Self-adaptive nanosecond driving protection method for high-voltage IGBT (Insulated Gate Bipolar Translator) power module

Abstract

The invention discloses a self-adaptive nanosecond driving protection method for a high-voltage IGBT power module. The method comprises the steps that data such as high-voltage side voltage, collector current, grid voltage and junction temperature are collected and preprocessed, and standardized feature vectors are obtained; extracting statistical features, frequency domain features and trend features to obtain a comprehensive feature vector; density clustering is carried out based on historical fault data, and an extreme value model is established by using generalized Pareto distribution to generate an optimized threshold value; constructing a time sequence feature matrix, predicting a fault probability by adopting a time sequence convolutional network and a long-short term memory network, and generating a protection strategy parameter; and generating an optimal driving waveform and a control signal, and updating parameters according to an actual response effect. According to the invention, predictive protection and nanosecond-level precise control of faults are realized, and the reliability and performance of the high-voltage IGBT module are remarkably improved.

Description

Technical Field

[0001] The present invention relates to IGBT technology, in particular to an adaptive nanosecond-level drive protection method for high-voltage IGBT power modules. Background Art

[0002] As a core device in modern power electronic systems, high-voltage IGBT power modules are widely used in fields such as high-speed railway traction, smart grid, and new energy power generation. With the development of power electronic technology towards high frequency and high power density, IGBT modules are facing a more complex working environment and more stringent performance requirements. During the nanosecond-level switching process, how to achieve precise drive control and reliable protection is related to the safe and stable operation of the entire system, and has important research significance and application value. Especially for high-voltage applications, problems such as overvoltage, overcurrent, and parasitic oscillation during the switching process are more prominent, which may lead to equipment damage or even system collapse. Therefore, it is necessary to achieve precise protection control on the nanosecond time scale to improve system reliability and extend equipment life.

[0003] Traditional IGBT drive protection methods mainly adopt methods such as overvoltage and overcurrent protection with fixed thresholds, soft turn-off technology, and active clamping circuits. Fixed-threshold protection is usually based on preset voltage and current limits, and once the detected parameters exceed the threshold, the protection action is triggered, lacking the adaptive ability to working condition differences; soft turn-off technology reduces the impact stress by reducing the switching speed, but often leads to an increase in switching losses; the active clamping circuit can effectively suppress overvoltage, but increases system complexity and cost. In addition, most of the existing protection methods adopt single-parameter monitoring or simple multi-parameter combinations, lacking comprehensive analysis and excavation of multi-dimensional features, and it is difficult to accurately identify complex fault modes. In terms of data processing, common filtering and denoising technologies are difficult to effectively process complex noise in a high-voltage environment, resulting in poor signal quality and affecting the accuracy of judgment.

[0004] At present, IGBT drive protection technology faces several key technical problems: First, most of the existing protection methods are post-fault protection, lacking effective fault prediction ability and unable to take preventive measures before a fault occurs; Second, fixed-threshold protection cannot adapt to the protection requirements under different working conditions, and it is easy to have over-protection or under-protection problems, affecting the system performance and safety; Third, the traditional drive waveform design is relatively simple, difficult to perform fine control for different switching stages, and unable to achieve the best balance between switching loss and EMI suppression; Fourth, the existing system lacks an effective closed-loop adaptive mechanism and cannot automatically optimize the control strategy according to the actual response effect, resulting in a decline in long-term operating performance; Fifth, in a high-noise environment, traditional signal processing methods are difficult to extract key features, especially on the nanosecond time scale, unable to capture small but critical parameter changes, resulting in some potential faults not being identified in time. These problems severely restrict the performance and reliable operation of high-voltage IGBT modules under extreme working conditions. Summary of the Invention

[0005] The object of the invention is to provide an adaptive nanosecond drive protection method for high-voltage IGBT power modules, in order to solve one of the above problems.

[0006] Technical solution: An adaptive nanosecond drive protection method for high-voltage IGBT power modules includes the following steps: Collect high-voltage side voltage data, collector current data, gate voltage data, junction temperature data, switch status signals and historical fault data sets, and preprocess the data to obtain standardized feature vectors; Extract statistical features, frequency domain features and trend features from the standardized feature vectors to obtain comprehensive feature vectors; Perform density clustering based on the historical fault data set, establish an extreme value model in combination with the generalized Pareto distribution, and generate an optimized threshold; Construct a time series feature matrix, and use a time series convolutional network and a long short-term memory network for processing to generate protection strategy parameters; Generate a drive control signal based on the protection strategy parameters, monitor the actual response effect and update the control parameters.

[0007] By combining a time series convolutional network and a long short-term memory network, and combining density clustering with the generalized Pareto distribution to establish an extreme value model, solve the problems of untimely protection and non-adaptive protection strategy of high-voltage IGBTs during high-speed switching. Achieve nanosecond-level drive protection response and improve the reliability and lifespan of IGBT modules.

[0008] According to one aspect of the present application, the preprocessing includes: Perform wavelet denoising on the high-voltage side voltage data, collector current data and gate voltage data to obtain denoised voltage, denoised current and denoised gate voltage; Perform moving average filtering on the junction temperature data to obtain the smoothed temperature; Calculate the voltage change rate of the noise-reduced voltage and the current change rate of the noise-reduced current; Perform Z-score standardization on the noise-reduced voltage, noise-reduced current, noise-reduced gate voltage, smoothed temperature, voltage change rate, and current change rate to obtain the standardized feature vector.

[0009] In data preprocessing, combining wavelet denoising and Z-score standardization, different preprocessing methods are used for different types of data (wavelet denoising for voltage / current, moving average filtering for temperature). Solve the problems of large noise in the original data and difficulty in comparing data with different dimensions. Improve the accuracy of subsequent feature extraction and analysis and reduce the misjudgment rate.

[0010] According to one aspect of the present application, the wavelet denoising process includes: Perform wavelet basis function decomposition on the high-side voltage data, collector current data, and gate voltage data to obtain the wavelet coefficient sequence; Calculate the energy distribution of the wavelet coefficient sequence to obtain the energy distribution sequence, and set the threshold according to the energy distribution sequence; Perform soft threshold processing on the wavelet coefficient sequence to obtain the processed coefficient sequence, and use the processed coefficient sequence for wavelet reconstruction to obtain the noise-reduced voltage, noise-reduced current, and noise-reduced gate voltage.

[0011] Through the adaptive wavelet threshold setting method based on energy distribution. Perform energy analysis on the wavelet coefficients and set the threshold accordingly, rather than using a fixed threshold. Solve the problem that the traditional fixed threshold denoising method may lose valid information. Effectively remove noise while retaining the signal characteristics and improve the signal quality.

[0012] According to one aspect of the present application, the steps of extracting statistical features and frequency domain features include: Based on the preset window length and sliding step size, segment the standardized feature vector to obtain the time series labeled data segments; Calculate the mean, variance, peak value, and valley value of the time series labeled data segments to obtain the statistical features; Perform fast Fourier transform on the time series labeled data segments and calculate the power spectral density to obtain the frequency domain features.

[0013] Through the multi-dimensional feature extraction method combining time domain and frequency domain features. Adopt the segmentation processing method with the preset window length and sliding step size to retain the time series information. Solve the problem that a single feature cannot comprehensively reflect the system state. Provide a more comprehensive system state representation and enhance the fault recognition ability.

[0014] According to one aspect of the present application, the steps of extracting trend features include: The empirical mode decomposition algorithm is used to process the standardized feature vector to obtain the signal trend component and the high-frequency component; All local extreme points in the standardized feature vector are identified to obtain an extreme point sequence; The cubic spline interpolation method is used to process the extreme point sequence to obtain the upper and lower envelope lines and the mean envelope line; The standardized feature vector is subtracted by the mean envelope line to obtain a candidate component, and iterative processing is performed until the intrinsic mode function condition is satisfied; The change rate of the signal trend component is calculated to obtain the trend feature vector.

[0015] The empirical mode decomposition method is used to extract the signal trend feature. The cubic spline interpolation method is used to construct the upper and lower envelope lines, and the intrinsic mode function condition is satisfied through iterative processing. The problem that traditional signal processing methods are difficult to effectively separate the trend component and the high-frequency component is solved. The system operation trend is captured more accurately, and potential faults can be discovered in advance.

[0016] According to one aspect of the present application, density clustering includes: The Euclidean distance between data points in the historical fault dataset is calculated, a density threshold is set, and high-density data points are identified to obtain a core point set; Data points that are density-reachable are merged to obtain an initial clustering result, noise points are removed to obtain an optimized clustering result, and a fault mode is generated; The geometric center and covariance matrix of the fault mode are calculated, a fault feature library is constructed, and feature distribution parameters are obtained.

[0017] The density-based fault data clustering method is applied to IGBT fault mode recognition. By identifying high-density data points to form a core point set and removing noise points to optimize the clustering result. The problems that traditional clustering methods are sensitive to noise and the clustering results are unstable are solved. A more accurate fault feature library is formed, and the fault mode recognition rate is improved.

[0018] According to one aspect of the present application, the steps for generating an optimized threshold include: Based on the feature distribution parameters, a probability density function of the generalized Pareto distribution is constructed; An extreme value model is established using the maximum likelihood estimation method, and the dynamic protection threshold and the warning threshold are calculated; The confidence index of the current working condition is calculated, and the threshold is adjusted according to the confidence index to obtain the optimized threshold.

[0019] Through the extreme value model and the dynamic threshold generation method based on the generalized Pareto distribution. The confidence index is introduced to dynamically adjust the threshold to achieve adaptive optimization. The problem that fixed thresholds cannot adapt to different working conditions is solved. The false alarm and missed alarm rates are reduced, and the accuracy and reliability of the protection strategy are improved.

[0020] According to one aspect of the present application, the steps of constructing a time-series feature matrix and generating protection strategy parameters include: Sort the most recent comprehensive feature vectors in chronological order, and after segmented alignment, reconstruct them into a time-series feature matrix; Perform one-dimensional convolution and pooling operations on the time-series feature matrix to obtain time-series features; Input the time-series features into a long short-term memory network to predict the fault probability and reliability index; Generate protection strategy parameters based on the fault probability and reliability index.

[0021] Apply the fusion of a time-series convolutional network and a long short-term memory network to fault prediction. Extract time-series features through convolution and pooling operations, and then use the LSTM network to predict the fault probability. Solve the problem that traditional fault prediction methods are difficult to handle complex time-series correlations. Improve the accuracy and lead time of fault prediction, and provide a more sufficient decision-making basis for drive protection.

[0022] According to one aspect of the present application, the steps of generating a drive control signal include: Calculate the optimal drive voltage and optimal drive current based on the protection strategy parameters; Determine the drive timing parameters and generate a drive waveform; Calculate the compensation coefficient and output the drive control signal.

[0023] An optimal drive parameter generation method based on protection strategy parameters. Convert the fault prediction result into specific drive voltage, current, and timing parameters. Solve the problem of mismatch between drive parameters and protection strategies. Achieve precise control of drive parameters, and improve the system response speed and stability.

[0024] According to one aspect of the present application, the steps of calculating the compensation coefficient include: Extract waveform distortion data from the historical fault dataset and perform spectral analysis to obtain the initial compensation coefficient; Calculate the working condition adjustment factor based on the current working condition feature vector; Perform weighted calculation on the initial compensation coefficient and the adjustment factor sequence to obtain the real-time compensation coefficient; Correct the real-time compensation coefficient based on a preset compensation reference value to obtain the compensation coefficient.

[0025] Through a compensation coefficient calculation method based on historical fault waveform distortion data. Combine the working condition feature vector to adjust the real-time compensation coefficient to achieve adaptive compensation. Solve the problem that the drive waveform is distorted due to external factors during actual use. Improve the quality of the drive signal, reduce waveform distortion, and enhance drive accuracy.

[0026] Beneficial effects: predictive protection against faults and nanosecond-level precise control are achieved, significantly improving the reliability and performance of high-voltage IGBT modules. Description of the Drawings

[0027] Figure 1 is the flowchart of the present invention.

[0028] Figure 2 is the flowchart for preprocessing the data of the present invention.

[0029] Figure 3 is the flowchart for wavelet denoising processing of the present invention.

[0030] Figure 4 is the flowchart for extracting statistical features and frequency-domain features of the present invention.

[0031] Figure 5 is the flowchart for extracting trend features of the present invention.

[0032] Figure 6 is the flowchart for density clustering of the present invention. Detailed Embodiments

[0033] As Figure 1 shown, an adaptive nanosecond-level drive protection method for high-voltage IGBT power modules Step S1: Collect high-voltage side voltage data, collector current data, gate voltage data, junction temperature data, switch state signals, and historical fault data sets. Perform wavelet denoising processing on the collected high-voltage side voltage data, collector current data, and gate voltage data to obtain denoised voltage, denoised current, and denoised gate voltage. Perform moving average filtering on the junction temperature data to obtain smoothed temperature. Calculate the voltage change rate of the denoised voltage to obtain the voltage change rate. Calculate the current change rate of the denoised current to obtain the current change rate. Perform Z-score standardization on the denoised voltage, denoised current, denoised gate voltage, smoothed temperature, voltage change rate, and current change rate to obtain a standardized feature vector.

[0034] Step S2: Based on a preset window length and sliding step size, extract statistical features and frequency-domain features from the standardized feature vector to obtain statistical features and frequency-domain features. Use the empirical mode decomposition algorithm to process the standardized feature vector to obtain a signal trend component and a high-frequency component. Calculate the change rate of the signal trend component to obtain a trend feature vector. Perform fusion processing on the statistical features, frequency-domain features, and trend feature vector to obtain a comprehensive feature vector.

[0035] Step S3: Conduct density clustering analysis based on the historical fault data set to obtain a fault feature library and fault modes, calculate the characteristic distribution parameters of the fault modes to obtain the characteristic distribution parameters, establish an extreme value model based on the generalized Pareto distribution and combine with the characteristic distribution parameters, calculate to obtain the dynamic protection threshold and the warning threshold, calculate the confidence index of the current working condition to obtain the confidence index, and adjust the dynamic protection threshold and the warning threshold according to the confidence index to obtain the optimized threshold.

[0036] Step S4: Construct a time-series feature matrix containing the most recent time-series data, use a time-series convolutional network to process the time-series feature matrix to obtain time-series features, calculate the correlation of the time-series features to obtain the feature correlation matrix, input the time-series features into a long short-term memory network for fault prediction to obtain the fault probability and the reliability index, and generate protection strategy parameters based on the fault probability and the reliability index.

[0037] Step S5: Calculate the optimal drive voltage and the optimal drive current based on the protection strategy parameters, determine the drive timing parameters, generate a drive waveform, calculate the compensation coefficient, output a drive control signal, monitor the actual response effect to obtain the actual response effect, update the control parameter library to obtain the control parameter library, and adjust the compensation coefficient according to the actual response effect.

[0038] Through the organic combination of data acquisition and preprocessing, multi-dimensional feature extraction, adaptive threshold generation, time-series deep learning, and nanosecond-level drive control, a complete set of adaptive nanosecond-level drive protection systems for high-voltage IGBT power modules is constructed. The overall technical solution forms a closed-loop adaptive control process of "perception - analysis - prediction - execution - optimization", realizing all-round, multi-level, and intelligent protection of high-voltage IGBT modules. At the perception level, high-quality operating data is obtained through methods such as wavelet denoising and moving average filtering; at the analysis level, the working state of the system is comprehensively characterized through the extraction and fusion of statistical features, frequency-domain features, and trend features; at the prediction level, combined with density clustering, generalized Pareto distribution, and deep learning methods, accurate fault prediction is realized; at the execution level, through the calculation of optimal drive parameters and the generation of complex drive waveforms, nanosecond-level precise control is achieved; at the optimization level, the control strategy is continuously adjusted through closed-loop feedback to continuously improve the system performance. Compared with traditional IGBT protection methods, this solution has significant advantages such as fast response speed (nanosecond level), strong adaptability, good predictability, and high protection accuracy. In high-voltage high-power applications, this solution can effectively prevent the damage of IGBTs caused by problems such as overvoltage, overcurrent, and sudden temperature changes during the switching process, significantly improving the system reliability; at the same time, by optimizing the drive strategy, the switching loss is reduced, the energy conversion efficiency is improved, and the EMI interference is reduced, providing technical support for the wide application of high-voltage IGBTs in fields such as smart grids, high-speed railways, and new energy power generation, with significant economic and social benefits.

[0039] According to one aspect of the present application, step S1 is specifically as follows: Step S11: Collect high-voltage side voltage data, collector current data, gate voltage data, junction temperature data, and the high and low level states of the switch status signal, and read the historical fault data set including fault types and fault times.

[0040] Step S12: Perform wavelet denoising processing on the collected high-voltage side voltage data, collector current data, and gate voltage data to obtain denoised voltage, denoised current, and denoised gate voltage. Perform moving average filtering processing on the collected junction temperature data to obtain smoothed temperature. Calculate the voltage change rate based on the denoised voltage, and calculate the current change rate based on the denoised current.

[0041] Step S13: Perform Z-score standardization processing on the denoised voltage, denoised current, denoised gate voltage, smoothed temperature, voltage change rate, and current change rate to generate a standardized feature vector.

[0042] By performing wavelet denoising processing on the high-voltage side voltage data, collector current data, and gate voltage data, performing moving average filtering processing on the junction temperature data, calculating the voltage change rate and the current change rate, and then performing Z-score standardization processing to form a standardized feature vector, high-quality data acquisition and preprocessing of the core operating parameters of the IGBT module are realized. Wavelet denoising processing can effectively remove electromagnetic interference and noise in the high-voltage environment and retain the effective components of the signal, especially the transient characteristic information in the nanosecond-level switching process; moving average filtering processing smooths the fluctuations of the junction temperature data and eliminates the random errors in the temperature sensor acquisition process, making the temperature data better reflect the true temperature state of the IGBT chip; the calculation of the voltage change rate and the current change rate reflects the dynamic characteristics in the switching process of the power module and is a key indicator for predicting overvoltage and overcurrent faults; Z-score standardization processing eliminates the scale differences between parameters with different dimensions, enabling the parameters to be compared and analyzed under a unified standard, laying a foundation for subsequent feature extraction and pattern recognition. In the high-voltage IGBT module, this step can obtain high-quality key operating parameters on the nanosecond time scale, provide data support for realizing accurate and rapid protection decisions, effectively prevent misjudgment or missed judgment caused by data distortion, and improve the reliability and safety of the system.

[0043] According to one aspect of the present application, step S2 is specifically as follows: Step S21: Select a preset window length and sliding step size, calculate the mean, variance, peak value, and valley value of the standardized feature vector within the window to obtain statistical features, and perform power spectral density analysis on the standardized feature vector within the window to obtain frequency domain features.

[0044] Step S22: Decompose the standardized feature vector using the empirical mode decomposition algorithm to obtain the signal trend component and the high-frequency component, calculate the change rate of the signal trend component, and obtain the trend feature vector.

[0045] Step S23: Perform feature fusion processing on the statistical features, frequency domain features, and trend feature vector, and output the comprehensive feature vector.

[0046] Based on the preset window length and sliding step size, statistical features and frequency domain features are extracted from the standardized feature vector, and the empirical mode decomposition algorithm is used to process the standardized feature vector. The change rate of the signal trend component is calculated to obtain the trend feature vector, and finally, they are fused into a comprehensive feature vector, realizing the multi-dimensional and all-round feature characterization of the IGBT module's working state. Statistical features reflect the distribution characteristics and change trends of data, and are particularly effective for identifying abnormal fluctuations and short-term shocks; frequency domain features reveal the frequency composition of signals through power spectral density analysis, and can detect periodic anomalies and harmonic interferences that are not easily found in time-domain analysis; empirical mode decomposition decomposes complex signals into trend components and high-frequency components, where the trend component reflects the long-term change law of the signal, and the high-frequency component corresponds to the fast-changing characteristics on a short time scale. This multi-dimensional feature extraction method can capture the subtle changes of the high-voltage IGBT module under different working states on the nanosecond time scale, especially the abnormal phenomena during the switching transient process, such as Miller plateau anomalies, switching spike currents, voltage oscillations, etc., which are precursors to faults. Through feature fusion processing, the feature information of different dimensions is organically integrated, forming a comprehensive description of the IGBT working state, providing rich feature basis for subsequent fault warning and protection, significantly improving the accuracy and timeliness of fault identification, and providing a solid foundation for nanosecond-level protection decision-making.

[0047] According to one aspect of the present application, step S3 is specifically as follows: Step S31: Perform density clustering analysis on the historical fault data set, construct a fault feature library, identify fault patterns, and calculate the feature distribution of the fault patterns to obtain feature distribution parameters.

[0048] Step S32: Establish an extreme value model based on the generalized Pareto distribution, calculate the dynamic protection threshold in combination with the feature distribution parameters, and calculate the warning threshold according to the preset safety factor and the dynamic protection threshold.

[0049] Step S33: Calculate the confidence index under the current working condition, dynamically adjust the dynamic protection threshold and the warning threshold according to the confidence index, and output the optimized threshold.

[0050] Based on the historical fault data set, density clustering analysis is carried out to obtain a fault feature library and fault modes. The characteristic distribution parameters are calculated and an extreme value model is established based on the generalized Pareto distribution. The threshold is adjusted by combining the confidence index of the current working condition, forming a set of adaptive dynamic protection thresholds and warning thresholds. The density clustering method can effectively identify different types of fault modes, especially suitable for dealing with non-spherical distributed data in high-dimensional feature spaces. Compared with traditional clustering methods such as K-means, it can more accurately capture the distribution characteristics of complex fault modes of high-voltage IGBT modules; the construction of the fault feature library accumulates the "experience" of the system and can be continuously enriched and improved as the operation time increases; the generalized Pareto distribution is a probability distribution specifically used for modeling extreme events, especially suitable for describing extreme fault phenomena such as overvoltage and overcurrent in high-voltage IGBT modules. The protection threshold established through extreme value theory has a solid statistical basis; the introduction of the confidence index takes into account the similarity between the current working condition and historical fault data, enabling the protection threshold to be adaptively adjusted according to changes in working conditions. This step realizes the transformation from "fixed threshold protection" to "dynamic threshold protection", can automatically adjust the protection sensitivity under different working conditions, maximize the performance of the IGBT module on the premise of ensuring safety, avoid overprotection or underprotection problems in traditional fixed threshold protection, and significantly improve the reliability and stability of high-voltage IGBT modules under extreme working conditions.

[0051] According to one aspect of the present application, step S4 is specifically as follows: Step S41: Construct a time series feature matrix based on the most recent time series data, process the time series feature matrix using a time series convolutional network, extract time series features, and calculate the correlation of the time series features to obtain a feature correlation matrix.

[0052] Step S42: Input the time series features into a long short-term memory network, predict the fault probability, and generate a reliability index.

[0053] Step S43: Based on the fault probability and the reliability index, generate protection strategy parameters including drive strength and response time.

[0054] Construct a time series feature matrix, obtain time series features after processing with a time series convolutional network, calculate the feature correlation matrix, and input the time series features into a long short-term memory network for fault prediction to generate protection strategy parameters, realizing predictive protection for potential faults of IGBT modules. The time series feature matrix retains the time dependence of the data and can capture the evolution law of parameters over time; the time series convolutional network effectively extracts local features and multi-scale patterns in the time series data through one-dimensional convolution and pooling operations. Compared with traditional filter and feature engineering methods, it can automatically learn more complex time series features; the feature correlation matrix reflects the mutual influence between different features and helps to understand the coupling relationship between system parameters; the long short-term memory network, with its gating mechanism and memory unit, can capture long-term dependence relationships and accurately predict the fault development trend of IGBT modules. This method combining deep learning and time series analysis can identify fault precursors within hundreds of nanoseconds to several microseconds before a fault occurs, providing valuable warning time for the system and achieving a leap from "post-fault protection" to "predictive protection". By generating protection strategy parameters including drive strength and response time, the system can formulate precise coping strategies for different types and severities of potential faults, minimizing the fault risk to the greatest extent while minimizing the impact of protection actions on the normal operation of the system, greatly improving the reliability and robustness of high-voltage IGBT modules under harsh working conditions.

[0055] According to one aspect of the present application, step S5 is specifically as follows: Step S51: Calculate the optimal drive voltage based on the protection strategy parameters, calculate the optimal drive current, and determine the drive timing parameters.

[0056] Step S52: Generate a drive waveform, calculate the compensation coefficient, and output a drive control signal.

[0057] Step S53: Monitor the actual response effect to obtain the actual response effect, update the control parameter library to obtain the control parameter library, and adjust the compensation coefficient according to the actual response effect.

[0058] Based on the protection strategy parameters, the optimal drive voltage and current are calculated, the drive timing parameters are determined, the drive waveform is generated, the compensation coefficient is calculated, the drive control signal is output, and the control parameter library and the compensation coefficient are updated according to the actual response effect, forming a closed-loop adaptive drive control system. The calculation of the optimal drive voltage and current takes into account the characteristics of the current working condition, and can suppress overshoot and oscillation while ensuring the switching speed; the determination of the drive timing parameters optimizes each time period during the switching process, achieving a balance between switching loss and EMI suppression; the generation of the drive waveform is no longer limited to simple square waves, but can form complex multi-level drive waveforms according to needs to more precisely control the switching process; the introduction of the compensation coefficient takes into account the influence of factors such as temperature and load on the drive effect, and offsets these effects through real-time adjustment. This closed-loop adaptive mechanism can continuously optimize the drive strategy according to the actual response effect, enabling the system to have the ability of "learning". As the running time increases, the control effect will become more and more precise. For high-voltage IGBT modules, this step realizes nanosecond-level precise drive control, significantly reduces the switching loss, suppresses EMI interference, reduces thermal stress, extends the device life, and at the same time improves the adaptability and reliability of the system under harsh working conditions, providing a solid guarantee for the efficient and stable operation of high-voltage high-power conversion systems.

[0059] According to one aspect of the present application, step S12 is specifically as follows: Step S121: Perform wavelet basis function decomposition on the high-side voltage data, collector current data, and gate voltage data to obtain a wavelet coefficient sequence, calculate the energy distribution of the wavelet coefficient sequence to obtain an energy distribution sequence, and set a threshold according to the energy distribution sequence to obtain a threshold sequence.

[0060] Step S122: Perform soft threshold processing on the wavelet coefficient sequence to obtain a processed coefficient sequence, and use the processed coefficient sequence for wavelet reconstruction to obtain the noise-reduced voltage, noise-reduced current, and noise-reduced gate voltage.

[0061] Step S123: Slide the junction temperature data according to a preset window length, calculate the weighted average value of the data in each window to obtain the smoothed temperature, perform differential operations on the noise-reduced voltage and noise-reduced current and divide by the sampling time interval to obtain the voltage change rate and current change rate respectively.

[0062] By performing wavelet basis function decomposition on the high-voltage side voltage data, collector current data, and gate voltage data, calculating the energy distribution setting threshold, performing soft threshold processing and reconstruction to obtain noise-reduced data, and at the same time performing sliding window weighted averaging on the junction temperature data and calculating the change rates of voltage and current, the refined preprocessing of the key parameters of the IGBT module is achieved. Wavelet basis function decomposition can decompose the signal into different frequency scales, which is particularly suitable for processing non-stationary signals containing short transients and discontinuities, such as the voltage and current waveforms during the IGBT switching process; the energy distribution calculation and threshold setting achieve adaptive noise recognition, distinguishing the effective components and noise components in the signal; compared with hard threshold processing, soft threshold processing can suppress noise while maintaining the continuity of the signal, avoiding unnatural breakpoints in the processed signal; the weighted averaging process of the junction temperature data assigns higher weights to recent temperature data, which not only smooths random fluctuations but also retains the temperature change trend; the calculation of the change rates of voltage and current captures the dynamic characteristics during the switching process and is a key indicator for identifying hard switching, soft switching, and their abnormal states. The combined application of these preprocessing techniques enables the system to obtain working parameter data with high signal-to-noise ratio in a strong interference environment, especially accurately capturing the rapid changes of voltage and current on the nanosecond time scale, laying a solid foundation for subsequent feature extraction and analysis. High-quality preprocessed data significantly improves the accuracy and timeliness of fault recognition, reduces the false alarm and missed alarm rates, provides a necessary condition for achieving high-precision nanosecond-level protection, and at the same time provides reliable data support for system performance optimization and fault analysis.

[0063] According to one aspect of the present application, step S21 is specifically as follows: Step S211: Segment the standardized feature vector according to a preset window length to generate feature data segments, perform overlapping movement on the feature data segments with a movement step size of a preset value to obtain an overlapping data segment sequence, and mark each data segment in the overlapping data segment sequence to obtain a time-series marked data segment.

[0064] Step S212: Calculate the arithmetic mean of each data in the time-series marked data segment to obtain a mean sequence, calculate the sum of the squared deviations of each data in the time-series marked data segment from the arithmetic mean and divide by the number of data to obtain a variance sequence, find the maximum value in the time-series marked data segment to obtain a peak sequence, find the minimum value in the time-series marked data segment to obtain a valley sequence, and combine the mean sequence, variance sequence, peak sequence, and valley sequence to obtain statistical features.

[0065] Step S213: Perform fast Fourier transform on the time-series marked data segment to obtain frequency components, calculate the square of the frequency components to obtain a power sequence, divide the power sequence by the number of sampling points to obtain a power density sequence, and perform normalization processing on the power density sequence to obtain frequency-domain features.

[0066] By selecting a preset window length and sliding step size, calculating the mean, variance, peak value, and valley value of the normalized feature vectors within the window to obtain statistical features, and performing power spectral density analysis on the normalized feature vectors within the window to obtain frequency-domain features, a dual characterization of the operating state of the IGBT module in the time domain and frequency domain is achieved. The preset of the window length balances the time resolution and statistical reliability, and the setting of the sliding step size ensures the continuity and overlap degree of feature extraction; the calculation of statistical features reflects the central tendency and dispersion degree of the data within the window. The mean represents the average level of the parameter, the variance measures the degree of data fluctuation, and the peak value and valley value capture extreme conditions; power spectral density analysis reveals the frequency composition of the signal and can detect periodic changes and frequency anomalies that are not easily noticed in time-domain analysis. This feature extraction method combining the time domain and frequency domain provides a more comprehensive description of the operating state of the IGBT module, especially having unique advantages for judging phenomena such as LC oscillation, parasitic oscillation, and Miller plateau anomaly during the switching process. In high-voltage and high-power applications, these abnormal phenomena are often precursors of equipment failures. By timely capturing these features, the system can take preventive measures before the failure develops to an irreversible stage. At the same time, this feature extraction method has high computational efficiency and is suitable for real-time processing in embedded systems, meeting the strict requirements of nanosecond-level protection for computational speed and providing the necessary feature support for implementing a fast-response intelligent protection strategy.

[0067] According to one aspect of the present application, step S22 is specifically as follows: Step S221: Identify all local extreme points in the normalized feature vector to obtain an extreme point sequence. Use the cubic spline interpolation method to interpolate the local maximum points in the extreme point sequence to obtain an upper envelope, interpolate the local minimum points in the extreme point sequence to obtain a lower envelope, and calculate the average value of the upper envelope and the lower envelope to obtain an average envelope.

[0068] Step S222: Subtract the average envelope from the normalized feature vector to obtain a candidate component. Determine whether the candidate component meets the conditions of the intrinsic mode function. If not, use the candidate component as a new input signal and repeat the processing process of step S221 to obtain a new candidate component until the candidate component meets the conditions of the intrinsic mode function to obtain a high-frequency component.

[0069] Step S223: Subtract the high-frequency component from the normalized feature vector to obtain a residual signal. Repeat the processing processes of step S221 and step S222 for the residual signal until the number of extreme points of the residual signal is less than a preset threshold. Use the finally obtained residual signal as a signal trend component, and calculate the change rate of the signal trend component to obtain a trend feature vector.

[0070] The empirical mode decomposition algorithm is used to decompose the standardized feature vector to obtain the signal trend component and the high-frequency component, calculate the change rate of the signal trend component, and realize the extraction of the essential characteristics of the IGBT module operating parameters. As an adaptive signal processing method, empirical mode decomposition does not depend on preset basis functions and can decompose according to the characteristics of the signal itself, especially suitable for processing non-linear and non-stationary signals; by identifying local extreme points, constructing upper and lower envelope lines and mean envelope lines, the obtained intrinsic mode functions through iterative decomposition reflect the oscillation modes of the signal at different time scales; the signal trend component represents the long-term change trend of the parameter, excluding the influence of short-term fluctuations and noise. In the protection of high-voltage IGBT modules, this decomposition method can effectively distinguish the transient fluctuations during normal switching processes from the abnormal fluctuations caused by abnormal states, improving the specificity of fault identification; the extraction of the trend feature vector provides an important basis for predicting the development trend of faults and can capture the early signs of potential faults. Compared with traditional time-frequency analysis methods such as short-time Fourier transform and wavelet transform, empirical mode decomposition has stronger adaptability and accuracy in processing non-stationary and non-linear signals, especially suitable for processing the parameter changes of IGBT modules under complex working conditions. The implementation of this step significantly enhances the system's ability to identify potential faults, lays the foundation for realizing predictive protection, and effectively prevents equipment damage and system failures caused by misjudgment or delayed judgment.

[0071] According to one aspect of the present application, step S31 is specifically as follows: Step S311: Calculate the Euclidean distance between each data point in the historical fault dataset and other data points to obtain a distance matrix, set a density threshold, count the number of data points within the specified distance range of each data point to obtain a density value sequence, and identify high-density data points according to the density value sequence to obtain a core point set.

[0072] Step S312: Centering on the points in the core point set, merge the data points that are density-reachable to obtain an initial clustering result, remove the noise points in the initial clustering result to obtain an optimized clustering result, and number each cluster in the optimized clustering result to obtain a fault mode.

[0073] Step S313: Calculate the geometric center of each cluster in the fault mode to obtain a center sequence, calculate the covariance matrix of each cluster to obtain a dispersion matrix, construct a fault feature library according to the center sequence and the dispersion matrix, and calculate the characteristic statistic of each cluster to obtain a characteristic distribution parameter.

[0074] Density clustering analysis is performed on the historical fault data set to construct a fault feature library, identify fault modes, calculate the characteristic distribution parameters of fault modes, and achieve automatic classification and feature extraction of IGBT module fault types. Density clustering analysis clusters based on the density distribution of data points, can identify clusters of arbitrary shapes, is insensitive to outliers, and is particularly suitable for dealing with complex distributions in high-dimensional feature spaces; by calculating the Euclidean distance and setting the density threshold, high-density data points are identified as core points, and different types of fault modes can be automatically discovered without pre-specifying the number of fault types; optimizing the removal of noise points in the clustering results improves the purity and reliability of fault mode recognition; the calculation of the geometric center and covariance matrix quantifies the central position and dispersion degree of each fault mode, providing a mathematical basis for the construction of the fault feature library. This data-driven fault feature extraction method overcomes the subjectivity and limitations of traditional expert experience-based fault judgment methods, can be continuously improved and updated with the accumulation of historical data, and adapts to changes in the working conditions and fault modes of IGBT modules. In high-voltage and high-power applications, this method can identify complex fault modes and fault correlations that are difficult to discover by traditional methods, provide a more comprehensive fault knowledge base for the system, improve the accuracy and comprehensiveness of fault diagnosis, provide solid data support for subsequent threshold setting and protection strategy formulation, and significantly enhance the system's adaptive learning ability and intelligent decision-making ability.

[0075] According to one aspect of the present application, step S32 is specifically as follows: Step S321: Based on the characteristic distribution parameters, construct a probability density function of the generalized Pareto distribution to obtain density function parameters, use the maximum likelihood estimation method to estimate the values of the density function parameters to obtain an estimated parameter sequence, and establish an extreme value model based on the estimated parameter sequence to obtain extreme value model parameters.

[0076] Step S322: Substitute the extreme value model parameters into a preset threshold calculation formula to obtain an initial threshold, correct the initial threshold in combination with the characteristic distribution of the current working condition to obtain a corrected threshold, and smooth the corrected threshold to obtain a dynamic protection threshold.

[0077] Step S323: Scale the dynamic protection threshold according to a preset safety factor to obtain an initial warning threshold, and calibrate the initial warning threshold based on the historical warning accuracy rate to obtain a warning threshold.

[0078] Construct the probability density function of the generalized Pareto distribution based on the characteristic distribution parameters, estimate the parameter values using the maximum likelihood estimation method, establish an extreme value model, and dynamically adjust the threshold in combination with the current working conditions, realizing the generation of an adaptive protection threshold for the high-voltage IGBT module. The generalized Pareto distribution is a statistical distribution specifically used to describe extreme value events and is particularly suitable for simulating extreme events such as overvoltage and overcurrent in power electronic systems. Its probability density function can accurately depict the occurrence law and probability characteristics of such events; the maximum likelihood estimation method infers the optimal values of the distribution parameters based on the observed data. Compared with methods such as moment estimation, it has higher statistical efficiency and consistency and can extract the maximum amount of information from limited historical fault samples; the parameters of the extreme value model reflect the behavioral characteristics of the system under extreme conditions and provide a mathematical basis for determining the protection threshold; correcting the threshold in combination with the characteristic distribution of the current working conditions enables the protection strategy to automatically adjust according to changes in the working conditions, avoiding both the risk of underprotection in harsh working conditions and overprotection in normal working conditions; the smoothing process eliminates the sudden change of the threshold and ensures the smooth transition of the protection strategy; the introduction of the safety factor provides an additional safety margin for the system, and the calibration based on the historical warning accuracy optimizes the sensitivity of the warning threshold and reduces the false alarm rate. This method for generating a threshold based on extreme value theory and dynamic adjustment enables the protection system of the IGBT module to automatically adjust the protection sensitivity under different working conditions, realizing the transformation from "fixed threshold protection" to "dynamic threshold protection", and greatly improving the accuracy and adaptability of the protection. In high-voltage and high-power application scenarios, this method can maximize the performance of the IGBT module on the premise of ensuring safety, effectively avoiding the problems of overprotection or underprotection in traditional fixed threshold protection, and significantly improving the reliability and robustness of the system under extreme and rapidly changing working conditions, providing key protection technology support for the application of high-voltage IGBT modules in fields such as smart grids, high-speed railways, and new energy power generation.

[0079] According to one aspect of the present application, step S41: Step S411: Sort the recent comprehensive feature vectors in chronological order to obtain a time series data sequence, segment the time series data sequence according to a preset sequence length to obtain a set of data segments, perform alignment processing on each data segment in the set of data segments to obtain an aligned data sequence, and reconstruct the aligned data sequence into a two-dimensional array to obtain a time series feature matrix.

[0080] Step S412: Perform a one-dimensional convolution operation on the time series feature matrix to obtain a convolution feature map, perform a pooling operation on the convolution feature map to obtain a pooled feature map, repeat the convolution and pooling operations to obtain a multi-layer pooled feature map, and flatten the multi-layer pooled feature map to obtain time series features.

[0081] Step S413: Calculate the Pearson correlation coefficients between the features of each dimension in the temporal features to obtain a correlation coefficient matrix, and perform normalization processing on the correlation coefficient matrix to obtain a feature correlation matrix.

[0082] Construct a temporal feature matrix based on the most recent temporal data, process it using a temporal convolutional network, extract temporal features, and calculate the feature correlation matrix, which realizes the efficient capture of the temporal evolution law of the IGBT module's working state. The construction of the temporal feature matrix retains the time sequence information of the data, and ensures the comparability of data at different time points through alignment processing; the temporal convolutional network can automatically learn local time patterns and feature combinations through one-dimensional convolutional operations. Compared with traditional temporal analysis methods, such as autoregressive models and moving average models, it has stronger feature extraction ability and non-linear expression ability; the pooling operation reduces the feature dimension, improves the calculation efficiency, and retains key feature information at the same time; the repeated application of multi-layer convolution and pooling forms a hierarchical feature expression, which can capture patterns and laws at different time scales; the feature correlation matrix quantifies the mutual relationship between different features and reveals the coupling mechanism between system parameters. In the protection of high-voltage IGBT modules, this deep learning-based temporal analysis method can learn typical patterns and laws of fault development from a large amount of historical data, especially those complex temporal relationships that are difficult to describe with simple rules, providing a more reliable model basis for fault prediction. The implementation of this step significantly improves the sensitivity and recognition accuracy of the system to fault precursors, and can predict potential risks within the time scale of hundreds of nanoseconds to several microseconds before the fault occurs, providing an adequate time window for the implementation of preventive protection measures.

[0083] According to one aspect of the present application, step S42: Step S421: Segment the temporal features according to a preset time step to obtain an input sequence, construct a memory cell state matrix and a hidden state matrix as the initial states of the long short-term memory network, and input the input sequence into the input gate of the long short-term memory network to obtain the input gate state.

[0084] Step S422: Calculate the forget gate output according to the input gate state, the memory cell state matrix and the hidden state matrix to obtain the forget gate state, calculate the output gate state to obtain the output gate state, and update the memory cell state matrix and the hidden state matrix to obtain a prediction result sequence.

[0085] Step S423: Perform normalization processing on the prediction result sequence to obtain a fault probability, calculate the variance of the prediction result sequence to obtain a prediction fluctuation degree, and calculate a reliability index based on the prediction fluctuation degree.

[0086] The time series features are input into the long short-term memory network to predict the fault probability and generate reliability indicators, realizing the probability prediction and reliability assessment of potential faults in IGBT modules. As a special type of recurrent neural network, the long short-term memory network can effectively learn long-term dependencies through the coordinated action of the input gate, forget gate, and output gate, solving the problems of gradient vanishing and gradient explosion in traditional recurrent neural networks; the design of the memory cell state matrix and hidden state matrix enables the network to "remember" important information in long sequences and use this information for prediction at the appropriate time; the input gate controls the entry of new information, the forget gate determines which historical information should be retained, and the output gate controls the influence of the current state on the output. This fine-grained gating mechanism enables the network to automatically learn what information should be remembered and what information can be forgotten, making it particularly suitable for processing the long-term evolution process of the IGBT module's working state; the calculation of the prediction fluctuation provides a quantitative assessment of the reliability of the prediction results, providing a basis for risk control in protection decisions. In high-voltage and high-power applications, this prediction method based on the long short-term memory network can capture complex fault precursors that are difficult to detect by traditional methods, such as small but continuous abnormal change trends in parameters, providing the system with a longer warning time and a more accurate risk assessment, enabling protection measures to be implemented at the most appropriate time point in the most appropriate way, minimizing interference with normal operation while ensuring safety, and significantly improving the reliability and stability of high-voltage IGBT modules under extreme operating conditions.

[0087] According to one aspect of the present application, step S51 is specifically as follows: Step S511: Extract the drive strength parameter from the protection strategy parameters to obtain the strength parameter, query the pre-stored drive voltage mapping table according to the strength parameter to obtain the mapped voltage value, and compensate the mapped voltage value in combination with the current working state to obtain the optimal drive voltage.

[0088] Step S512: Calculate the initial drive current according to the optimal drive voltage and load characteristics, correct the initial drive current considering the influence of parasitic parameters to obtain the corrected drive current, and perform clipping processing on the corrected drive current to obtain the optimal drive current.

[0089] Step S513: Calculate the switching timing based on the optimal drive voltage and optimal drive current to obtain the initial timing parameters, adjust the initial timing parameters according to the temperature and load conditions to obtain the adjusted timing parameters, and perform jitter optimization on the adjusted timing parameters to obtain the drive timing parameters.

[0090] Based on the protection strategy parameters, the optimal drive voltage is calculated, the optimal drive current is calculated, and the drive timing parameters are determined, realizing precise control of the IGBT module switching process. The drive strength parameter is extracted from the protection strategy parameters and the drive voltage mapping table is queried, and compensation is performed in combination with the current working state to realize the adaptive calculation of the optimal drive voltage; the initial drive current is calculated according to the optimal drive voltage and load characteristics, corrected considering the influence of parasitic parameters, and limited to ensure that the drive current can not only meet the requirements of fast switching but also will not cause overshoot and oscillation; the switching timing is calculated based on the optimal drive voltage and current and adjusted according to the temperature and load conditions, realizing precise control of each time period of the switching process. This drive control method with multi-parameter collaborative optimization overcomes the limitations of the traditional fixed-parameter drive method, can flexibly adjust the drive strategy according to different working conditions and protection requirements, and maximizes the switching performance on the premise of ensuring safety. In the application of high-voltage IGBT modules, this precise drive control can effectively suppress overvoltage, overcurrent, and oscillation during the switching process, reduce switching losses, reduce EMI interference, extend the device life, and improve system efficiency; at the same time, by optimizing the switching timing and jitter control, the thermal stress and mechanical stress during the switching process are reduced, further enhancing the reliability and stability of the system. Especially in high-frequency and high-voltage application scenarios, this nanosecond-level precise drive control provides the necessary conditions for the IGBT module to exert its ultimate performance, meeting the strict requirements of modern power electronic systems for high efficiency, high reliability, and high power density.

[0091] According to one aspect of the present application, step S52 is specifically as follows: Step S521: Generate a reference waveform based on the optimal drive voltage and the optimal drive current to obtain a reference drive waveform, adjust the rise time and fall time of the reference drive waveform according to the drive timing parameters to obtain an adjusted drive waveform, and smooth the conversion inflection point of the adjusted drive waveform to obtain a drive waveform.

[0092] Step S522: Calculate an initial compensation coefficient based on the waveform distortion data in the historical fault dataset, dynamically adjust the initial compensation coefficient according to the current working state to obtain a real-time compensation coefficient, and compare the real-time compensation coefficient with a preset compensation reference to obtain a compensation coefficient.

[0093] Step S523: Perform a convolution operation on the drive waveform and the compensation coefficient to obtain a compensated drive waveform, perform quantization processing on the compensated drive waveform to obtain a quantized drive signal, and convert the quantized drive signal into a format recognizable by the controller to obtain a drive control signal.

[0094] Through generating waveforms based on optimal drive parameters, calculating compensation coefficients, and performing signal processing, refined drive control of high-voltage IGBT power modules is achieved. First, a reference waveform is generated according to the optimal drive voltage and drive current, and the rise / fall time is precisely adjusted based on drive timing parameters. Then, the conversion inflection points are smoothed to solve the high-frequency oscillation and EMI problems in traditional square-wave drives. Second, waveform distortion information is extracted from historical fault data to calculate the initial compensation coefficient, which is dynamically adjusted in real time according to the current working conditions, forming a closed-loop adaptive compensation mechanism that effectively overcomes the influence of factors such as temperature, load changes, and device aging on the drive effect. Finally, the drive waveform and the compensation coefficient are fused through convolution operations, followed by quantization processing and format conversion to generate drive control signals that can be directly executed by the controller. This refined drive technology can precisely control the switching process of IGBTs on a nanosecond time scale, significantly reduce switching losses, suppress voltage / current spikes and parasitic oscillations, relieve switching stress, reduce the EMI interference level, and at the same time improve the anti-interference ability and environmental adaptability of the system. In high-voltage high-power applications, this technology not only extends the service life of IGBT modules, but also improves energy conversion efficiency, reduces electromagnetic interference, provides key technical guarantees for the reliable operation of high-power conversion equipment, and is particularly suitable for scenarios with extremely high requirements for switching performance and system reliability, such as smart grids, high-speed railway traction systems, and new energy power generation equipment.

[0095] According to one aspect of the present application, step S212 is specifically as follows: Step S2121: Separate the timing marker data segment according to preset feature dimensions to obtain a sequence of feature components. Calculate the number of data for each sequence in the sequence of feature components to obtain a data count value. Accumulate the data in the sequence of feature components to obtain a data accumulation value. Divide the data accumulation value by the corresponding data count value to obtain a mean sequence.

[0096] Step S2122: Calculate the deviation of each data point based on the sequence of feature components and the mean sequence to obtain a deviation sequence. Square each data in the deviation sequence to obtain a squared deviation sequence. Accumulate the data in the squared deviation sequence to obtain the sum of squared deviations. Divide the sum of squared deviations by the corresponding data count value to obtain a variance sequence.

[0097] Step S2123: Scan each sequence in the sequence of feature components, record the position of the maximum value of each sequence to obtain a peak position sequence, extract data from the original sequence according to the peak position sequence to obtain a peak sequence, record the position of the minimum value of each sequence to obtain a valley position sequence, and extract data from the original sequence according to the valley position sequence to obtain a valley sequence.

[0098] Step S2124: Arrange the mean sequence, variance sequence, peak sequence, and valley sequence in a preset feature order to obtain a sorted feature sequence, perform normalization processing on the sorted feature sequence to obtain a normalized feature sequence, and reconstruct the normalized feature sequence into a feature vector to obtain statistical features.

[0099] By separating the time-series marked data segments according to feature dimensions, calculating the mean, variance, peak, and valley values, and performing normalization processing to form a statistical feature vector, a comprehensive characterization of the statistical characteristics of the IGBT module's working state is achieved. The separation of the feature component sequences enables the system to independently analyze the statistical characteristics of each parameter, improving the pertinence and accuracy of feature extraction; the calculation of the data count value and cumulative value provides a basis for mean calculation, ensuring the accuracy of the statistical results; the calculation process of the deviation sequence, squared deviation sequence, and sum of squared deviations details each step of variance calculation, guaranteeing the mathematical rigor of the statistical features; the recording of the peak position sequence and valley position sequence not only preserves the extreme value data but also retains the time position information where the extreme values occur, providing an important basis for subsequent time-series analysis; feature sorting and normalization processing ensure that different features are compared and fused on a unified scale. This detailed statistical feature extraction method can describe the distribution characteristics and variation laws of the IGBT module's working parameters from multiple perspectives, especially having unique advantages in identifying abnormal fluctuations, mutations, and trend changes of parameters. In high-voltage high-power applications, these statistical features can effectively distinguish the normal working state from the abnormal state, providing reliable data support for fault diagnosis and early warning; at the same time, the calculation process of the statistical features is simple and efficient, suitable for real-time implementation in embedded systems, meeting the strict requirements of nanosecond-level protection for calculation speed, and providing the necessary feature basis for implementing a fast-response intelligent protection strategy.

[0100] According to one aspect of the present application, step S312 is specifically as follows: Step S3121: Calculate the neighborhood radius of each point in the core point set to obtain a neighborhood radius sequence, calculate the density reachable points of each core point based on the neighborhood radius sequence to obtain a reachable point set, and perform connectivity analysis on the reachable point set to obtain a connectivity region sequence.

[0101] Step S3122: Merge the data points of each connectivity region in the connectivity region sequence to obtain a clustering subset sequence, calculate the number of data points in each subset in the clustering subset sequence to obtain a clustering scale sequence, and merge the clustering subset sequence to obtain an initial clustering result.

[0102] Step S3123: Calculate the density distribution of each cluster in the initial clustering result to obtain a density distribution sequence, calculate a density threshold based on the density distribution sequence to obtain a noise threshold, and mark the data points with a density lower than the noise threshold as noise points to obtain a noise point set.

[0103] Step S3124: remove the noise point set from the initial clustering result to obtain an optimized clustering result, sort the clusters in the optimized clustering result according to their sizes to obtain a sorted clustering sequence, and number the clusters in the sorted clustering sequence in sequence to obtain a fault mode.

[0104] By calculating the neighborhood radius and density reachable points of the core points, connectivity analysis is performed, connected areas are merged, noise points are removed, and fault modes are obtained by sorting and numbering, thus realizing accurate clustering of high-voltage IGBT module fault data and automatic identification of fault modes. The calculation of the neighborhood radius is based on the local density characteristics of data distribution, which can adaptively determine the clustering range of different areas, and is more adaptable to the distribution characteristics of different fault modes than the traditional fixed radius method; the determination of density reachable points establishes the connection relationship between data points and forms a density-based connectivity graph; connectivity analysis identifies a set of interconnected data points based on graph theory methods, and can discover fault clusters of any shape, not limited to spherical or ellipsoidal shapes; data point merging and clustering scale calculation provide quantitative information of clustering results, which helps to evaluate the frequency and importance of each fault mode; the calculation of density distribution sequence and noise threshold realizes the automatic identification of abnormal points and improves the purity and reliability of clustering results; sorting and numbering establish a systematic classification system for fault modes. This density-based clustering method is particularly suitable for processing complex distribution data in high-dimensional feature space, and can effectively identify various fault modes in IGBT modules, including common faults and rare faults. In practical applications, this method can automatically discover and classify different types of fault modes from massive operating data without pre-specifying the number of fault types. It has strong self-learning capabilities and the potential to discover new fault modes. At the same time, by eliminating noise points and optimizing clustering results, this method improves the accuracy and reliability of fault mode recognition, providing a solid data foundation for subsequent threshold setting and protection strategy formulation. In the protection system of high-voltage IGBT modules, this high-precision fault mode recognition capability is the prerequisite for implementing differentiated protection strategies, enabling the system to take the best response measures for different fault types, significantly improving the pertinence and effectiveness of protection, and providing key technical support for improving the overall reliability and safety of the system.

[0105] According to one aspect of the present application, step S522 is specifically: Step S5221: extract waveform distortion data from the historical fault data set to obtain a distortion data sequence, perform spectrum analysis on the distortion data sequence to obtain a distortion spectrum, calculate the amplitude of each frequency component based on the distortion spectrum to obtain a frequency amplitude sequence, and calculate the initial compensation coefficient based on the frequency amplitude sequence.

[0106] Step S5222: Obtain the load characteristics of the current working state to get load parameters, obtain the current temperature conditions to get temperature parameters, obtain the current working frequency to get frequency parameters, and construct a working condition feature vector based on the load parameters, temperature parameters, and frequency parameters.

[0107] Step S5223: Calculate the working condition adjustment factor based on the working condition feature vector to obtain an adjustment factor sequence, perform weighted calculation on the initial compensation coefficient and the adjustment factor sequence to obtain a weighted compensation coefficient, and perform normalization processing on the weighted compensation coefficient to obtain a real-time compensation coefficient.

[0108] Step S5224: Read the preset compensation reference to obtain a compensation reference value, compare the real-time compensation coefficient with the compensation reference value to obtain a deviation sequence, and correct the real-time compensation coefficient based on the deviation sequence to obtain a compensation coefficient.

[0109] By extracting waveform distortion data from historical fault data, performing spectrum analysis, and calculating the compensation coefficient in combination with the current working condition characteristics, precise compensation for the driving waveform of the IGBT module is achieved. The extraction and spectrum analysis of the distortion data sequence reveal the frequency characteristics and amplitude distribution of waveform distortion under different working conditions, providing a data basis for the calculation of the compensation coefficient; the calculation of the frequency amplitude sequence quantifies the intensity of each frequency component, enabling the compensation to be precisely adjusted for the main distortion components; the acquisition of load parameters, temperature parameters, and frequency parameters comprehensively considers the main external factors affecting the driving effect, providing a guarantee for the adaptability of the compensation; the construction of the working condition feature vector organically integrates these influencing factors to form a comprehensive description of the current working state; the calculation of the working condition adjustment factor realizes the mapping from the working condition characteristics to the compensation coefficient, enabling the compensation to be automatically adjusted according to the working condition changes; the weighted calculation and normalization processing ensure the rationality and smoothness of the compensation; the comparison and correction with the compensation reference ensure the stability and reliability of the compensation. This adaptive waveform compensation method based on historical data and current working conditions can effectively offset the influence of factors such as parasitic parameters, temperature changes, and load fluctuations on the driving effect, realizing precise control of the driving waveform. In the application of high-voltage IGBT modules, this precise compensation can significantly improve the switching characteristics, reduce waveform distortion and oscillation, lower the switching loss, suppress EMI interference, improve the system efficiency and reliability; especially under extreme working conditions and high-frequency working conditions, precise waveform compensation provides a strong guarantee for the safe and stable operation of the IGBT module, extends the device service life, improves the system reliability, creates conditions for the high-performance operation of high-voltage high-power conversion systems, and meets the strict requirements of modern power electronic systems for high efficiency, high reliability, and high power density.

[0110] In summary, for the problem of lacking effective fault prediction ability and being unable to take preventive measures before a fault occurs.

[0111] A complete fault prediction system is established, including: constructing a time series feature matrix, using a time series convolutional network (step S41) to extract time series features, and capturing the evolution law of parameters over time. The time series features are input into a long short-term memory network (step S42) for fault prediction to obtain the fault probability and reliability index. The prediction fluctuation degree is obtained by calculating the variance of the prediction result sequence to evaluate the reliability of the prediction result (step S423). Through the method of combining deep learning and time series analysis, the fault precursor can be identified, realizing the transformation from "post-fault protection" to "predictive protection". By generating protection strategy parameters including drive strength and response time, the system can take preventive measures within hundreds of nanoseconds to several microseconds before the actual occurrence of a fault.

[0112] Aiming at the problem that the fixed-threshold protection cannot meet the protection requirements under different working conditions, an adaptive dynamic threshold protection is realized, including: performing density clustering analysis based on the historical fault data set to construct a fault feature library and identify fault patterns; establishing an extreme value model based on the generalized Pareto distribution to calculate the dynamic protection threshold and warning threshold. Calculate the confidence index of the current working condition, dynamically adjust the threshold, and correct the initial threshold in combination with the feature distribution of the current working condition.

[0113] The transformation from "fixed-threshold protection" to "dynamic-threshold protection" is realized, which can automatically adjust the protection sensitivity according to the change of working conditions, avoiding overprotection or underprotection problems.

[0114] Aiming at the problem that it is difficult to finely control different switching stages in the existing technology. Extract the drive strength parameter from the protection strategy parameters, calculate the optimal drive voltage and optimal drive current, adjust the timing parameters according to the temperature and load conditions, perform jitter optimization on the adjusted timing parameters, and generate the optimal drive waveform for the current working condition, not limited to a simple square wave. Introduce a compensation coefficient to consider various influencing factors, and calculate the amplitude of each frequency component through spectrum analysis. No longer limited to a simple square wave drive, but can form complex multi-stage drive waveforms according to needs, more precisely control different stages in the switching process, and achieve the balance between switching loss and EMI suppression.

[0115] Aiming at the problem of lack of an effective closed-loop adaptive mechanism, by monitoring the actual response effect, updating the control parameter library, adjusting the compensation coefficient according to the actual response effect, calibrating the initial warning threshold based on the historical warning accuracy rate, calculating the adjustment factor sequence according to the actual working condition, and performing weighted calculation on the initial compensation coefficient and the adjustment factor sequence; correcting the real-time compensation coefficient based on the deviation sequence. It can continuously optimize the drive strategy according to the actual response effect, making the system have the "learning" ability, and the control effect will be more and more precise as the running time increases.

[0116] Traditional signal processing methods are difficult to extract key features, especially on the nanosecond time scale. Wavelet denoising is performed on the high-voltage side voltage data, collector current data, and gate voltage data. The wavelet coefficient sequence is obtained by decomposing with wavelet basis functions, the energy distribution is calculated to set a threshold, and soft threshold processing is carried out. The empirical mode decomposition algorithm is used to process the standardized feature vectors to obtain the signal trend component and the high-frequency component. Statistical features, frequency domain features, and trend features are extracted to form a comprehensive feature vector. It is possible to obtain working parameter data with a high signal-to-noise ratio in a strong interference environment, especially accurately capture the rapid changes in voltage and current on the nanosecond time scale, extract key feature information, and provide a reliable basis for fault identification.

[0117] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. An adaptive nanosecond level driving protection method for a high voltage IGBT power module, characterized in that: The following steps are involved: Collect high-voltage side voltage data, collector current data, gate voltage data, junction temperature data, switch state signal and historical fault data set, and pre-process the data to obtain a standardized feature vector; Extract statistical features, frequency domain features and trend features from the standardized feature vector to obtain a comprehensive feature vector; Based on the historical fault data set, density clustering is performed, and an extreme value model is established in combination with the generalized Pareto distribution to generate an optimized threshold value; Construct a time series feature matrix, use a time series convolutional network and a long short-term memory network to process it, and generate protection strategy parameters; Generate drive control signals based on protection strategy parameters, monitor actual response effects and update control parameters.

2. The method according to claim 1, characterized in that Data preprocessing includes: Perform wavelet denoising on the high-voltage side voltage data, collector current data and gate voltage data to obtain noise-reduced voltage, noise-reduced current and noise-reduced gate voltage; The junction temperature data is processed by moving average filtering to obtain smooth temperature; Calculate the voltage change rate of the noise reduction voltage and the current change rate of the noise reduction current; The noise reduction voltage, noise reduction current, noise reduction gate voltage, smoothing temperature, voltage change rate and current change rate are subjected to Z-score normalization to obtain a normalized feature vector.

3. The method according to claim 2, characterized in that Wavelet denoising processing includes: Decomposing the high-voltage side voltage data, collector current data and gate voltage data by wavelet basis function to obtain a wavelet coefficient sequence; Calculate the energy distribution of the wavelet coefficient sequence to obtain an energy distribution sequence, and set a threshold value according to the energy distribution sequence; The wavelet coefficient sequence is subjected to soft threshold processing to obtain a processing coefficient sequence, and the processing coefficient sequence is used to perform wavelet reconstruction to obtain the noise reduction voltage, noise reduction current and noise reduction gate voltage.

4. The method according to claim 1, characterized in that The steps of extracting statistical features and frequency domain features include: Based on the preset window length and sliding step size, the standardized feature vector is segmented to obtain time series labeled data segments; Calculate the mean, variance, peak and valley values ​​of the time series marked data segment to obtain statistical features; Perform fast Fourier transform on the time-series labeled data segment, calculate the power spectrum density, and obtain the frequency domain features.

5. The method according to claim 1, characterized in that The steps to extract trend features include: The standardized eigenvector is processed by the empirical mode decomposition algorithm to obtain the signal trend component and high-frequency component; Identify all local extreme points in the standardized feature vector to obtain an extreme point sequence; The cubic spline interpolation method is used to process the extreme point sequence to obtain the upper and lower envelopes and the mean envelope; The standardized eigenvector is subtracted from the mean envelope to obtain the candidate component, and the process is iterated until the intrinsic mode function condition is met; The rate of change of the signal trend component is calculated to obtain the trend feature vector.

6. The method according to claim 1, characterized in that Density clustering includes: Calculate the Euclidean distance between data points in the historical fault data set, set the density threshold, identify high-density data points and obtain the core point set; The data points with reachable density are merged to obtain the initial clustering result, and the noise points are removed to obtain the optimized clustering result and generate the fault mode; The geometric center and covariance matrix of the fault mode are calculated, a fault feature library is constructed, and characteristic distribution parameters are obtained.

7. The method according to claim 6, characterized in that The steps to generate the optimized thresholds include: Construct the probability density function of generalized Pareto distribution based on characteristic distribution parameters; The extreme value model is established using the maximum likelihood estimation method, and the dynamic protection threshold and warning threshold are calculated; The confidence index of the current working condition is calculated, and the threshold is adjusted according to the confidence index to obtain the optimized threshold.

8. The method according to claim 1, characterized in that The steps of constructing the time series characteristic matrix and generating protection strategy parameters include: Sort the most recent comprehensive feature vectors in chronological order, align the segments, and reconstruct them into a temporal feature matrix; Perform one-dimensional convolution and pooling operations on the time series feature matrix to obtain time series features; Input the time series features into the long short-term memory network to predict the failure probability and reliability index; Generate protection strategy parameters based on failure probability and reliability indicators.

9. The method according to claim 1, characterized in that The steps of generating a drive control signal include: Calculate the optimal driving voltage and optimal driving current based on the protection strategy parameters; Determine driving timing parameters and generate driving waveform; Calculate the compensation coefficient and output the drive control signal.

10. The method according to claim 9, characterized in that The steps to calculate the compensation coefficient include: Extract waveform distortion data from historical fault data sets and perform spectrum analysis to obtain initial compensation coefficients; Calculate the operating condition adjustment factor based on the current operating condition characteristic vector; Perform weighted calculation on the initial compensation coefficient and the adjustment factor sequence to obtain the real-time compensation coefficient; The real-time compensation coefficient is corrected based on a preset compensation reference value to obtain a compensation coefficient.

Images