Power module junction temperature estimation method and device based on convolutional neural network, and medium

Abstract

The invention relates to the technical field of power modules, in particular to a power module junction temperature estimation method and device based on a convolutional neural network and a medium, and the method comprises the steps: obtaining the current working condition data of a target power module, the current working condition data comprising temperature data, operation data and electrical data; inputting the current working condition data into a pre-trained convolutional neural network model, and extracting feature vectors corresponding to the temperature data, the operation data and the electrical data; and carrying out multi-modal fusion on the feature vectors, and outputting a junction temperature estimated value through a regression network. The high-precision and real-time estimation of the junction temperature of the power module can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power modules, in particular to a power module junction temperature estimation method, device and medium based on a convolutional neural network. BACKGROUND

[0002] With the rapid development of the new energy vehicle industry, the thermal management problem of power electronic modules has become increasingly prominent. As the core component of the motor controller, the power module integrates IGBT (Insulated Gate Bipolar Transistor), MOSFET and other power devices, and bears the heavy responsibility of power conversion under high frequency, high voltage and high current. During the switching process of the device, a large amount of heat is inevitably generated. If the heat is not dissipated in time or sufficiently, the junction temperature (i.e. the core junction temperature) inside the device will rise sharply, which may lead to performance degradation, or even thermal runaway and permanent damage, directly threatening the safety and reliability of electric vehicles. Therefore, real-time and accurate monitoring of the junction temperature of the power module has become a core technical demand to ensure the safe and stable operation of the vehicle.

[0003] Traditional junction temperature monitoring methods rely on external sensing devices such as thermistors and infrared thermometers for indirect measurement, which generally have response lags, large measurement deviations, and complex sensor layout. More notably, some methods rely on structural destructive testing or high-cost experimental platforms, which are difficult to adapt to the actual scenarios of long-term and dynamic operation of the vehicle, and have limited practicality.

[0004] In recent years, with the rise of artificial intelligence and deep learning technology, the use of neural networks to model the operating state of power devices and estimate the junction temperature based on real-time operating data has gradually become a research hotspot. However, existing solutions have obvious limitations: most methods rely only on a single data source (such as sampled voltage, current or heat sink temperature) to build a model, ignoring the synergistic relationship between multiple sources of information, resulting in insufficient robustness of the model under complex conditions, weak adaptability to different scenarios, and difficulty in meeting the actual application requirements. SUMMARY

[0005] The purpose of the present application is to provide a power module junction temperature estimation method, device and medium based on a convolutional neural network, which can realize high-precision and real-time estimation of the junction temperature of the power module.

[0006] To achieve the above purpose, the technical solution adopted by the present application is as follows: In a first aspect, the present application discloses a power module junction temperature estimation method based on a convolutional neural network, which comprises: obtaining current operating condition data of a target power module, the current operating condition data including temperature data, operating data and electrical data; The current working condition data is input into a pre-trained convolutional neural network model to extract feature vectors corresponding to the temperature data, the operation data and the electrical data; the feature vectors are fused in multiple modes, and a junction temperature estimation value is output through a regression network.

[0007] Further, the temperature data of the target power module is acquired by infrared thermal image acquisition, thermal simulation or thermocouple array; The operation data includes current, voltage, switching frequency, motor speed, motor torque, PWM duty ratio, shell temperature and cooling liquid operation parameters. The electrical data includes three-phase current, bus current and bus voltage.

[0008] Further, the current working condition data is input into a pre-trained convolutional neural network model, and the current working condition data is preprocessed and time stamp alignment is completed by marking the sampling time of the current working condition data, so that the input current working condition data has consistency in the time dimension.

[0009] Further, the preprocessing specifically includes temperature data preprocessing, operation data preprocessing and electrical data preprocessing. The temperature data preprocessing includes: representing the temperature data by a temperature distribution chart, adjusting the temperature distribution chart to a preset size, converting the temperature distribution chart into a grayscale chart, mapping the pixel value range from [0, 255] to [0, 1], then standardizing the image, and finally performing data enhancement by at least one of flipping, rotating and adding noise. The operation data preprocessing includes: filling in missing values by using mean, median or previous value, then normalizing and standardizing, and finally removing outliers. The electrical data preprocessing includes: cropping or filling each piece of time series electrical data to unify the time series electrical data to a preset length; then independently normalizing each electrical data channel; then removing the direct current component in the electrical data, smoothing the electrical data by using a sliding window or median filtering, and finally filtering and denoising, removing or correcting abnormal sampling points.

[0010] Further, the feature vector corresponding to the temperature data is extracted by a convolutional neural network, specifically including: The input temperature data is convolved by a convolutional layer to extract local spatial features; The feature map after convolution is down-sampled by a pooling layer; The multi-dimensional convolution features after multiple rounds of convolution and down-sampling are spliced and converted into a one-dimensional feature vector, i.e. the feature vector corresponding to the temperature data is obtained.

[0011] Further, the feature vector corresponding to the operation data is extracted by a multi-layer perception, specifically including: the operation data collected is linearly mapped from an input space to a high-dimensional feature space through a fully connected layer, and the features output by the previous layer are further linearly mapped through the stacking of multiple hidden layers, so that the feature vector corresponding to the operation data is output; after each layer of linear mapping, a non-linear transformation is introduced through an activation function.

[0012] Further, the feature vector corresponding to the electrical data is extracted by a convolutional neural network, specifically including: The convolutional layer slides on the waveform of the electrical data through a convolution kernel in the time dimension, performs weighted calculation on continuous sampling points, captures waveform changes in a local time window, extracts local dynamic information related to the junction temperature, and introduces a non-linear transformation through an activation function after the convolution operation; The feature sequence after the convolution is down-sampled through a pooling layer, After multiple rounds of convolution and down-sampling, the electrical data is compressed into a fixed-length feature vector, i.e., the feature vector corresponding to the electrical data.

[0013] Further, the convolutional neural network model is trained based on a sample data set, and the sample data set includes historical working condition data of the power module and junction temperature data corresponding to the historical working condition data.

[0014] In a second aspect, the application discloses a power module junction temperature estimation device based on a convolutional neural network, which includes: An acquisition module is configured to acquire current working condition data of a target power module, and the current working condition data includes temperature data, operation data and electrical data; An estimation module is configured to input the current working condition data into a pre-trained convolutional neural network model, extract feature vectors corresponding to the temperature data, the operation data and the electrical data, perform multi-modal fusion on the feature vectors, and output a junction temperature estimation value through a regression network.

[0015] In a third aspect, the application discloses a storage medium having a computer program stored thereon, and the computer program is executed by a processor to implement the steps of the power module junction temperature estimation method based on the convolutional neural network.

[0016] The application has the following unexpected beneficial effects: The power module junction temperature estimation method based on the convolutional neural network fuses three types of heterogeneous information of temperature data, operation data and electrical data, and fully utilizes the complementarity of multi-source data. Compared with traditional single data source modeling, multi-modal data can more comprehensively reflect the heating state of the power module, reduce the estimation error caused by the deviation or absence of single data, and make the model still maintain high precision under complex working conditions, and the robustness is significantly improved. Moreover, the convolutional neural network model has efficient feature extraction capability, and can be deployed on edge devices through model optimization to meet real-time requirements. The real-time output of the junction temperature estimation value can provide decision basis for intelligent thermal management of the electric vehicle power system, timely adjust the cooling strategy, avoid damage of the device due to overheating, and improve the running safety and reliability of the vehicle. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the specific embodiments of the present application or the technical solutions in the prior art, the following will briefly introduce the drawings needed to be used in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present application.

[0018] Figure 1 A flowchart of the power module junction temperature estimation method based on the convolutional neural network provided by the present application is shown; Figure 2 A structure diagram of the power module junction temperature estimation device based on the convolutional neural network provided by the present application is shown. DETAILED DESCRIPTION

[0019] The embodiments of the present application will be described below with reference to the drawings and preferred embodiments, and those skilled in the art can easily understand other advantages and effects of the present application from the content disclosed in the specification. The present application can also be implemented or applied by different specific embodiments, and the details in the specification can be modified or changed based on different views and applications without departing from the spirit of the present application. It should be understood that the preferred embodiments are only for illustrating the present application, but not for limiting the protection scope of the present application.

[0020] In an embodiment, referring to Figure 1 The power module junction temperature estimation method based on the convolutional neural network includes: Obtaining current working condition data of a target power module, the current working condition data including temperature data, operation data and electrical data; Inputting the current working condition data into a pre-trained convolutional neural network model, extracting feature vectors corresponding to the temperature data, the operation data and the electrical data; performing multi-modal fusion on the feature vectors, and outputting a junction temperature estimation value through a regression network.

[0021] Compared with traditional single data source modeling, the multi-modal data can more comprehensively reflect the heating state of the power module, reduce the estimation error caused by the deviation or absence of single data, make the model still maintain high precision under complex working conditions, and significantly improve the robustness. Moreover, the convolutional neural network model has efficient feature extraction capability, and can be deployed on edge devices through model optimization, meeting the real-time requirement. The real-time output of the junction temperature estimation value can provide decision basis for intelligent thermal management of the electric vehicle power system, timely adjust the cooling strategy, avoid damage of the device due to overheating, and improve the running safety and reliability of the vehicle.

[0022] Moreover, the traditional junction temperature monitoring relies on external sensors such as thermistors and infrared thermographs, and has the problems of complex layout and high maintenance cost, and some methods need destructive testing or high-cost experimental platform. The method is based on the data such as running data and electrical data collected by the existing controller and sampling circuit, combined with non-invasive data such as infrared thermal image, without the need for additional special temperature sensors, reducing the hardware cost and vehicle integration difficulty, and being more suitable for large-scale mass production application.

[0023] As a preferred embodiment of the present application, the temperature data of the target power module is acquired by infrared thermal image acquisition, thermal simulation or thermocouple array.

[0024] The infrared thermal image acquisition can non-invasively acquire the two-dimensional thermal distribution image of the surface of the power module under different loads, temperatures and cooling conditions, intuitively reflect the temperature difference of different regions, such as the local hot spot of the chip and the packaging structure, without damaging the module structure, and is suitable for real-time monitoring scene; The thermal simulation software (ANSYS or COMSOL) can construct the thermal distribution under extreme working conditions or complex structures in a virtual environment, and make up for the blank of limit scene data difficult to reproduce in physical experiments. Moreover, the thermal simulation software can accurately output the aligned data of the thermal distribution map + junction temperature label, and provide high-quality labeled samples for convolutional neural network model training.

[0025] The thermocouple array acquires temperature data through a regular two-dimensional grid, balancing between cost and accuracy, which can reflect the spatial distribution characteristics of temperature, similar to a simplified thermal map, and is easier to integrate into mass production equipment than an infrared thermal imager.

[0026] It should be noted that, whether it is the thermal map of the infrared thermal imager under different loads, temperatures and cooling conditions, or the typical working condition thermal distribution constructed by the simulation software, the binding of temperature data with specific working conditions is emphasized, so that the temperature information is not only isolated values, but also can be linked with operation data and electrical data to help the model learn the internal rules of working condition-heat generation-temperature distribution more accurately when training the convolutional neural network model.

[0027] The operation data includes current, voltage, switching frequency, motor speed, motor torque, PWM duty ratio, shell temperature and cooling liquid operation parameters. Among them, current, voltage and switching frequency determine switching loss and conduction loss (core heat source); motor speed and torque reflect load intensity and affect the dynamic change of loss; shell temperature and cooling liquid parameters (such as flow and temperature) reflect the heat dissipation condition and are the key variables of temperature conduction; PWM duty ratio is related to the switching frequency and conduction time of power devices, which directly affects the transient heating characteristics. The inclusion of these operation data enables the model to fully capture the whole chain process of input energy, loss heating and heat dissipation efficiency, avoiding estimation deviation caused by missing key working condition parameters.

[0028] The electrical data includes three-phase current, bus current and bus voltage. These are the most core electrical characteristics in the operation of power modules, among which three-phase current directly corresponds to the conduction current of power devices and is positively correlated with conduction loss; bus voltage determines the amplitude of switching loss, which is approximately proportional to the square of voltage; accurate collection of the two can enable the model to directly lock the core variables of loss calculation, reduce irrelevant data interference and improve the efficiency and pertinence of feature extraction.

[0029] The preferred embodiment ensures that multi-modal data can fully and accurately reflect the heating state of the power module by clearly defining the diversified collection path of temperature data and refining the core composition of operation data and electrical data, and improves the model efficiency by matching the structure of data and convolutional neural network, while taking into account the scene adaptation of high-precision demand and low-cost production, providing solid technical support for the landing application of the whole junction temperature estimation method.

[0030] As a preferred embodiment of the present application, the method further comprises: after inputting the current working condition data into the pre-trained convolutional neural network model, pre-processing the current working condition data, and completing timestamp alignment by marking the sampling time of the current working condition data, so that the input current working condition data has consistency in time dimension.

[0031] The preprocessing of current working condition data (such as temperature data, operation data, electrical data, etc.) can reduce the interference information in the original data, such as sensor fluctuations, sampling errors, etc., ensure that the data set input to the convolutional neural network is more pure, avoid the interference of invalid data on model training, and enable the model to more accurately capture core features, thereby improving the stability and consistency of junction temperature estimation.

[0032] By marking the sampling time to align the time dimension, the time difference problem that may exist in the collection process of multi-modal data is solved, enabling the model to learn the causal chain in the dynamic process and significantly improving the time dimension accuracy of junction temperature estimation. Moreover, power modules face transient changes in actual applications, such as current mutations during sudden acceleration and deceleration. The multi-source data after timestamp alignment can record the dynamic process of working condition mutation-loss change-temperature response. The convolutional neural network model can learn transient features such as the hysteresis of junction temperature and the time difference between temperature peak and load peak, enabling the junction temperature estimation to not only be suitable for steady-state conditions but also accurately respond to dynamic scenarios such as sudden load changes and start-stop cycles, ensuring estimation accuracy under complex conditions.

[0033] At the same time, the preprocessed standardized data (combined with timestamps) reduces the risk of model overfitting caused by non-uniform data formats and chaotic time sequences. This enables the convolutional neural network model to access more standardized and more realistic time-sequenced multi-source data in production scenarios, thereby maintaining stable estimation results for different vehicle models and different use scenarios in mass production applications.

[0034] As a preferred embodiment of the present application, the preprocessing specifically includes temperature data preprocessing, operation data preprocessing, and electrical data preprocessing. The temperature data preprocessing includes: representing the temperature data with a temperature distribution chart, adjusting the temperature distribution chart to a preset size, converting the temperature distribution chart to a grayscale chart, mapping the pixel value range from [0, 255] to [0, 1], then standardizing the image, and finally performing data augmentation by at least one of flipping, rotating, and adding noise.

[0035] Adjusting the temperature distribution chart to a preset size and converting it to a grayscale chart unifies the format and dimension of the input image, such as 64x64 pixels, avoiding the disorder of convolutional layer feature extraction caused by different image sizes. Mapping the pixel value from [0, 255] to [0, 1] and standardizing it eliminates the measurement range differences of different devices, making the temperature data comparable at the same scale and reducing the interference of data magnitude differences on model weight updates. For example, the standardization calculation formula is: standardized value = (original value - mean) / standard deviation, where the mean is set to 0 and the standard deviation is set to 1.

[0036] By flipping (horizontal / vertical), rotating (such as ±15°), adding noise (such as Gaussian noise), and the like, the diversity of the temperature distribution sample is artificially expanded, so that the model is finally exposed to more abundant temperature and working condition associated samples in the training and estimation process, overfitting is avoided, and stable output can still be maintained in actual application under non-ideal collection conditions.

[0037] The operation data preprocessing includes: filling in missing values by using mean, median or previous value, then normalizing and standardizing, and finally removing outliers.

[0038] The missing data is filled in by using mean, median or previous value, such as torque data loss caused by CAN bus transmission interruption, to avoid destroying the time sequence correlation between working condition change and heating response due to data breakage, and to ensure that the model can learn the temperature evolution law under continuous working conditions. For example: for the slowly changing shell temperature, the previous value can be used to fill in to retain the temperature change trend; for example, for the large fluctuation PWM duty cycle, the median value can be used to fill in to reduce the interference of extreme values.

[0039] Normalization maps operation data of different magnitudes (such as rotational speed unit rpm and cooling liquid flow unit L / min) to the same interval, avoiding large numerical parameters dominating gradient updates in model training; Outliers are removed to reduce the misleading of "dirty data" to model learning, so that the model focuses on normal heating characteristics under real working conditions. For example, values exceeding 3σ are removed or pruned, and σ is the standard deviation.

[0040] The electrical data preprocessing includes: cropping or filling in each time sequence electrical data to make the time sequence electrical data uniform to a preset length; then independently normalizing each electrical data channel; then removing the direct current component in the electrical data, smoothing the electrical data by using a sliding window or median filtering, and finally filtering and denoising, removing or correcting abnormal sampling points.

[0041] The time sequence data is cropped or filled to a preset length, such as uniform sampling sequence within 1000ms, to ensure that the electrical data under different working conditions are comparable in time dimension, facilitating the model to capture transient processes; Each channel is independently normalized to retain the relative difference of each phase current, which is used as a key feature reflecting local heating unevenness of the module.

[0042] After removing the direct current component, the electrical data more accurately reflects the alternating loss. By sliding window smoothing and filtering and denoising, the fundamental wave and low-frequency fluctuation characteristics of current and voltage are highlighted. Abnormal sampling points are corrected to ensure that the physical correlation between electrical data and actual loss is not destroyed, so that the model can accurately learn the mapping relationship between electrical parameters, loss and temperature.

[0043] As a preferred embodiment of the present application, the feature vector corresponding to the temperature data is extracted by a convolutional neural network, specifically comprising: The input temperature data is convolved by a convolutional layer to extract local spatial features; The feature map after convolution is down-sampled by a pooling layer; The multi-dimensional convolutional features after multiple rounds of convolution and down-sampling are spliced and converted into a one-dimensional feature vector, i.e. the feature vector corresponding to the temperature data is obtained.

[0044] The core information of temperature data (such as infrared thermal image, two-dimensional grid data of thermocouple array) lies in spatial distribution features, such as hot spot position of chip area, temperature gradient change, and temperature difference between different power devices. The convolutional layer can automatically capture these key information by sliding the convolution kernel to extract features in the local area. Exemplarily, the local area is a 3x3 or 5x5 pixel block. This local feature extraction method is highly consistent with the physical distribution law of the temperature field, and can retain more detailed information compared to traditional artificial features.

[0045] The pooling layer down-samples the feature map after convolution, which compresses the data dimension while retaining the key features of the local area. On the one hand, it can reduce redundant information, reduce model calculation, and improve running efficiency; on the other hand, it can enhance the translation invariance of the features, making the model more fault-tolerant to the collection bias of temperature data.

[0046] After the multi-dimensional convolutional features are spliced and converted into a one-dimensional vector, a structured temperature feature vector is formed, which matches the one-dimensional feature vector format of the operating data and electrical data, providing convenience for subsequent multi-modal fusion and ensuring that different modal data are calculated in the same dimension. Moreover, under the premise of retaining core spatial features, the two-dimensional temperature distribution information is compressed into a compact one-dimensional vector, which not only avoids information loss, but also reduces the computational burden of the subsequent regression network, balancing accuracy and efficiency.

[0047] As a preferred embodiment of the present application, the feature vector corresponding to the operating data is extracted by a multi-layer perceptron, specifically comprising: the collected operating data is linearly mapped from the input space to a high-dimensional feature space by a fully connected layer, and the features output by the previous layer are further linearly mapped by stacking multiple hidden layers, and the feature vector corresponding to the operating data is output; after each linear mapping, a non-linear transformation is introduced by an activation function.

[0048] The relationship between operating data (such as motor speed, torque, cooling liquid flow, etc.) and power module junction temperature is not simply linear, for example, the influence of torque increase at high speed on junction temperature may present nonlinear changes due to different cooling liquid flows. The fully connected layer maps the original operating data from a low-dimensional input space to a high-dimensional feature space, which can capture the cross-correlation between data, and compared with the traditional linear model, can more comprehensively depict the complex mapping relationship between operating parameters and junction temperature.

[0049] Each layer of linear mapping is followed by an activation function (such as ReLU, Sigmoid) to introduce a nonlinear transformation, in order to enhance the expression ability, so that the model can learn the nonlinear features in the operating data, solve the dynamic characteristics of the operating data in a wide operating condition range, and avoid feature distortion caused by linear assumption.

[0050] The stacking of multiple hidden layers realizes the hierarchical extraction of features, and the shallow hidden layer focuses on the basic features of a single parameter, such as the numerical range of the speed and the fluctuation frequency of the torque; the deep hidden layer integrates the shallow features to form high-order abstract features. This hierarchical feature is highly matched with the influence mechanism of the operating data on the junction temperature, so that the extracted features have more physical meaning, such as the "high current" + "high speed" leading to "high heat" phenomenon.

[0051] The multi-layer perceptron compresses the original high-dimensional operating data (which may include multiple parameters) into a low-dimensional but information-intensive feature vector through layer-by-layer mapping and transformation, which not only retains the key influencing factors, but also reduces redundant data, reduces the computational cost of subsequent multi-modal fusion, and avoids the influence of dimension disaster on the performance of the model. At the same time, the one-dimensional feature vector output by the multi-layer perceptron is uniform with the feature vector format of the temperature data and the electrical data, which ensures that the three types of data can be fused in the same dimension, avoids the fusion obstacles caused by format differences, and makes the complementarity of multi-modal data fully play. Moreover, the operating data is easily disturbed by the environment in the actual production scene, and the nonlinear fitting ability and hierarchical feature extraction mechanism of the multi-layer perceptron make it have stronger fault tolerance to such disturbances.

[0052] As a preferred embodiment of the present application, the feature vector corresponding to the electrical data is extracted by a convolutional neural network, which specifically includes: The convolutional layer slides on the waveform of the electrical data through the convolution kernel in the time dimension, performs weighted calculation on the continuous sampling points, captures the waveform changes in the local time window, extracts the local dynamic information related to the junction temperature, and introduces a nonlinear transformation through an activation function after convolution operation; The feature sequence after convolution is down-sampled by the pooling layer, After multiple rounds of convolution and down-sampling, the electrical data is compressed into a fixed-length feature vector, i.e. the feature vector corresponding to the electrical data.

[0053] The core features of electrical data are the timing change rules, such as the width of the current pulse, the frequency of the voltage fluctuation, the amplitude of the transient peak, and the like, which are directly related to the switching loss and conduction loss of the power module, and further affect the junction temperature.

[0054] The convolution layer slides on the waveform through the time dimension convolution kernel to perform weighted calculation on the continuous sampling points, and can capture the dynamic information in the local time window. This extraction method is highly consistent with the physical mechanism of the electrical parameters affecting the junction temperature, and can retain more dynamic loss information compared with the simple average value or peak value.

[0055] The convolution operation is followed by an activation function (such as ReLU, LeakyReLU) to introduce a nonlinear transformation, so that the model can learn these nonlinear correlations to enhance the model expression.

[0056] The pooling layer performs down-sampling on the feature sequence after convolution, which compresses the data dimension while retaining the key dynamic features.

[0057] Electrical data is usually a long time series sampled at high frequency, and direct input into the model will cause a sharp increase in computational load. Multiple rounds of convolution and pooling compress the data layer by layer to convert long time series data into fixed length feature vectors, which greatly reduces the data dimension under the premise of retaining key information, reduces the model inference time, and meets the real-time requirements of edge devices (such as vehicle controllers).

[0058] As a preferred embodiment of the present application, the convolutional neural network model is trained based on a sample data set, and the sample data set includes historical working condition data of the power module and junction temperature data corresponding to the historical working condition data.

[0059] In this preferred embodiment, the historical working condition data also includes temperature data, operation data and electrical data, and the multi-modal fusion mechanism of the convolutional neural network model integrates spatial features (temperature data), time series features (electrical data) and numerical features (operation data), so that the model can learn more essential heating rules. At the same time, the convolutional neural network model adopts modular design, such as independent temperature, operation and electrical feature extraction branches, which facilitates the addition of new data modalities or optimization of branch structure in the later stage, improves the adaptability of the junction temperature estimation method to different power module types and different vehicle types, and has stronger generalization ability and expansion capability.

[0060] In addition, the trained convolutional neural network model has high efficient feature extraction capability, and can be deployed on edge devices such as vehicle industrial computers and microcontrollers through model optimization to meet real-time requirements.

[0061] Further, after the multi-modal neural network, i.e., the convolutional neural network model, is designed, model training and evaluation are key steps to ensure its performance and generalization ability. A large amount of data needs to be prepared for training, 70% of which is used as a training set for parameter updating, 15% of which is used as a validation machine for hyperparameter tuning, and the last 15% is used as a test set for performance evaluation and generalization testing.

[0062] In the training process, the loss function is selected as mean square error (MSE) for judgment, the optimizer is selected as Adam, the Batch Size is selected as 64, the Epochs is selected as 100-300, the learning rate is selected as 1e-3, the Dropout is selected as 0.2-0.4 in the training configuration parameter. At the same time, the Early Stopping strategy is used, that is, if the validation set Loss does not decrease for 10 consecutive epochs, the training is stopped.

[0063] The evaluation indicators include MAE / RMSE and R2 parameters, etc., wherein MAE is the mean absolute error, the smaller the value is, the better, RMSE is the root mean square error, which reflects the fluctuation of the error, and R2 is the determination coefficient, the closer the value is to 1, the better.

[0064] The trained convolutional neural network model can be deployed on an industrial computer or an edge device, such as NVIDIA Jetson Nano or Xavier NX; or can be deployed on a microcontroller plus an external infrared acquisition device, such as STM32 plus an external infrared module. The optimization directions include using lightweight CNN, model pruning and quantization, multi-modal module asynchronous updating (reducing the pressure of processing all frames) and other directions.

[0065] Further, the estimated results are combined with time series trend analysis algorithms, such as sliding window average and local outlier factor (LOF), to realize junction temperature abnormal fluctuation detection and trigger fault warning.

[0066] Further, the convolutional neural network model adopts a modular construction method, which facilitates the addition of new modalities (such as structural stress and sound signal) or the replacement of existing branch model structures in the later stage, improving the flexibility and universality of the method.

[0067] Further, a graphical visualization interface of the junction temperature estimation result is integrated on the edge device side, and a heat map and temperature curve are output through a small OLED screen, HMI or upper computer port.

[0068] In an embodiment, the present application discloses a power module junction temperature estimation device based on a convolutional neural network, as shown in Figure 2 The junction temperature estimation device 10 includes: The acquisition module 11 is configured to acquire current working condition data of the target power module, wherein the current working condition data comprises temperature data, operation data and electrical data; The estimation module 12 is configured to input the current working condition data into a pre-trained convolutional neural network model, extract a feature vector corresponding to the temperature data, the operation data and the electrical data, perform multi-modal fusion on the feature vector, and output a junction temperature estimation value through a regression network.

[0069] In an embodiment, the application discloses a storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps of the power battery insulation fault protection method.

[0070] In several embodiments provided in the present application, it should be understood that the disclosed apparatus and method can be implemented in other manners. The described apparatus embodiment is merely schematic, for example, the division of the units is merely a logical function division, and there can be another division manner in actual implementation, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed coupling, or direct coupling or communication connection between the components can be indirect coupling or communication connection through some interfaces, devices or units, and can be electrical, mechanical or other forms.

[0071] The units described as separate components can or can not be physically separate, and the components displayed as units can or can not be physical units; they can be located in one place, or distributed on a plurality of network units; and some or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.

[0072] In addition, each functional unit in each embodiment of the present application can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be realized in the form of hardware, or in the form of hardware plus software functional unit.

[0073] Those skilled in the art can understand that all or part of the steps of the above-mentioned method embodiments can be completed by program instruction related hardware, and the foregoing program can be stored in a computer readable storage medium, and the program is executed to perform the steps of the above-mentioned method embodiments; and the foregoing storage medium includes mobile storage equipment, read only memory (Read Only Memory, ROM), magnetic disc or optical disc and various storage program codes.

[0074] Alternatively, the above-mentioned integrated units of the present application, if realized in the form of software function modules and sold or used as independent products, can also be stored in a computer-readable storage medium. Based on such understanding, the technical solutions of the present application or the parts that make contributions to the related art can be embodied in the form of a software product, which is stored in a storage medium and includes a number of instructions to make an electronic device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the methods described in the various embodiments of the present application. The aforementioned storage medium includes mobile storage devices, ROM, magnetic discs or optical discs, and various media that can store program codes.

[0075] The above is only an embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for estimating the junction temperature of a power module based on a convolutional neural network, characterized in that, include: Obtain the current operating condition data of the target power module, which includes temperature data, operating data, and electrical data; The current operating condition data is input into a pre-trained convolutional neural network model to extract feature vectors corresponding to temperature data, operating data, and electrical data. The feature vectors are fused in multiple modes, and the estimated junction temperature is output through a regression network.

2. The power module junction temperature estimation method based on convolutional neural networks according to claim 1, characterized in that: Temperature data of the target power module is acquired using infrared thermal imaging, thermal simulation, or thermocouple arrays. The operating data includes current, voltage, switching frequency, motor speed, motor torque, PWM duty cycle, housing temperature, and coolant operating parameters; The electrical data includes three-phase current, bus current, and bus voltage.

3. The power module junction temperature estimation method based on convolutional neural networks according to claim 1, characterized in that, Also includes: After the current operating condition data is input into a pre-trained convolutional neural network model, the current operating condition data is preprocessed, and timestamp alignment is completed by marking the sampling time of the current operating condition data, so that the input current operating condition data has consistency in the time dimension.

4. The power module junction temperature estimation method based on convolutional neural networks according to claim 3, characterized in that: The preprocessing specifically includes temperature data preprocessing, operational data preprocessing, and electrical data preprocessing; The temperature data preprocessing includes: representing the temperature data with a temperature distribution map, adjusting the temperature distribution map to a preset size, converting the temperature distribution map into a grayscale image, mapping the pixel value range from [0,255] to [0,1], standardizing the image, and finally performing data enhancement by at least one of flipping, rotating, and adding noise. The preprocessing of the operational data includes: filling missing values ​​with the mean, median or previous value, then normalizing and standardizing, and finally removing outliers; The electrical data preprocessing includes: trimming or padding each segment of time-series electrical data to unify the time-series electrical data to a preset length; then performing independent normalization for each electrical data channel; then removing the DC component from the electrical data; using a sliding window or median filter to smooth the electrical data; and finally filtering and denoising to remove or correct abnormal sampling points.

5. The power module junction temperature estimation method based on convolutional neural networks according to claim 1, characterized in that, The feature vector corresponding to the temperature data is extracted using a convolutional neural network, specifically including: Local spatial features are extracted by performing convolution operations on the input temperature data through convolutional layers. The convolutional feature map is downsampled using a pooling layer; The multidimensional convolutional features, after multiple rounds of convolution and downsampling, are concatenated and transformed into a one-dimensional feature vector, thus obtaining the feature vector corresponding to the temperature data.

6. The power module junction temperature estimation method based on convolutional neural networks according to claim 1, characterized in that, The feature vectors corresponding to the running data are extracted by a multilayer perceptron. Specifically, the collected running data is linearly mapped from the input space to a high-dimensional feature space through a fully connected layer. Then, through the stacking of multiple hidden layers, the features output by the previous layer are further linearly mapped to obtain the feature vectors corresponding to the running data. After each linear mapping, a nonlinear transformation is introduced through an activation function.

7. The power module junction temperature estimation method based on convolutional neural networks according to claim 1, characterized in that, The feature vector corresponding to the electrical data is extracted using a convolutional neural network, specifically including: The convolutional layer slides the convolutional kernel on the waveform of the electrical data in the time dimension, performs weighted calculations on continuous sampling points, captures waveform changes within a local time window, extracts local dynamic information related to junction temperature, and introduces nonlinear transformation through an activation function after the convolution operation. The convolutional feature sequence is downsampled using a pooling layer. After multiple rounds of convolution and downsampling, the electrical data is compressed into a fixed-length feature vector, which is the feature vector corresponding to the electrical data.

8. The power module junction temperature estimation method based on convolutional neural networks according to claim 1, characterized in that: The convolutional neural network model is trained based on a sample dataset, which includes historical operating data of the power module and junction temperature data corresponding to the historical operating data.

9. A power module junction temperature estimation device based on a convolutional neural network, characterized in that, include: The acquisition module is used to acquire the current operating condition data of the target power module, which includes temperature data, operating data and electrical data; The estimation module is used to input the current operating condition data into a pre-trained convolutional neural network model and extract feature vectors corresponding to temperature data, operating data and electrical data; The feature vectors are fused in multiple modes, and the estimated junction temperature is output through a regression network.

10. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the power module junction temperature estimation method based on a convolutional neural network as described in any one of claims 1 to 7.

Images