A method, system, medium, and equipment for fault diagnosis of press-fit IGBTs

Abstract

This invention relates to the field of fault diagnosis of marine flexible converter valves, and discloses a fault diagnosis method, system, medium, and equipment for press-fit IGBTs. The method includes: real-time monitoring data of the press-fit IGBT's electrical, thermal, and physical dimensions, acquired in real time, being preprocessed to obtain dual-modal data; inputting the dual-modal data into a trained deep learning model for multi-dimensional feature extraction and filtering, and dynamically aligning, adaptively fusing, and optimizing the filtered multi-dimensional features to output a target feature vector that accurately represents the operating state; calculating the probability of each state category and diagnostic confidence based on the target feature vector, and using a confidence threshold to identify normal states and subdivided faults; integrating multi-dimensional diagnostic information based on the confidence threshold determination results to output comprehensive and standardized diagnostic results, providing a clear basis for operation and maintenance decisions, and using operation and maintenance feedback data to optimize the deep learning model.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology for offshore flexible converter valves, and in particular to a fault diagnosis method, system, medium, and equipment for a press-fit type IGBT feature fusion method for offshore wind power flexible DC converter valves. Background Technology

[0002] Press-fit IGBTs, with their advantages of low parasitic inductance, excellent short-circuit failure characteristics, and double-sided heat dissipation, have become the core power devices for offshore flexible converter valves. As key equipment for offshore wind power grid connection and inter-regional submarine cable power transmission, offshore flexible converter valves operate for extended periods in extreme marine environments characterized by high salt spray, high humidity, strong electromagnetic interference, and intense vibration. Their operational reliability directly determines the safety, stability, and power supply continuity of the offshore power system. However, press-fit IGBTs have a multi-layered heterogeneous structure, including the chip, ceramic substrate, and metal electrodes. During operation, they must simultaneously withstand the combined effects of cyclic thermal stress, mechanical vibration, electrical stress, and the corrosive marine environment, making them highly susceptible to unique failure modes such as fretting wear, micro-ablation of contact surfaces, spring fatigue, gate oxide layer degradation, and increased contact resistance due to corrosion. These failure modes initially manifest as weak parameter drifts, difficult to detect using conventional monitoring methods, and are prone to sudden device failures later on. This can not only lead to converter valve shutdowns but also cause huge economic losses due to the limited space for offshore maintenance, high repair costs, and long cycles, seriously threatening the safe and stable operation of the offshore power system.

[0003] Existing fault diagnosis methods for press-fit IGBTs suffer from the following shortcomings: 1) Limited feature extraction: These methods primarily focus on electrical parameters such as on-state voltage drop and collector current, neglecting the mechanical and thermal characteristics related to the press-fit IGBT packaging structure. This makes it difficult to comprehensively characterize the evolution process of its unique failure modes. 2) Insufficient anti-interference capability: The operating environment of offshore flexible converter valves is characterized by strong electromagnetic interference and sensor signal drift caused by salt spray corrosion. Electrical signals are easily distorted by interference, and feature extraction from a single signal source is prone to misjudgment. Furthermore, traditional deep learning models are sensitive to complex noise in the marine environment, further reducing diagnostic accuracy and making them unsuitable for extreme offshore conditions. 3) Weak early fault diagnosis capability: Existing methods mainly target mid-to-late stage faults, lacking sufficient feature capture capability for weak faults such as fretting wear and initial degradation of the gate oxide layer, making early warning difficult. 4) Poor model adaptability: Existing deep learning methods based on stacked noise reduction autoencoders have low hyperparameter optimization efficiency and weak generalization ability, and do not optimize the network structure for the multi-physics coupling characteristics of press-fit IGBTs. Summary of the Invention

[0004] In response to the extreme operating environment of offshore flexible converter valves, the purpose of this invention is to provide a fault diagnosis method, system, medium, and equipment for press-fit IGBTs, which can integrate multi-dimensional features, has strong anti-interference capabilities, and can accurately identify early faults, thereby filling the gap in existing technologies and ensuring the safe and reliable operation of offshore flexible converter valves and the stable power supply of offshore power systems.

[0005] To achieve the above objectives, in a first aspect, the technical solution adopted by the present invention is as follows: a fault diagnosis method for press-fit IGBTs, comprising: real-time monitoring data of the electrical, thermal, and physical dimensions of the press-fit IGBT acquired in real time, obtaining dual-modal data after data preprocessing; inputting the dual-modal data into a trained deep learning model, extracting and filtering multi-dimensional features, and dynamically aligning, adaptively fusing, and optimizing the filtered multi-dimensional features to output a target feature vector that accurately represents the operating state; calculating the probability of each state category and the diagnostic confidence based on the target feature vector, and using a confidence threshold to determine the identification of normal state and subdivided faults; integrating multi-dimensional diagnostic information according to the confidence threshold determination result to output full and standardized diagnostic results, providing a clear basis for operation and maintenance decisions, and using operation and maintenance feedback data inversely for deep learning model optimization.

[0006] Furthermore, data preprocessing includes: data cleaning, noise reduction, standardization, and bimodal reconstruction;

[0007] The dual-modal reconstruction process is as follows: the standardized one-dimensional time-series data of various press-fit IGBTs are retained for time-series feature extraction; for signals rich in frequency domain features, two-dimensional reconstruction is completed through short-time Fourier transform, which is converted into a two-dimensional time-frequency diagram to obtain two-dimensional time-frequency data of press-fit IGBTs, so as to present the signal energy distribution in different spatiotemporal dimensions for spatial feature extraction.

[0008] Furthermore, the bimodal data is input into the trained deep learning model to extract and filter multi-dimensional features, including:

[0009] For the input one-dimensional timing data of the press-fit IGBT, the recurrent neural network structure is used to capture the temporal correlation and dynamic change characteristics of the data, and automatically extract the key information features reflecting the operating status of the press-fit IGBT in the time domain, so as to effectively characterize whether there are potential faults in the press-fit IGBT.

[0010] The extracted features are screened, and redundant and invalid features are eliminated based on the feature importance evaluation index. The feature importance is determined by calculating the correlation coefficient between each time-series feature and the fault category of the press-fit IGBT. Features with an absolute value of correlation coefficient greater than a set threshold are retained, thereby realizing the screening of time-domain features of the press-fit IGBT.

[0011] For the input two-dimensional time-frequency data of the press-fit IGBT, the spatial distribution characteristics and frequency domain energy characteristics of the data are captured by convolution operation, and key information features in the spatial domain and frequency domain are automatically extracted to effectively reflect the potential characteristics of the press-fit IGBT fault.

[0012] After feature extraction is completed, the feature importance evaluation mechanism is used to select features that contribute more than a set threshold to the identification of faults in press-fit IGBTs by combining the inherent correlation between spatial domain and frequency domain features, and redundant information is removed.

[0013] After extracting and filtering the time domain, spatial domain, and frequency domain features of the press-fit IGBT, the effective features filtered from each dimension are initially integrated to form a multi-dimensional feature set for subsequent feature fusion.

[0014] Furthermore, the selected multi-dimensional features are dynamically aligned, adaptively fused, and optimized for dimensionality reduction to output a target feature vector that accurately represents the operating state, including:

[0015] By using the feature mapping mechanism of deep learning networks, various features are first mapped to the same feature space to achieve dynamic alignment of multi-dimensional features;

[0016] An adaptive fusion strategy is adopted to deeply integrate the aligned features. The weight allocation mechanism strengthens the feature weights that are strongly correlated with the fault state of the press-fit IGBT and weakens the influence of redundant and interference features.

[0017] The deeply integrated features are weighted and summed to obtain a fused feature vector. The integrated features are then subjected to dimensionality reduction optimization to output the target feature vector.

[0018] Furthermore, based on the target feature vector, the probability of each state category and the diagnostic confidence level are calculated. A confidence threshold is then used to determine the identification of normal states and subdivided faults, including:

[0019] The target feature vector is transformed into a linear feature vector corresponding to the fault category and normal state by high-dimensional feature mapping and nonlinear transformation through the fully connected layer of the deep learning network. Then, the probability quantization of each state category is achieved through normalization.

[0020] Based on the probability quantification results of each state category, the confidence level of the current diagnosis result is calculated, and the confidence level is the maximum value among the probabilities of all state categories.

[0021] Using confidence level as the judgment standard, when the confidence level reaches or exceeds the set threshold, it is judged as a clear diagnostic result and the corresponding operating status is output; when the confidence level is lower than the set threshold, it is judged as a suspected state, triggering an early warning signal to prompt maintenance personnel to strengthen component operation monitoring.

[0022] Furthermore, the high-dimensional target feature vector is transformed into a linear feature vector corresponding to the fault category and normal state, and then the probability quantization of each state category is achieved through normalization processing, specifically:

[0023] The linear feature vectors output by the fully connected layer are transformed into numerical values ​​that conform to the probability distribution law. A normalization function is introduced to process the linear feature vectors and output the probability values ​​of each fault category and normal state.

[0024] The normalization function is expressed as follows:

[0025]

[0026] In the formula, Let be the probability of the k-th class state. Let be the linear feature value corresponding to the k-th state, and C be the total number of state categories.

[0027] Furthermore, based on the confidence threshold determination results, multi-dimensional diagnostic information is integrated to output comprehensive and standardized diagnostic results, providing a clear basis for operation and maintenance decisions, including:

[0028] Based on the confidence level determination, the final fault diagnosis result is output;

[0029] Based on the fault diagnosis results, the operating status category, the probability distribution of all categories, and the diagnostic confidence level are extracted. Based on the operating status category, the current normal operating status or specific sub-fault type of the press-fit IGBT is clearly output. Based on the probability distribution of all categories, the probability values ​​corresponding to all status categories are output, intuitively presenting the confidence distribution of the model's identification. Based on the diagnostic confidence level, the confidence level of the core judgment index is output simultaneously, clearly reflecting the reliability of the current diagnostic results.

[0030] Based on the results of feature extraction and screening using deep learning networks, key feature parameters with the highest similarity to the current operating state are extracted. Through the precise output of key feature parameters, the specific location, severity, and potential causes of the fault are directly identified.

[0031] Secondly, the technical solution adopted by this invention is as follows: a fault diagnosis system for press-fit IGBTs, comprising: a data preprocessing module, which preprocesses real-time monitoring data of the electrical, thermal, and physical dimensions of press-fit IGBTs to obtain dual-modal data; a data analysis and feature fusion module, which inputs the dual-modal data into a trained deep learning model, extracts and filters multi-dimensional features, and performs dynamic alignment, adaptive fusion, and dimensionality reduction optimization on the filtered multi-dimensional features to output a target feature vector that accurately represents the operating state; a confidence-driven fault identification module, which calculates the probability of each state category and the diagnostic confidence based on the target feature vector, and uses a confidence threshold to determine the identification of normal state and subdivided faults; and a diagnostic result output module, which integrates multi-dimensional diagnostic information based on the confidence threshold determination result to output full and standardized diagnostic results, providing a clear basis for operation and maintenance decisions, and using operation and maintenance feedback data to optimize the deep learning model.

[0032] Thirdly, the technical solution adopted by the present invention is: a computer-readable storage medium for storing one or more programs, wherein the one or more programs include instructions, which, when executed by a computing device, cause the computing device to perform any of the methods described above.

[0033] Fourthly, the technical solution adopted by the present invention is: a computing device comprising: one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs include instructions for performing any of the methods described above.

[0034] The present invention has the following advantages due to the adoption of the above technical solutions:

[0035] This invention overcomes the limitations of single-dimensional feature extraction, integrating multi-dimensional electrical, thermal, and physical features to comprehensively characterize the evolution process of failure modes unique to press-fit IGBTs. It employs a parameter-specific denoising and multi-modal data fusion strategy to enhance the model's resistance to electromagnetic interference and signal drift, adapting to extreme offshore operating conditions. Leveraging the multi-scale feature extraction capabilities of deep learning, it strengthens the capture of early, weak fault features, enabling early fault warnings. Optimizing the deep learning network structure, combined with end-to-end training and adaptive optimization, improves the model's hyperparameter optimization efficiency and generalization ability, adapting to multi-physics coupling characteristics. Overall, it significantly improves the accuracy and reliability of fault diagnosis for press-fit IGBTs, reduces offshore operation and maintenance costs, and ensures the stable operation of offshore power systems. Attached Figure Description

[0036] Figure 1 This is a flowchart of the fault diagnosis method for press-fit IGBTs in an embodiment of the present invention;

[0037] Figure 2 This is a structural diagram of the crimped IGBT fault diagnosis system in an embodiment of the present invention. Detailed Implementation

[0038] To address the shortcomings of existing fault diagnosis methods for press-fit IGBTs, this invention provides a fault diagnosis method, system, medium, and equipment for press-fit IGBTs. First, multi-dimensional monitoring data of the press-fit IGBTs is preprocessed to output standardized data suitable for the adaptation model. Second, a data analysis module automatically extracts and filters multi-dimensional features. Subsequently, a feature fusion module optimizes and outputs the target feature vector. Then, a confidence-driven fault identification module accurately distinguishes between normal states and subdivided faults. Finally, a complete diagnostic result is output and linked with the operation and maintenance feedback optimization model to form a closed-loop support, ensuring the safe operation of the converter valve.

[0039] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.

[0040] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0041] In one embodiment of the present invention, a fault diagnosis method for press-fit IGBTs is provided. In this embodiment, as shown... Figure 1 As shown, the method includes the following steps:

[0042] 1) The real-time monitoring data of the electric, thermal and physical dimensions of the press-fit IGBT are obtained in real time and then processed to obtain dual-modal data. The data preprocessing includes data cleaning, noise reduction, standardization and dual-modal reconstruction.

[0043] 2) Input the dual-modal data into the trained deep learning model, and automatically extract and filter multi-dimensional features through the temporal and spatial feature extraction submodule in the deep learning model; and dynamically align, adaptively fuse and optimize the filtered multi-dimensional features through the feature fusion module, and output the target feature vector that can accurately represent the running state.

[0044] 3) Calculate the probability and diagnostic confidence of each state category of the press-fit IGBT based on the target feature vector, and use the confidence threshold to determine the accurate identification of normal state and subdivided faults, avoiding false positives and false negatives. The states of the press-fit IGBT include normal conduction, normal turn-off, abnormal conduction, abnormal turn-off, short circuit, and open circuit.

[0045] 4) Based on the confidence threshold judgment results, integrate multi-dimensional diagnostic information to output full and standardized diagnostic results, provide clear basis for operation and maintenance decisions, and use operation and maintenance feedback data to optimize deep learning models, forming a closed-loop support system for the whole process to ensure the safe and stable operation of offshore flexible converter valves.

[0046] In step 1) above, monitoring data transmitted in real time from the front-end multi-source data acquisition unit is received. The acquisition frequency is synchronized with the front-end, covering three core parameters during the operation of the press-fit IGBT: electrical parameters, thermal parameters, and physical parameters. Immediately after receiving the data, a preliminary verification is performed to quickly identify data transmission anomalies. An immediate retransmission command is triggered for abnormal data to ensure the integrity of the received data. Simultaneously, the acquisition timestamps of each parameter are recorded to achieve time-series alignment, providing accurate time-dimensional support for subsequent time-series feature extraction.

[0047] In this embodiment, to address the issue of outliers and missing values ​​that are prone to occur in the real-time monitoring data of press-fit IGBTs, a highly adaptable real-time cleaning strategy is adopted to prevent invalid and distorted data from entering subsequent processes. Specifically, outlier removal uses the 3σ criterion, calculating the mean μ and standard deviation σ of each parameter in real time, identifying data exceeding the [μ-3σ, μ+3σ] range as outliers and removing them immediately. Missing value completion employs linear interpolation. For sporadic missing data within a short period, linear fitting is performed based on adjacent valid data points to quickly fill in the missing values, ensuring the continuity of data sequence. For data missing for extended periods, in addition to linear interpolation completion, a sensor status warning is simultaneously triggered, prompting maintenance personnel to check the corresponding sensor's operating status to prevent continuous data loss due to sensor malfunction.

[0048] In this embodiment, considering the complex operating environment of the press-fit IGBT and the susceptibility of monitoring data to electromagnetic interference, mechanical vibration noise, and cooling system noise, a parameter-specific denoising strategy is adopted instead of the traditional uniform denoising method. This ensures effective denoising while preserving subtle fault characteristics to the greatest extent possible. Specifically, wavelet threshold denoising is used for vibration and voltage / current signals. A db4 wavelet basis function is selected, and a three-level wavelet decomposition is set. Through adaptive threshold calculation, high-frequency electromagnetic interference and random noise are effectively filtered out, while retaining the fault characteristic frequency components in the signal. A five-point moving average filtering method is used for temperature signals. The average value of the current data point and the two adjacent data points is calculated in real time as the filtered output. This smooths temperature fluctuations while avoiding the loss of temperature change trend characteristics due to over-filtering. Median filtering is used for press-fit pressure signals to quickly filter out instantaneous impact noise and ensure the stability of pressure data.

[0049] In this embodiment, due to the significant differences in the dimensions and numerical ranges of different types of monitoring parameters, directly inputting them into a deep learning model would lead to slow model training convergence and imbalanced feature weights, affecting diagnostic accuracy. Therefore, this embodiment standardizes all cleaned and denoised data using a min-max normalization method, mapping each parameter data to the [0,1] interval in real time. The standardization formula is as follows: ,in For the data before standardization, For standardized data, , These are the historical maximum and minimum valid values ​​of the parameter, respectively.

[0050] In this embodiment, after standardization, all parameter data are on the same numerical order of magnitude, which not only adapts to the input requirements of deep learning models but also accelerates the model inference process and improves real-time diagnostic efficiency. To adapt to the temporal and spatial feature extraction requirements of dual-branch deep learning models, this embodiment adds a data reconstruction stage after data standardization to form dual-modal input data. This differs from the existing single-dimensional input method and can fully explore the temporal and spatial frequency domain features of multi-source data. Specifically, the dual-modal reconstruction process in data preprocessing is as follows: the standardized one-dimensional time-series data of various compression-type IGBTs is retained as the input of the temporal feature extraction branch for temporal feature extraction; for signals rich in frequency domain features, two-dimensional reconstruction is completed through short-time Fourier transform, transforming them into two-dimensional time-frequency maps to obtain two-dimensional time-frequency data of compression-type IGBTs, presenting the signal energy distribution in different spatiotemporal dimensions, providing sufficient feature information for spatial feature extraction. After dual-modal data reconstruction, format adaptation is completed and the data is output to the multi-source data input layer of the model, realizing the entire process of real-time data preprocessing and input.

[0051] Among them, signals rich in frequency domain characteristics refer to timing data located within the set switching frequency range of the press-fit IGBT. Since frequency domain characteristics refer to the Fourier series decomposition of a single bit timing signal, the fault characteristics of timing data in the frequency domain are correlated with the characteristics near the switching frequency of the IGBT. For example, if the fundamental frequency of the timing signal is 50Hz and the switching frequency is 600Hz, then the characteristics at the 12th harmonic are important characteristics.

[0052] In step 2) above, the preprocessed dual-modal data is received, and the data format is automatically adapted to the input interface of the deep learning model to ensure smooth data input to the corresponding feature extraction submodule. Specifically, one-dimensional time-series data is directly input to the time-series feature extraction submodule to capture the time-domain variations of parameters such as voltage, current, and temperature during the operation of the press-fit IGBT; two-dimensional time-frequency data is input to the spatial feature extraction submodule to mine the spatial and frequency domain features of the press-fit IGBT's vibration and voltage signals, achieving multi-dimensional parallel data processing and improving the feature extraction efficiency of the press-fit IGBT.

[0053] Specifically, the bimodal data is input into the trained deep learning model to extract and filter multi-dimensional features, including the following steps:

[0054] 2.1.1) For the input one-dimensional timing data of the press-fit IGBT, the recurrent neural network structure is used to capture the temporal correlation and dynamic change characteristics of the data, and automatically extract the key information features reflecting the operating status of the press-fit IGBT in the time domain, so as to effectively characterize whether the press-fit IGBT has potential faults, such as thermal failure, poor contact, etc.

[0055] Among them, the key information features reflecting the operating status of press-fit IGBTs in the time domain include the trend features, peak features, and periodic features of the operating status of press-fit IGBTs.

[0056] During feature extraction, nonlinear transformation capabilities are introduced through activation functions to enhance the fitting effect of submodules on the complex timing features of compression-type IGBTs. Commonly used activation functions include... The expression is:

[0057]

[0058] 2.1.2) The extracted features are screened, and redundant and invalid features are eliminated based on the feature importance evaluation index. The feature importance is determined by calculating the correlation coefficient between each time-series feature and the fault category of the press-fit IGBT. Features with an absolute value of correlation coefficient greater than a set threshold are retained to achieve the screening of time-domain features of the press-fit IGBT and improve the pertinence of subsequent fault diagnosis.

[0059] Among them, the correlation coefficient The calculation formula is:

[0060]

[0061] in, Let x be the covariance of a certain timing characteristic x of a press-fit IGBT and the fault category y. Let x be the variance of the time series feature. Let y be the variance of the fault category.

[0062] 2.1.3) For the input two-dimensional time-frequency data of the press-fit IGBT, the spatial distribution characteristics and frequency domain energy characteristics of the data are captured by convolution operation, and key information features in the spatial domain and frequency domain are automatically extracted to effectively reflect the potential characteristics of the press-fit IGBT fault, such as fretting wear of the press-fit IGBT and damage to the gate oxide layer.

[0063] Key information features in the spatial and frequency domains include: texture features and gradient features in the spatial domain, and characteristic frequencies and energy peaks in the frequency domain.

[0064] The core formula for convolution is:

[0065]

[0066] in, For two-dimensional time and frequency data input of press-fit IGBTs, For convolution kernel, The bias term is represented by M and N, which are the height and width of the convolution kernel, respectively. This is the feature map output by the convolution operation. Through multiple convolution operations and pooling, the dimensionality of the feature map is gradually reduced, enhancing key feature information related to faults in compression-type IGBTs. The core calculation formula for the pooling operation is:

[0067]

[0068] in, For pooling step size, For the input features within the pooling window, This is the pooling output feature.

[0069] 2.1.4) After feature extraction is completed, the feature importance evaluation mechanism is used to automatically select features that contribute more than a set threshold to the identification of IGBT faults by combining the inherent correlation between spatial domain and frequency domain features. Redundant information is eliminated to ensure the effectiveness and conciseness of the features, laying the foundation for subsequent accurate diagnosis.

[0070] 2.1.5) After extracting and filtering the time domain, spatial domain and frequency domain features of the press-fit IGBT, the effective features filtered from each dimension are initially integrated to form a multi-dimensional feature set for subsequent feature fusion.

[0071] In this embodiment, two types of core features are received: time-domain features captured by the time-series feature extraction submodule (such as trends, peak values, and periodic features of parameters like voltage, current, and temperature), and spatial and frequency-domain features mined by the spatial feature extraction submodule (such as texture, gradient features, characteristic frequencies, and energy peaks of vibration signals). Upon receipt, the two types of features are automatically validated for dimensionality and screened for validity, eliminating invalid features and redundant information caused by data interference or feature extraction bias. This ensures the reliability and relevance of the input features, laying the foundation for subsequent fusion calculations.

[0072] In step 2) above, the selected multi-dimensional features are dynamically aligned, adaptively fused, and optimized for dimensionality reduction to output a target feature vector that accurately represents the running state. This includes the following steps:

[0073] 2.2.1) Considering the dimensional differences and intrinsic correlations of features in the time domain, spatial domain, and frequency domain, the feature mapping mechanism of deep learning networks is used to first map various features (i.e., time domain, spatial domain, and frequency domain features in the multi-dimensional feature set) to the same feature space, thereby achieving dynamic alignment of multi-dimensional features and eliminating the problem of dimensional mismatch.

[0074] The core formula for feature mapping is:

[0075]

[0076] In the formula, For the first Feature vectors mapped from class features (time domain / spatial domain / frequency domain features); This is the mapping weight matrix for the corresponding features; This is the original feature vector; For mapping bias terms Nonlinear activation functions are used to enhance the capabilities of feature maps.

[0077] 2.2.2) An adaptive fusion strategy is adopted to deeply integrate the aligned features. The weight allocation mechanism strengthens the feature weights that are strongly correlated with the fault state of the press-fit IGBT and weakens the influence of redundant and interference features.

[0078] Specifically:

[0079]

[0080] In the formula, For the updated feature weights, As the current weight, Let L be the learning rate, and L be the fault diagnosis loss function. This represents the loss gradient corresponding to the weights.

[0081] 2.2.3) The deeply integrated features are weighted and summed to obtain the fused feature vector. The integrated features are then subjected to dimensionality reduction optimization to remove redundant information generated during the fusion process, so as to output the target feature vector after dimensionality reduction optimization.

[0082] The formula for calculating the fused feature vector is as follows:

[0083]

[0084] In the formula, The fused feature vector is denoted as n, which is the number of feature categories (here, n=3, corresponding to time domain, spatial domain, and frequency domain features), ensuring that the fused features can accurately capture the essential characteristics of various faults.

[0085] In this embodiment, the dimensionality reduction operation uses linear projection, and the formula is as follows:

[0086]

[0087] In the formula, The final output is the target feature vector. For the dimension-reduced projection matrix, This is the fused feature vector.

[0088] In this embodiment, the target feature vector has fully integrated the complementary information of the multi-dimensional operating characteristics of the press-fit IGBT, which can comprehensively and accurately characterize the differences in the characteristics of the normal operating state of the component and various subdivided faults. This provides core support for subsequent fault category probability calculation and confidence determination, effectively improving the accuracy and efficiency of the entire fault diagnosis process and ensuring the engineering practicality of fault diagnosis of key components of the converter valve.

[0089] The entire process is fully automated, requiring no manual intervention in feature design, extraction threshold setting, or other operations. This effectively improves the automation level and processing efficiency of the fault diagnosis process for key components of the converter valve. At the same time, through comprehensive mining of multi-dimensional features, it accurately captures the characteristic differences of various faults in press-fit IGBTs, providing a solid guarantee for the accurate identification of subsequent fault types.

[0090] In step 3) above, the target feature vector is received. This vector deeply integrates multi-dimensional effective features from the time, spatial, and frequency domains of the press-fit IGBT operating data, enabling a comprehensive and accurate characterization of the component's operating status. The system automatically performs format adaptation and dimensional verification between the target feature vector and the fault classification module's input interface, ensuring that the feature vector can be smoothly input into the classification model for subsequent calculations. Simultaneously, it removes potentially invalid information from the vector, guaranteeing the integrity and validity of the input data and laying the foundation for subsequent probability calculations.

[0091] Specifically, the probability of each state category and the diagnostic confidence level are calculated based on the target feature vector, and the identification of normal state and subdivided fault is achieved by using the confidence level threshold. This includes the following steps:

[0092] 3.1) The target feature vector is subjected to high-dimensional feature mapping and nonlinear transformation through the fully connected layer of the deep learning network, and the high-dimensional target feature vector is transformed into a linear feature vector corresponding to the fault category and normal state. Then, the probability quantization of each state category is achieved through normalization processing.

[0093] The core formula for feature mapping operation in fully connected layers is:

[0094]

[0095] In the formula, The linear feature vector output by the fully connected layer. This is the weight matrix of the fully connected layer. The input is the target feature vector. This is a bias term.

[0096] In this embodiment, the high-dimensional target feature vector is transformed into a linear feature vector corresponding to the fault category and normal state. Then, the probability quantification of each state category is achieved through normalization processing. Specifically, the linear feature vector output by the fully connected layer is transformed into a value that conforms to the probability distribution law. A normalization function is introduced to process the linear feature vector and output the probability value of each fault category and normal state.

[0097] The normalization function is expressed as follows:

[0098]

[0099] In the formula, Let k be the probability of the k-th type of state (a certain type of fault or normal state). Let be the linear feature value corresponding to the k-th state, and C be the total number of state categories (including normal states and various subdivided faults). This formula can ensure that the sum of the probabilities of all state categories is 1, thus achieving reasonable quantification and comparison of probabilities.

[0100] 3.2) Based on the probability quantification results of each state category, calculate the confidence level of the current diagnosis result. The confidence level is the maximum value among all state category probabilities, that is, the probability value corresponding to the most likely operating state.

[0101] The confidence level is:

[0102]

[0103] In the formula, For the confidence level of the diagnostic results, to These are the probability values ​​corresponding to each state category.

[0104] 3.3) Using confidence level as the judgment standard, when the confidence level reaches or exceeds the set threshold, it is judged as a clear diagnostic result and the corresponding operating status is output; when the confidence level is lower than the set threshold, it is judged as a suspected state, triggering an early warning signal to prompt maintenance personnel to strengthen component operation monitoring, avoid misjudgment or missed judgment due to weak features, complex operating conditions or interference signals, and improve the reliability of diagnostic results.

[0105] In step 4) above, based on the judgment results of the confidence-driven fault identification module, multi-dimensional diagnostic information is integrated to output full and standardized diagnostic results. This provides maintenance personnel with clear and explicit fault location and handling basis, constructing a closed-loop support system for the entire process of fault early warning, diagnosis, handling, and feedback, ensuring the scientific and efficient nature of maintenance decisions for key components of the converter valve. The focus is on the full output of information regarding the operating status of press-fit IGBTs.

[0106] Specifically, based on the confidence threshold determination results, multi-dimensional diagnostic information is integrated to output comprehensive and standardized diagnostic results, providing a clear basis for operation and maintenance decisions. This includes the following steps:

[0107] 4.1) Output the final fault diagnosis result based on the confidence level determination result.

[0108] If the diagnosis is definitive, the system simultaneously outputs the corresponding status category information, confidence level value, and probability distribution for each status category, allowing maintenance personnel to intuitively grasp the component's operating status and fault risks. If the status is suspected, the system outputs suspected warning information and probability distribution for each status, providing maintenance personnel with targeted monitoring directions. Simultaneously, the diagnosis results are fed back to the entire fault diagnosis system, achieving closed-loop control of the diagnosis process. This provides precise support for subsequent fault location, component repair, and other maintenance decisions, effectively improving the engineering practicality and reliability of fault diagnosis for key components of the converter valve.

[0109] 4.2) Based on the fault diagnosis results, comprehensively sort out and extract the operating status categories, the probability distribution of all categories, and the diagnostic confidence level;

[0110] Based on the operational status category, it can be clearly identified whether the output pressure-type IGBT is currently in normal operation or a specific fault type, accurately distinguishing typical faults such as thermal failure, fretting wear, poor contact, and gate oxide layer damage, ensuring that maintenance personnel can quickly grasp the core operational status of the component.

[0111] Based on the probability distribution of all categories, the system outputs the probability values ​​corresponding to all state categories, intuitively presenting the credibility distribution of the model's recognition, and providing data support for operations and maintenance personnel to judge state uncertainty.

[0112] Based on the diagnostic confidence level, the confidence level of the core judgment indicators is output synchronously, which clearly reflects the reliability of the current diagnostic results and provides a basis for the priority determination of subsequent operation and maintenance actions.

[0113] 4.3) Synchronously output key feature parameters with strong correlation. Based on the results of feature extraction and screening by deep learning network, key feature parameters with the highest similarity to the current operating state are extracted. Through the accurate output of key feature parameters, the specific location, severity and potential causes of the fault are directly pointed to, which greatly reduces the difficulty and time spent by maintenance personnel in fault location.

[0114] Among them, the key characteristic parameters with the highest similarity to the current operating state can accurately characterize the essential characteristics of the fault or the core operating rules of the normal state, covering multiple dimensions of core indicators of press-fit IGBTs, including electrical, thermal, and physical aspects. For example, for thermal failure faults, it outputs related characteristic parameters such as chip junction temperature and cooling system heat exchange efficiency; for poor contact faults, it outputs core characteristic parameters such as forward voltage drop fluctuation and conduction current stability; and for normal operating conditions, it outputs steady-state characteristic indicators of parameters in various dimensions.

[0115] In this embodiment, a complete closed-loop support system is constructed. The output of full diagnostic results provides maintenance personnel with clear fault handling guidelines. Maintenance personnel can combine state categories, probability distributions, confidence levels, and key characteristic parameters to quickly formulate targeted maintenance plans, fault handling procedures, and subsequent monitoring strategies, achieving efficient fault handling and hazard elimination. Simultaneously, the maintenance handling results are back-synchronized to the deep learning model training and optimization stage, providing practical engineering data support for model parameter iteration and improved recognition accuracy, continuously optimizing fault diagnosis performance. For suspected states that are not clearly defined, early warning information is output and the front-end monitoring module is linked to strengthen targeted data collection, effectively improving the engineering practicality and long-term reliability of fault diagnosis for key components of the converter valve, ensuring the safe and stable operation of the high-voltage direct current transmission system.

[0116] In one embodiment of the present invention, such as Figure 2 As shown, a crimp-type IGBT fault diagnosis system is provided, which includes:

[0117] The data preprocessing module takes the real-time monitoring data of the electric, thermal and physical dimensions of the press-fit IGBT and preprocesses it to obtain dual-modal data.

[0118] The data analysis and feature fusion module takes bimodal data as input to the trained deep learning model, extracts and filters multi-dimensional features, and performs dynamic alignment, adaptive fusion and dimensionality reduction optimization on the filtered multi-dimensional features, outputting a target feature vector that can accurately represent the running state.

[0119] The confidence-driven fault identification module calculates the probability of each state category and the diagnostic confidence based on the target feature vector, and uses the confidence threshold to identify normal state and subdivided faults.

[0120] The diagnostic results output module integrates multi-dimensional diagnostic information based on the confidence threshold to output full and standardized diagnostic results, providing a clear basis for operation and maintenance decisions, and using operation and maintenance feedback data to optimize deep learning models.

[0121] In the above embodiments, data preprocessing includes: data cleaning, noise reduction, standardization, and bimodal reconstruction;

[0122] The dual-modal reconstruction process is as follows: the standardized one-dimensional time-series data of various press-fit IGBTs are retained for time-series feature extraction; for signals rich in frequency domain features, two-dimensional reconstruction is completed through short-time Fourier transform, which is converted into a two-dimensional time-frequency diagram to obtain two-dimensional time-frequency data of press-fit IGBTs, so as to present the signal energy distribution in different spatiotemporal dimensions for spatial feature extraction.

[0123] In the above embodiments, the bimodal data is input into the trained deep learning model to extract and filter multi-dimensional features, including:

[0124] For the input one-dimensional timing data of the press-fit IGBT, the recurrent neural network structure is used to capture the temporal correlation and dynamic change characteristics of the data, and automatically extract the key information features reflecting the operating status of the press-fit IGBT in the time domain, so as to effectively characterize whether there are potential faults in the press-fit IGBT.

[0125] The extracted features are screened, and redundant and invalid features are eliminated based on the feature importance evaluation index. The feature importance is determined by calculating the correlation coefficient between each time-series feature and the fault category of the press-fit IGBT. Features with an absolute value of correlation coefficient greater than a set threshold are retained, thereby realizing the screening of time-domain features of the press-fit IGBT.

[0126] For the input two-dimensional time-frequency data of the press-fit IGBT, the spatial distribution characteristics and frequency domain energy characteristics of the data are captured by convolution operation, and key information features in the spatial domain and frequency domain are automatically extracted to effectively reflect the potential characteristics of the press-fit IGBT fault.

[0127] After feature extraction is completed, the feature importance evaluation mechanism is used to select features that contribute more than a set threshold to the identification of faults in press-fit IGBTs by combining the inherent correlation between spatial domain and frequency domain features, and redundant information is removed.

[0128] After extracting and filtering the time domain, spatial domain, and frequency domain features of the press-fit IGBT, the effective features filtered from each dimension are initially integrated to form a multi-dimensional feature set for subsequent feature fusion.

[0129] In the above embodiments, the selected multi-dimensional features are dynamically aligned, adaptively fused, and optimized for dimensionality reduction to output a target feature vector that accurately represents the operating state, including:

[0130] By using the feature mapping mechanism of deep learning networks, various features are first mapped to the same feature space to achieve dynamic alignment of multi-dimensional features;

[0131] An adaptive fusion strategy is adopted to deeply integrate the aligned features. The weight allocation mechanism strengthens the feature weights that are strongly correlated with the fault state of the press-fit IGBT and weakens the influence of redundant and interference features.

[0132] The deeply integrated features are weighted and summed to obtain a fused feature vector. The integrated features are then subjected to dimensionality reduction optimization to output the target feature vector.

[0133] In the above embodiments, the probability of each state category and the diagnostic confidence level are calculated based on the target feature vector, and the identification of normal state and subdivided fault is achieved by determining the confidence level threshold, including:

[0134] The target feature vector is transformed into a linear feature vector corresponding to the fault category and normal state by high-dimensional feature mapping and nonlinear transformation through the fully connected layer of the deep learning network. Then, the probability quantization of each state category is achieved through normalization.

[0135] Based on the probability quantification results of each state category, the confidence level of the current diagnosis result is calculated, and the confidence level is the maximum value among the probabilities of all state categories.

[0136] Using confidence level as the judgment standard, when the confidence level reaches or exceeds the set threshold, it is judged as a clear diagnostic result and the corresponding operating status is output; when the confidence level is lower than the set threshold, it is judged as a suspected state, triggering an early warning signal to prompt maintenance personnel to strengthen component operation monitoring.

[0137] In the above embodiments, the high-dimensional target feature vector is transformed into a linear feature vector corresponding to the fault category and normal state, and then the probability quantization of each state category is achieved through normalization processing, specifically:

[0138] The linear feature vectors output by the fully connected layer are transformed into numerical values ​​that conform to the probability distribution law. A normalization function is introduced to process the linear feature vectors and output the probability values ​​of each fault category and normal state.

[0139] The normalization function is expressed as follows:

[0140]

[0141] In the formula, Let be the probability of the k-th class state. Let be the linear feature value corresponding to the k-th state, and C be the total number of state categories.

[0142] In the above embodiments, based on the confidence threshold determination results, multi-dimensional diagnostic information is integrated to output full and standardized diagnostic results, providing a clear basis for operation and maintenance decisions, including:

[0143] Based on the confidence level determination, the final fault diagnosis result is output;

[0144] Based on the fault diagnosis results, the operating status category, the probability distribution of all categories, and the diagnostic confidence level are extracted. Based on the operating status category, the current normal operating status or specific sub-fault type of the press-fit IGBT is clearly output. Based on the probability distribution of all categories, the probability values ​​corresponding to all status categories are output, intuitively presenting the confidence distribution of the model's identification. Based on the diagnostic confidence level, the confidence level of the core judgment index is output simultaneously, clearly reflecting the reliability of the current diagnostic results.

[0145] Based on the results of feature extraction and screening using deep learning networks, key feature parameters with the highest similarity to the current operating state are extracted. Through the precise output of key feature parameters, the specific location, severity, and potential causes of the fault are directly identified.

[0146] The system provided in this embodiment is used to execute the above-described method embodiments. For specific processes and details, please refer to the above embodiments, which will not be repeated here.

[0147] In one embodiment of the present invention, a computing device is provided. This computing device can be a terminal and may include a processor, a communication interface, a memory, a display screen, and an input device. The processor, communication interface, and memory communicate with each other via a communication bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. When the computer programs are executed by the processor, they implement the methods provided in the above embodiments. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The communication interface is used for wired or wireless communication with an external terminal; wireless communication can be implemented using existing technology. The display screen can be a liquid crystal display (LCD) or an electronic ink display. The input device can be a touch layer covering the display screen, or buttons, a trackball, or a touchpad mounted on the casing of the computing device, or an external keyboard, touchpad, or mouse. The processor can call logical instructions in the memory to execute the methods provided in the above embodiments.

[0148] In one embodiment of the present invention, a computer program product is provided, the computer program product including a computer program stored on a non-transitory computer-readable storage medium, the computer program including program instructions, and when the program instructions are executed by a computer, the computer is able to perform the methods provided in the above-described method embodiments.

[0149] In one embodiment of the present invention, a non-transitory computer-readable storage medium is provided, which stores server instructions that cause a computer to perform the methods provided in the above embodiments.

[0150] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A press-pack IGBT fault diagnosis method for a flexible HVDC converter valve for offshore wind power, characterized by, include: The real-time monitoring data of the electric, thermal and physical dimensions of the press-fit IGBT are obtained in real time and then processed to obtain dual-modal data. The dual-modal data is input into the trained deep learning model for multi-dimensional feature extraction and filtering. The filtered multi-dimensional features are then dynamically aligned, adaptively fused, and optimized for dimensionality reduction, outputting a target feature vector that accurately characterizes the operating state. This includes: for the input one-dimensional time-series data of the press-fit IGBT, a recurrent neural network structure is used to capture the temporal correlation and dynamic change characteristics of the data, automatically extracting key information features reflecting the operating state of the press-fit IGBT in the time domain to effectively characterize whether there are potential faults in the press-fit IGBT; the extracted features are filtered, and redundant and invalid features are removed based on feature importance evaluation metrics. Feature importance is determined by calculating the correlation coefficient between each time-series feature and the fault category of the press-fit IGBT, retaining relevant features. Features with absolute values ​​of coefficients greater than a set threshold are used to filter time-domain features of press-fit IGBTs. For the input two-dimensional time-frequency data of press-fit IGBTs, convolution operations are used to capture the spatial distribution and frequency-domain energy characteristics of the data, automatically extracting key information features in both the spatial and frequency domains to effectively reflect potential fault characteristics of press-fit IGBTs. After feature extraction, a feature importance evaluation mechanism, combined with the inherent correlation between spatial and frequency domain features, filters out features whose contribution to press-fit IGBT fault identification is higher than a set threshold, eliminating redundant information. After completing the extraction and filtering of time-domain, spatial-domain, and frequency-domain features of press-fit IGBTs, the effective features filtered from each dimension are initially integrated to form a multi-dimensional feature set for subsequent feature fusion. The process involves dynamically aligning, adaptively fusing, and optimizing the multi-dimensional features selected from the data to output a target feature vector that accurately represents the operating state. This includes: firstly, using the feature mapping mechanism of a deep learning network, mapping various features to the same feature space to achieve dynamic alignment of multi-dimensional features; secondly, employing an adaptive fusion strategy to deeply integrate the aligned features, strengthening the feature weights strongly correlated with the fault state of the press-fit IGBT through a weight allocation mechanism while weakening the influence of redundant and interfering features; thirdly, weighting and summing the deeply integrated features to obtain a fused feature vector, and then optimizing the integrated features to output the target feature vector after dimensionality reduction. Based on the target feature vector, the probability of each state category and the diagnostic confidence level are calculated. The identification of normal state and subdivided fault is realized by using the confidence level as the judgment standard. When the confidence level reaches or exceeds the set threshold, it is judged as a clear diagnostic result and the corresponding operating status is output. When the confidence level is lower than the set threshold, it is judged as a suspected state, triggering an early warning signal to prompt maintenance personnel to strengthen the monitoring of component operation. Based on the confidence threshold determination results, multi-dimensional diagnostic information is integrated to output full and standardized diagnostic results, providing a clear basis for operation and maintenance decisions, and the operation and maintenance feedback data is used in reverse to optimize the deep learning model.

2. The crimp-type IGBT failure diagnosis method according to claim 1, characterized by, Data preprocessing includes: data cleaning, noise reduction, standardization, and bimodal reconstruction; The dual-modal reconstruction process is as follows: the standardized one-dimensional time-series data of various press-fit IGBTs are retained for time-series feature extraction; for signals rich in frequency domain features, two-dimensional reconstruction is completed through short-time Fourier transform, which is converted into a two-dimensional time-frequency diagram to obtain two-dimensional time-frequency data of press-fit IGBTs, so as to present the signal energy distribution in different spatiotemporal dimensions for spatial feature extraction.

3. The crimp-type IGBT failure diagnosis method according to claim 1, wherein Based on the target feature vector, the probability of each state category and the diagnostic confidence level are calculated. A confidence threshold is then used to determine the identification of normal states and subdivided faults, including: The target feature vector is transformed into a linear feature vector corresponding to the fault category and normal state by high-dimensional feature mapping and nonlinear transformation through the fully connected layer of the deep learning network. Then, the probability quantization of each state category is achieved through normalization. Based on the probability quantification results of each state category, the confidence level of the current diagnosis result is calculated, and the confidence level is the maximum value among the probabilities of all state categories.

4. The crimp-type IGBT failure diagnosis method according to claim 3, characterized by, The high-dimensional target feature vector is transformed into a linear feature vector corresponding to the fault category and normal state. Then, the probability quantization of each state category is achieved through normalization. Specifically: The linear feature vectors output by the fully connected layer are transformed into numerical values ​​that conform to the probability distribution law. A normalization function is introduced to process the linear feature vectors and output the probability values ​​of each fault category and normal state. The normalization function is expressed as follows: wherein is the probability of the kth class of states, is the linear eigenvalue corresponding to the kth class of states, and C is the total number of state classes.

5. The fault diagnosis method for press-fit IGBTs as described in claim 1, characterized in that, Based on the confidence threshold determination results, multi-dimensional diagnostic information is integrated to output comprehensive and standardized diagnostic results, providing a clear basis for operation and maintenance decisions, including: Based on the confidence level determination, the final fault diagnosis result is output; Based on the fault diagnosis results, the operating status category, the probability distribution of all categories, and the diagnostic confidence level are extracted; based on the operating status category, the current normal operating status or specific sub-fault type of the press-fit IGBT is clearly output; based on the probability distribution of all categories, the probability values ​​corresponding to all status categories are output, intuitively presenting the confidence distribution of the model's identification; based on the diagnostic confidence level, the confidence level of the core judgment index is output simultaneously, clearly reflecting the reliability of the current diagnostic results. Based on the results of feature extraction and screening using deep learning networks, key feature parameters with the highest similarity to the current operating state are extracted. Through the precise output of key feature parameters, the specific location, severity, and potential causes of the fault are directly identified.

6. A crimp-type IGBT fault diagnosis system, used to implement the crimp-type IGBT fault diagnosis method as described in any one of claims 1 to 5, characterized in that, include: The data preprocessing module takes the real-time monitoring data of the electric, thermal and physical dimensions of the press-fit IGBT and preprocesses it to obtain dual-modal data. The data analysis and feature fusion module takes bimodal data as input to the trained deep learning model, extracts and filters multi-dimensional features, and performs dynamic alignment, adaptive fusion and dimensionality reduction optimization on the filtered multi-dimensional features, outputting a target feature vector that can accurately represent the running state. The confidence-driven fault identification module calculates the probability of each state category and the diagnostic confidence based on the target feature vector, and uses the confidence threshold to identify normal state and subdivided faults. The diagnostic results output module integrates multi-dimensional diagnostic information based on the confidence threshold to output full and standardized diagnostic results, providing a clear basis for operation and maintenance decisions, and using operation and maintenance feedback data to optimize deep learning models.

7. A computer-readable storage medium for storing one or more programs, characterized in that, The one or more programs include instructions that, when executed by a computing device, cause the computing device to perform any of the methods described in claims 1 to 5.

8. A computing device, characterized in that, include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs including instructions for performing any of the methods described in claims 1 to 5.

Images