the rank of the shell temperature data under different working conditions at each location, and calculate the Spearman correlation coefficient of the shell temperature data under different working conditions at any two locations. S34. Sort the Pearson correlation coefficient or Spearman correlation coefficient of the case temperature data under different operating conditions at any two locations from smallest to largest. From the case temperature data under different operating conditions at any two locations corresponding to the smallest Pearson correlation coefficient or the smallest Spearman correlation coefficient, select the case temperature data that meets the condition of including the chip center position and the edge position of at least one insulated gate bipolar transistor, and use it as the case temperature distribution characteristic parameter.
5. The IGBT module state assessment method based on substrate shell temperature distribution and neural network model according to claim 4, characterized in that, The formula for calculating the Pearson correlation coefficient of shell temperature data under different operating conditions at any two locations is as follows: in, This represents the Pearson correlation coefficient of shell temperature data under different operating conditions at any two locations. , Representing positions respectively ,Location In working condition Below the shell temperature data, Indicates position The average shell temperature data under all operating conditions Indicates position The average shell temperature data under all operating conditions.
6. The IGBT module state assessment method based on substrate shell temperature distribution and neural network model according to claim 5, characterized in that, The formula for calculating the Spearman correlation coefficient of shell temperature data under different operating conditions at any two locations is as follows: in, This represents the Spearman correlation coefficient of shell temperature data under different operating conditions at any two locations. This indicates that any two positions are in the working condition. The difference in rank of the lower shell temperature data. Indicates the total number of operating conditions. Indicates rank or order.
7. The IGBT module state assessment method based on substrate shell temperature distribution and neural network model according to claim 6, characterized in that, The calculation formula for the artificial neural network model is: in, This represents the output of the input layer of an artificial neural network model. This represents the output of the hidden layer in an artificial neural network model. This represents the output of the output layer of the artificial neural network model. This represents the activation function of the hidden layer. This represents the activation function of the output layer. This represents the weight matrix of the hidden layer. This represents the bias matrix of the hidden layer. This represents the weight matrix of the output layer. This represents the bias matrix of the output layer. This indicates transpose.
8. The IGBT module state assessment method based on substrate shell temperature distribution and neural network model according to claim 7, characterized in that, Step S3 specifically includes: S31. Use the case temperature distribution characteristic parameters and ambient temperature as raw data, and use chip loss, chip junction temperature and solder aging level data as labeled data. Divide the labeled raw data into training set and validation set according to the proportion. S32. Set the hyperparameters to be optimized and the hyperparameter space, and set the objective function to evaluate the performance of the artificial neural network model. The input of the objective function is the combination of hyperparameters, and the output is the mean squared error under the validation set. S33. Randomly sample the hyperparameters to be optimized to generate several sets of hyperparameter combinations. For each set of hyperparameters, input the training set into the artificial neural network model for training. After training, input the validation set into the artificial neural network model for validation. Generate the mean squared error of the validation set under all hyperparameters and use it as the current dataset. S34. Fit a Gaussian process regression model using the current dataset, and obtain the predicted mean and variance under each hyperparameter combination by establishing the probability distribution of the objective function. S35. Based on the predicted mean and variance of each hyperparameter combination, a uniform grid is generated in the hyperparameter space. Each grid point corresponds to a set of hyperparameter combinations. The acquisition function value of each grid point is calculated. The grid point corresponding to the maximum acquisition function value is selected as the updated hyperparameter combination. S36. Train the artificial neural network model under the updated hyperparameter combination. After training, input the validation set into the artificial neural network model for validation, generate the mean squared error of the validation set under the updated hyperparameter combination, and add it to the current dataset to generate the updated dataset. S37. Determine whether the current iteration number is greater than or equal to the maximum iteration number or whether the mean square error of the validation set for N consecutive iterations has not decreased. If so, proceed to step S38. Otherwise, use the updated dataset to fit and update the Gaussian process regression model. By establishing the probability distribution of the objective function, obtain the predicted mean and variance under each hyperparameter combination, and proceed to step S35. S38. The combination of hyperparameters corresponding to the minimum mean square error in the updated dataset is taken as the optimal combination of hyperparameters, and the artificial neural network model trained using the optimal combination of hyperparameters is taken as the optimal trained artificial neural network model.
9. The IGBT module state assessment method based on substrate shell temperature distribution and neural network model according to claim 8, characterized in that, The hyperparameters to be optimized include learning rate, number of neurons, and batch size.
10. The IGBT module state assessment method based on substrate shell temperature distribution and neural network model according to claim 9, characterized in that, The formula for calculating the acquisition function value of each grid point is: in, Represents grid points The value of the acquisition function, This represents the optimal value of the current objective function. Represents grid points The predicted mean, Indicates the exploration coefficient. Indicates intermediate variables. The cumulative distribution function represents the standard normal distribution. Represents grid points The predicted variance is below. This represents the probability density function of the standard normal distribution.