Method and device for estimating service life of power module
By combining electrical and thermal operating parameters with a recurrent neural network model, accurate prediction of power module junction temperature and lifetime estimation are achieved, solving the problem that the influence of preceding operating conditions is not considered in the existing technology, and improving the accuracy and reliability of lifetime prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHENZHI AUTOMOBILE TECHNOLOGY GROUP CO LTD CHONGQING INNOVATION RESEARCH BRANCH
- Filing Date
- 2026-01-26
- Publication Date
- 2026-05-08
AI Technical Summary
Existing power module lifetime prediction methods fail to effectively consider the impact of preceding operating conditions on temperature, resulting in insufficient junction temperature accuracy, which in turn affects the accuracy of lifetime prediction results.
A power module lifetime estimation method based on a recurrent neural network model is adopted. By acquiring electrical and thermal operating condition parameters, the junction temperature increment is output using a pre-trained junction temperature prediction model. Combined with time period division and lifetime calculation, the correlation between historical operating conditions and current junction temperature is captured, avoiding errors caused by disordered accumulation.
It significantly improves the accuracy and reliability of power module lifetime estimation, reduces estimation bias caused by trend confounding, and the results better reflect the true loss status of the power module.
Smart Images
Figure CN121995141A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power module technology, and specifically to a method and apparatus for estimating the lifespan of a power module. Background Technology
[0002] Power modules are integrated power devices used in the motor controllers of new energy vehicles, primarily converting high-voltage direct current (DC) into three-phase alternating current (AC). During the AC-DC conversion process, the semiconductor wafers, crucial functional components of the power module, generate significant heat, causing their temperature to rise. Studies have shown that for every 10°C increase in wafer temperature, the failure risk of the power module increases by approximately 10%, and its lifespan is correspondingly shortened. By monitoring the junction temperature of the power module and employing appropriate lifespan estimation methods, the lifespan of the power module can be effectively determined, improving driving safety. Therefore, power module junction temperature estimation methods have significant practical importance.
[0003] Existing power module lifespan prediction methods mainly involve estimating the power module junction temperature under actual driving conditions using a thermal resistance model, and then using this junction temperature in conjunction with the power module lifespan curve to obtain the corresponding power module lifespan. However, these methods only consider the power module junction temperature at the current operating point and do not account for the influence of preceding operating conditions on the power module temperature. Therefore, the accuracy of the obtained junction temperature is insufficient, leading to deviations in the lifespan prediction results. Summary of the Invention
[0004] The purpose of this invention is to provide a method and apparatus for estimating the lifespan of power modules, which can improve the accuracy of power module lifespan estimation results.
[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0006] In a first aspect, the present invention discloses a method for estimating the lifetime of a power module, comprising:
[0007] Acquire the electrical and thermal parameters of the power module during its operating cycle;
[0008] The electrical and thermal parameters are input into a pre-trained junction temperature prediction model, which outputs the predicted junction temperature increment values for several sampling points within the operating cycle. The junction temperature prediction model is trained based on a historical power module sample dataset. Each sample in the sample dataset includes the electrical and thermal parameters of the power module and its corresponding simulated or measured junction temperature increment values.
[0009] The operating cycle is divided into several consecutive time periods. The temperature change trend of all sampling points in each time period is consistent, and the change trends between adjacent time periods are different.
[0010] The lifespan of the power module in each time period is calculated based on the junction temperature increment. The lifespans of each time period are added together to obtain the lifespan of the power module in the entire operating cycle.
[0011] Furthermore, the formula for calculating the lifespan of the power module within each time period is: In the formula, For the first The lifespan of the power module within a specific time period. For the first The sum of the junction temperature increments at all sampling points within a given time period. and It is a constant.
[0012] Furthermore, the electrical operating parameters include operating current, operating voltage, duty cycle, switching frequency, and power factor, while the thermal operating parameters include fluid temperature, fluid flow rate, and ambient temperature.
[0013] Furthermore, the junction temperature prediction model is a recurrent neural network model.
[0014] Furthermore, the recurrent neural network model is pre-trained in the following manner:
[0015] Obtain a historical power module sample dataset generated by three-dimensional thermal-electric coupling simulation or bench test. The historical power module sample dataset includes time series data under multiple operating conditions. The sample at each time step includes electrical operating parameters, thermal operating parameters, and junction temperature increments corresponding to the electrical and thermal operating parameters.
[0016] The historical power module sample dataset is divided into a training set and a validation set according to a preset ratio.
[0017] A recurrent neural network model is constructed with the goal of minimizing the mean square error between the predicted junction temperature increment and the actual junction temperature increment. The recurrent neural network model is trained using a training set, and the training process is monitored based on the validation set until the loss function value on the validation set is lower than a preset threshold.
[0018] Furthermore, if the predicted value of the junction temperature increment is greater than or equal to zero, the characteristic state of the corresponding sampling point is marked as a heating state; otherwise, the characteristic state of the corresponding sampling point is marked as a cooling state; when the characteristic states of two adjacent sampling points are different, the latter sampling point is determined as the inflection point of the junction temperature change and is used as the starting point of the new time period.
[0019] In a second aspect, the present invention discloses a power module lifetime estimation device, comprising:
[0020] The data acquisition module is used to acquire the electrical and thermal parameters of the power module.
[0021] The junction temperature prediction module is used to call a pre-trained recurrent neural network model, input the electrical condition parameters and thermal condition parameters into the pre-trained junction temperature prediction model, and output the predicted value of the junction temperature increment at several sampling points within the operating cycle.
[0022] The time division module is used to determine the temperature change trend based on the predicted junction temperature increment of each sampling point, and divide the operating cycle into several consecutive time periods, with the temperature change trend of all sampling points being consistent within each time period.
[0023] The lifespan calculation module is used to calculate the lifespan of the power module in each time period based on the junction temperature increment, and to sum the lifespans of each time period to obtain the lifespan of the power module in the entire operating cycle.
[0024] This invention offers the following advantages: By acquiring continuous electrical and thermal operating parameters within the operating cycle, and utilizing a junction temperature prediction model to output junction temperature increments at several sampling points, this invention captures the correlation between historical operating conditions and the current junction temperature. This makes the junction temperature calculation more closely reflect the actual temperature variation patterns during power module operation, providing an accurate data foundation for lifetime estimation. Furthermore, by dividing the operating cycle into continuous time periods with consistent temperature change trends, it avoids lifetime calculation errors caused by the disordered accumulation of junction temperature increments under different trends. Segmented calculations can separately match the junction temperature change characteristics of corresponding stages, and then the overall lifetime is obtained through accumulation, significantly reducing estimation bias caused by mixed trends, making the results more reflective of the true wear and tear status of the power module. Attached Figure Description
[0025] Figure 1 A flowchart illustrating a power module lifetime estimation method provided in an embodiment of this application is shown.
[0026] Figure 2 A schematic diagram of the structure of a power module lifetime estimation device provided in an embodiment of this application is shown. Detailed Implementation
[0027] The embodiments of the present invention will be described below with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are only for illustrating the present invention and not for limiting the scope of protection of the present invention.
[0028] In one embodiment, see Figure 1 As shown, this application provides a method for estimating the lifetime of a power module, which includes:
[0029] Acquire the electrical and thermal parameters of the power module during its operating cycle;
[0030] The electrical and thermal parameters are input into a pre-trained junction temperature prediction model, which outputs the predicted junction temperature increment values for several sampling points within the operating cycle. The junction temperature prediction model is trained based on a historical power module sample dataset. Each sample in the sample dataset includes the electrical and thermal parameters of the power module and its corresponding simulated or measured junction temperature increment values.
[0031] The operating cycle is divided into several consecutive time periods. The temperature change trend of all sampling points in each time period is consistent, and the change trends between adjacent time periods are different.
[0032] The lifespan of the power module in each time period is calculated based on the junction temperature increment. The lifespans of each time period are added together to obtain the lifespan of the power module in the entire operating cycle.
[0033] According to actual tests, the power module lifetime estimation method of the recurrent neural network provided by this invention can improve the prediction accuracy by about 5%.
[0034] Specifically, the formula for calculating the cumulative lifespan is: In the formula, The lifespan of the power module during the operating cycle, For the first running cycle The lifespan of the power module within a specific time period. This represents the total number of time periods.
[0035] This embodiment acquires continuous electrical and thermal operating parameters within the operating cycle and uses a junction temperature prediction model to output junction temperature increments at several sampling points. This captures the correlation between historical operating conditions and the current junction temperature, making the junction temperature calculation more closely match the actual temperature variation patterns of the power module during operation, thus providing an accurate data foundation for lifetime estimation. Furthermore, by dividing the operating cycle into continuous time periods with consistent temperature change trends, it avoids lifetime calculation errors caused by the disordered accumulation of junction temperature increments under different trends. Segmented calculations can separately match the junction temperature change characteristics of corresponding stages, and then the overall lifetime is obtained by accumulation, significantly reducing estimation bias caused by mixed trends and making the results more reflective of the true wear and tear of the power module.
[0036] In a preferred embodiment of this application, the lifespan of the power module within each time period is calculated as follows: In the formula, For the first The lifespan of the power module within a specific time period. For the first The sum of the junction temperature increments at all sampling points within a given time period. and It is a constant.
[0037] In this preferred embodiment, and To ensure that the constants obtained through power device life tests accurately reflect the actual lifespan degradation patterns of power modules, the calculation formulas are used. Based on the cumulative calculation of time-series junction temperature increments, a combination of data-driven approach and mechanism derivation is achieved, avoiding the problem of pure theoretical models being disconnected from actual operating conditions, and making the life assessment results more valuable for engineering reference.
[0038] In a preferred embodiment of this application, the electrical parameters include operating current, operating voltage, duty cycle, switching frequency, and power factor. Operating current and operating voltage are key variables determining the conduction losses of the power module. Duty cycle and switching frequency directly affect switching losses; a higher switching frequency results in more switching operations per unit time, leading to more significant accumulated losses. The power factor is associated with additional losses due to reactive power. These five parameters together constitute a complete evaluation dimension for the heat generation intensity of the power module, avoiding errors in heat generation estimation caused by omitting key electrical parameters.
[0039] The thermal parameters include fluid temperature, fluid flow rate, and ambient temperature. Fluid temperature and flow rate directly determine the heat exchange efficiency of the heat dissipation medium, while ambient temperature serves as the reference temperature for the heat dissipation process. Higher ambient temperatures result in smaller temperature differences and lower heat dissipation efficiency. These three parameters provide a comprehensive description of the power module's heat dissipation conditions, resolving the problem of inaccurate heat dissipation estimation caused by traditional methods that only focus on ambient temperature and ignore the dynamic characteristics of the heat dissipation medium.
[0040] In this preferred embodiment, the combination of electrical and thermal parameters neither includes redundant parameters unrelated to junction temperature changes nor omits core variables, ensuring that the parameters input to the model are all high-value features for junction temperature prediction, thereby improving the accuracy of junction temperature increment calculation from the source.
[0041] The electrical and thermal parameters are synchronously collected by sensors or a control system at a fixed sampling period to form a multidimensional time series arranged in chronological order. Before being input into the pre-trained junction temperature prediction model, the time series is divided into several consecutive sampling points, each of which contains all electrical and thermal parameters at the same time.
[0042] In a preferred embodiment of this application, the junction temperature prediction model is a recurrent neural network model.
[0043] When a power module operates in a new energy vehicle motor controller, its junction temperature change depends not only on the current operating conditions but also on the heat accumulation effect of historical operating conditions. For example, a high-current operating condition in the previous moment may have caused the wafer temperature to rise. If the heat has not been sufficiently dissipated, the junction temperature may continue to rise even if the current decreases in the current moment. Recurrent neural network models possess inherent temporal memory capabilities, enabling them to transmit information about the operating conditions from the previous moment to the current moment's junction temperature increment prediction process through the hidden layer states. Specifically, during model inference, The hidden state generated after processing the input data at time step using the recurrent neural network model will be used as... Contextual information for calculating junction temperature increments at all times, thereby enabling The predicted junction temperature at any given time naturally reflects The method addresses the thermal accumulation effect over time. Compared to traditional thermal resistance models that estimate junction temperature based solely on instantaneous operating conditions, this method more realistically reproduces the dynamic thermal behavior of power modules, significantly improving the accuracy of junction temperature increment estimation and thus enhancing the reliability of lifetime prediction.
[0044] In a preferred embodiment of this application, the recurrent neural network model is pre-trained in the following manner:
[0045] Obtain a historical power module sample dataset generated by three-dimensional thermal-electric coupling simulation or bench test. The historical power module sample dataset includes time series data under multiple operating conditions. The sample at each time step includes electrical operating parameters, thermal operating parameters, and junction temperature increments corresponding to the electrical and thermal operating parameters.
[0046] The historical power module sample dataset is divided into a training set and a validation set according to a preset ratio.
[0047] A recurrent neural network model is constructed with the goal of minimizing the mean square error between the predicted junction temperature increment and the actual junction temperature increment. The recurrent neural network model is trained using a training set, and the training process is monitored based on the validation set until the loss function value on the validation set is lower than a preset threshold.
[0048] For example, the training process of the recurrent neural network model is as follows:
[0049] The parameter composition of the input layer of the recurrent neural network model was determined. The input layer contains eight time-series parameters: operating current, operating voltage, duty cycle, switching frequency, power factor, fluid temperature, fluid flow rate, and ambient temperature. Specifically: the operating current covers the operating range of the motor controller, from 0A to 500A; the operating voltage is selected from typical operating voltages for new energy vehicles, including 200V, 350V, and 450V levels; the duty cycle, switching frequency, and power factor are all derived from bench test data to accurately reflect actual operating conditions.
[0050] A stratified random sampling method was used to construct the training dataset to ensure a balanced distribution of operating conditions across different voltage levels. Specifically, a total of 120 time-series samples were collected and generated, evenly divided into three categories according to voltage level: 40 samples for 200V, 40 samples for 350V, and 40 samples for 450V. The corresponding chip junction temperature time series for each sample was obtained through three-dimensional thermo-electric coupling simulation. Subsequently, the time gradient derivative of the junction temperature series was performed to obtain the junction temperature increment between adjacent sampling points, and this junction temperature increment was used as the output target value of the recurrent neural network model.
[0051] The recurrent neural network model structure is constructed as follows: the input layer contains eight input nodes, corresponding to the eight input parameters mentioned above; the output layer is a single neuron that outputs the junction temperature increment at the next time step. The number of hidden layers and the number of neurons in each layer of the recurrent neural network model were determined through trial and error, and finally, two hidden layers were selected, with five neurons in each layer.
[0052] To optimize model performance, comparative experiments were conducted using different activation functions. Models were trained using Sigmoid, Tanh, and ReLU activation functions, and the fitting accuracy was evaluated on the same training set (100 sets) and validation set (20 sets). The results are shown in Table 1.
[0053] Table 1. Accuracy statistics of models using different activation functions
[0054]
[0055] Preferably, the ReLU method is used as the activation function for training the recurrent neural network model.
[0056] In a preferred embodiment of this application, if the predicted value of the junction temperature increment is greater than or equal to zero, the characteristic state of the corresponding sampling point is marked as a heating state; otherwise, the characteristic state of the corresponding sampling point is marked as a cooling state; when the characteristic states of two adjacent sampling points are different, the latter sampling point is determined as the inflection point of the junction temperature change and is used as the starting point of a new time period.
[0057] The positive or negative value of the predicted junction temperature increment directly reflects the core thermal state of the power module, indicating whether heat generation is greater than heat dissipation or vice versa. Compared to traditional segmentation based on fixed time intervals, such as dividing a time period into 10-second intervals, this marking method skips irrelevant time dimensions and directly uses the junction temperature change trend as the basis for segmentation. For example, when the power module enters a continuous heating state due to a sudden increase in operating current, it can be continuously marked as heating state, avoiding the forced splitting of the same thermal trend by fixed segmentation. When the cooling system starts, it can be immediately marked as cooling state, accurately reflecting the thermal state switch caused by enhanced heat dissipation.
[0058] By comparing the characteristic states of adjacent sampling points, the inflection point of junction temperature change can be directly located. This determination method is based on the abrupt change characteristics of thermal state, which is more realistic than relying on the absolute value threshold of junction temperature. For example, if the junction temperature exceeds 80°C to determine the inflection point, when the junction temperature of the power module rises from 75°C to 78°C but the increment changes from positive to negative, this method can identify it as an inflection point. In contrast, the threshold method would miss the critical segment point because the temperature threshold is not reached, resulting in disordered accumulation of junction temperature increments during the heating and cooling stages in subsequent lifetime calculations, thus causing errors.
[0059] For example, the heating state is represented by the number 1, and the cooling state is represented by the number 0.
[0060] In one embodiment, see Figure 2 As shown, this application discloses a power module lifetime estimation device. The estimation device 10 includes a data acquisition module 11, a junction temperature prediction module 12, a time division module 13, and a lifetime calculation module 14.
[0061] The data acquisition module 11 is used to acquire the electrical and thermal parameters of the power module.
[0062] The junction temperature prediction module 12 is used to call a pre-trained recurrent neural network model, input the electrical condition parameters and thermal condition parameters into the pre-trained junction temperature prediction model, and output the predicted values of the junction temperature increment at several sampling points within the operating cycle.
[0063] The time division module 13 is used to determine the temperature change trend based on the predicted junction temperature increment of each sampling point, and divide the operating cycle into several continuous time periods, with the temperature change trend of all sampling points being consistent in each time period.
[0064] The lifespan calculation module 14 is used to calculate the lifespan of the power module in each time period based on the junction temperature increment, and to accumulate the lifespan of each time period to obtain the lifespan of the power module in the entire operating cycle.
[0065] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments, and various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention. Furthermore, the embodiments of the present invention and the features thereof can be combined with each other unless otherwise specified.
Claims
1. A method for estimating the lifespan of a power module, characterized in that, include: Acquire the electrical and thermal parameters of the power module during its operating cycle; The electrical and thermal parameters are input into the pre-trained junction temperature prediction model, and the predicted junction temperature increment values of several sampling points within the operating cycle are output. The junction temperature prediction model is trained based on a historical power module sample dataset. Each sample in the sample dataset includes the electrical and thermal parameters of the power module and its corresponding simulated or measured values of the junction temperature increment. The operating cycle is divided into several consecutive time periods. The temperature change trend of all sampling points in each time period is consistent, and the change trends between adjacent time periods are different. The lifespan of the power module in each time period is calculated based on the junction temperature increment. The lifespans of each time period are added together to obtain the lifespan of the power module in the entire operating cycle.
2. The power module lifespan estimation method according to claim 1, characterized in that: The formula for calculating the lifespan of the power module within each time period is: In the formula, For the first The lifespan of the power module within a specific time period. For the first The sum of the junction temperature increments at all sampling points within a given time period. and It is a constant.
3. The power module lifespan estimation method according to claim 1, characterized in that: The electrical operating parameters include operating current, operating voltage, duty cycle, switching frequency, and power factor, while the thermal operating parameters include fluid temperature, fluid flow rate, and ambient temperature.
4. The power module lifetime estimation method according to claim 1, characterized in that: The junction temperature prediction model is a recurrent neural network model.
5. The power module lifetime estimation method according to claim 4, characterized in that: The recurrent neural network model is pre-trained in the following manner: Obtain a historical power module sample dataset generated by three-dimensional thermal-electric coupling simulation or bench test. The historical power module sample dataset includes time series data under multiple operating conditions. The sample at each time step includes electrical operating parameters, thermal operating parameters, and junction temperature increments corresponding to the electrical and thermal operating parameters. The historical power module sample dataset is divided into a training set and a validation set according to a preset ratio. A recurrent neural network model is constructed with the goal of minimizing the mean square error between the predicted junction temperature increment and the actual junction temperature increment. The recurrent neural network model is trained using a training set, and the training process is monitored based on the validation set until the loss function value on the validation set is lower than a preset threshold.
6. The power module lifetime estimation method according to claim 1, characterized in that: If the predicted junction temperature increment is greater than or equal to zero, the characteristic state of the corresponding sampling point is marked as a heating state; otherwise, the characteristic state of the corresponding sampling point is marked as a cooling state. When the characteristic states of two adjacent sampling points are different, the latter sampling point is determined as the inflection point of junction temperature change and is used as the starting point of a new time period.
7. A power module lifespan estimation device, characterized in that, include: The data acquisition module is used to acquire the electrical and thermal parameters of the power module. The junction temperature prediction module is used to call a pre-trained recurrent neural network model, input the electrical condition parameters and thermal condition parameters into the pre-trained junction temperature prediction model, and output the predicted value of the junction temperature increment at several sampling points within the operating cycle. The time division module is used to determine the temperature change trend based on the predicted junction temperature increment of each sampling point, and divide the operating cycle into several consecutive time periods, with the temperature change trend of all sampling points being consistent within each time period. The lifespan calculation module is used to calculate the lifespan of the power module in each time period based on the junction temperature increment, and to sum the lifespans of each time period to obtain the lifespan of the power module in the entire operating cycle.