IGBT health state high-robustness diagnosis method and system, medium and equipment
By arranging temperature sensors on the IGBT module substrate, constructing and normalizing the shell temperature matrix, and combining the finite element model and convolutional neural network, the problem of difficult detection of internal degradation of IGBT modules is solved, and a highly robust and accurate health status diagnosis is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XI AN JIAOTONG UNIV
- Filing Date
- 2025-12-31
- Publication Date
- 2026-05-08
AI Technical Summary
IGBT modules are prone to degradation and failure under unstable operating conditions, leading to wind power generation system failures, but their internal degradation is difficult to detect directly.
By arranging temperature sensors on the outer surface of the substrate directly below each chip in the IGBT module, a shell temperature matrix is constructed, preprocessed and normalized, and combined with a finite element thermo-mechanical coupling model and a convolutional neural network, intelligent diagnosis of shell temperature distribution characteristics is achieved.
It improves the robustness and accuracy of IGBT health status diagnosis, effectively monitors uneven degradation of the substrate solder layer, reduces errors caused by changes in operating conditions, and is suitable for online diagnosis.
Smart Images
Figure CN121997121A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of IGBT module testing technology, and in particular to a robust intelligent diagnostic method, system, medium, and device for IGBT health status based on case temperature distribution characteristics. Background Technology
[0002] Wind power generation, as one of the most important forms of new energy, has attracted much attention. In wind power generation systems, the wind power converter plays a core role in connecting the wind turbine to the power grid and is an indispensable component of the entire system. The core component of the wind power converter is the IGBT power module. However, under unstable operating conditions, IGBTs are prone to degradation and failure, leading to system-wide malfunctions. Since IGBTs typically operate in a fully encapsulated manner, internal degradation is difficult to detect directly. Therefore, identifying key state parameters is crucial for IGBT health status diagnosis.
[0003] The information disclosed in the background section is only for enhancing the understanding of the background of this invention, and therefore may contain information that does not constitute prior art known to those skilled in the art. Summary of the Invention
[0004] This invention provides a robust intelligent diagnostic method, system, medium, and device for IGBT health status based on shell temperature distribution characteristics. The shell temperature distribution characteristics are used as a state parameter, and the relative change value is obtained through normalization processing to reduce the error caused by changes in operating conditions. Finally, a neural network is used to realize a robust intelligent diagnostic of the health status.
[0005] A robust intelligent diagnostic method for IGBT health status based on shell temperature distribution characteristics includes:
[0006] Temperature sensors are arranged on the outer surface of the substrate directly below multiple chips of the IGBT module to collect case temperature data at each measuring point and construct a case temperature matrix to characterize the spatial distribution characteristics of the case temperature of the IGBT module.
[0007] The shell temperature matrix is preprocessed by subtracting the ambient temperature from the temperature at each measuring point to obtain the temperature rise matrix. Then, based on the parallel half-bridge structure in the IGBT module, the temperature rise values of the corresponding parallel branches are cross-differentiated and divided by the sum of the temperature rises of the corresponding branches to construct a normalized diagnostic temperature matrix, so as to eliminate the influence of ambient temperature and power loss fluctuations.
[0008] A three-dimensional finite element thermo-mechanical coupling model of the IGBT module was established to simulate the changes in heat transfer performance of the substrate solder layer under different degradation states, and a case temperature simulation dataset including healthy, early degradation, severe degradation and failure states was generated.
[0009] An inverter experimental platform was built, and the case temperature data under real operating conditions was collected. After Kalman filtering and sliding window averaging, an experimental test dataset was constructed.
[0010] A convolutional neural network model is trained based on the simulation dataset. The network contains at least three two-dimensional convolutional layers and fully connected layers. The input is a normalized diagnostic temperature matrix, and the output is an IGBT health status category, realizing intelligent classification of the type and degree of uneven degradation of the solder layer.
[0011] In the aforementioned robust intelligent diagnostic method for IGBT health status based on case temperature distribution characteristics, the dimension of the case temperature matrix corresponds to the number and layout of power chips in the IGBT module, and the measurement point is located on the substrate directly below the IGBT chip or FWD chip.
[0012] In the aforementioned robust intelligent diagnostic method for IGBT health status based on shell temperature distribution characteristics, the normalized diagnostic temperature matrix is constructed as follows: Measurement points A, B, C, and D correspond to the sum of temperature rises of the low-side IGBT, low-side FWD, high-side FWD, and high-side IGBT, respectively; the relative temperature difference between adjacent branches is calculated.
[0013]
[0014] In the formula, ΔTA equals the sum of ΔT1, ΔT3 and ΔT5, ΔTB equals the sum of ΔT2, ΔT4 and ΔT6, ΔTC equals the sum of ΔT7, ΔT9 and ΔT11, and ΔTD equals the sum of ΔT8, ΔT10 and ΔT12.
[0015] In the aforementioned robust intelligent diagnostic method for IGBT health status based on case temperature distribution characteristics, the degradation status of the substrate solder layer includes: healthy state, early degradation with an area reduction of 20%-35%, severe degradation with an area reduction of 35%-50%, and failure state with an area reduction of at least 50%, and the degradation form is unevenly distributed.
[0016] In the aforementioned robust intelligent diagnostic method for IGBT health status based on shell temperature distribution characteristics, the temperature data is processed by Kalman filtering, and noise is suppressed by using a sliding window to take the average value.
[0017] In the aforementioned robust intelligent diagnostic method for IGBT health status based on shell temperature distribution characteristics, the convolutional neural network does not contain pooling layers. The input layer receives a normalized temperature matrix of size 2×2 or 3×4, extracts spatial features through multiple two-dimensional convolutional layers, and finally outputs the health status classification result by a fully connected layer.
[0018] In the aforementioned robust intelligent diagnostic method for IGBT health status based on shell temperature distribution characteristics, random noise within the range of ±0.01℃ to ±0.1℃ is superimposed on the simulation data to simulate actual measurement errors.
[0019] A system for implementing the method includes:
[0020] The temperature sensing unit is used to arrange multiple temperature sensors on the substrate directly below the IGBT chip.
[0021] The data acquisition and preprocessing unit is used to acquire shell temperature data and perform ambient temperature compensation and normalized differential processing.
[0022] The finite element simulation module is used to build IGBT models and generate noisy multi-state shell temperature datasets.
[0023] A convolutional neural network diagnostic model is used to receive a normalized temperature matrix and output a health status classification result.
[0024] The display and alarm unit is used to visualize diagnostic results and issue warnings when degradation is detected.
[0025] A computer storage medium including computer instructions that, when run on a computer, cause the computer to perform the method.
[0026] An electronic device, the electronic device comprising:
[0027] Memory, processor, and computer programs stored in memory and executable on the processor, wherein,
[0028] The processor implements the method when executing the program.
[0029] Compared with existing technologies, this invention has the following advantages: This invention proposes using the case temperature distribution characteristics as a state parameter. Sensors are arranged on the substrate directly below each IGBT chip. By measuring the temperature at the substrate, a case temperature matrix is formed, which can characterize the case temperature distribution characteristics. Normalization of the case temperature matrix reduces errors caused by changes in operating conditions. Robustness is improved through intelligent diagnosis based on neural networks. Attached Figure Description
[0030] Various other advantages and benefits of the present invention will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiments below. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. Furthermore, the same reference numerals denote the same parts throughout the drawings.
[0031] In the attached diagram:
[0032] Figure 1 This is a schematic diagram of the degradation cycle test of the substrate solder layer of the present invention;
[0033] Figure 2 This is a schematic diagram of the IGBT of the present invention;
[0034] Figure 3 This is a physical image of the IGBT of this invention;
[0035] Figure 4 This is a schematic diagram of the 12 shell temperature measurement points of the present invention;
[0036] Figure 5 This is a schematic diagram of the finite element model of the IGBT module of the present invention;
[0037] Figure 6 This is a schematic diagram of the temperature distribution of the IGBT module of the present invention;
[0038] Figure 7 This is a schematic diagram of the IGBT module case temperature according to the present invention;
[0039] Figure 8 This is a schematic diagram of the degradation of the solder layer on the substrate of the present invention;
[0040] Figure 9 This is a schematic diagram of Kalman filtering and sliding window averaging in this invention. Figure 9 In the middle (a), the temperature curve is Kalman filtered. Figure 9 (b) is a schematic diagram of the sliding window averaging method;
[0041] Figure 10 This is a schematic diagram of the CNN framework of the present invention;
[0042] Figure 11 This is a schematic diagram illustrating the training and recognition accuracy of the present invention.
[0043] The present invention will be further explained below with reference to the accompanying drawings and embodiments. Detailed Implementation
[0044] Specific embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While specific embodiments of the invention are shown in the drawings, it should be understood that the invention may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.
[0045] It should be noted that certain terms are used in the specification and claims to refer to specific components. Those skilled in the art will understand that different terms may be used to refer to the same component. This specification and claims do not distinguish components based on differences in terminology, but rather on differences in function. The terms "comprising" or "including" used throughout the specification and claims are open-ended and should be interpreted as "comprising but not limited to." The following descriptions are preferred embodiments for carrying out the invention; however, these descriptions are for the purpose of understanding the general principles of the specification and are not intended to limit the scope of the invention. The scope of protection of this invention is determined by the appended claims.
[0046] To facilitate understanding of the embodiments of the present invention, further explanations and descriptions will be provided below with reference to the accompanying drawings and specific embodiments. The accompanying drawings do not constitute a limitation on the embodiments of the present invention.
[0047] like Figures 1 to 11 As shown, the robust intelligent diagnostic method for IGBT health status based on shell temperature distribution characteristics includes the following steps:
[0048] Temperature sensors are arranged on the outer surface of the substrate directly below multiple chips of the IGBT module to collect case temperature data at each measuring point and construct a case temperature matrix to characterize the spatial distribution characteristics of the case temperature of the IGBT module.
[0049] The shell temperature matrix is preprocessed by subtracting the ambient temperature from the temperature at each measuring point to obtain the temperature rise matrix. Then, based on the parallel half-bridge structure in the IGBT module, the temperature rise values of the corresponding parallel branches are cross-differentiated and divided by the sum of the temperature rises of the corresponding branches to construct a normalized diagnostic temperature matrix, so as to eliminate the influence of ambient temperature and power loss fluctuations.
[0050] A three-dimensional finite element thermo-mechanical coupling model of the IGBT module was established to simulate the changes in heat transfer performance of the substrate solder layer under different degradation states, and a case temperature simulation dataset including healthy, early degradation, severe degradation and failure states was generated.
[0051] An inverter experimental platform was built, and case temperature data under real operating conditions was collected. After Kalman filtering and sliding window averaging, an experimental test dataset was constructed; the measured dataset was used as the test dataset. The simulation dataset was used to train the model, and the measured dataset was used to test the model.
[0052] A convolutional neural network model is trained based on the simulation dataset. The network contains at least three two-dimensional convolutional layers and fully connected layers. The input is a normalized diagnostic temperature matrix, and the output is an IGBT health status category, realizing intelligent classification of the type and degree of uneven degradation of the solder layer.
[0053] It's understandable that the trained model achieves the required generalization ability. Deploying the trained model, using the test data as input, enables intelligent classification of the type and degree of solder layer non-uniform degradation. Naturally, the test data will be processed into the same format as the dataset used to train the model.
[0054] In a preferred embodiment of the IGBT health status highly robust intelligent diagnostic method based on case temperature distribution characteristics, the dimension of the case temperature matrix corresponds to the number and layout of power chips in the IGBT module, and the measurement point is located on the substrate directly below the IGBT chip or FWD chip.
[0055] In a preferred embodiment of the highly robust intelligent diagnostic method for IGBT health status based on shell temperature distribution characteristics, the normalized diagnostic temperature matrix is constructed as follows:
[0056] The four sets of measuring points A, B, C, and D correspond to the sum of temperature rises of the low-side IGBT, low-side FWD, high-side FWD, and high-side IGBT, respectively; calculate the relative temperature difference between adjacent branches:
[0057]
[0058] In the formula, ΔT1, ΔT2, ΔT3, ..., ΔT12 represent the difference between the measured temperature and the ambient temperature for numbers 1-12, respectively; ΔTA equals the sum of ΔT1, ΔT3 and ΔT5; ΔTB equals the sum of ΔT2, ΔT4 and ΔT6; ΔTC equals the sum of ΔT7, ΔT9 and ΔT11; and ΔTD equals the sum of ΔT8, ΔT10 and ΔT12.
[0059] In a preferred embodiment of the IGBT health status high robust intelligent diagnostic method based on case temperature distribution characteristics, the substrate solder layer degradation status includes: healthy status, early degradation with an area reduction of 20%-35%, severe degradation with an area reduction of 35%-50%, and failure status with an area reduction of at least 50%, and the degradation form is non-uniformly distributed.
[0060] In a preferred embodiment of the IGBT health status highly robust intelligent diagnostic method based on shell temperature distribution characteristics, the temperature data is processed by Kalman filtering, and noise is suppressed by using a sliding window to take the average value.
[0061] In a preferred embodiment of the IGBT health status highly robust intelligent diagnostic method based on shell temperature distribution characteristics, the convolutional neural network does not contain pooling layers. The input layer receives a normalized temperature matrix of size 2×2 or 3×4, extracts spatial features through multiple two-dimensional convolutional layers, and finally outputs the health status classification result by a fully connected layer.
[0062] In a preferred embodiment of the IGBT health status highly robust intelligent diagnostic method based on shell temperature distribution characteristics, random noise within the range of ±0.01℃ to ±0.1℃ is superimposed on the simulation data to simulate actual measurement errors.
[0063] A system for implementing the method includes:
[0064] The temperature sensing unit is used to arrange multiple temperature sensors on the substrate directly below the IGBT chip.
[0065] The data acquisition and preprocessing unit is used to acquire shell temperature data and perform ambient temperature compensation and normalized differential processing.
[0066] The finite element simulation module is used to build IGBT models and generate noisy multi-state shell temperature datasets.
[0067] A convolutional neural network diagnostic model is used to receive a normalized temperature matrix and output a health status classification result.
[0068] The display and alarm unit is used to visualize diagnostic results and issue warnings when degradation is detected.
[0069] A computer storage medium including computer instructions that, when run on a computer, cause the computer to perform the method.
[0070] An electronic device, the electronic device comprising:
[0071] Memory, processor, and computer programs stored in memory and executable on the processor, wherein,
[0072] The processor implements the method when executing the program.
[0073] In one embodiment, a commercial IGBT (FF150R12ME3G) is used as an example.
[0074] 1. Conduct accelerated aging tests to clarify the degradation mechanism of IGBTs.
[0075] During operation, IGBTs are subjected to thermal cycling stress, which can lead to varying degrees of degradation in the baseplate solder layer (BS). Due to differences in the solder layer application process during production and the thermal coupling effect inherent in IGBTs, this degradation typically manifests as uneven degradation.
[0076] To clarify the degradation mechanism of the IGBT substrate solder layer under thermal cycling stress during operation, a temperature cycling accelerated degradation experiment was conducted on brand-new IGBTs. The experiment settings were: temperature cycling from -55℃ to 150℃, with each cycle lasting 15 minutes. The experimental results showed that before cycling, due to differences in the manufacturing process, voids existed in the substrate solder layer. After 100 cycles, there was no significant degradation of the substrate solder layer. After 300 cycles, the substrate solder layer exhibited uneven degradation. Measurements of the solder layer area showed that the left solder layer degraded to 80% of its healthy state, the middle solder layer to 84%, and the right solder layer to 81%. After 400 cycles, the uneven degradation worsened further, with the left solder layer being the most severely degraded, reaching 70%.
[0077] 2. The shell temperature matrix is obtained by arranging sensors.
[0078] Since IGBT heat transfer is from top to bottom, deterioration of the intermediate substrate solder layer will affect the IGBT's heat transfer performance. Figure 2 As shown. Since the temperature of an IGBT chip is difficult to measure directly, this patent proposes using the temperature of the substrate directly beneath the chip, i.e., the case temperature, as a state parameter. Since there are multiple IGBT chips, such as... Figure 3 As shown. Therefore, the case temperatures of multiple chips can form a matrix, called the case temperature matrix. For the FF150R12ME3G model IGBT in this patent, which has 12 chips, the case temperature measurement points of the 12 chips are as follows... Figure 4 As shown.
[0079] 3. By normalizing the shell temperature matrix, errors caused by changes in operating conditions are reduced, and robustness is improved.
[0080] (1) Construction of shell temperature matrix
[0081] The two selected temperature measurement points are located on the outer surface of the substrate directly below each chip, such as... Figure 4 As shown. Therefore, the shell temperature matrix T case It can be expressed as the following equation
[0082]
[0083] Wherein, T1 / T3 / T5 represent the case temperature at measurement points 1 / 3 / 5 located below the low-side IGBT chip. T2 / T4 / T6 represent the case temperature at measurement points 2 / 4 / 6 located below the low-side FWD chip. T7 / T9 / T11 represent the case temperature at measurement points 7 / 9 / 11 located below the high-side FWD chip. T8 / T10 / T12 represent the case temperature at measurement points 8 / 10 / 12 located below the high-side IGBT chip.
[0084] (2) Shell temperature matrix preprocessing
[0085] To eliminate the influence of operating conditions, the shell temperature matrix is preprocessed. First, the shell temperature is subtracted from the ambient temperature to remove the influence of the ambient temperature, as expressed by the following equation.
[0086]
[0087]
[0088] Where Ti represents the shell temperature at measurement point i, and Th is the ambient temperature. Therefore, the shell temperature matrix Tcase can be converted into ΔTc_h.
[0089] Secondly, the column vectors of ΔTc_h are cross-subtracted to obtain information on the type and degree of degradation. The current of the parallel chips in the three parallel half-bridges is evenly distributed, and the power loss of the parallel chips is the same. The result of the subtraction is divided by the sum of the original values to remove the influence of power loss, thereby constructing a temperature matrix ΔTin that can be used for diagnosis.
[0090]
[0091] In the formula, ΔTA equals the sum of ΔT1, ΔT3, and ΔT5. ΔTB equals the sum of ΔT2, ΔT4, and ΔT6. ΔTC equals the sum of ΔT7, ΔT9, and ΔT11. ΔTD equals the sum of ΔT8, ΔT10, and ΔT12.
[0092] 4. Establish an IGBT finite element model, conduct simulations based on solder layer degradation simulations, and construct a dataset.
[0093] (1) Construct a finite element model based on the IGBT structural dimensions and material parameters, such as Figure 5 As shown.
[0094] Simulations were conducted, and the results showed that the surface temperature of the IGBT and FWD chips was higher than that of other surfaces. Furthermore, due to thermal coupling, the surface temperature of the IGBT chip located in the central part was higher than that of other IGBT chips. Figure 6 As shown. Shell temperature as Figure 7As shown, the shell temperature in the central region is higher than that in other regions. This uneven shell temperature distribution can serve as a monitoring indicator of uneven degradation.
[0095] (2) Training dataset setup
[0096] First, a training dataset is constructed using simulation data for training the neural network.
[0097] H, E, S, and F represent healthy, early degradation (20% area reduction), severe degradation (35% area reduction), and failure (50% area reduction), respectively. The degradation type of the substrate solder layer is set according to the actual testing requirements, such as... Figure 8 As shown.
[0098] Since the actual measurement used a PT1000 sensor with a temperature error of ±0.1℃, and the measured data underwent Kalman filtering, the final filtered temperature error was assumed to be less than ±0.1℃. Therefore, noise data ranging from ±0.01℃ to ±0.1℃ was added to the simulation data to simulate interference that might be encountered in actual measurements.
[0099] (3) Test dataset settings
[0100] An inverter experimental platform was built with an operating frequency of 50Hz, IGBT switching frequency of 10kHz, bus voltage of 250V, inductance of 2.5mH, and load resistance of 0.25Ω. Twelve PT1000 temperature sensors were used to measure the case temperature. The PT1000 sensors were connected to a multi-channel temperature acquisition unit with a sampling frequency of 10Hz. The case temperature takes approximately 5 minutes to reach steady state; to obtain more experimental data, the temperature acquisition time was set to 25 minutes. To reduce the impact of noise, Kalman filtering was applied to the acquired data, and the filtered temperature data was then averaged using a sliding window method. Figure 9 As shown, construct the experimental dataset.
[0101] 5. Train a neural network based on the dataset to achieve intelligent diagnosis of IGBT health status.
[0102] A neural network is constructed, specifically a convolutional neural network (CNN). The CNN consists of three convolutional layers and one fully connected layer. The first three convolutional layers use 3×3 kernels with a stride of 1, increasing the number of kernels by 8, 16, and 32 respectively. A ReLU activation function is used after each layer to introduce non-linearity. The final fully connected layer contains four output units, corresponding to four health state categories, and uses the SoftMax activation function to output the probability distribution for each category.
[0103] Set the hyperparameters and loss function for training. Use Adam as the optimizer, set the learning rate to 0.001, and the number of training epochs to 1000. Set the loss function to the mean squared error between the model's output class and the true class. Also, set an early stopping condition: stop training when the mean squared error is less than 1e-4.
[0104] Since the shell temperature matrix is a two-dimensional matrix, a CNN algorithm is chosen to classify different types and degrees of thermal degradation. A typical CNN consists of convolutional layers, pooling layers, and fully connected layers. Convolutional layers are used to extract features, pooling layers are used to reduce computational burden, and classification results are derived by fully connected layers. Because the input temperature matrix ΔTin is relatively small, pooling layers are not included in the proposed CNN framework. Three two-dimensional convolutional layers with fully connected layers are constructed, as follows: Figure 10 As shown.
[0105] Simulated data was used for training, and experimental data was used for testing and recognition. Due to the dispersed nature of training and recognition, each group underwent 5 training iterations, with 5 corresponding recognition iterations. The training and recognition results under different noise conditions are shown below. Figure 11 As shown, within the error range of ±0.01℃ to ±0.1℃, the training accuracy decreases with increasing noise; the experimental data recognition accuracy increases, and the dispersion decreases. At low noise levels, different category features are clearly distinguished, but the data cannot encompass all experimental data features, hence the training accuracy is high, but the recognition accuracy is low. Conversely, at high noise levels, the opposite is true. Regardless of the noise level, although the training accuracy decreases, the recognition accuracy can still reach an objective level of over 88%. This verifies the practicality of the proposed method.
[0106] Furthermore, this invention uses case temperature distribution characteristics as an indirect state parameter to effectively represent unmeasurable internal degradation. IGBT modules are typically fully packaged, and the junction temperature of their core chips is difficult to measure directly. Traditional single-point case temperature or average temperature cannot reflect local degradation information. This invention innovatively arranges sensors on the substrate directly below multiple chips to acquire spatially distributed case temperature data and construct a case temperature matrix. Since degradation of the substrate solder layer changes the local thermal resistance, leading to a decrease in heat transfer performance and an increase in case temperature in the corresponding area, the spatial distribution difference of case temperature can sensitively reflect the location and degree of uneven degradation of the solder layer. This method cleverly utilizes the measurable external temperature field to invert the internal material state, achieving non-invasive monitoring of package-level faults. A two-level normalization processing strategy is proposed to significantly improve the operational robustness of the diagnostic method. In actual operation, the power loss, ambient temperature, cooling conditions, and other operating conditions of IGBTs change frequently, causing overall case temperature drift and seriously interfering with health status assessment. This invention proposes a two-step normalization mechanism: first, by subtracting the ambient temperature Th, the influence of external temperature changes is eliminated; second, by utilizing the cross-difference and sum normalization of temperature rise between parallel branches, the power common-mode variation caused by load current fluctuations is effectively eliminated. This processing transforms absolute temperature into relative thermal distribution characteristics, making the diagnostic results mainly dependent on the asymmetry of the module's internal thermal resistance rather than the operating power, thereby significantly improving the algorithm's stability and generalization ability under varying operating conditions. A multi-state dataset integrating simulation and experiment is constructed to address the problem of scarce real aging data. The IGBT aging process is slow, long-term accelerated testing is costly, and it is difficult to obtain a sufficient number of complete lifecycle samples for machine learning training. This invention establishes a high-fidelity finite element thermo-mechanical coupling model to accurately simulate the shell temperature response under different solder layer degradation areas (e.g., 20%, 35%, 50% reduction) and non-uniform distribution patterns. Random noise is superimposed on the simulation data based on the actual sensor accuracy (±0.1℃) to generate a large number of labeled "quasi-realistic" training samples. This method overcomes the spatiotemporal limitations of experimental data acquisition, ensuring the diversity and coverage of the training set, providing a sufficient learning foundation for the neural network, and avoiding the risk of overfitting to a single experimental condition. A lightweight CNN architecture is designed to achieve efficient and accurate end-to-end health status classification. Shell temperature distribution is essentially a two-dimensional spatial feature map, and traditional machine learning methods (such as SVM and BP neural networks) struggle to effectively extract its spatial correlations. This invention employs a convolutional neural network (CNN) to automatically extract local pattern features (such as high-temperature zones in the center and abnormal temperature differences at the edges) from the temperature matrix using convolutional kernels, making it particularly suitable for identifying regional degradation of the solder layer. To address the small input matrix size (e.g., 3×4), pooling layers are omitted to retain all spatial information. A three-layer convolutional + fully connected structure is adopted, reducing model complexity while maintaining classification accuracy. This makes it suitable for deployment on embedded or edge computing platforms, meeting the real-time requirements of online diagnostics.To verify the practicality of the closed-loop design guarantee method and achieve a leap from theory to application, this invention not only relies on simulation training but also collects test data through a real inverter experimental platform to independently verify the trained model. Experimental results show that after introducing Kalman filtering and sliding window averaging, even though the training accuracy decreases with increasing noise, the model's recognition accuracy for experimental data remains stable at over 88%, demonstrating the strong robustness of the method in real-world noise environments. This closed-loop design process of "simulation modeling → data augmentation → model training → experimental verification" ensures that the diagnostic method is not only theoretically sound but also possesses practical engineering application value.
[0107] Although embodiments of the present invention have been described above in conjunction with the accompanying drawings, the present invention is not limited to the specific embodiments and application fields described above. The specific embodiments described above are merely illustrative and instructive, and not restrictive. Those skilled in the art can make many other forms based on the guidance of this specification and without departing from the scope of protection of the claims of the present invention, and all of these are within the scope of protection of the present invention.
Claims
1. A robust intelligent diagnostic method for IGBT health status based on shell temperature distribution characteristics, characterized in that, Includes the following steps: Temperature sensors are arranged on the outer surface of the substrate directly below multiple chips of the IGBT module to collect case temperature data at each measuring point and construct a case temperature matrix to characterize the spatial distribution characteristics of the case temperature of the IGBT module. The shell temperature matrix is preprocessed by subtracting the ambient temperature from the temperature at each measuring point to obtain the temperature rise matrix. Then, based on the parallel half-bridge structure in the IGBT module, the temperature rise values of the corresponding parallel branches are cross-differentiated and divided by the sum of the temperature rises of the corresponding branches to construct a normalized diagnostic temperature matrix, so as to eliminate the influence of ambient temperature and power loss fluctuations. A three-dimensional finite element thermo-mechanical coupling model of the IGBT module was established to simulate the changes in heat transfer performance of the substrate solder layer under different degradation states, and a case temperature simulation dataset including healthy, early degradation, severe degradation and failure states was generated. An inverter experimental platform was built, and the case temperature data under real operating conditions was collected. After Kalman filtering and sliding window averaging, an experimental test dataset was constructed. A convolutional neural network model is trained based on the simulation dataset. The network contains at least three two-dimensional convolutional layers and fully connected layers. The input is a normalized diagnostic temperature matrix, and the output is an IGBT health status category, realizing intelligent classification of the type and degree of uneven degradation of the solder layer.
2. The robust intelligent diagnostic method for IGBT health status based on shell temperature distribution characteristics according to claim 1, characterized in that, Preferably, the dimensions of the case temperature matrix correspond to the number and layout of the power chips in the IGBT module, and the measurement points are located on the substrate directly below the IGBT chip or FWD chip.
3. The robust intelligent diagnostic method for IGBT health status based on shell temperature distribution characteristics according to claim 1, characterized in that, The normalized diagnostic temperature matrix is constructed as follows: The four sets of measuring points A, B, C, and D correspond to the sum of temperature rises of the low-side IGBT, low-side FWD, high-side FWD, and high-side IGBT, respectively; calculate the relative temperature difference between adjacent branches: ； In the formula, ΔTA equals the sum of ΔT1, ΔT3 and ΔT5, ΔTB equals the sum of ΔT2, ΔT4 and ΔT6, ΔTC equals the sum of ΔT7, ΔT9 and ΔT11, and ΔTD equals the sum of ΔT8, ΔT10 and ΔT12.
4. The robust intelligent diagnostic method for IGBT health status based on shell temperature distribution characteristics according to claim 1, characterized in that, The degradation state of the substrate solder layer includes: healthy state, early degradation with an area reduction of 20%-35%, severe degradation with an area reduction of 35%-50%, and failure state with an area reduction of at least 50%, and the degradation form is unevenly distributed.
5. The robust intelligent diagnostic method for IGBT health status based on shell temperature distribution characteristics according to claim 1, characterized in that, Temperature data are processed by Kalman filtering, and noise is suppressed by using a sliding window to take the average value.
6. The robust intelligent diagnostic method for IGBT health status based on shell temperature distribution characteristics according to claim 1, characterized in that, The convolutional neural network does not contain pooling layers. The input layer receives a normalized temperature matrix of size 2×2 or 3×4, extracts spatial features through multiple two-dimensional convolutional layers, and finally outputs the health status classification result by a fully connected layer.
7. The robust intelligent diagnostic method for IGBT health status based on shell temperature distribution characteristics according to claim 1, characterized in that, Random noise within the range of ±0.01℃ to ±0.1℃ is superimposed on the simulation data to simulate actual measurement errors.
8. A system for implementing the method according to any one of claims 1-7, characterized in that, It includes: The temperature sensing unit is used to arrange multiple temperature sensors on the substrate directly below the IGBT chip. The data acquisition and preprocessing unit is used to acquire shell temperature data and perform ambient temperature compensation and normalized differential processing. The finite element simulation module is used to build IGBT models and generate noisy multi-state shell temperature datasets. A convolutional neural network diagnostic model is used to receive a normalized temperature matrix and output a health status classification result; The display and alarm unit is used to visualize diagnostic results and issue warnings when degradation is detected.
9. A computer storage medium, characterized in that, The storage medium includes computer instructions that, when executed on a computer, cause the computer to perform the method as described in any one of claims 1-7.
10. An electronic device, characterized in that, The electronic device includes: Memory, processor, and computer programs stored in memory and executable on the processor, wherein, When the processor executes the program, it implements the method as described in any one of claims 1-7.