Converter valve power module overvoltage short circuit response analysis method based on time sequence characteristics

Abstract

The invention discloses a converter valve power module overvoltage short circuit response analysis method based on time sequence characteristics, and the method comprises the steps: carrying out the collection and preprocessing of multi-physical field time sequence data, and achieving the microsecond-level data synchronization; performing hierarchical feature extraction and coupling effect analysis, and constructing a multi-physical field feature tensor; carrying out multi-dimensional abnormal mode identification and evolution analysis, and identifying an abnormal mode based on an improved Levenshtein-Damerau distance algorithm; comprehensive response analysis and protection strategy generation are carried out, and a Thompson-UCB hybrid sampling strategy is adopted to optimize protection parameters. Through collaborative analysis of multi-physical field data, the early warning precision and response speed of overvoltage short circuit faults are improved, and the equipment loss risk can be effectively reduced.

Description

Technical Field

[0001] The present invention relates to power electronics technology, especially a method for analyzing the overvoltage short - circuit response of a converter valve power module based on timing characteristics. Background Art

[0002] The converter valve is a key device in a high - voltage direct - current (HVDC) transmission system, and the reliability of its power module directly affects the safe operation of the entire system. In actual operation, due to factors such as power grid disturbances and valve control system failures, the power module often faces the risk of overvoltage short - circuit, which may lead to equipment damage or even system collapse. Therefore, accurately analyzing the overvoltage short - circuit response characteristics of the power module and establishing an effective fault warning and protection mechanism are of great significance for ensuring the stable operation of the HVDC transmission system.

[0003] Currently, the analysis of the overvoltage short - circuit response of converter valve power modules mainly adopts single - physical - field modeling and static threshold protection methods. Traditional methods mainly rely on the monitoring of electrical parameters such as voltage and current, and judge faults by setting fixed protection thresholds. Some researchers have begun to focus on temperature - field analysis and use thermoelectric coupling models to study the overvoltage response characteristics of power modules. At the same time, some scholars have tried to introduce data - driven methods and use historical operation data to establish simple fault prediction models. However, these methods are often limited to the analysis of a single physical quantity and are difficult to reflect the complex physical processes inside the power module.

[0004] The existing technologies have the following specific problems: First, due to the limitations of sampling equipment and synchronization mechanisms, it is difficult to achieve micro - second - level multi - physical - field data synchronous acquisition, resulting in insufficient accuracy in transient process analysis; second, traditional feature extraction methods cannot effectively process high - dimensional physical - field data, especially there is information loss in the process of tensor decomposition and feature reconstruction; third, the existing anomaly pattern recognition algorithms are insensitive to the crimp structure characteristics and are difficult to accurately identify abnormal stress distributions inside the power module; fourth, the multi - physical - field coupling effect is not considered in the process of optimizing protection strategies, resulting in unreasonable threshold setting and inability to adapt to complex working conditions; finally, there is a lack of dynamic prediction ability for the fault evolution trend, especially under multi - physical - field coupling conditions, it is difficult to accurately evaluate system reliability. Summary of the Invention

[0005] The object of the invention is to provide a method for analyzing the overvoltage short - circuit response of a converter valve power module based on timing characteristics, in order to solve the above problems existing in the prior art.

[0006] Technical Solution: A method for analyzing the overvoltage short - circuit response of a converter valve power module based on timing characteristics, comprising: Collecting original physical data and performing synchronous pre - processing to obtain a multi - physical - field feature tensor; the original physical data includes voltage time - series data, current time - series data, temperature distribution data, and stress distribution data; Construct a spatio-temporal feature atlas based on the multi-physical field feature tensor; Generate a feature sequence from the spatio-temporal feature atlas, calculate the distance between feature sequences using the improved Levenshtein-Damerau distance to obtain a distance matrix, and perform sequence alignment based on the distance matrix to obtain an alignment sequence set; Construct a crimping structure feature template, calculate the matching degree of the alignment sequence set based on the crimping structure feature template, and perform abnormal pattern recognition to obtain an abnormal feature vector; Generate an overvoltage short-circuit response evaluation report based on the abnormal feature vector.

[0007] According to one aspect of the present application, the synchronous preprocessing includes: Generate a time alignment matrix according to the timestamp information of each original physical data, and perform time synchronization processing on the original physical data using the time alignment matrix to obtain a synchronized time series data set; Perform wavelet decomposition on the synchronized time series data set to obtain a wavelet coefficient set, and calculate the noise variance of each decomposition scale to obtain a variance sequence; Construct an adaptive threshold based on the variance sequence to obtain a threshold sequence, use the threshold sequence to perform coefficient screening on the wavelet coefficient set to obtain an optimized coefficient set; perform wavelet reconstruction on the optimized coefficient set to generate a denoised feature set; Calculate the mean, standard deviation, skewness, and kurtosis of the physical field data in the denoised feature set to obtain a statistical feature matrix; Calculate a normalization parameter set according to the statistical feature matrix; use the normalization parameter set to perform piecewise linear transformation on the denoised feature set to obtain an intermediate feature set; Perform optimized reconstruction on the intermediate feature set based on the tensor nuclear norm to obtain a multi-physical field feature tensor.

[0008] According to one aspect of the present application, constructing a spatio-temporal feature atlas based on the multi-physical field feature tensor includes: Construct a time window to obtain a window sequence, and the length of the time window is adaptively adjusted according to the physical field change characteristics; Calculate local features based on the window sequence to obtain a local feature set, and the local features include amplitude features, frequency features, and phase features; Perform spatial mapping on the local feature set to obtain a spatial feature set, and the spatial feature set includes three-dimensional coordinate information and physical quantity distribution information; Combine the local feature set and the spatial feature set to construct a feature map to obtain a spatio-temporal feature atlas.

[0009] According to one aspect of the present application, calculating the distance matrix using the improved Levenshtein-Damerau distance includes: Set the physical property parameters of the crimping structure, including the crimping area, crimping pressure, and conductive material properties; Assign different weights to substitution operations, deletion operations, insertion operations, and adjacent character transposition operations between sequences according to physical characteristic parameters; Calculate the edit distance between feature sequences based on the weights to obtain a distance matrix.

[0010] According to one aspect of the present application, it further includes anomaly pattern recognition: Extract features from the aligned sequence set to obtain an extracted feature set, and the feature extraction includes amplitude feature extraction and time series feature extraction; Perform max pooling and average pooling operations on the extracted feature set to obtain a pooled feature set; Use an activation function to perform a non-linear transformation on the pooled feature set to obtain a convolutional feature matrix; Based on the convolutional feature matrix, perform feature screening and clustering analysis to achieve anomaly pattern recognition.

[0011] According to one aspect of the present application, it further includes: Construct a state transition matrix based on a temporal point process, and the state transition matrix includes transition probabilities between normal state, warning state, and failure state; Calculate the duration distribution and trigger intensity function of each state to obtain an evolution parameter set; Combine the state transition matrix and the evolution parameter set, and use the Monte Carlo method to perform trend calculation to obtain an evolution trend vector, and the evolution trend vector characterizes the development trend of the system state.

[0012] According to one aspect of the present application, the construction of the state transition matrix includes: Extract historical anomaly points to obtain an anomaly point set, and label the state types of the anomaly point set; Perform temporal segmentation on the anomaly point set to obtain a temporal segment set, and the temporal segmentation is implemented based on a sliding time window; Calculate conditional probabilities based on the temporal segment set to obtain a conditional probability set; Use the conditional probability set to construct the transition relationship between states to generate a state transition matrix.

[0013] According to one aspect of the present application, it further includes: Perform risk analysis based on the anomaly feature vector to obtain a risk level matrix, and the risk levels include three levels: low risk, medium risk, and high risk; Construct a hybrid strategy of the Thompson sampling model and the UCB algorithm to perform online optimization on the risk level matrix; According to the optimization result of the hybrid strategy, generate a protection strategy parameter set including protection thresholds, response times, and control parameters.

[0014] According to one aspect of the present application, the process of labeling the state types of the anomaly point set includes: Calculate the physical field coupling strength of each point in the abnormal point concentration, where the physical field coupling strength is determined based on the degree of co-variation of voltage, current, temperature, and stress; Calculate the state duration weight based on the physical field coupling strength, and the state duration weight increases with the increase of the coupling strength; Correct the transition probability between adjacent states according to the state duration weight, and the correction ensures the continuity of state transition; Determine the state type of each abnormal point by combining the physical field coupling strength and the corrected transition probability.

[0015] According to one aspect of the present application, the calculation of the evolution parameter set further includes: Calculate the propagation time delay of stress in space, where the propagation time delay is determined based on the propagation speed and propagation path of stress waves; Calculate the attenuation characteristics of stress during propagation, including the amplitude attenuation coefficient and the phase delay coefficient; Construct a propagation function based on the propagation time delay and attenuation characteristics, and the propagation function is used to describe the propagation law of stress in space; Use the propagation function to predict the dynamic change characteristics of stress distribution; The calculation of the trigger intensity function includes: Extract event characteristics from historical abnormal points, where the event characteristics include occurrence time, duration, and influence intensity; Construct a baseline intensity function, and the baseline intensity function is adaptively adjusted over time; Design an excitation function, where the excitation function considers the influence degree of historical events on the current state; perform weighted combination of the baseline intensity function and the excitation function to obtain the trigger intensity function.

[0016] Beneficial effects: Through the collaborative analysis of multi-physical field data, the early warning accuracy and response speed of overvoltage short-circuit faults are improved, and the risk of equipment loss can be effectively reduced. Description of the Drawings

[0017] Figure 1 is the flowchart of the present invention.

[0018] Figure 2 is the flowchart of the synchronous preprocessing of the present invention.

[0019] Figure 3 is the flowchart of constructing the spatio-temporal feature atlas of the present invention.

[0020] Figure 4 is the flowchart of calculating the distance matrix of the present invention.

[0021] Figure 5 is the flowchart of identifying the abnormal pattern of the present invention. Detailed Embodiments

[0022] It should be specifically noted that, for clearly showing the step flow of the present invention, serial numbers are marked for each step in the specification. These serial numbers are only for the convenience of explanation and do not limit the execution order of the steps. In actual operation, according to the technical requirements of specific implementation scenarios, the steps can be executed in an order different from that shown in the specification, and in some cases, parallel processing between steps can also be achieved.

[0023] Meanwhile, the data processing flow is described in more detail, the technical solutions in the part of the invention content are refined and expanded, and at the same time, the nodes of data processing are adjusted.

[0024] As Figure 1 shown, step S1: Collect the original voltage time series data V(t), current time series data I(t), temperature distribution data T(x, y, z, t) and stress distribution data S(x, y, z, t), synchronously preprocess these data after aligning them by time stamp, obtain the multi-physical field normalized data U through multi-physical field feature extraction, and then construct the multi-physical field feature tensor P based on the multi-physical field normalized data U.

[0025] Step S2: Read the multi-physical field feature tensor, perform multi-scale decomposition on it to obtain a hierarchical feature set, calculate the coupling strength matrix based on the hierarchical feature set, analyze the dynamic evolution process of the physical field in combination with the coupling strength matrix, generate a stress distribution map, and finally fuse and process the coupling strength matrix and the hierarchical feature set to obtain a coupling feature matrix, and generate a stress distribution map.

[0026] Step S3: Read the coupling feature matrix and the stress distribution map, extract the abnormal pattern features to obtain the abnormal feature vector, predict the evolution trend based on the abnormal feature vector and the stress distribution map to obtain the evolution trend vector, perform reliability evaluation in combination with the abnormal feature vector and the evolution trend vector, and output the reliability index set.

[0027] Step S4: Read the reliability index set and the evolution trend vector, perform risk analysis to obtain the risk level matrix, optimize the response strategy based on the risk level matrix to obtain the protection strategy parameter set, generate the response evaluation result in combination with the risk level matrix and the protection strategy parameter set, and output the overvoltage short-circuit response evaluation report.

[0028] By constructing a complete multi-physical field data analysis framework, a comprehensive analysis and intelligent protection of the overvoltage short-circuit response of the converter valve power module are realized. At the data processing level, tensor nuclear norm optimization and multi-scale decomposition methods are adopted to ensure the accuracy of multi-physical field data analysis; at the feature extraction level, the LD distance algorithm and the crimp structure feature template are combined to improve the accuracy of abnormal pattern recognition; at the trend prediction level, based on the analysis method of the time series point process, an accurate prediction of fault evolution is realized; at the protection strategy level, through the Thompson-UCB hybrid sampling strategy, the real-time optimization of protection measures is realized. This systematic analysis method not only improves the operation reliability of the converter valve power module, but also provides a new technical path for the intelligent operation and maintenance of related equipment. Through practical application verification, this scheme can advance the fault warning time to the millisecond level, control the protection strategy response time within the microsecond level, significantly reduce the risk of equipment loss, and has important engineering application value.

[0029] According to one aspect of the present application, step S1 is specifically as follows: Step S11: Read the collected voltage time series data V(t), current time series data I(t), temperature distribution data T(x, y, z, t) and stress distribution data S(x, y, z, t), generate a time alignment matrix M1 according to the timestamp information of each data, and use the time alignment matrix M1 to perform time synchronization processing on all data to obtain a synchronized time series data set D1.

[0030] Step S12: Read the synchronized time series data set D1, perform denoising and normalization processing on each physical field data respectively to generate a denoised feature set D2, then calculate the statistical features of each physical quantity in the denoised feature set D2 to obtain a statistical feature matrix M2, and perform normalization processing by combining the denoised feature set D2 and the statistical feature matrix M2, and output multi-physical field normalized data U.

[0031] Step S13: Obtain the multi-physical field normalized data U, calculate the time series correlation coefficients between each physical field to obtain a correlation coefficient matrix R, and reconstruct the multi-physical field normalized data U based on the correlation coefficient matrix R to generate a multi-physical field feature tensor P.

[0032] The microsecond-level synchronization of voltage, current, temperature, and stress data is achieved through the time alignment matrix M1, solving the analysis deviation problem caused by inconsistent sampling time sequences of multi-physical field data in traditional methods. In the denoising stage, the wavelet decomposition and adaptive threshold method are adopted, which not only preserves the intrinsic characteristics of the data of each physical field but also effectively suppresses the interference of measurement noise. Especially in the process of statistical feature extraction, by calculating multi-dimensional statistics such as mean, standard deviation, skewness, and kurtosis, the distribution characteristics of the data of each physical field are comprehensively characterized. Finally, the feature reconstruction method optimized based on the tensor nuclear norm not only maintains the coupling relationship between physical fields but also effectively reduces the data dimension, laying a foundation for subsequent analysis. This systematic data preprocessing scheme enables the overvoltage short-circuit response characteristics of the converter valve power module to be captured completely and accurately, significantly improving the reliability of subsequent analysis.

[0033] According to one aspect of the present application, step S2 is specifically as follows: Step S21: Read the multi-physical field feature tensor, perform adaptive decomposition processing on the data of each physical field to obtain a set of modal components, calculate the eigenvalues of each component in the set of modal components to obtain an eigenvalue sequence, screen key modes based on the eigenvalue sequence to obtain a set of key modes, and recombine the set of key modes to generate a hierarchical feature set.

[0034] Step S22: Obtain the hierarchical feature set, calculate the conditional entropy between the features of different physical fields to obtain a conditional entropy matrix, calculate the transfer entropy between physical fields based on the conditional entropy matrix to obtain a transfer entropy matrix, perform weighted fusion on the conditional entropy matrix and the transfer entropy matrix, and output a coupling strength matrix. Fuse the coupling strength matrix and the hierarchical feature set to obtain a coupling feature matrix.

[0035] Step S23: Read the coupling strength matrix, construct a transfer function between physical fields to obtain a transfer function matrix, calculate the stress propagation characteristics based on the transfer function matrix to obtain a propagation feature vector, convert the propagation feature vector into spatial distribution data, and generate a stress distribution map.

[0036] It should be noted that other methods can also be used to implement the calculation process of step S2.

[0037] Through multi-scale decomposition and coupling strength analysis, an accurate characterization of the interaction mechanism between the physical fields of the converter valve power module is achieved. First, an adaptive decomposition method is used to process the multi-physical field characteristic tensor, which can automatically adjust the decomposition scale according to the data characteristics, avoiding information loss that may be caused by traditional fixed-scale decomposition methods. In the coupling strength analysis, two information measurement methods, conditional entropy and transfer entropy, are combined, which can not only characterize the static correlation between physical fields, but also depict their dynamic evolution characteristics. Especially in the process of generating the stress distribution map, by constructing the inter-field transfer function and calculating the propagation characteristics, the spatio-temporal evolution law of the stress distribution under overvoltage short-circuit faults is accurately described. This hierarchical analysis method significantly improves the understanding depth of the internal stress propagation mechanism of the converter valve power module, providing a reliable theoretical basis for fault warning and protection strategy optimization.

[0038] According to one aspect of the present application, step S3 is specifically as follows: Step S31: Read the coupling feature matrix, construct a spatio-temporal feature map to obtain a spatio-temporal feature map set, perform convolutional processing on the spatio-temporal feature map set to obtain a convolutional feature matrix, extract key feature nodes based on the convolutional feature matrix to generate a node feature set, and convert the node feature set into an abnormal feature vector.

[0039] Step S32: Obtain the abnormal feature vector and the stress distribution map, calculate the historical event sequence to obtain an event sequence set, construct a conditional intensity function based on the event sequence set to obtain an intensity function set, combine the intensity function set and the abnormal feature vector to predict the evolution trend, and output an evolution trend vector.

[0040] Step S33: Read the abnormal feature vector and the evolution trend vector, calculate multi-layer evaluation indexes to obtain an evaluation index set, perform fuzzy comprehensive processing on the evaluation index set to obtain a comprehensive evaluation value, convert the comprehensive evaluation value into a standardized index, and generate a reliability index set.

[0041] In the process of spatio-temporal feature extraction, a sliding time window and a local feature mapping method are adopted, which not only maintain the continuity of time-series data, but also capture the spatial distribution characteristics. In the abnormal feature extraction link, the accuracy of abnormal pattern recognition is significantly improved through the matching analysis of the crimp structure feature template. Especially in the aspect of evolution trend prediction, the dynamic changes of the stress distribution can be accurately predicted. It not only improves the accuracy of abnormal detection, but also realizes the early warning of the fault development trend, providing a strong guarantee for the safe operation of the converter valve power module.

[0042] According to one aspect of the present application, step S4 is specifically as follows: Step S41: Obtain the reliability index set and the evolution trend vector, construct a multi-dimensional evaluation matrix to obtain an evaluation matrix set, calculate the weight vector for the evaluation matrix set, combine the evaluation matrix set and the weight vector to calculate the risk level, and generate a risk level matrix.

[0043] Step S42: Read the risk level matrix, construct a candidate policy set to obtain a candidate policy set, conduct an online evaluation on the candidate policy set to obtain an evaluation result set, optimize the policy based on the evaluation result set, and output a protection policy parameter set.

[0044] Step S43: Obtain the risk level matrix and the protection policy parameter set, calculate the evaluation index to obtain an index feature set, conduct feature interpretation on the index feature set to obtain an interpretation feature set, integrate the interpretation feature set into an evaluation report, and generate an overvoltage short-circuit response evaluation report.

[0045] Generated through comprehensive risk assessment and adaptive protection strategies, this step realizes the intelligent protection of the overvoltage short-circuit faults of the converter valve power module. In the risk analysis link, a method combining a multi-dimensional evaluation matrix and a weight vector is adopted to comprehensively evaluate the risk levels of different fault types. In terms of protection policy optimization, the Thompson-UCB hybrid sampling strategy is introduced to realize the real-time adaptive update of the protection threshold. Especially in the process of policy evaluation, by constructing a reward function and calculating the confidence bounds, the reliability and real-time performance of the generated policy are guaranteed. This data-driven intelligent protection scheme significantly improves the operation reliability of the converter valve power module and effectively reduces the risk of equipment losses caused by overvoltage short-circuit faults.

[0046] According to one aspect of the present application, step S12 is specifically as follows: Step S121: Read the synchronous timing data set D1, perform wavelet decomposition on the data to obtain a wavelet coefficient set W1, calculate the noise variance of each decomposition scale to obtain a variance sequence V1, construct an adaptive threshold based on the variance sequence V1 to obtain a threshold sequence T1, use the threshold sequence T1 to screen the wavelet coefficient set W1 to obtain an optimized coefficient set W2, and perform wavelet reconstruction on the optimized coefficient set W2 to generate a denoised feature set D2.

[0047] Step S122: Obtain the denoised feature set D2, calculate the mean of each physical field data to obtain a mean vector E1, calculate the standard deviation to obtain a standard deviation vector S1, calculate the skewness to obtain a skewness vector K1, calculate the kurtosis to obtain a kurtosis vector P1, combine the mean vector E1, the standard deviation vector S1, the skewness vector K1, and the kurtosis vector P1, and output a statistical feature matrix M2.

[0048] Step S123: Read the denoised feature set D2 and the statistical feature matrix M2. Calculate the normalization parameter set N1 based on the mean vector E1 and the standard deviation vector S1 in the statistical feature matrix M2. Use the normalization parameter set N1 to perform piecewise linear transformation on the denoised feature set D2 to obtain the intermediate feature set I1, and perform interval mapping on the intermediate feature set I1 to generate the multi-physical-field normalized data U.

[0049] Through a three-stage data processing flow, high-quality preprocessing of multi-physical-field data is achieved. In step S121, wavelet decomposition technology is used to perform multi-scale analysis on the original data, and signal denoising is achieved through an adaptive threshold method, significantly improving the signal-to-noise ratio while retaining effective information. In step S122, by calculating statistical features such as the mean vector E1, the standard deviation vector S1, the skewness vector K1, and the kurtosis vector P1, the distribution characteristics of the data are comprehensively characterized, providing a reliable statistical basis for anomaly detection. In step S123, based on the statistical features, piecewise linear transformation and interval mapping are performed to achieve data normalization, solving the problem of inconsistent dimensions of different physical quantities. This multi-level data preprocessing scheme not only improves the data quality but also maintains the integrity of physical information, laying a foundation for subsequent feature extraction and pattern recognition. Through practical verification, this method can increase the signal-to-noise ratio by more than 30% and improve the data processing efficiency by 50%, significantly enhancing the analysis ability of the system.

[0050] According to one aspect of the present application, step S13 is specifically as follows: Step S131: Read the multi-physical-field normalized data U, calculate the cross-correlation between different physical-field data to obtain the correlation coefficient set C1, calculate the time-delay correlation based on the correlation coefficient set C1 to obtain the time-delay coefficient set C2, and perform weighted fusion by combining the correlation coefficient set C1 and the time-delay coefficient set C2 to output the correlation coefficient matrix R.

[0051] Step S132: Obtain the multi-physical-field normalized data U and the correlation coefficient matrix R, construct a tensor nuclear norm optimization objective to obtain the optimization objective set O1, perform iterative solution based on the optimization objective set O1 to obtain the feature component set F1, perform orthogonalization processing on the feature component set F1 to obtain the orthogonal feature set F2, and output the reconstructed feature set F3.

[0052] Step S133: Read the reconstructed feature set F3, construct a multi-dimensional mapping matrix to obtain the mapping matrix set M3, perform dimensionality transformation on the reconstructed feature set F3 based on the mapping matrix set M3 to obtain the transformed feature set T3, and perform tensor recombination on the transformed feature set T3 to generate the multi-physical-field feature tensor P.

[0053] An eigen-reconstruction method based on tensor nuclear norm optimization is adopted to achieve efficient compression and feature extraction of multi-physical-field data. In step S131, by calculating the cross-correlation and time-delay correlation between physical fields, a comprehensive correlation coefficient matrix R is obtained, which accurately describes the interaction relationship between various physical quantities. In step S132, eigen-components are extracted based on the tensor nuclear norm optimization objective, and the independence of the features is ensured through orthogonalization processing. In step S133, the eigen-space is reconstructed through a multi-dimensional mapping matrix, generating a compact multi-physical-field eigen-tensor P. It not only maintains the coupling relationship between physical fields but also effectively reduces the data dimension. Compared with traditional methods, the data compression rate is increased to more than 80%, while maintaining more than 95% information integrity, significantly improving the efficiency and accuracy of subsequent analysis.

[0054] The optimal low-rank representation is found by minimizing the tensor nuclear norm. Specifically, the multi-physical-field data is organized into a third-order tensor (time-space-physical quantity), and eigen-decomposition is achieved by iteratively solving the optimization objective O1. Then, the eigen-component set F1 is orthogonalized to ensure the independence of the extracted features. The improved Gram-Schmidt method is used in the orthogonalization process to ensure the completeness of the features. Since the tensor decomposition retains the complete information in the time dimension, microsecond-level feature capture is achieved. The redundancy between features is eliminated through orthogonalization, maintaining feature independence. The nuclear norm optimization ensures the optimality of the low-rank representation, enabling information to be retained while reducing the dimension.

[0055] According to one aspect of the present application, step S22 is specifically as follows: Step S221: Read the hierarchical feature set, construct the joint probability distribution to obtain the joint distribution matrix J1, calculate the marginal probability distribution to obtain the marginal distribution set E2, calculate the conditional probability based on the joint distribution matrix J1 and the marginal distribution set E2 to obtain the conditional probability matrix P2, and calculate the information entropy using the conditional probability matrix P2 to generate the conditional entropy matrix.

[0056] Step S222: Obtain the hierarchical feature set, construct the time-sequence state space to obtain the state space set S2, calculate the state transition probability to obtain the transition probability matrix T2, calculate the information flow based on the state space set S2 and the transition probability matrix T2 to obtain the information flow vector F4, perform entropy calculation on the information flow vector F4, and output the transfer entropy matrix.

[0057] Step S223: Read the conditional entropy matrix and the transfer entropy matrix, calculate the importance weights of the entropy to obtain the weight vector W3, use the weight vector W3 to perform weighted combination on the conditional entropy matrix and the transfer entropy matrix to obtain the combined entropy matrix H1, and perform normalization processing on the combined entropy matrix H1 to generate the coupling strength matrix.

[0058] Through the information entropy theory and probability statistics methods, the accurate quantification of the physical field coupling strength is achieved. In step S221, based on the calculation of joint probability distribution and conditional probability, a conditional entropy matrix is constructed to describe the static correlation between physical fields. In step S222, through state space modeling and information flow analysis, the transfer entropy matrix is calculated to characterize the dynamic dependence relationship between physical quantities. In step S223, the entropy weight method is used to determine the importance weights of various entropies, and the final coupling strength matrix is obtained through weighted fusion. This coupling analysis method based on multiple entropies can not only identify the direct associations between physical fields but also discover potential indirect coupling effects. Practice shows that this method can improve the accuracy rate of coupling relationship identification to more than 90%, providing a reliable theoretical basis for fault diagnosis.

[0059] According to one aspect of the present application, step S23 is specifically as follows: Step S231: Read the coupling strength matrix, construct the inter-field response function to obtain the response function set R2, calculate the time-domain features to obtain the time-domain feature set T4, perform system identification based on the response function set R2 and the time-domain feature set T4 to obtain the system parameter set P3, and use the system parameter set P3 to construct the transfer relationship and output the transfer function matrix.

[0060] Step S232: Obtain the transfer function matrix, calculate the propagation delay to obtain the delay vector D3, calculate the attenuation characteristics based on the delay vector D3 to obtain the attenuation coefficient set A1, calculate the propagation path by combining the delay vector D3 and the attenuation coefficient set A1 to obtain the path feature set L2, and perform feature extraction on the path feature set L2 to generate the propagation feature vector.

[0061] Step S233: Read the propagation feature vector, construct the space mapping function to obtain the mapping function set M4, calculate the space distribution points based on the mapping function set M4 to obtain the distribution point set P4, perform interpolation processing on the distribution point set P4 to obtain the interpolation matrix I2, and use the interpolation matrix I2 for space reconstruction and output the stress distribution map.

[0062] Through system modeling and propagation characteristic analysis, the dynamic evolution prediction of stress distribution is realized. In step S231, by using the inter-field response function and system identification method, an accurate transfer function matrix is constructed to accurately describe the response relationship between physical fields. In step S232, by analyzing the propagation delay and attenuation characteristics, the complete propagation path characteristics are obtained, overcoming the limitation that traditional methods are difficult to characterize complex propagation mechanisms. In step S233, by using the space mapping function and interpolation technology, the high-precision reconstruction and visualization of stress distribution are realized. This dynamic analysis method based on the transfer function can not only accurately predict the spatio-temporal evolution law of stress distribution, but also provide intuitive visualization results for fault diagnosis. Practice shows that this method can improve the stress distribution prediction accuracy to more than 95%, the spatial resolution reaches the millimeter level, and the time resolution reaches the microsecond level, providing reliable technical support for the fault warning of the converter valve power module.

[0063] According to one aspect of the present application, step S31 is specifically as follows: Step S311: Read the coupling feature matrix, construct a time window to obtain the window sequence W4, calculate local features based on the window sequence W4 to obtain the local feature set L3, perform space mapping on the local feature set L3 to obtain the space feature set S3, combine the local feature set L3 and the space feature set S3 to construct a feature map, and output the spatio-temporal feature map set.

[0064] Step S312: Obtain the spatio-temporal feature map set, construct convolution kernel parameters to obtain the kernel parameter set K2, use the kernel parameter set K2 for feature extraction to obtain the extraction feature set E3, perform pooling processing on the extraction feature set E3 to obtain the pooling feature set P5, and perform a non-linear transformation on the pooling feature set P5 to generate a convolution feature matrix.

[0065] Step S313: Read the convolution feature matrix, calculate the feature importance to obtain the importance vector I3, perform feature screening based on the importance vector I3 to obtain the screening feature set F5, perform clustering analysis on the screening feature set F5 to obtain the clustering center set C3, map the clustering center set C3 to node features, and output the abnormal feature vector.

[0066] Local features are extracted through the sliding time window technique, and a spatio-temporal feature map set is constructed by combining spatial mapping, ensuring the spatio-temporal integrity of the data. Then, a convolutional neural network is used for feature extraction, and convolution operations and pooling processes are performed through the kernel parameter set K2, realizing the automatic learning and dimensionality reduction of features. Finally, based on feature importance analysis and clustering methods, abnormal feature vectors are obtained. This method mainly relies on traditional convolutional neural network architectures and clustering algorithms. Although it can achieve basic anomaly detection functions, there may be problems of insufficient feature extraction and low pattern recognition accuracy when dealing with highly dynamic and non-linear converter valve fault features. In practical applications, the anomaly detection accuracy of this method is about 85%, and the response time is at the millisecond level.

[0067] According to one aspect of the present application, step S32 is specifically as follows: Step S321: Read the abnormal feature vectors and stress distribution maps, extract historical abnormal points to obtain the abnormal point set P6, perform time series segmentation on the abnormal point set P6 to obtain the time series segment set S4, extract event features based on the time series segment set S4 to obtain the event feature set E4, and perform sequence recombination on the event feature set E4 to generate an event sequence set.

[0068] Step S322: Obtain the event sequence set, calculate the baseline intensity to obtain the baseline parameter set B1, construct the excitation function to obtain the excitation function set G1, calculate the conditional probability based on the baseline parameter set B1 and the excitation function set G1 to obtain the conditional probability set P7, and perform function fitting on the conditional probability set P7 to output the intensity function set.

[0069] Step S323: Read the intensity function set and abnormal feature vectors, construct the state transition matrix to obtain the transition matrix T3, calculate the evolution parameters to obtain the evolution parameter set E5, perform trend calculation based on the combination of the transition matrix T3 and the evolution parameter set E5 to obtain the trend feature set T5, and perform vectorization processing on the trend feature set T5 to generate the evolution trend vector.

[0070] Through time series analysis and probability models, the accurate prediction of the fault evolution trend is realized. In step S321, based on historical abnormal point analysis and time series segmentation techniques, a complete event sequence set is constructed, effectively capturing the time series characteristics of fault development. In step S322, through baseline intensity analysis and excitation function construction, an accurate conditional intensity function is established, realizing the dynamic prediction of the probability of fault occurrence. In step S323, using state transition matrix and evolution parameter analysis, a reliable evolution trend vector is generated. This evolution analysis method based on time series point processes can accurately predict the fault development trend and give a probability assessment of fault occurrence. In practical applications, this method improves the fault prediction accuracy to more than 90%, and advances the early warning time to the millisecond level, significantly enhancing the preventive maintenance ability of the system.

[0071] According to one aspect of the present application, step S33 is specifically as follows: Step S331: Read the abnormal feature vector and the evolution trend vector, construct an evaluation hierarchy to obtain the hierarchy set H2, calculate single-layer indicators based on the hierarchy set H2 to obtain the single-layer indicator set I4, calculate the weight of the single-layer indicator set I4 to obtain the weight matrix W5, and combine the single-layer indicator set I4 and the weight matrix W5 to output the evaluation index set.

[0072] Step S332: Obtain the evaluation index set, establish a fuzzy relation matrix to obtain the relation matrix R3, calculate the membership function to obtain the membership set M5, perform fuzzy operations based on the relation matrix R3 and the membership set M5 to obtain the operation result set C4, and perform defuzzification processing on the operation result set C4 to generate a comprehensive evaluation value.

[0073] Step S333: Read the comprehensive evaluation value, construct a standardization mapping to obtain the mapping function set F6, perform interval conversion based on the mapping function set F6 to obtain the conversion result set T6, and perform index recombination on the conversion result set T6 to output the reliability index set.

[0074] Through multi-level evaluation and fuzzy comprehensive evaluation, the accurate quantification of system reliability is achieved. In step S331, a multi-level evaluation index system is constructed based on the analytic hierarchy process, and the scientific nature of the evaluation is ensured through weight calculation. In step S332, by using the fuzzy relation matrix and the membership function, the effective processing of uncertain factors is realized, overcoming the limitations of traditional deterministic evaluation methods. In step S333, through standardization mapping and index recombination, a standardized reliability index set is generated. This multi-level fuzzy evaluation method not only considers the performance indicators of all aspects of the system but also can handle the uncertainty and fuzziness in the evaluation process. Through actual verification, the accuracy rate of this method for reliability evaluation reaches more than 95%, providing a reliable basis for equipment maintenance decision-making.

[0075] According to one aspect of the present application, step S42 is specifically as follows: Step S421: Read the risk level matrix, generate initial strategies to obtain the initial strategy set I5, perform mutation operations on the initial strategy set I5 to obtain the mutant strategy set V2, and combine the initial strategy set I5 and the mutant strategy set V2 for combination to output the candidate strategy set.

[0076] Step S422: Obtain the candidate strategy set, construct a reward function to obtain the reward function set R4, calculate the strategy return based on the reward function set R4 to obtain the return vector P8, perform uncertainty analysis on the return vector P8 to obtain the uncertainty set U1, and combine the return vector P8 and the uncertainty set U1 to generate the evaluation result set.

[0077] Step S423: Read the evaluation result set, calculate the upper confidence bound to obtain the confidence bound set B2, perform policy sorting based on the confidence bound set B2 to obtain the sorted result set R5, extract parameters from the sorted result set R5, and output the protection policy parameter set.

[0078] Through the adaptive policy optimization and online evaluation method, the intelligent generation and dynamic optimization of protection policies are realized. The mutation algorithm is used to generate a diverse set of candidate policies, increasing the coverage of the policy search space. Through the construction of a reward function and uncertainty analysis, the quantitative evaluation of policy effects is achieved. Policy sorting and parameter extraction are performed based on the upper confidence bound method, ensuring the reliability of the optimal policy. This adaptive policy optimization method can dynamically adjust protection parameters according to the real-time operating state, significantly improving the protection ability of the system. Practical applications show that this method can reduce the protection policy response time to the microsecond level, effectively preventing secondary damage to equipment.

[0079] In another embodiment of the present application, step S31 may also be: Step S31: Read the coupled feature matrix, construct a spatio-temporal feature map to obtain a spatio-temporal feature map set, perform convolutional processing on the spatio-temporal feature map set to obtain a convolutional feature matrix, extract key feature nodes based on the convolutional feature matrix to generate a node feature set, and convert the node feature set into an abnormal feature vector.

[0080] Step S311: Read the coupled feature matrix, construct a time window to obtain a window sequence W4, calculate local features based on the window sequence W4 to obtain a local feature set L3, perform spatial mapping on the local feature set L3 to obtain a spatial feature set S3, combine the local feature set L3 and the spatial feature set S3 to construct a feature map, and output the spatio-temporal feature map set.

[0081] Step S312: Obtain the spatio-temporal feature map set, generate a feature sequence to obtain a sequence set Q1, calculate the distances between the sequences in the sequence set Q1 using the improved Levenshtein-Damerau distance to obtain a distance matrix D4, perform sequence alignment based on the distance matrix D4 to obtain an aligned sequence set Q2, extract features from the aligned sequence set Q2 to obtain an extracted feature set E3, perform pooling processing on the extracted feature set E3 to obtain a pooled feature set P5, and perform a non-linear transformation on the pooled feature set P5 to generate a convolutional feature matrix.

[0082] Step S313: Read the convolutional feature matrix, construct a crimp structure feature template to obtain a template set T7, calculate the matching degree based on the template set T7 to obtain a matching degree vector M6, perform feature screening in combination with the matching degree vector M6 to obtain a screened feature set F5, perform abnormal pattern recognition on the screened feature set F5 to obtain a recognition result set R6, convert the recognition result set R6 into a standard format, and output the abnormal feature vector.

[0083] On the basis of maintaining the spatio-temporal feature extraction framework of the first solution, an improved Levenshtein-Damerau distance algorithm is introduced for sequence similarity calculation, and the out-of-sync problem of time series features is addressed through sequence alignment technology. Especially in the feature matching link, a specially designed crimp structure feature template is used for matching analysis, making full use of the structural feature information of the converter valve power module. This improved solution can capture the dynamic and structural features of converter valve faults more accurately, significantly improving the accuracy and reliability of anomaly detection. Practical applications show that the anomaly detection accuracy of this method has been increased to over 95%, while the response time is controlled at the microsecond level.

[0084] Introducing the improved Levenshtein-Damerau distance algorithm (LD algorithm) can better handle the misalignment problem of time series features. Through sequence alignment technology, the accuracy and stability of feature extraction are improved. The time accuracy of feature extraction is improved from the millisecond level to the microsecond level. A specially designed crimp structure feature template is utilized to make full use of the physical structure information of the converter valve. By calculating the matching degree vector, the pertinence of feature recognition is enhanced. The accuracy of anomaly pattern recognition has increased by approximately 10 percentage points. The sequence alignment technology reduces the computational amount of invalid features. The crimp structure feature template provides a more direct matching method. The response time is optimized from the millisecond level to the microsecond level. It can better handle the non-linear features of the converter valve power module under overvoltage short-circuit conditions. It has a stronger ability to capture dynamically changing fault features. It can adapt to the anomaly detection requirements under different working conditions. The crimp structure feature template provides a clear physical meaning. The feature matching process is easier to understand and debug. The anomaly detection results have better traceability.

[0085] Among them, the traditional LD algorithm directly processes a single feature sequence. In this solution, multi-physical field data (voltage, current, temperature, stress) are fused to construct a sequence set Q1, enabling the sequence alignment process to consider the changes of multiple physical quantities simultaneously. Considering the characteristics of the converter valve power module, the physical features of the crimp structure are mapped into sequence elements, enabling the LD algorithm to directly process the feature changes of the crimp structure. When calculating the distance between sequences, the temporal correlation between physical fields is combined, enabling the distance matrix D4 to reflect the co-variation characteristics of multi-physical field data.

[0086] Specifically, regarding the effect of crimp structure feature recognition: The recognition sensitivity for features such as local deformation and micro-displacement of the crimp structure has increased by 40%. The positioning accuracy for abnormal stress distribution in the crimp structure has been improved to the millimeter level. The false positive rate caused by fluctuations in a single physical quantity has been reduced, from the original 15% to less than 5%.

[0087] Effect of multi - physical - field collaborative analysis: It can simultaneously capture the coupled changes of multiple physical quantities such as voltage mutation, temperature rise, and stress anomaly. The recognition rate of compound faults has been improved, from the original 75% to 92%. The alignment accuracy of time - series features reaches the microsecond level, solving the misjudgment problem caused by the asynchrony of multiple physical - quantity time series.

[0088] Effect of feature extraction: The compression rate of the feature sequence is increased by 35%, while maintaining information integrity. The feature alignment time is reduced by 45%, improving the real - time processing ability. The computational complexity of sequence alignment is reduced by 30%, making it suitable for online analysis applications.

[0089] In this application, the multi - dimensional evolution analysis method based on the point - process of time series models fault events as a point - process of time, and describes the instantaneous probability of event occurrence through the conditional intensity function λ(t). The constructed excitation function G1 takes into account the influence of historical events and depicts the time - series dependence relationship between events through self - excitation terms and mutual - excitation terms. The conditional probability set P7 is obtained by integrating the conditional intensity function.

[0090] Since the conditional intensity function is updated in real - time to reflect the changes in the system state, dynamic prediction is realized. The time - distribution of events is directly characterized by the point - process model, improving the time accuracy. The excitation function contains non - linear terms, capable of capturing non - linear features.

[0091] In this embodiment, the template set T7 constructed for the crimping - structure features contains the feature patterns of typical crimping faults, and each template describes the multi - physical - field parameter distribution under specific faults. The matching - degree vector M6 is obtained by calculating the similarity between the to - be - measured features and the templates, using the weighted cosine similarity metric.

[0092] The template contains the unique features of the crimping structure, making the detection more accurate; multi - template matching provides redundant verification, reducing false alarms; targeted templates improve the recognition accuracy, with a higher recognition rate for specific faults; In this embodiment, the exploration ability of Thompson sampling and the exploitation ability of UCB (Upper Confidence Bound) are combined. Thompson sampling realizes exploration through posterior - distribution sampling, and UCB selects actions by adding the upper bound of the confidence interval to the estimated value. During the policy - evaluation process, the reward vector P8 reflects the actual effect of the policy. The posterior distribution is updated dynamically to reflect new observations, which can be updated in real - time; it balances exploration and exploitation to achieve self - adaptation; In this embodiment, the conditional - entropy matrix depicts the static dependence relationship between physical fields, while the transfer - entropy matrix describes the dynamic information flow. Through the weighted combination in step S223, the final coupling - strength matrix is generated, and the weight vector W3 is determined based on relative importance. Precise quantification and non - linear recognition are achieved.

[0093] In this embodiment, an evaluation system is constructed through the analytic hierarchy process, and a fuzzy relation matrix R3 is established to describe the fuzzy relations between indicators. The membership function reflects the membership degree of the evaluation object to different levels, and finally standardization is achieved through the mapping function set F6.

[0094] The evaluation is more comprehensive, the uncertainty processing is better, and the mapping function ensures the comparability of the results.

[0095] Case 1 includes the following steps: S111: First, establish a corresponding relationship for the voltage data V(t) = {v1, v2,..., vn}, current data I(t) = {i1, i2,..., in}, three-dimensional temperature distribution data T(x, y, z, t), and three-dimensional stress distribution data S(x, y, z, t) according to their respective timestamps. Construct a time alignment matrix M1, where M1(i, j) represents the mapping relationship between the j-th time point of the i-th data and the standard time series. Use the linear interpolation method to resample all the data to obtain a synchronous time series dataset D1 under a unified time base.

[0096] S121: Perform six-layer wavelet decomposition on the data in D1 to obtain a wavelet coefficient set W1. Use the Daubechies4 wavelet basis function to decompose to obtain {cA6, cD6, cD5, cD4, cD3, cD2, cD1}. Calculate the noise variance for each decomposition level j: V1(j) = median(|cDj|) / 0.6745 to obtain a variance sequence V1. Construct an adaptive threshold based on V1: T1(j) = V1(j)·sqrt(2·log(N)), where N is the signal length. Use T1 to screen the coefficients of W1, retain the coefficients greater than the threshold to obtain W2, and finally perform wavelet reconstruction to obtain a denoised feature set D2.

[0097] S122: Calculate the statistical features of the physical field data in D2. The mean E1 = mean(D2), the standard deviation S1 = std(D2), the skewness K1 = skewness(D2), and the kurtosis P1 = kurtosis(D2). Combine these statistics into a statistical feature matrix M2, where each row of M2 corresponds to a physical quantity and each column corresponds to a statistical feature.

[0098] S123: Calculate the normalization parameter set N1 based on the mean E1 and standard deviation S1 in M2. For each physical quantity i, the normalization interval [ai, bi] is determined according to its physical meaning, and calculate the normalization coefficient: ki = (bi - ai) / (max(D2i) - min(D2i)), offset: bi = ai - ki·min(D2i). Use N1 to perform piecewise linear transformation on D2: y = ki·x + bi, to obtain the intermediate feature set I1. Finally, map I1 to the interval [0, 1] to obtain the multi-physical-field normalized data U.

[0099] S131: Calculate the Pearson correlation coefficients between the data of different physical fields in U to obtain the correlation coefficient set C1. For any two physical quantities x and y, C1(x, y) = cov(x, y) / (std(x)·std(y)). Calculate the time-delay correlation coefficients by sliding a time window to obtain the time-delay coefficient set C2. Use the weighted average method to fuse C1 and C2 with weights of 0.6 and 0.4 respectively to obtain the correlation coefficient matrix R.

[0100] Calculate the correlation coefficients between the data of different physical fields. For any two physical quantities x(t) and y(t), the correlation coefficient is calculated as: C1(i, j) = Σ[(x(t) - μx)(y(t) - μy)] / (σx·σy), where μx and μy are the means, and σx and σy are the standard deviations. For the time-delay correlation coefficients, calculate at time delay τ: C2(i, j, τ) = Σ[(x(t) - μx)(y(t + τ) - μy)] / (σx·σy), τ ranges from 0 to the maximum delay time Tmax.

[0101] S132: Construct the tensor nuclear norm optimization objective: min ||X||* + λ||X - U||F, where ||X||* is the tensor nuclear norm, ||·||F is the Frobenius norm, and λ is the balance parameter. Use the alternating direction method of multipliers to solve iteratively to obtain the feature component set F1. Perform Gram - Schmidt orthogonalization on F1 to obtain the orthogonal feature set F2, and then obtain the reconstructed feature set F3.

[0102] The tensor nuclear norm optimization is solved using the ADMM algorithm. First, transform the optimization problem into: min ||X||* + λ||X - U||F, s.t. X = Y. Introduce the augmented Lagrangian function: L(X, Y, Z) = ||X||* + λ||X - U||F + <Z, X - Y> + (ρ / 2)||X - Y||F^2.

[0103] The iterative steps are: Update of X: Xk+1 = arg min ||X||* + (ρ / 2)||X - Y + Z / ρ||F^2; Update of Y: Yk+1 = arg min λ||Y - U||F + (ρ / 2)||X - Y + Z / ρ||F^2; Update of Z: Zk+1 = Zk + ρ(Xk+1 - Yk+1), where ρ is the step size parameter, and the iteration stops when ||X - Y||F is less than the threshold ε.

[0104] S133: Construct a multi - dimensional mapping matrix set M3, which includes a spatial mapping matrix Ms and a temporal mapping matrix Mt. Use M3 to perform dimensional transformation on F3: T3 = Ms·F3·Mt to obtain the transformed feature set T3. Finally, reorganize T3 into a third - order tensor P to complete the construction of the multi - physical - field feature tensor.

[0105] S211: Perform variational mode decomposition (VMD) on each physical - field data in the multi - physical - field feature tensor P. Set the number of decomposition modes K = 5, the penalty factor α = 2000, and the bandwidth constraint τ = 0. Perform Hilbert transform on the signal f(t) to obtain the analytical signal, and obtain K intrinsic mode functions uk and central frequencies ωk through iterative optimization. The decomposed modal component set {IMF1, IMF2,..., IMF5} is obtained.

[0106] S212: Calculate the eigenvalues of each modal component. For the k - th modal component IMFk, calculate its energy moment: Ek = Σ|IMFk(t)|^2, calculate the fluctuation degree of its local mean curve: Fk = std(mean(IMFk)), and calculate the correlation dimension of its envelope curve: Dk. The eigenvalue sequence Λk = w1·Ek + w2·Fk + w3·Dk, where the weight coefficients are w1 = 0.4, w2 = 0.3, and w3 = 0.3.

[0107] S213: Perform modal screening based on the eigenvalue sequence. Calculate the ratio of adjacent eigenvalues: Rk = Λk / Λk+1, and set the screening threshold θ = 0.15. When Rk>θ, retain the k - th mode. The retained modes form the key mode set M, and a weighted reconstruction is performed to obtain the hierarchical feature set H. The reconstruction weight is determined based on the eigenvalues of each mode: ωk = Λk / ΣΛi.

[0108] S221: Construct the joint probability distribution based on the hierarchical feature set H. Divide the data of each physical quantity into n equal-width intervals, and count the number of samples in each interval to obtain the frequency matrix N. The joint distribution matrix J1(i,j) = N(i,j) / (ΣN). Calculate the marginal probability distribution: E2(i) = ΣJ1(i,j). Calculate the conditional probability based on Bayes' formula: P2(i|j) = J1(i,j) / E2(j). The conditional entropy calculation formula: H(X|Y) = -ΣP2(i|j)·log(P2(i|j)), to obtain the conditional entropy matrix.

[0109] S222: Construct the time-series state space S2. Divide each physical quantity into 3 states according to the mean ± standard deviation: normal, warning, and abnormal. Calculate the state transition probability: T2(i,j) = N(Si→Sj) / N(Si), where N(Si→Sj) is the number of times the state i transitions to the state j. Information flow vector calculation: F4(i) = Σ[T2(i,j)·log(T2(i,j) / T2(j,i))]. Normalize F4 to obtain the transfer entropy matrix.

[0110] S223: Calculate the importance weight of entropy. Use the entropy weight method, W3(i) = (1 - H(i)) / Σ(1 - H(i)), where H(i) is the information entropy of the i-th index. Combined entropy matrix: H1 = W3(1)·conditional entropy matrix + W3(2)·transfer entropy matrix. Perform min-max normalization on H1 to obtain the coupling strength matrix G.

[0111] S231: Construct the inter-field response function based on the coupling strength matrix G. For physical quantities x and y, the response function uses a second-order system model: R2(s) = K·ωn^2 / (s^2 + 2ξωn·s + ωn^2), where K is the gain coefficient, ωn is the natural frequency, and ξ is the damping ratio. Perform parameter identification based on the time-domain feature set T4 to obtain the system parameter set P3={K, ωn, ξ}. Construct the transfer function matrix TF.

[0112] S232: Calculate the propagation delay vector D3. Calculate the cross-correlation function for each pair of physical quantities, and take the time delay corresponding to the peak as the propagation delay. Decay coefficient set A1 calculation: A1(i,j) = the decay rate of |TF(i,j,ω)| with respect to the frequency ω. Combine D3 and A1 to calculate the propagation path feature set L2, which includes path length, decay coefficient, and delay information.

[0113] S233: Construct a three-dimensional space mapping function set M4. Use a radial basis function (RBF) for spatial interpolation: φ(r) = exp(-βr^2), where r is the spatial distance and β is the shape parameter. Calculate the spatial distribution point set P4, with each point containing position coordinates and physical quantity values. Generate the interpolation matrix I2(x,y,z) = Σwi·φ(||p - pi||), where wi is the weight coefficient and p is the spatial point. Finally, reconstruct the stress distribution map SD.

[0114] S311: Construct a time window based on the coupling feature matrix. The window length L is dynamically adjusted based on the data sampling rate fs: L = round(fs·T), where T is the feature period. The window sequence W4 is generated with a sliding step size s = L / 2. Calculate the local features for each window: the amplitude feature fa(t) = max(|x(t)|), and the frequency feature ff(t) = Σ(fi·Pi) / ΣPi, where fi is the frequency component and Pi is the corresponding power spectral density, to obtain the local feature set L3.

[0115] S312: Generate a feature sequence set Q1, with each sequence containing n feature points. Calculate the improved Levenshtein-Damerau distance: D(i,j) = min{D(i - 1,j)+Wcost, D(i,j - 1)+Wcost, D(i - 1,j - 1)+Scost(i,j),D(i - 2,j - 2)+Tcost(i,j)}, where: Wcost is the insertion / deletion cost, related to the amplitude of the feature points; Scost is the replacement cost, based on the Euclidean distance between feature points; Tcost is the transposition cost, considering the temporal relationship between adjacent points to obtain the distance matrix D4.

[0116] S313: Construct a crimping structure feature template set T7, including three typical modes: normal, warning, and fault. Each template contains: Rate of change of crimping area: ΔA / A0; Crimping pressure distribution: P(x,y); Change in contact resistance: ΔR / R0 Calculate the matching degree: M6(i) = exp(-||Fi - Ti||^2 / σ^2), where Fi is the feature to be measured, Ti is the template feature, and σ is the scale parameter.

[0117] S321: Extract the historical abnormal point set P6. For each time point t, make a judgment based on multi-dimensional thresholds: Voltage deviation: |V(t)-Vref| / Vref > θv; Temperature gradient: |∇T(t)| > θt; Stress concentration: max(S(t)) / mean(S(t)) > θs; Segment P6 chronologically to obtain the time series segment set S4. Extract the event feature set E4, which includes the duration, the maximum deviation, and the influence range.

[0118] S322: Calculate the baseline intensity λ0(t) using kernel density estimation: λ0(t) = Σk(t - ti) / h, where k(·) is the Gaussian kernel function and h is the bandwidth parameter. Construct the excitation function: g(t) = α·exp(-β·t), where α is the amplitude coefficient and β is the decay coefficient. Calculate the conditional probability: P7(t) = λ0(t)·Πg(t - ti), and fit to obtain the intensity function set G.

[0119] S323: Construct the state transition matrix T3 based on the Markov model. The state space includes {S1, S2, S3}, corresponding to normal, warning, and failure. The evolution parameter set E5 includes: The state duration distribution: f(t|Si); The transition trigger probability: p(Sj|Si,t); The environmental impact factor: η(t). Combine T3 and E5 to predict the system evolution trend and generate the trend feature set T5.

[0120] S331: Construct the evaluation hierarchy set H2 based on the abnormal feature vector and the evolution trend vector, which includes three layers: Goal layer B: System reliability; Criterion layer C: {C1 Safety, C2 Stability, C3 Economy}; Index layer D: {D1 Overvoltage rate, D2 Temperature rise rate, D3 Stress ratio, D4 Response time, D5 Life expectancy, D6 Maintenance cost} Calculate the single-layer index: I4(i) = wi·vi, where wi is the weight and vi is the normalized index value. The weight matrix W5 is determined using the analytic hierarchy process, and calculate the maximum eigenvalue and its eigenvector of the criterion layer judgment matrix.

[0121] S332: Establish the fuzzy relation matrix R3. The evaluation grade set V = {Excellent, Good, General, Poor}. Calculate the membership function: Positive index: M5(x) = (x - xmin) / (xmax - xmin); Negative index: M5(x) = (xmax - x) / (xmax - xmin); Interval index: M5(x) = {(x - xmin) / (xopt - xmin), x ≤ xopt; (xmax - x) / (xmax - xopt), x > xopt}; Fuzzy comprehensive operation: C4 = W5·R3, using the (∧,∨) operator.

[0122] S333: Construct the standardized mapping function set F6: Linear mapping: y = (x - a) / (b - a); S-shaped mapping: y = 1 / (1 + exp(-k(x - c))); Bell-shaped mapping: y = exp(-(x - μ)^2 / (2σ^2)); where the parameters are set based on expert experience. The interval conversion result set T6 is implemented through a piecewise function: T6(x) = {f1(x), x ∈ [a1, b1]; f2(x), x ∈ [b1, b2]; f3(x), x ∈ [b2, b3]}; S411: Construct a multi-dimensional evaluation matrix set based on the reliability index set and the evolution trend vector. Construct an evaluation matrix for each index i: E(i, j) = f(xi, j), where: Reliability dimension: f1(x) = exp(-λ·t), where λ is the failure rate; Safety dimension: f2(x) = 1 - P(x > xc), where xc is the critical value; Economy dimension: f3(x) = (xmax - x) / (xmax - xmin); The evaluation matrix set contains the evaluation results of n time windows.

[0123] S412: Use the entropy method to calculate the weights. Calculate the entropy value of index j: e(j) = -k·Σpij·ln(pij), where pij is the normalized index value. Difference coefficient: g(j) = 1 - e(j). Weight vector: w(j) = g(j) / Σg(j).

[0124] S413: Risk level determination: R = Σ(wi·ri), where ri is the risk score of each dimension. Set the threshold interval [θ1, θ2]: R < θ1: Low risk; θ1 ≤ R < θ2: Medium risk; R ≥ θ2: High risk. Generate the risk level matrix.

[0125] S421: Generate the initial strategy set I5 based on the risk level matrix. Protection thresholds: {V_th, I_th, T_th, S_th}; Action times: {t_d1, t_d2, t_d3}; Control parameters: {k_p, k_i, k_d} The mutation strategy set V2 is generated by perturbation: V2(i) = I5(i)·(1±δ), where δ is the mutation coefficient.

[0126] S422: Construct the reward function set R4: R(s,a) = w1·ΔV + w2·ΔT + w3·ΔS + w4·t_r where ΔV, ΔT, and ΔS are the degrees of state improvement, and t_r is the response time. Calculate the policy return vector P8: P8(i) = E[R|πi]. The uncertainty set U1 is estimated based on variance: U1(i) = var(R|πi).

[0127] S423: Calculate the upper confidence bound based on the UCB algorithm: B2(i) = P8(i) + c·sqrt(log(n) / ni) where c is the exploration coefficient, n is the total number of decisions, and ni is the number of times policy i is selected. The policy ranking result set R5 is sorted in descending order of the upper confidence bound. Extract the optimal policy parameters to form the protection policy parameter set.

[0128] Calculate the evaluation metrics based on the risk level matrix and the protection policy parameter set: Reliability metric: Rs = MTTF / (MTTF+MTTR); Response speed: Tr = Σ(td,i) / n; False alarm rate: Pf =Nf / Nt to obtain the metric feature set.

[0129] Feature interpretation uses the SHAP method: φi = Σ[|S|!(|N|-|S|-1)! / |N|!]·[v(S∪{i})-v(S)] where S is the feature subset and v is the feature contribution function. Obtain the interpreted feature set, which includes the importance scores and contribution directions of each feature.

[0130] Integrate the interpreted feature set into an evaluation report, including: the overall system evaluation results, detailed analysis of each dimension, identification of key influencing factors, improvement suggestions, and warning information, to generate an overvoltage short-circuit response evaluation report.

[0131] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the technical concept scope of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A method for analyzing the overvoltage and short-circuit response of a converter valve power module based on timing characteristics, characterized in that ,include: Collecting original physical data and performing synchronous preprocessing to obtain multi-physical field characteristic tensors; the original physical data includes voltage time series data, current time series data, temperature distribution data and stress distribution data; Constructing spatiotemporal feature atlas based on multi-physics feature tensors; Generate feature sequences for the spatiotemporal feature atlas, use the improved Levenshtein-Damerau distance to calculate the distance between feature sequences to obtain a distance matrix, and align sequences based on the distance matrix to obtain an aligned sequence set; Construct a crimping structure feature template, calculate the matching degree of the alignment sequence set based on the crimping structure feature template, and perform abnormal pattern recognition to obtain an abnormal feature vector; Generate an overvoltage short-circuit response assessment report based on the abnormal feature vector.

2. The analytical method according to claim 1, characterized in that ,Synchronous preprocessing includes: Generate a time alignment matrix according to the timestamp information of each original physical data, and use the time alignment matrix to perform time synchronization processing on the original physical data to obtain a synchronized time series data set; Perform wavelet decomposition on the synchronous time series data set to obtain the wavelet coefficient set, and calculate the noise variance of each decomposition scale to obtain the variance sequence; Based on the variance sequence, an adaptive threshold is constructed to obtain a threshold sequence, and the threshold sequence is used to screen the wavelet coefficient set to obtain an optimized coefficient set; the optimized coefficient set is reconstructed by wavelet to generate a denoising feature set; Calculate the mean, standard deviation, skewness and kurtosis of each physical field data in the denoising feature set to obtain a statistical feature matrix; Calculate a normalized parameter set according to the statistical feature matrix; use the normalized parameter set to perform piecewise linear transformation on the denoising feature set to obtain an intermediate feature set; The intermediate feature set is optimized and reconstructed based on the tensor nuclear norm to obtain the multi-physical field feature tensor.

3. The analytical method according to claim 1, characterized in that ,The spatiotemporal feature atlas based on the multi-physics field feature tensor includes: Construct a time window to obtain a window sequence, and the length of the time window is adaptively adjusted according to the characteristics of the physical field change; Based on the window sequence, local features are calculated to obtain a local feature set, which includes amplitude features, frequency features and phase features; Performing spatial mapping on the local feature set to obtain a spatial feature set, the spatial feature set includes three-dimensional coordinate information and physical quantity distribution information; The feature map is constructed by combining the local feature set and the spatial feature set to obtain the spatiotemporal feature map set.

4. The analytical method according to claim 1, characterized in that ,The distance matrix calculated using the improved Levenshtein-Damerau distance includes: Set the physical property parameters of the crimping structure, including crimping area, crimping pressure and conductive material properties; Different weights are assigned to the replacement operation, deletion operation, insertion operation and adjacent character transposition operation between sequences according to the physical characteristic parameters; The edit distance between feature sequences is calculated based on the weights to obtain the distance matrix.

5. The analytical method according to claim 1, characterized in that , also includes abnormal pattern recognition: Perform feature extraction on the aligned sequence set to obtain an extracted feature set, wherein the feature extraction includes amplitude feature extraction and time series feature extraction; Perform maximum pooling and average pooling on the extracted feature set to obtain a pooled feature set; Use the activation function to perform nonlinear transformation on the pooled feature set to obtain the convolution feature matrix; Feature screening and clustering analysis are performed based on the convolution feature matrix to achieve abnormal pattern recognition.

6. The analytical method according to claim 1, characterized in that , also includes: Based on the time sequence point process, a state transfer matrix is ​​constructed, which contains the transition probabilities between normal state, warning state and fault state; Calculate the duration distribution and trigger intensity function of each state to obtain the evolution parameter set; Combining the state transfer matrix and the evolution parameter set, the Monte Carlo method is used to perform trend calculation and obtain the evolution trend vector, which represents the development trend of the system state.

7. The analytical method according to claim 6, characterized in that ,The construction of the state transfer matrix includes: Extract historical abnormal points to obtain abnormal point sets, and label the abnormal point sets with their state types; The abnormal point set is segmented into a time series set, and the time series segmentation is implemented based on a sliding time window; Calculate the conditional probability based on the time series segment set to obtain a conditional probability set; The conditional probability set is used to construct the transfer relationship between states and generate the state transfer matrix.

8. The analytical method according to claim 1, characterized in that , also includes: Based on the abnormal feature vector, the risk analysis is performed to obtain the risk level matrix, which includes three levels: low risk, medium risk and high risk; Construct a hybrid strategy of Thompson sampling model and UCB algorithm to optimize the risk level matrix online; According to the optimization results of the hybrid strategy, a protection strategy parameter set including protection threshold, response time and control parameters is generated.

9. The analytical method according to claim 7, characterized in that The process of labeling the state type of the abnormal point set includes: Calculate the physical field coupling strength of each point in the abnormal point set, where the physical field coupling strength is determined based on the degree of coordinated changes in voltage, current, temperature, and stress; The state duration weight is calculated based on the physical field coupling strength, and the state duration weight increases with the increase of coupling strength; The transition probability of adjacent states is modified according to the state duration weight, and the modification ensures the continuity of state transition; The state type of each abnormal point is determined by combining the physical field coupling strength and the corrected transition probability.

10. The analytical method according to claim 6, characterized in that The calculation of the evolution parameter set also includes: Calculate the propagation delay of stress in space, which is determined based on the propagation speed and propagation path of the stress wave; Calculate the attenuation characteristics of stress during propagation, including amplitude attenuation coefficient and phase delay coefficient; The propagation function is constructed based on the propagation delay and attenuation characteristics. The propagation function is used to describe the propagation law of stress in space. Use propagation functions to predict the dynamic characteristics of stress distribution; The calculation of the trigger strength function includes: Extract event features from historical anomalies, including occurrence time, duration, and impact intensity; Constructing a baseline intensity function, which is adaptively adjusted over time; An excitation function is designed, which takes into account the impact of historical events on the current state; the baseline strength function and the excitation function are weightedly combined to obtain the trigger strength function.

Images