A fault automatic diagnosis method and system for an inverse conducting IGBT intelligent power module

Abstract

The application relates to the technical field of power electronic device diagnosis, and discloses a fault automatic diagnosis method and system for an inverse-conducting IGBT intelligent power module. The method comprises the following steps: acquiring multi-source monitoring data of the power module during operation, wherein the multi-source monitoring data comprises a gate voltage waveform, a collector current waveform and a shell temperature change curve; then, a dynamic feature extraction model is established, time domain and frequency domain joint analysis is performed on the multi-source monitoring data, and a feature parameter set is generated; subsequently, a fault feature space is constructed, the feature parameter set is mapped to a high-dimensional space to form a feature vector distribution graph; then, a self-adaptive clustering algorithm is used to divide the feature vector distribution graph into regions, and an abnormal feature aggregation area is identified; finally, a pattern matching engine is used to compare the abnormal feature aggregation area with a preset fault feature library, and a fault type identification result is output. The method can integrate multi-source data to realize dynamic feature extraction and self-adaptive fault identification, and improves the real-time performance and accuracy of fault diagnosis.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power electronic device diagnosis, in particular to a fault automatic diagnosis method and system for reverse-conducting IGBT intelligent power module. BACKGROUND

[0002] In the power electronic system, the reverse-conducting IGBT intelligent power module as the core power conversion unit is widely used in new energy power generation, industrial frequency conversion, rail transit and smart grid fields, and its operation stability directly determines the reliability of the whole system. With the increasing requirements of power density and switching frequency in application scenarios, the reverse-conducting IGBT intelligent power module is long-term under the working condition of high voltage, large current and complex thermal cycle, which is easy to cause faults such as damage of gate oxide layer, leakage between collector and emitter, and increase of packaging thermal resistance. If these faults are not diagnosed and handled in time, it may lead to performance degradation of the module, system shutdown, equipment burning or even safety accidents.

[0003] Current fault diagnosis methods for reverse-conducting IGBT intelligent power modules mostly rely on traditional offline detection or single parameter monitoring. Offline detection requires the module to be disassembled from the system, which not only consumes time and effort, but also cannot capture dynamic fault characteristics in the module running process, making it difficult to meet the real-time diagnosis demand. Single parameter monitoring focuses on a certain index such as case temperature or collector current, ignoring the correlation characteristics between multiple physical quantities under different fault modes. For example, gate failure may cause distortion of gate voltage waveform and overshoot of collector current at the same time, and monitoring a single parameter may lead to misjudgment or omission. In addition, in the existing diagnosis methods, feature extraction is mostly limited to a single dimension in time domain analysis or frequency domain analysis, which is difficult to fully represent the fault information in multi-source monitoring data, resulting in insufficient pertinence of the generated feature parameters; the fault recognition link often uses fixed threshold or traditional clustering algorithm, which cannot adapt to the feature drift caused by working condition fluctuation in the module running process, and has low recognition accuracy for abnormal feature clusters under complex working conditions.

[0004] The fault comparison of traditional diagnosis methods mostly relies on simple feature matching rules preset by artificial, lacks systematic fault feature library support, has poor adaptability to new fault modes, and cannot realize accurate identification of fault types. These problems lead to obvious short boards in real-time performance, accuracy and adaptability of the existing reverse-conducting IGBT intelligent power module fault diagnosis methods, which are difficult to meet the high requirements of modern power electronic systems for fault early warning and health management, and an automatic diagnosis method is urgently needed to integrate multi-source monitoring data, realize dynamic feature extraction and adaptive fault recognition, so as to improve the efficiency and reliability of module fault diagnosis and ensure the stable operation of power electronic system. SUMMARY

[0005] The application aims to provide a fault automatic diagnosis method and system for reverse-conducting IGBT intelligent power modules to solve the problems in the background art.

[0006] To achieve the above-mentioned purpose, the application provides a fault automatic diagnosis method for reverse-conducting IGBT intelligent power modules, which comprises the following steps:

[0007] Obtaining multi-source monitoring data during power module operation, wherein the multi-source monitoring data comprises gate voltage waveform, collector current waveform and shell temperature change curve;

[0008] Establishing a dynamic feature extraction model, performing time domain and frequency domain joint analysis on the multi-source monitoring data, and generating a feature parameter set;

[0009] Constructing a fault feature space, mapping the feature parameter set to a high-dimensional space to form a feature vector distribution map;

[0010] Dividing the feature vector distribution map into regions by using an adaptive clustering algorithm, and identifying an abnormal feature aggregation area;

[0011] Comparing the abnormal feature aggregation area with a preset fault feature library through a pattern matching engine, and outputting a fault type identification result.

[0012] Preferably, the step of establishing a dynamic feature extraction model comprises:

[0013] Extracting rising edge and falling edge features of the gate voltage waveform, and calculating voltage change rate during switching process;

[0014] Analyzing turn-on and turn-off characteristics of the collector current waveform, and extracting current overshoot amplitude and oscillation frequency;

[0015] Monitoring the slope change of the shell temperature change curve, and calculating thermal resistance dynamic response parameter;

[0016] Combining the voltage change rate, current overshoot amplitude, oscillation frequency and thermal resistance dynamic response parameter to form the feature parameter set.

[0017] Preferably, the step of constructing a fault feature space comprises:

[0018] Standardizing the feature parameter set to eliminate dimensional differences;

[0019] Projecting the standardized feature parameter to a three-dimensional space by using a nonlinear dimension reduction method;

[0020] Marking the boundary of normal working condition area in the three-dimensional space according to historical fault data distribution rule;

[0021] The distance between the current feature vector and the boundary of the normal operating region is calculated to generate a feature vector distribution map.

[0022] Preferably, the adaptive clustering algorithm comprises:

[0023] The number of cluster centers is automatically determined according to the density distribution of the feature vectors.

[0024] A dynamic neighborhood radius parameter is set to adjust the boundary of the clustering region.

[0025] The feature vector dispersion of each clustering region is calculated to screen abnormal feature clusters with dispersion exceeding a threshold.

[0026] Preferably, the comparison by the pattern matching engine comprises:

[0027] A multi-level fault feature tree is established, which contains three types of fault modes: device aging, gate failure and thermal runaway.

[0028] The geometric feature parameters of the abnormal feature cluster are extracted, including area, shape factor and center position.

[0029] The geometric feature parameters are compared with standard patterns in the fault feature tree for similarity calculation.

[0030] The fault type corresponding to the standard pattern with the highest similarity is selected as the recognition result.

[0031] Preferably, the establishment of the multi-level fault feature tree comprises:

[0032] Typical fault sample data are collected to extract common features and form basic fault patterns.

[0033] According to the fault development process, the severity level is divided to establish the fault pattern evolution path.

[0034] Feature weight coefficients are set to distinguish the influence degree of main features and secondary features.

[0035] Preferably, the method further comprises:

[0036] The fault feature library is updated in real time, and when a new abnormal feature cluster is detected:

[0037] The feature parameters and operating environment data of the new abnormal feature cluster are recorded.

[0038] The hazard level of the new abnormal feature cluster is evaluated by an expert system.

[0039] The confirmed new fault mode is supplemented to the fault feature tree.

[0040] Preferably, the real-time updating of the fault feature library comprises:

[0041] A feature variation monitoring window is set to track the long-term evolution trend of the feature parameters;

[0042] A feature drift detection mechanism is established to identify the gradual change in the parameter distribution pattern;

[0043] When significant feature drift is detected, a fault feature library reconstruction process is triggered.

[0044] Preferably, the method further comprises constructing a fault prediction model based on the evolution trend of the feature parameters:

[0045] The degradation rate of the key feature parameters is calculated; the remaining useful life is predicted; and preventive maintenance recommendations are generated.

[0046] Preferably, the method further comprises an intelligent power module fault automatic diagnosis system for reverse-conducting IGBT, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the computer program, implements the steps of the above-mentioned intelligent power module fault automatic diagnosis method for reverse-conducting IGBT.

[0047] Compared with the prior art, the present application has the following advantages:

[0048] The intelligent power module fault automatic diagnosis method for reverse-conducting IGBT provided by the present application effectively solves the deficiencies of traditional diagnosis methods in real-time, accuracy and adaptability by integrating multi-source monitoring data and constructing a systematic diagnosis process, providing a more optimal solution for fault diagnosis of reverse-conducting IGBT intelligent power modules. In the data acquisition link, the method acquires multi-source monitoring data including gate voltage waveform, collector current waveform and shell temperature change curve, rather than relying on a single parameter, which can fully capture the electrical and thermal characteristic information during module operation, avoid missing fault features due to limited monitoring dimensions, provide a more comprehensive foundation for subsequent diagnosis from the data source, and enable the diagnosis process to cover the multi-physical quantity correlation characteristics of different fault modes of the module, reducing the misjudgment or omission caused by one-sided information.

[0049] In the feature extraction link, the method establishes a dynamic feature extraction model to perform time and frequency domain joint analysis on the multi-source monitoring data, rather than being limited to single-dimensional feature mining, which can more fully extract the fault information contained in the monitoring data. Time domain analysis can capture dynamic transient features such as voltage change rate and current overshoot amplitude, frequency domain analysis can identify implicit features such as oscillation frequency, and the combination of the two generates a feature parameter set that is more targeted and representative, which can accurately reflect the subtle changes in the module operating state, provide a high-quality feature basis for subsequent fault feature space construction and anomaly recognition, and avoid a decrease in diagnosis accuracy due to insufficient feature representation.

[0050] In the fault feature space construction and abnormality identification link, the method maps the feature parameter set to a high-dimensional space to form a feature vector distribution graph, and adopts an adaptive clustering algorithm for region division instead of relying on a fixed threshold or a traditional clustering manner. The adaptive clustering algorithm can automatically determine the number of clustering centers and adjust the neighborhood radius parameter according to the density distribution of the feature vector, effectively adapt to the feature drift caused by the working condition fluctuation in the module running process, accurately divide the normal and abnormal feature regions, accurately identify the abnormal feature aggregation area, avoid misjudgment of the abnormal area caused by the working condition change, and improve the adaptability to complex working conditions.

[0051] In the fault type identification link, the method compares the abnormal feature aggregation area with a preset fault feature library through a pattern matching engine instead of relying on a simple matching rule preset by artificial. The systematic fault feature library can provide standardized comparison basis for different fault modes, ensure the normativity and accuracy of fault type identification, and at the same time, the architecture of the preset fault feature library also provides a basis for the expansion of subsequent fault modes, so that the method has scalability when facing new faults. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 A working principle diagram of the reverse-IGBT intelligent power module fault automatic diagnosis method described in the application;

[0053] Figure 2 A working principle diagram established for a dynamic feature extraction model;

[0054] Figure 3 A working principle diagram of fault feature space construction;

[0055] Figure 4 An adaptive clustering filtering diagram. DETAILED DESCRIPTION

[0056] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the application.

[0057] Please refer to Figure 1The application provides a kind of reverse-conducting IGBT intelligent power module fault automatic diagnosis method and system, the method includes power module operation, real-time acquisition multi-source monitoring data by sensor array, these data cover grid voltage waveform, collector current waveform and shell temperature variation curve and other key parameters;Dynamic feature extraction model is established, and multi-source monitoring data is carried out time domain and frequency domain joint analysis, time domain analysis focuses on the time series characteristics of waveform, and frequency domain analysis utilizes fast fourier transform to extract harmonic component, to generate comprehensive feature parameter set;Fault feature space is built, and feature parameter set is mapped to high-dimensional space by standardization and nonlinear dimension reduction technique, and form visual feature vector distribution diagram, the distribution diagram can clearly reflect the relevance between parameters;Adaptive clustering algorithm is used to intelligently divide the area of feature vector distribution diagram, and the algorithm adjusts clustering boundary according to vector density, and identifies abnormal feature aggregation area, and these areas correspond to potential fault mode;Abnormal feature aggregation area is compared with preset fault feature library by mode matching engine, and fault feature library includes the data of historical fault mode, and engine calculates similarity and outputs fault type identification result, such as device aging or grid failure, to realize automatic diagnosis.

[0058] Example 1: see Figure 2, the implementation process of the dynamic feature extraction model begins with the fine processing of the gate voltage waveform, the core of which is to capture the subtle changes in the switching transient process. The system captures the voltage signal through high-voltage differential probes deployed on the gate drive circuit at a very high sampling rate, which is usually set to be able to distinguish the precision of nanosecond-level pulse front, thereby ensuring the authenticity of the original data. The original voltage waveform collected is first subjected to digital filtering to suppress high-frequency noise interference, and then a sliding window algorithm is used to identify the key turning points of the waveform, i.e. the starting point of the rising edge and the ending point of the falling edge. After accurately defining these time boundaries, the voltage rate of change in the switching process is calculated. This calculation is not a simple end-point difference, but a linear fitting of the waveform edge based on the least squares method, and the slope obtained is taken as the quantitative indicator of the rate of change. This method can effectively weaken the error caused by random fluctuations and reflect the average rate of voltage change. The analysis of the collector current waveform focuses on its dynamic response characteristics. A Rogowski coil or Hall effect current sensor is coupled to the main power circuit to obtain the current signal in a non-invasive manner. The analysis of the conduction characteristics of the current waveform focuses on the interval from zero to steady-state value, and the absolute value of the rise time and the smoothness of the rising edge are extracted. The analysis of the off characteristics is more complex, and attention needs to be paid to the current tail phenomenon and the current spike that may occur at the moment of switching off. The extraction of the current overshoot amplitude is achieved by comparing the theoretical current value at the off time with the actual peak current collected, and its calculation involves the zero-crossing detection of the first derivative of the current waveform. The extraction of the oscillation frequency requires local spectral analysis of the current waveform after switching off, usually using short-time Fourier transform or wavelet transform to identify the dominant oscillation frequency component and its decay characteristics. These parameters together reveal the stress state of the power switching device under hard switching or soft switching conditions.

[0059] The monitoring of the shell temperature change curve relies on high-precision thermocouples or thermistors attached to the module substrate or the surface of the shell. The collection frequency of the temperature signal is much lower than that of the electrical signal, but the long-term trend of the data contains key information. The calculation of the thermal resistance dynamic response parameter requires correlating the shell temperature change with the power loss data at the same time, which can be calculated from the collected voltage and current waveforms. By analyzing the response curve of the shell temperature under a certain power step change, a first-order thermal network model is used to fit the temperature rise and fall process, and the thermal resistance and heat capacity parameters that characterize the heat dissipation capacity are extracted. The time constant index of the thermal resistance dynamic response parameter is an indicator of the speed of the module temperature response, which is obtained from the process of fitting a first-order thermal network model from the shell temperature change curve. The index is specifically the product of the thermal resistance value and the heat capacity value obtained from the model fitting. During the model fitting process, the thermal resistance and heat capacity are adjusted through an optimization algorithm to make the theoretical temperature rise curve output by the model best match the measured curve. At this time, the product of the determined thermal resistance and heat capacity is the time constant. The index realizes the quantification of the thermal inertia of the module. The larger the index value, the slower the temperature response, which helps to evaluate the thermal behavior of the module under different power change frequencies. The dynamic parameter can better reflect the thermal behavior of the module under actual working conditions than the static thermal resistance value. The key step in forming the feature parameter set is to effectively combine the feature parameters extracted from different physical quantities. The combination process is not simply splicing, but needs to consider the dimensional differences and physical meanings between parameters. First, the voltage change rate, current overshoot amplitude, oscillation frequency, and thermal resistance dynamic response parameter are normalized and preprocessed to convert them into dimensionless relative values. The reference values for normalization come from the rated values specified in the device data manual or the statistical average values under a large number of historical normal working conditions. Subsequently, appropriate weight coefficients are assigned to each parameter according to its sensitivity to failure. The weight distribution of feature parameter fusion refers to assigning a coefficient value that reflects the relative importance of the voltage change rate, current overshoot amplitude, and other different feature parameters. The weight distribution is based on the sensitivity of each parameter to the discrimination of failure modes. The sensitivity is determined by analyzing the historical fault data set to calculate the deviation of each parameter before and after the failure from the normal state. Parameters with large and stable deviation obtain higher weights. In the technical solution, the normalized feature parameter values are multiplied by their corresponding weight coefficients, and then all the weighted parameter values are summed or combined into a comprehensive feature index. This process highlights the contribution of key parameters in the overall feature. For example, the abnormality of the voltage change rate may directly indicate a gate drive failure and thus be assigned a higher weight, while the thermal resistance change is relatively slow and its weight may be lower. The final feature parameter set is a structured data vector that serves as a digital fingerprint of the power module health status.

[0060] At the hardware implementation level, the above feature extraction function is usually integrated in a dedicated signal conditioning and processing unit, which contains multi-channel synchronous acquisition circuit, high-performance floating-point digital signal processor and static memory for temporary storage of waveform data. The signal conditioning circuit is responsible for amplifying, level shifting and anti-aliasing filtering of the raw signal output by the sensor, ensuring the quality of the signal sent to the ADC. The embedded software running on the digital signal processor implements the various feature extraction algorithms described above. The software adopts modular design, encapsulating voltage analysis, current analysis and temperature analysis as independent task modules, which are scheduled by a real-time operating system to ensure the timeliness of processing. The extracted feature parameter set is uploaded to the main control unit through a communication interface (such as CAN bus or Ethernet) for subsequent analysis. For example, under different conditions such as module startup, light load operation or overload, the normal range of some feature parameters (such as current overshoot amplitude) may be different. Therefore, a working condition recognition module is pre-set in the system. This module classifies the current working condition according to the average current and switching frequency and other operating parameters. The threshold parameters of the working condition recognition module are the boundary values used to classify the current operating state, such as current or power thresholds to distinguish between light load, rated load and overload. These parameters are obtained by analyzing the historical operating data of the device and combining its rated specifications. The specific method is to statistically analyze the typical distribution range of average current and switching frequency under different load conditions, and select boundary values that can clearly distinguish different working conditions. In the technical solution, this parameter is used for real-time comparison: comparing the currently measured average current and switching frequency with the pre-set threshold parameter to determine the current working condition category and select the corresponding normal reference range for feature parameter evaluation. This dynamic adjustment mechanism enhances the robustness and accuracy of the feature extraction model in different application scenarios, enabling subsequent fault diagnosis to be based on more realistic context information.

[0061] Example 2: see Figure 3, the process of constructing the fault feature space starts with the standardization preprocessing of the feature parameter set, which aims to eliminate the numerical difference caused by different physical dimensions, so that the subsequent analysis is not affected by the absolute numerical size of the parameters. The Z-score standardization method is used to linearly transform each feature parameter, calculate its deviation from the mean value of the sample set, and measure it in units of standard deviation. After processing, each feature parameter will conform to the standard normal distribution with a mean of zero and a standard deviation of one. This processing allows parameters such as voltage rate of change and current overshoot amplitude, which originally have different dimensions, to be compared and operated on the same scale. After standardization, the nonlinear dimension reduction stage is entered. The t-distributed Stochastic Neighbor Embedding (t-SNE) algorithm is chosen as the main tool in this stage. This algorithm is particularly suitable for handling the local structure of high-dimensional data. In the implementation process, a suitable perplexity parameter needs to be set. This parameter is essentially a smoothing measure of the number of neighbors considered by the algorithm for each point. Through an iterative optimization process, the probability distribution of similarity between data points in high-dimensional and low-dimensional spaces is continuously adjusted to minimize the KL divergence between them. Finally, the feature parameter set of dozens or even more dimensions is projected into a three-dimensional visualization space. Each point in this three-dimensional space uniquely corresponds to an original feature vector, and the relative distance between points reflects the similarity between the original features. On the basis of the three-dimensional projection space, the boundary of the normal operating condition area needs to be marked. This work relies on the accumulation of a large amount of historical normal operating data. First, collect the feature vectors generated by the system during long-term operation under known normal conditions. These normal data are also processed by standardization and dimension reduction and projected into the same three-dimensional space. Then, a density-based spatial clustering method is used to analyze the distribution of these normal data points. The algorithm automatically identifies the most densely populated area of normal data points and generates a minimum convex hull surface that can wrap most of the normal points. This surface is the boundary of the normal operating condition area. Any new data point that falls outside the boundary may indicate an abnormal state.

[0062] The process of calculating the distance of the current feature vector from the normal operating region boundary involves spatial geometry operations. For each newly generated feature vector point in a three-dimensional space, the system calculates the shortest Euclidean distance from the point to the boundary surface of the normal region. The sign of the distance value is used to distinguish the position relationship of the point: a positive sign indicates that the point is outside the normal region, and a negative sign indicates that the point is inside the normal region. The absolute value of the distance reflects the degree of abnormality or normality. The spatial position and boundary distance information of all current feature vectors together form a dynamic feature vector distribution map, which is updated in real time as new data continuously flows in. Parameter tuning of the nonlinear dimensionality reduction algorithm is an important step. The learning rate parameter in the t-SNE algorithm needs to be adjusted according to the data size to prevent the optimization process from falling into a local optimum. The setting of the number of iterations needs to balance between calculation accuracy and efficiency. Usually, hundreds to thousands of iterations are needed to obtain stable projection results. For large-scale data sets, hierarchical sampling methods can be used for preprocessing to improve the efficiency of dimensionality reduction calculation. The accuracy of the normal region boundary marker directly affects the sensitivity and specificity of fault detection. A too loose boundary will increase the false negative rate, and a too strict boundary may produce false positives. Therefore, the tightness of the boundary needs to be adjusted according to the actual application scenario. During the initial operation of the system, the normal region boundary can be set relatively loose. As normal operation data accumulates, the boundary range is gradually tightened. This gradual adjustment strategy can improve the adaptability of the system during the initial deployment.

[0063] The generation of the feature vector distribution map is not a one-time task, but requires a continuous updating mechanism. The system maintains a sliding time window, retaining only the feature vector data from the recent period for the generation of the distribution map. This ensures that the distribution map reflects the latest state of the system while controlling memory usage. The visualization of the distribution map can help operations personnel intuitively understand the health status of the system and provide a visual reference for fault diagnosis. The construction process of the entire fault feature space is designed as a closed-loop system. As new normal operation data accumulates, the normal region boundary and the feature vector distribution map are updated accordingly. This adaptive mechanism enables the system to automatically adjust the normal state benchmark as the device ages, avoiding false alarms due to natural device aging while maintaining the sensitivity of fault detection. In engineering implementation, the construction of the fault feature space is usually completed by an independent calculation module. This module receives the feature parameter set from the feature extraction module, performs standardization, dimensionality reduction, boundary calculation, and distance measurement, and finally outputs the feature vector distribution map with distance labels. The distribution map data is sent to the subsequent clustering analysis module for further processing.

[0064] The implementation of the adaptive clustering algorithm is based on the density distribution of feature vectors in space to automatically determine the number of cluster centers, which identifies potential core areas by analyzing the local concentration of points, and the calculation of density distribution uses kernel density estimation technology to perform weighted smoothing on the neighborhood of each feature vector point using a Gaussian kernel function, thereby obtaining a continuous density surface, and the bandwidth parameter of the kernel function is adaptively adjusted according to the overall size of the data set to avoid over-smoothing or under-smoothing problems, and the process of automatically determining the number of cluster centers relies on the detection of density peaks, and the algorithm scans the entire density surface to identify points that are significantly higher in local density than surrounding points and have sufficient distance from higher density points as candidate centers, and the distance threshold is dynamically set by multiplying the average distance between data points by a proportion factor to ensure that the selection of center points is neither too dense nor too sparse, and this automatic mechanism eliminates the subjectivity of manually setting the number of clusters and adapts to changes in feature vector distribution under different operating conditions.

[0065] Setting a dynamic neighborhood radius parameter is a key step in the clustering process, and the neighborhood radius directly affects the boundary division of the clustering area and the sensitivity of abnormal point identification, and the initial radius value is calculated based on the global distribution characteristics of the feature vectors, for example, using the median distance between all pairs of points as a reference benchmark, and a dynamic adjustment mechanism is implemented by monitoring real-time changes in the data stream, for example, when new feature vectors continuously flow in, the system will periodically recalculate the average density of the vectors, and if the density increases, the radius will be appropriately reduced to capture finer local structures, and vice versa, the radius will be expanded to avoid misjudging normal points as isolated points, and the dynamic update of the radius uses a sliding window method, and the window size is synchronized with the data sampling frequency to ensure the timeliness of the adjustment, in addition, the radius parameter is also associated with the stability of the cluster center, and if the center point position changes too much in continuous iterations, the system will trigger a radius recalibration to maintain the consistency of the cluster.

[0066] A quantitative indicator is introduced to measure the dispersion of feature vectors within the cluster when calculating the dispersion of feature vectors in each cluster region, and the dispersion formula is defined as:

[0067]

[0068] Where: represents the dispersion value of the cluster region , and the larger the value, the more dispersed the distribution of feature vectors in the region; is the total number of feature vectors contained in the cluster ; represents the th feature vector in the cluster ; is the center vector of the cluster , which is obtained by calculating the arithmetic mean of all feature vectors in the region; Euclidean norm of a vector, used to measure the geometric distance between vectors. The calculation of dispersion is performed immediately after the initial formation of clusters, and each cluster region calculates its own value independently. The vector data required for calculation is obtained from the real-time cache to ensure the timeliness of the results.

[0069] The screening of abnormal feature clusters with dispersion exceeding the threshold value relies on a preset threshold mechanism, the size of the threshold is determined according to the statistical distribution of cluster dispersion in historical normal operation data, for example, taking the 95th percentile of the dispersion value in the normal state as the baseline, the system will periodically update the threshold to reflect the impact of device aging or environmental changes, the screening process traverses all identified cluster regions, compares the value of each region with the current threshold, if the dispersion of a region is significantly higher than the threshold, it will be marked as an abnormal cluster, the geometric features of the abnormal region such as area and shape will be extracted for subsequent analysis, in addition, the trend of dispersion change is also monitored, if the dispersion of a region continues to rise within a continuous time window, the system will increase its abnormal priority to ensure that potential faults can be discovered early.

[0070] The implementation of the clustering algorithm is embedded in a dedicated data processing unit, which is equipped with a multi-core processor to parallel process the calculation task of high-dimensional feature vectors, an incremental update strategy is adopted during the algorithm iteration process, when a new feature vector arrives, the system does not recalculate the entire cluster but only adjusts the boundaries of the affected region, which greatly improves the calculation efficiency to meet the real-time requirement, the calculation of density distribution uses spatial index structure such as KD tree to accelerate neighborhood query, so that the processing of large-scale vector data can be completed within milliseconds, the parameter adjustment of dynamic neighborhood radius is realized through closed-loop control, the feedback signal comes from the evaluation indicators of cluster quality such as silhouette coefficient, to ensure that the radius value is always in the optimal range. The identification results of abnormal feature clusters are combined with the learning mechanism of the system, when a new abnormal pattern appears, the algorithm will record its dispersion characteristics and occurrence context, these information are used to optimize the threshold setting and radius parameter, forming a self-improving diagnostic ability, the design of the entire adaptive clustering process focuses on robustness, for example, by introducing a noise tolerance mechanism to avoid misjudgment caused by temporary interference, the visualization tool of cluster boundaries allows operation and maintenance personnel to intuitively check the vector distribution, providing reference for algorithm tuning, finally output the list of abnormal cluster regions and their attribute data, seamlessly connecting the next stage of pattern matching engine.

[0071] Referring to Figure 4In the automatic fault diagnosis process of the reverse conducting IGBT intelligent power module, adaptive clustering is the key step to identify the abnormal feature aggregation area. This figure shows the distribution of clustering dispersion and the abnormal screening results: the horizontal axis is the clustering dispersion, and the vertical axis is the number of clusters. The algorithm takes "dispersion threshold 2.5" as the boundary, the light gray area on the left is the "normal clustering area (dispersion < threshold)", which reflects the aggregation state of the feature vector when the module is running normally; the dark gray area on the right is the "abnormal feature aggregation area (dispersion > threshold)", which represents the clustering population deviating from the normal distribution. Through this density-based dispersion analysis and dynamic threshold screening, the adaptive clustering algorithm can automatically distinguish between normal and abnormal feature areas, solving the problem that traditional methods cannot adapt to the fluctuations of working conditions, providing reliable abnormal area input for the subsequent pattern matching engine to accurately identify fault types, and ensuring high recognition accuracy of abnormal features of the module under complex working conditions.

[0072] Example 4: The comparison operation of the pattern matching engine is based on a multi-level fault feature tree, which classifies known faults into three main branches: device aging, gate failure and thermal runaway. Each branch is further subdivided into more specific fault modes, for example, the device aging branch may include gate oxide degradation, bond wire fatigue, etc. subcategories, while the gate failure branch covers drive resistance abnormalities, Miller capacitance changes, etc. specific problems. The construction of the fault feature tree is not a static process, but is derived from the deep analysis of a large number of historical fault samples, and forms standardized reference templates by mining the common patterns of feature parameters when different faults occur. Extracting the geometric feature parameters of the abnormal feature aggregation area is the premise of accurate comparison. The system first obtains the spatial contour information of the abnormal area from the output of the adaptive clustering algorithm. Take an abnormal aggregation area labeled as "region Alpha" as an example, its geometric features are quantified to form a set of calculable data. These parameters include the projection area of the region in the three-dimensional feature space, the shape factor (such as the ratio of the contour perimeter to the area) that describes the shape complexity, and the precise coordinates of the region's centroid in space, see Table 1.

[0073] Table 1: Geometric feature parameters of abnormal aggregation area region Alpha

[0074]

[0075] After obtaining the quantitative geometric features of the abnormal region, the pattern matching engine starts the similarity calculation process, which compares the geometric feature parameters of the abnormal region with the pre-stored standard fault patterns in the fault feature tree item by item. The similarity calculation is not a simple Euclidean distance measurement, but a weighted cosine similarity algorithm. This algorithm assigns different weight coefficients to different geometric feature parameters. For example, the region center position may have a higher weight than the region area because it better reflects the essential characteristics of the fault. The calculation process first vectorizes the feature vector of the abnormal region and the feature vector of the standard fault pattern, then calculates the cosine value of the included angle between the two vectors in the direction, multiplies it by the corresponding weight coefficient matrix, and finally obtains a comprehensive similarity score between 0 and 1. The closer the score is to 1, the higher the matching degree.

[0076] The establishment process of the fault feature tree reflects a deep understanding of the fault evolution law. When collecting typical fault sample data, the system not only records the data snapshot at the moment of fault occurrence, but also traces the data sequence during the fault development process. By analyzing the trend of parameter changes over time, it extracts common evolution characteristics and solidifies them as basic fault patterns. For example, for gradual faults such as gate oxide layer degradation, the characteristics not only include the final parameter anomaly, but also the rate and trajectory of parameter drift. According to the severity of the fault development process, the feature tree establishes a clear hierarchical division within it, distinguishing different stages of the same type of fault, such as early warning, mid-development, and late severity, and defining the evolution path between stages. When constructing the feature tree, the system also sets dynamic weight coefficients for each feature parameter. The allocation of weight coefficients is based on the discriminant ability of the parameter in distinguishing different fault patterns. The discriminant ability is evaluated by statistical methods such as information gain or chi-square test, ensuring that the main features play a decisive role in the matching process.

[0077] The execution of the pattern matching engine is accompanied by a decision logic. When the similarity scores of an abnormal area with multiple standard patterns in the fault feature tree are all high, the engine does not simply choose the highest score. Instead, a confidence evaluation mechanism is introduced, which checks the gap between the second highest score and the highest score. If the gap is less than a preset threshold, it is determined that the matching result is ambiguous, and a more in-depth data analysis request is triggered, such as requiring the system to trace the evolution history of the abnormal area in the past period of time, or cross-verify with other sensor data (such as the real-time working current and environmental temperature of the module). This design effectively reduces the risk of misjudgment. The entire pattern matching process is designed as an interactive loop. After the matching result is output, the system allows experienced operation and maintenance personnel to review and confirm the result. If the operation and maintenance personnel believe that the matching result is biased based on experience, they can feed back the data of the abnormal area and the final fault type judgment as new samples to the system. The system will use these high-quality samples confirmed by humans to perform incremental learning on the fault feature tree, optimizing existing standard patterns or adding new fault patterns. This mechanism enables the fault feature library to evolve and become more accurate as actual operation experience accumulates.

[0078] Example 5: The mechanism of real-time updating of the fault feature library is the core of the system to maintain long-term diagnostic accuracy. This mechanism tracks the long-term evolution trend of key parameters by setting a feature change monitoring window. The monitoring window usually uses a sliding time window mechanism, and the window size is dynamically adjusted according to the data update frequency and the real-time requirements of the application scenario. For example, in a wind power converter application, the system may maintain a data window of three months in length. The feature parameter data within the window is organized into a time series and trend fitting is performed. The key parameters to be tracked include the baseline value of the gate voltage change rate, the average value of the collector current overshoot amplitude, and the slow drift amount of the shell temperature resistance. By performing linear regression or more complex seasonal decomposition on these sequences, the system can capture the gradual changes of parameters over time or switching times. This trend tracking provides a direct basis for determining whether a device has entered an aging stage. The feature drift detection mechanism is established to identify the gradual changes in the distribution pattern of parameters. The core of this mechanism is to continuously compare the statistical distribution of the current feature parameters with the baseline distribution established during the initial learning stage of the system. Concepts from statistical process control are used, such as calculating the Wasserstein distance or maximum mean difference between the current feature vector set and the historical baseline set to quantify the degree of distribution drift. The system will calculate these statistics periodically (e.g., every week) and compare them with the control limits. When the statistics continuously exceed the control limits, it indicates that the basic distribution of the parameters has changed significantly. This change may be due to the performance degradation of the power module itself, or it may be caused by long-term changes in the operating environment (such as ambient temperature and load conditions). After triggering the drift detection, the system will start the root cause analysis process to try to associate the drift with specific work log events (such as continuous overload operation and cooling system efficiency decline records).

[0079] When significant feature drift is confirmed, the system will trigger the reconstruction process of the fault feature library. Reconstruction is not simply replacing old data with new data, but a cautious incremental learning process. First, the system creates a backup of the current fault feature tree. Then, the existing classification model is fine-tuned using recent feature data (such as data within a certain period before and after the drift is detected). During the fine-tuning process, new data is given a higher weight, but valuable generalization knowledge from the original model is preserved through regularization techniques. For the fault feature tree itself, reconstruction may involve adjusting the boundary thresholds of the normal operating condition region, updating the feature weight coefficients used in fault pattern matching, or adding new branches under the corresponding nodes of the feature tree after confirming a new fault evolution path. The entire reconstruction process is carried out in a sandbox environment. Only after verifying that the new model performs better than the old model, an online switch is performed.

[0080] Building a failure prediction model is a natural extension of the feature parameter evolution trend, which focuses on calculating the degradation rate of key feature parameters, such as analyzing the slow rising trend of the gate threshold voltage, fitting its change curve and calculating the voltage increment per thousand hours of operation time as the degradation rate of the parameter, when predicting the remaining useful life, the model combines the current parameter value, its degradation rate and the preset failure threshold, uses the proportional risk model or the physical-based degradation model to calculate the expected time from the current state to functional failure, when generating preventive maintenance recommendations, the system not only considers the predicted remaining life, but also comprehensively evaluates the key degree of the module in the system, the historical maintenance cost and the current load situation, outputs the operational recommendations, such as "it is recommended to check the gate drive circuit during the next planned shutdown" or "module preventive replacement needs to be arranged within the next two months".

[0081] When the system detects a new type of abnormal feature aggregation area, the update process enters an interactive mode that requires human intervention. The system will record all feature parameter raw data and corresponding operating environment data, such as DC bus voltage, output current frequency, radiator inlet temperature, etc., within a certain period of time before and after the appearance of the new type of aggregation area. These data are packaged into a case to be analyzed and submitted to the domain engineer through the expert system interface. The built-in evaluation module of the expert system provides preliminary analysis, such as comparing early signs of known failures to give a possible hazard level assessment of the new abnormality (such as "potential risk", "need attention", "high risk"). The engineer confirms the system evaluation based on his own experience and feeds back the final conclusion to the system. The confirmed new failure mode and its feature parameters are assigned a unique mode code and then officially supplemented to the corresponding level of the fault feature tree. The complete context information of the first discovery and confirmation of the mode is also recorded. The update of the feature library is not a passive data accumulation, but an active, evidence-based knowledge accumulation process. The introduction of the failure prediction model upgrades the diagnosis from "post-judgment" to "pre-warning", and the discovery and confirmation mechanism of new modes ensures that the system can adapt to unknown failure types. This design makes the intelligent diagnostic system not a static tool, but a learning system that can evolve with the equipment and accumulate experience.

[0082] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting; it is not intended to exclude myriad other embodiments of the present application that other inventors can develop based on the same general inventive concepts embodied by the described embodiments. That is, although the present application is described in terms of particular embodiments and implementations, it is to be understood that the terminology used is for the purpose of descriptive clarity and that it is intended to be limited only by the words recited in the appended claims. It is to be understood that the terms such as first and second, etc., merely are used to differentiate one from another without necessarily implying or requiring any actual relationship or order between them. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0083] While the embodiments of the application have been shown and described herein, it is to be understood that the application is not limited to these embodiments. Rather, many modifications, changes, substitutions, and alterations can be made thereto without departing from the spirit and scope of the present application as defined by the appended claims and their equivalents.

Claims

1. An automatic fault diagnosis method for reverse-conducting IGBT intelligent power modules, characterized in that, include: Acquire multi-source monitoring data during the operation of the power module, including gate voltage waveform, collector current waveform, and case temperature change curve; A dynamic feature extraction model is established to perform joint time-domain and frequency-domain analysis on the multi-source monitoring data, generating a set of feature parameters. Construct a fault feature space and map the set of feature parameters to a high-dimensional space to form a feature vector distribution map; An adaptive clustering algorithm is used to divide the feature vector distribution map into regions and identify areas of anomalous feature clusters. The abnormal feature clusters are compared with a preset fault feature database using a pattern matching engine, and the fault type identification result is output.

2. The automatic fault diagnosis method for a reverse-conducting IGBT intelligent power module according to claim 1, characterized in that, The establishment of the dynamic feature extraction model includes: The rising and falling edge features of the gate voltage waveform are extracted, and the voltage change rate during the switching process is calculated. Analyze the conduction and turn-off characteristics of the collector current waveform to extract the current overshoot amplitude and oscillation frequency; Monitor the slope change of the shell temperature change curve and calculate the dynamic response parameters of thermal resistance; The voltage change rate, current overshoot amplitude, oscillation frequency, and thermal resistance dynamic response parameters are combined to form a set of characteristic parameters.

3. The automatic fault diagnosis method for a reverse-conducting IGBT intelligent power module according to claim 2, characterized in that, The construction of the fault feature space includes: The set of feature parameters is standardized to eliminate dimensional differences; A nonlinear dimensionality reduction method is used to project the standardized feature parameters onto a three-dimensional space. Based on the distribution patterns of historical fault data, the boundaries of the normal operating condition area are marked in three-dimensional space; Calculate the distance between the current feature vector and the boundary of the normal operating condition area, and generate a feature vector distribution map.

4. The automatic fault diagnosis method for a reverse-conducting IGBT intelligent power module according to claim 3, characterized in that, The adaptive clustering algorithm employed includes: The number of cluster centers is automatically determined based on the feature vector density distribution. Set the dynamic neighborhood radius parameter to adjust the clustering region boundaries; Calculate the feature vector dispersion of each cluster region and filter out abnormal feature clusters with dispersion exceeding the threshold.

5. The automatic fault diagnosis method for a reverse-conducting IGBT intelligent power module according to claim 4, characterized in that, The comparison performed using a pattern matching engine includes: A multi-level fault feature tree is established, which includes three types of fault modes: device aging, gate failure, and thermal runaway. Extract the geometric feature parameters of the abnormal feature cluster region, including region area, shape factor, and center position; The similarity between the geometric feature parameters and the standard patterns in the fault feature tree is calculated. The fault type corresponding to the standard pattern with the highest similarity was selected as the identification result.

6. The automatic fault diagnosis method for a reverse-conducting IGBT intelligent power module according to claim 5, characterized in that, The establishment of the multi-level fault feature tree includes: Collect typical fault sample data and extract common features to form basic fault patterns; Based on the failure development process, severity levels are classified, and failure mode evolution paths are established; Set feature weight coefficients to distinguish the influence of primary and secondary features.

7. The automatic fault diagnosis method for a reverse-conducting IGBT intelligent power module according to claim 6, characterized in that, The method further includes: The fault feature database is updated in real time, and when a new cluster of abnormal features is detected: Record the characteristic parameters and operating environment data of new anomaly clusters; The hazard level of newly emerging anomalous feature clusters is assessed using an expert system. The newly identified fault modes are added to the fault feature tree.

8. The automatic fault diagnosis method for a reverse-conducting IGBT intelligent power module according to claim 7, characterized in that, The real-time updating of the fault feature database includes: Set up a feature change monitoring window to track the long-term evolution trend of feature parameters; Establish a feature drift detection mechanism to identify gradual changes in parameter distribution patterns; When a significant feature drift is detected, the fault feature library reconstruction process is triggered.

9. The automatic fault diagnosis method for a reverse-conducting IGBT intelligent power module according to claim 8, characterized in that, The method further includes: constructing a fault prediction model based on the evolution trend of feature parameters. Calculate the degradation rate of key characteristic parameters; predict remaining service life; generate preventative maintenance recommendations.

10. An automatic fault diagnosis system for a reverse-conducting IGBT intelligent power module, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the automatic fault diagnosis method for a reverse-conducting IGBT intelligent power module as described in any one of claims 1 to 9.

Images