A method and system for evaluating the safety status of electric power equipment, a device and a readable storage medium

Abstract

The application discloses a kind of power equipment safety state evaluation method and system, equipment, readable storage medium, it is related to power equipment evaluation technical field, including, acquisition power equipment communication interaction data stream, carry out time synchronization and outlier elimination, construct multimodal data matrix;With abnormal probability score and abnormal probability threshold value are evaluated, obtain abnormal state flag, and generate security event data package in combination with preliminary abnormal type and power equipment communication interaction data stream;Using LiNGAM causal discovery algorithm, abnormal event time series in security event data package is non-gaussianity test and time series lag analysis, and directed acyclic causal graph is constructed;According to real-time equipment load rate and communication quality index, the causal strength of directed acyclic causal graph is dynamically adjusted, and dynamic causal graph is output;The application improves the accuracy and adaptability of abnormal propagation analysis, finally enhances the real-time performance and reliability of power equipment safety evaluation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power equipment evaluation, and in particular to a power equipment safety state evaluation method and system, equipment and a readable storage medium. BACKGROUND

[0002] With the continuous improvement of the digitalization and intelligentization level of the power system, the state perception, risk identification and early warning evaluation of power equipment have become an important link to ensure the safe and stable operation of the power grid. At present, the method of evaluating the state of power equipment based on equipment operating parameters and communication data has gradually become a research hotspot. Among them, a common technical path is to use various sensors to collect voltage, current, frequency, temperature and other electrical and environmental parameters, combined with expert rule library or traditional machine learning models such as support vector machine (SVM) and random forest, to identify abnormal operating conditions or potential fault trends in the equipment operation. Some research further introduces an event-driven mechanism to classify and time sequence track abnormal events, thereby constructing a state evaluation system based on a rule atlas. This method has certain practicality in equipment health degree evaluation, abnormality identification and preliminary classification, and has obtained certain application foundation, especially in the detection scene of single abnormality or local fault.

[0003] In the face of complex working conditions, multi-source data fusion and unclear abnormal causal relationship, the existing evaluation method still has certain limitations in the correlation analysis and evolution mechanism modeling of abnormal events. Most current methods often focus on the static identification and label classification of abnormal points, but lack modeling of the potential causal relationship between abnormal events, especially the inability to accurately mine the interaction between electrical data and non-electrical data and its influence on the fault propagation path. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides a power equipment safety state evaluation method to solve the problems of difficulty in modeling the causal relationship between abnormal events and difficulty in identifying the influence on the fault propagation path.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] In a first aspect, the present application provides a power equipment safety state evaluation method, which comprises collecting power equipment communication interaction data stream, performing time synchronization and outlier elimination, and constructing a multi-modal data matrix;

[0008] inputting the multi-modal data matrix into a Transformer model, extracting the harmonic energy proportion of electrical quantity data and the statistical features of non-electrical quantity data, and outputting abnormal probability scores and preliminary abnormal types;

[0009] The abnormal probability score is evaluated with an abnormal probability threshold to obtain an abnormal state flag, and a security event data packet is generated by combining the preliminary abnormal type and the power equipment communication interaction data stream;

[0010] The abnormal event time series in the security event data packet is subjected to non-Gaussianity test and time series lag analysis by using the LiNGAM causal discovery algorithm, and a directed acyclic causal graph is constructed.

[0011] According to the real-time device load rate and the communication quality index, the causal strength of the directed acyclic causal graph is dynamically adjusted, and a dynamic causal graph is output.

[0012] The dynamic causal graph is combined with the device account data and the real-time running state of the device to perform path backtracking analysis, evaluate the fault propagation depth and node contribution degree, and generate a power equipment safety evaluation report.

[0013] As a preferred scheme of the power equipment safety state evaluation method, the time synchronization and outlier elimination are performed, and a multi-modal data matrix is constructed, and the specific steps are as follows,

[0014] The power equipment communication interaction data stream is resampled to obtain synchronous communication data; the power equipment communication interaction data stream includes electrical quantity data and non-electrical quantity data;

[0015] The mean and standard deviation of the electrical quantity data are calculated, and the abnormal data exceeding the normal interval range are eliminated;

[0016] The non-electrical quantity data is plotted as a box plot, and the abnormal data exceeding the upper and lower quartile range is eliminated;

[0017] The cleaned electrical quantity data, the cleaned non-electrical quantity data and the synchronous communication data are aligned and normalized according to the time stamp, and are spliced into a multi-dimensional matrix according to the column to construct a multi-modal data matrix.

[0018] As a preferred scheme of the power equipment safety state evaluation method, the output abnormal probability score and preliminary abnormal type are output, and the specific steps are as follows,

[0019] The multi-modal data matrix is subjected to standardization processing, and the electrical quantity data is subjected to fast Fourier transform to extract the harmonic energy proportion of the electrical quantity data;

[0020] The mean and standard deviation of the non-electrical quantity data are used to extract the volatility and stability features of the non-electrical quantity data, and statistical features are obtained;

[0021] The harmonic energy proportion and the statistical characteristics are spliced into a comprehensive feature vector, input into a Transformer model, mapped to a high-dimensional feature space through an embedding layer to obtain a unified feature representation, the dependency relationship of the unified feature representation is extracted by using a self-attention mechanism, and an abnormal probability score is output through a full connection layer, the abnormal probability score is classified by a Softmax layer to output a preliminary abnormal type.

[0022] As a preferred scheme of the power equipment safety state evaluation method, the specific steps of generating the safety event data packet are as follows,

[0023] The historical abnormal probability score statistical distribution is obtained, the optimal segmentation point is selected through the ROC curve, and the real-time running state of the equipment is dynamically adjusted in combination with the real-time running state of the equipment to set the abnormal probability threshold value.

[0024] The abnormal probability score is compared with the abnormal probability threshold value, when the abnormal probability score is higher than the abnormal probability threshold value, it is judged as an abnormal state, otherwise it is judged as a normal state, and the corresponding abnormal state flag is generated;

[0025] The abnormal state flag is associated with the preliminary abnormal type, the preliminary abnormal type corresponding to each abnormal state is obtained, and the time period to which the abnormal source data belongs is marked;

[0026] All data segments in the time period to which the abnormal source data belongs are extracted from the power equipment communication interaction data stream as abnormal data sources;

[0027] The abnormal state flag, the preliminary abnormal type, the abnormal probability score and the abnormal data source are structured and packaged to generate a safety event data packet.

[0028] As a preferred scheme of the power equipment safety state evaluation method, the specific steps of constructing the directed acyclic causal graph are as follows,

[0029] The abnormal source data corresponding to the abnormal event time period is extracted from the safety event data packet to form an abnormal event time sequence, input into a multi-node parallel computing architecture, and subjected to multi-dimensional non-Gaussianity test to output non-Gaussian time series data, and subjected to feature extraction and dynamic time lag identification to obtain optimized lag time series data;

[0030] The LiNGAM causal discovery algorithm is used to perform linear non-Gaussian modeling and topological sorting on the optimized lag time series data, to infer the direct causal path between variables, and output a causal relationship candidate edge set;

[0031] The causal relationship candidate edge set is drawn into a directed acyclic causal graph according to the lag order.

[0032] As a preferred scheme of the power equipment safety state evaluation method, wherein: the output dynamic causal graph has the following specific steps,

[0033] The real-time load rate and communication quality indicators of the power equipment are collected, normalized, standardized performance data is formed, and a weighted linear combination method is used to calculate the causal strength adjustment coefficient to correct the initial strength value of each causal edge in the directed acyclic causal graph;

[0034] The corrected initial strength value is assigned to the corresponding causal edge of the directed acyclic causal graph, and the dynamic causal graph is output.

[0035] As a preferred scheme of the power equipment safety state evaluation method, wherein: the power equipment safety evaluation report is generated, and the specific steps are as follows,

[0036] The nodes and causal edges in the dynamic causal graph are associated and integrated with the equipment account data and the real-time running state of the equipment to form a comprehensive data structure, and the abnormal nodes are extracted. Taking the abnormal node as the starting point, the causal path is traversed in reverse along the dynamic directed causal edge, path backtracking analysis is performed, and the fault propagation depth is output;

[0037] Based on the causal edge weight in the backtracking path and the real-time running state of the equipment, the contribution degree of each node to the fault propagation is evaluated, and the node contribution degree is obtained;

[0038] The fault propagation depth, node contribution degree, equipment account data and real-time running state of the equipment are summarized to generate a power equipment safety evaluation report.

[0039] In a second aspect, the present application provides a power equipment safety state evaluation system, comprising a data acquisition module for acquiring power equipment communication interaction data stream, performing time synchronization and outlier elimination, and constructing a multi-modal data matrix;

[0040] The feature extraction module is used for inputting the multi-modal data matrix into the Transformer model, extracting the harmonic energy proportion of the electrical quantity data and the statistical features of the non-electrical quantity data, and outputting the abnormal probability score and the preliminary abnormal type;

[0041] The anomaly determination module is used for evaluating the abnormal probability score and the abnormal probability threshold to obtain an abnormal state flag, and combining the preliminary abnormal type and the power equipment communication interaction data stream to generate a safety event data packet;

[0042] The relationship analysis module is used for adopting the LiNGAM causal discovery algorithm to perform non-Gaussianity test and time series lag analysis on the abnormal event time series in the safety event data packet, and constructing a directed acyclic causal graph;

[0043] A strength adjustment module is configured to dynamically adjust the causal strength of the directed acyclic causal graph according to the real-time device load rate and the communication quality index, and output a dynamic causal graph.

[0044] A fault assessment module is configured to combine the dynamic causal graph with the device account data and the real-time running state of the device, perform path backtracking analysis, assess the fault propagation depth and node contribution degree, and generate a power device safety assessment report.

[0045] In a third aspect, the present application provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and wherein the computer program is executed by the processor to implement any step of the power device safety state assessment method according to the first aspect of the present application.

[0046] In a fourth aspect, the present application provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement any step of the power device safety state assessment method according to the first aspect of the present application.

[0047] The present application has the following beneficial effects: by constructing a directed acyclic causal graph, the causal relationship between abnormal events is modeled, the propagation path and the chronological relationship between abnormalities are revealed, the logical clarity and the analysis depth of the power device fault diagnosis are enhanced; by introducing the device load rate and the communication quality index to dynamically adjust the causal strength, a dynamic causal graph is constructed, the causal relationship can reflect the changes in the running state in real time, the accuracy and the adaptability of the abnormal propagation analysis are improved, and the real-time performance and the reliability of the power device safety assessment are ultimately enhanced. BRIEF DESCRIPTION OF DRAWINGS

[0048] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0049] Fig. 1 The flowchart of the power device safety state assessment method.

[0050] Fig. 2 The schematic diagram of the power device safety state assessment system.

[0051] Fig. 3 The flowchart of the multi-modal data matrix construction.

[0052] Fig. 4 The flowchart of the dynamic causal graph generation. DETAILED DESCRIPTION

[0053] In order to make the above objectives, characteristics and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0054] In the following description, a large number of specific details are set forth in order to facilitate a thorough understanding of the present application, but the present application can also be implemented in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the concept of the present application, therefore the present application is not limited to the specific embodiments disclosed below.

[0055] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the present application. "In one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an embodiment that is independent or alternative to other embodiments.

[0056] Reference Figs. 1-4 For one embodiment of the present application, the embodiment provides a power equipment safety state evaluation method, comprising the following steps:

[0057] S1, collect power equipment communication interaction data stream, perform time synchronization and outlier elimination, and construct a multi-modal data matrix.

[0058] It should be noted that the power equipment communication interaction data stream includes electrical quantity data and non-electrical quantity data;

[0059] The electrical quantity data refers to the numerical measurement data such as voltage, current, frequency, active power, reactive power, power factor and the like directly related to power transmission collected from the power equipment;

[0060] The non-electrical quantity data refers to the data such as control signals, state information, alarm information, communication message content, function code sequence and the like which are not directly related to electrical characteristics but can reflect the behavior characteristics of the equipment during the operation of the equipment.

[0061] S1.1, resample the power equipment communication interaction data stream to obtain synchronous communication data.

[0062] It should be noted that the time stamps in the power equipment communication interaction data stream are divided by a uniform sampling interval, the electrical quantity data and the non-electrical quantity data are resampled by linear interpolation method according to the sampling frequency (for example, 1 hertz (1 time per second)), the electrical quantity data and the non-electrical quantity data at the missing time points are filled or interpolated, and the sampling step and the time stamp identification of the electrical quantity data and the non-electrical quantity data are kept consistent in the time dimension, and finally the synchronous communication data is output.

[0063] S1.2, calculate the mean and standard deviation of the electrical quantity data, and eliminate the abnormal data exceeding the normal interval range.

[0064] It should be noted that the mean and standard deviation of the electrical quantity data are calculated by using the sliding window statistical method according to the time sequence, and the statistical analysis is performed on all sampling points of each electrical quantity data during the calculation process to obtain the corresponding mean and standard deviation. According to the statistical results, it is determined whether each sampling point belongs to abnormal data, and the specific judgment condition is whether the sampling value is outside the normal interval range of three times the standard deviation above and below the mean. The sampling points outside the normal interval range are marked as abnormal and are excluded from the electrical quantity data, and the cleaned electrical quantity data is output;

[0065] It should be noted that the sampling value of the electrical quantity data is obtained by collecting the original measurement values of voltage, current and other parameters of the power equipment at consecutive time points.

[0066] S1.3, box plot is drawn for non-electrical quantity data, and abnormal data outside the range of upper and lower quartiles is removed.

[0067] It should be noted that the non-electrical quantity data is analyzed item by item according to the time sequence, the first quartile and the third quartile of the non-electrical quantity data are calculated by using the sorting statistical method, and the abnormal threshold interval of the non-electrical quantity data is determined by using the limit rule of the box plot, that is, the values below a certain range below the first quartile and above the third quartile are used as the abnormal threshold. All sampling points exceeding the abnormal threshold are determined as abnormal data, which are removed from the non-electrical quantity data set to ensure the stability and effectiveness of the non-electrical quantity data, and the cleaned non-electrical quantity data is output.

[0068] S1.4, the cleaned electrical quantity data, the cleaned non-electrical quantity data and the synchronous communication data are aligned by time stamp and normalized, and are spliced into a multi-dimensional matrix by column to construct a multi-modal data matrix.

[0069] It should be noted that the cleaned electrical quantity data, the cleaned non-electrical quantity data and the synchronous communication data are aligned by time stamp to ensure consistency in the time dimension, and the cleaned electrical quantity data, the cleaned non-electrical quantity data and the synchronous communication data after alignment are normalized respectively, and are standardized to a unified range to eliminate the influence of different data magnitudes. According to the time stamp as the index, the column is spliced to form a multi-dimensional matrix, and finally a multi-modal data matrix is formed.

[0070] S2, input the multi-modal data matrix into the Transformer model, extract the harmonic energy proportion of the electrical quantity data and the statistical features of the non-electrical quantity data, and output the abnormal probability score and the preliminary abnormal type.

[0071] It should be noted that:

[0072] The Transformer model is a deep neural network structure based on the combination of multi-head self-attention mechanism and feedforward neural network, which is composed of embedding layer, position encoding layer, multi-layer multi-head self-attention mechanism, residual connection and normalization structure. The dependence between each feature dimension in the input sequence is extracted by stacking the encoder layer to form a unified feature representation. Due to the good feature representation ability and parallel processing structure of the Transformer model, and the coding result can be connected to multiple output layers, different output branches can be constructed to realize the multi-task decoupling processing of the unified feature representation.

[0073] In the prior art, the Transformer model is widely used for simultaneous regression prediction and classification discrimination tasks. The multi-output structure is usually composed of a shared encoding part and independent task branches, which correspond to different loss functions. After extracting features from the unified input, multiple prediction results are output simultaneously. Therefore, on the basis of maintaining the original structure advantages of the Transformer model, the abnormal probability score and the preliminary abnormal type can be output simultaneously in one forward propagation by connecting the fully connected regression layer and the Softmax classification layer respectively.

[0074] S2.1, standardize the multi-modal data matrix, and use fast Fourier transform to extract the harmonic energy proportion of the electrical quantity data.

[0075] It should be noted that all data in the multi-modal data matrix is standardized to eliminate the influence of different data magnitudes and ensure analysis in a unified scale. Fast Fourier transform (FFT) is used for electrical quantity data to convert time domain signals into frequency domain signals, decompose each frequency component in the electrical quantity data, and calculate the energy proportion of each frequency component. The expression is:

[0076] ;

[0077] In the formula, represents the energy proportion of the frequency component at the position, represents the energy of the frequency component corresponding to the frequency component, represents the index of the frequency component;

[0078] According to the energy proportion of each frequency component, the frequency position corresponding to the first harmonic is determined, and the integer multiple frequency positions of the first harmonic are identified as harmonic frequencies in turn. The energy proportions of these harmonic frequencies are added or used as indicators respectively, thereby specifically reflecting the contribution of each order harmonic to the overall signal in the electrical quantity data, and finally obtaining the harmonic energy proportion of the electrical quantity data.

[0079] It should be noted that the first harmonic is the lowest frequency component in the frequency domain signal obtained by performing fast Fourier transform (FFT) on the electrical quantity data.

[0080] S2.2, the mean and standard deviation of the non-electric quantity data are used to extract the volatility and stability characteristics of the non-electric quantity data, and statistical characteristics are obtained.

[0081] It should be noted that the volatility of the non-electric quantity data is quantified by the standard deviation value of the non-electric quantity data in the selected time window. The larger the standard deviation value, the greater the change range of the non-electric quantity data, the stronger the volatility, and the greater the fluctuation of the non-electric quantity data. The stability of the non-electric quantity data is calculated by calculating the ratio of the mean value to the standard deviation, that is, the quotient value of the mean value and the standard deviation, as a stability index. The larger the quotient value, the smaller the fluctuation of the non-electric quantity data relative to the mean value, and the non-electric quantity data tends to be stable. Based on the standard deviation of the non-electric quantity data and the ratio of the mean value to the standard deviation, the volatility and stability characteristics are extracted, and finally the statistical characteristics are obtained.

[0082] S2.3, the harmonic energy proportion and the statistical characteristics are spliced into a comprehensive feature vector, which is input into the Transformer model, mapped to a high-dimensional feature space through an embedding layer, and a unified feature representation is obtained. The self-attention mechanism is used to extract the dependency relationship of the unified feature representation, and the abnormal probability score is output through the fully connected layer. The abnormal probability score is classified through the Softmax layer, and the preliminary abnormal type is output.

[0083] It should be noted that the harmonic energy proportion and the statistical characteristics are spliced in the order of dimensions to form a comprehensive feature vector.

[0084] The comprehensive feature vector is input into the Transformer model, and the embedding layer in the Transformer model is used to map the comprehensive feature vector to a high-dimensional feature space to obtain a unified feature representation.

[0085] The unified feature representation calculates the dependency relationship between each dimension of the comprehensive feature vector through the self-attention mechanism in the Transformer model, identifies the high-weight feature dimension in the overall comprehensive feature vector, and updates the weight of the high-weight feature dimension to enhance the abnormal discrimination ability.

[0086] The unified feature representation processed by the self-attention mechanism is input into the fully connected layer in the Transformer model, and the abnormal probability score corresponding to the comprehensive feature vector is calculated based on the weight parameter, and the expression is:

[0087] ;

[0088] In the formula, is the abnormal probability score, is the weight matrix of the fully connected layer in the Transformer model, is the unified feature representation, is the bias vector of the fully connected layer in the Transformer model;

[0089] The Softmax layer in the Transformer model is inputted with the anomaly probability score, and the classification probability value corresponding to each anomaly type is calculated, expressed as:

[0090] ;

[0091] In the formula, is the classification probability value corresponding to the i-th anomaly type, is the anomaly probability score of the i-th anomaly type, is the result of exponential operation on the anomaly probability score of the i-th anomaly type, is the anomaly probability score of the i-th anomaly type, is the result of exponential operation on the anomaly probability score of the i-th anomaly type, is the anomaly probability score of the i-th anomaly type, is the result of exponential operation on the anomaly probability score of the i-th anomaly type, is the anomaly probability score of the i-th anomaly type, is the result of exponential operation on the anomaly probability score of the i-th anomaly type, is the index of all anomaly types traversed in the Softmax normalization process, is the index of the anomaly type currently requiring calculation of the Softmax output, represents the total number of anomaly types; According to the classification probability of each anomaly type output by the Softmax layer, the anomaly type corresponding to the maximum classification probability is selected as the preliminary anomaly type, and the final output result is the anomaly probability score and the preliminary anomaly type.

[0092] It should be noted that the steps of training the Transformer model are as follows: input the labeled comprehensive feature vector into the Transformer model, map it to a high-dimensional space through the embedding layer, and then learn the dependency between features using the multi-head self-attention mechanism; then, a multi-task joint training strategy is adopted to simultaneously optimize the anomaly probability score (mean square error loss) output by the fully connected layer and the anomaly type classification (cross-entropy loss) of the Softmax layer, and the Transformer model parameters are iteratively updated through backpropagation and the Adam optimizer (adaptive moment estimation optimizer), and finally a Transformer model capable of simultaneously predicting the anomaly probability and type is obtained.

[0093]

[0094] ​S3, evaluate the anomaly probability score and the anomaly probability threshold, obtain the anomaly state flag, and combine the preliminary anomaly type and the power equipment communication interaction data stream to generate a security event data packet.

[0095] S3.1, statistically distribute the historical anomaly probability score, select the optimal segmentation point through the ROC curve, and dynamically adjust in combination with the real-time collected device real-time running state to set the anomaly probability threshold.

[0096] It should be noted that by collecting the anomaly probability score data generated during the past operation of the power equipment, frequency statistics and distribution analysis are performed according to the numerical range to obtain the statistical distribution of the historical anomaly probability score. The receiver operating characteristic curve (ROC curve) method is used to determine the optimal segmentation point of the anomaly probability score as the initial value of the anomaly probability threshold by calculating the change relationship between the true positive rate and the false positive rate.

[0097] According to the initial value of the anomaly probability threshold, in combination with the real-time collected device real-time running state, the anomaly probability threshold is dynamically corrected through a weighted adjustment method, so that the anomaly probability threshold can reflect the current running conditions and environmental changes of the device, and finally output the anomaly probability threshold for anomaly determination.

[0098] It should also be noted that the device real-time running state includes load rate, temperature, running time, and communication quality of the device, and other multiple indicators reflecting the current running conditions of the device.

[0099] S3.2, compare the anomaly probability score with the anomaly probability threshold. When the anomaly probability score is higher than the anomaly probability threshold, it is determined that the state is abnormal, otherwise it is determined that the state is normal, and the corresponding anomaly state flag is generated.

[0100] It should be noted that the anomaly probability score and the anomaly probability threshold are compared by value, and the running state of the power equipment is determined by judging whether the anomaly probability score exceeds the anomaly probability threshold.

[0101] When the anomaly probability score is greater than the anomaly probability threshold, it is determined that the power equipment is in an abnormal state.

[0102] When the anomaly probability score is less than or equal to the anomaly probability threshold, it is determined that the power equipment is in a normal state, and according to the determination result, the corresponding anomaly state flag is generated.

[0103] S3.3, associate the anomaly state flag with the preliminary anomaly type, correspond each anomaly state to the preliminary anomaly type, and mark the time period to which the anomaly source data belongs.

[0104] It should be noted that matching each abnormal state flag with the corresponding preliminary abnormal type as a unique identifier ensures the establishment of a one-to-one correspondence; then, according to the abnormal probability score and the time stamp information corresponding to the abnormal state flag, the specific time period to which each abnormal state corresponds to the preliminary abnormal type is marked, and the start and end time range of the abnormal occurrence is determined; the specific operation includes traversing all abnormal state flags, binding the preliminary abnormal type corresponding to each abnormal state flag as an attribute, and extracting all data in the time interval corresponding to the abnormal state from the power equipment communication interaction data stream through the time stamp index, to complete the marking of the abnormal source data time period.

[0105] S3.4, extracting all data segments in the time period of the abnormal source data from the power equipment communication interaction data stream as the abnormal data source.

[0106] It should be noted that the start and end time points of the time period of the abnormal source data identified by the abnormal state flag are used to determine the time range of the time period of the abnormal source data; then, the power equipment communication interaction data stream is filtered using the time stamp, and the complete continuous data segment containing all electrical quantity data and non-electrical quantity data in the time period of the abnormal source data is extracted, and the continuous data segment containing the electrical quantity data and non-electrical quantity data of the abnormal event extracted is taken as the abnormal data source.

[0107] S3.5, structurally encapsulating the abnormal state flag, the preliminary abnormal type, the abnormal probability score and the abnormal data source to generate a security event data packet.

[0108] It should be noted that a unified information structure is created for each abnormal event, and the abnormal state flag is used as a unique identifier; the corresponding preliminary abnormal type is bound with the abnormal state flag to ensure complete classification information; the abnormal probability score is added to the structure as a quantitative indicator of the abnormal event to reflect the severity of the abnormality; the abnormal data source extracted from the power equipment communication interaction data stream is completely integrated into the structured information to maintain the original continuity and timing characteristics of the abnormal behavior; after the above steps, all associated information is summarized as a security event data packet.

[0109] S4, using LiNGAM causal discovery algorithm (linear non-Gaussian acyclic model causal discovery algorithm) to perform non-Gaussianity test and timing lag analysis on the abnormal event time series in the security event data packet, and constructing a directed acyclic causal graph.

[0110] S4.1, extracting the abnormal source data corresponding to the abnormal event time period from the security event data packet to form an abnormal event time series, inputting the multi-node parallel computing architecture, performing multi-dimensional non-Gaussianity test, outputting non-Gaussian time series data, and performing feature extraction and dynamic timing lag identification to obtain optimized lag time series data.

[0111] It should be noted that the safety event data packet is scanned according to the abnormal state flag, the start and end time points of the abnormal state are identified, and the abnormal event time period is determined, so as to accurately locate the time range of the abnormal source data;

[0112] According to the time range, the abnormal source data containing electrical quantity data and non-electrical quantity data is extracted from the safety event data packet to form an abnormal event time sequence;

[0113] The abnormal event time sequence is input into a multi-node parallel computing architecture, and the multi-node distributed processing capability is used to simultaneously process multi-dimensional time sequence data; through a multi-dimensional non-Gaussianity test method, the symmetry, peak value and tail characteristics of the distribution of each variable (such as voltage, current, communication state, etc.) in the abnormal event time sequence are observed to judge the high-order statistical characteristics (such as skewness, kurtosis) of each variable, and the abnormal event time sequence is decomposed by independent component analysis (ICA) to separate the components (such as components with obvious asymmetry or peak characteristics) deviating from the Gaussian distribution of statistical characteristics, and then the characteristic components conforming to the non-Gaussian distribution are screened out by hypothesis testing (such as Jarque-Bera test (Chinese name: Jarque-Bera normality test)), and the noise or interference data close to the Gaussian distribution is removed, and the non-Gaussian time sequence data is output;

[0114] The non-Gaussian time sequence data is subjected to feature extraction, including time sequence statistical characteristics and frequency domain characteristics and other abnormal discrimination characteristics; in combination with a dynamic time sequence lag identification method, the time sequence corresponding relationship of different variable change trends is investigated, the order of appearance of the abnormal discrimination characteristics is analyzed, the reasonable lag step between causal variables is determined, and the pseudo-lag correlation is excluded by a causal verification method (such as Granger test (Chinese name: Granger causality test)), to obtain optimized lag time sequence data;

[0115] It should also be noted that the multi-node parallel computing architecture is a distributed computing architecture that improves data processing efficiency by dividing the computing task into multiple sub-tasks and distributing them to multiple computing nodes for parallel execution; the construction of the multi-node parallel computing architecture includes configuring multiple computing devices with network interconnection capability, building a distributed computing framework supporting task scheduling and resource management, such as Apache Hadoop (Apache Hadoop (big data distributed computing framework)) or Apache Spark (big data high-speed computing engine), coordinating the task allocation and data communication between nodes through a scheduler, and realizing efficient multi-dimensional time sequence data parallel processing capability.

[0116] S4.2, using LiNGAM causal discovery algorithm, performing linear non-Gaussian modeling and topology sorting on the optimized lag time sequence data, inferring the direct causal path between variables, and outputting a candidate edge set of causal relationship.

[0117] It should be noted that the optimized lag time series data is represented as a multi-dimensional time series matrix, each column corresponding to an electrical quantity data or non-electrical quantity data variable, and each row corresponding to an observation value at the same time step; then it is assumed that each variable is a linear combination of several other variables plus an error term with a non-Gaussian distribution, and a structural equation model reflecting the causal relationship between power equipment parameters is established, in which each variable is represented as a weighted combination of other variables plus a non-Gaussian noise term, to describe the causal dependence relationship between power equipment parameters; then, the independent component analysis method in the LiNGAM causal discovery algorithm is used to separate the source signals of the optimized lag time series data, and the linear influence coefficients between the latent variables are obtained;

[0118] After modeling is completed, the topological sorting mechanism in the LiNGAM causal discovery algorithm is used to sort all variables in a loop-free directed graph, and the direct causal direction between variables is determined, and finally the non-zero causal path is extracted based on the linear influence coefficient matrix to generate a causal relationship candidate edge set.

[0119] S4.3, the causal relationship candidate edge set is drawn as a directed acyclic causal graph according to the lag order.

[0120] It should be noted that according to the time index information of each variable in the optimized lag time series data, the time sequence of each pair of variables in the causal relationship candidate edge set is determined to ensure that the causal relationship direction strictly points from the variable at the earlier time step to the variable at the later time step; then, all causal relationship candidate edges that meet the time consistency are connected in turn to build a graph structure represented by nodes and directed edges representing causal impact paths; during the construction process, the causal edges that may form loops are removed to ensure that the graph structure meets the loop-free property; finally, the sorted causal path set is visualized as a directed acyclic causal graph.

[0121] S5, according to the real-time device load rate and communication quality index, dynamically adjust the causal strength of the directed acyclic causal graph, and output a dynamic causal graph.

[0122] S5.1, collect the real-time load rate and communication quality index of the power equipment, perform normalization processing to form standardized performance data, and use a weighted linear combination method to calculate a causal strength adjustment coefficient to correct the initial strength value of each causal edge in the directed acyclic causal graph.

[0123] It should be noted that the real-time load rate and communication quality index are normalized in the interval, respectively, to form standardized performance data under a unified dimension; then the standardized performance data is taken as input, based on the index weight parameter, and the causal strength adjustment coefficient corresponding to each power equipment is calculated through the weighted linear combination method, and the expression is:

[0124] ;

[0125] In the formula, represents the causal strength adjustment coefficient of the first causal strength adjustment coefficient of the power equipment of the first represents the weight coefficient of the normalized real-time load rate, represents the causal strength adjustment coefficient of the first normalized real-time load rate of the power equipment of the first represents the weight coefficient of the normalized communication quality index, represents the causal strength adjustment coefficient of the first normalized communication quality index of the power equipment of the first

[0126] After obtaining the causal strength adjustment coefficient, the initial strength value of the original causal edge is proportionally corrected according to the causal strength adjustment coefficient of the power equipment involved in each causal edge in the directed acyclic causal graph, so that the causal edge strength value can more accurately reflect the actual influence degree of the current running state of the power equipment on the causal relationship strength;

[0127] It should be noted that the real-time load rate of the power equipment is obtained by collecting the ratio of the active load to the rated load during the running of the power equipment, and the normalized real-time load rate of the power equipment is obtained after interval normalization processing of the real-time load rate of the power equipment;

[0128] The communication quality original index of the power equipment is obtained by collecting performance parameters such as packet loss rate, time delay and jitter generated by the power equipment in the communication process, and the normalized communication quality index of the power equipment is obtained after interval normalization processing of the communication quality original index of the power equipment.

[0129] S5.2, assign the corrected initial strength value to the corresponding causal edge of the directed acyclic causal graph, and output the dynamic causal graph.

[0130] It should be noted that the corrected causal edge strength value is sequentially assigned to all causal edges corresponding to the causal path between the power equipment variables in the directed acyclic causal graph, forming a directed acyclic causal graph structure with dynamic weights, and finally outputting the dynamic causal graph.

[0131] S6, combine the dynamic causal graph with the equipment account data and the real-time running state of the equipment, perform path backtracking analysis, evaluate the fault propagation depth and node contribution degree, and generate a power equipment safety evaluation report.

[0132] S6.1, associate and integrate the nodes and causal edges in the dynamic causal graph with the equipment account data and the real-time running state of the equipment, form a comprehensive data structure, and extract abnormal nodes, take the abnormal nodes as the starting point, traverse the causal path along the dynamic directed causal edge in reverse, perform path backtracking analysis, and output the fault propagation depth.

[0133] It should be noted that all nodes and causal edges in the dynamic causal graph are respectively aligned and logically matched with the field level of the power equipment account data and the real-time running state of the power equipment, forming a comprehensive data structure containing causal structure, device static attributes and device dynamic indicators;

[0134] In the comprehensive data structure, according to the abnormal state flag recorded in the real-time running state of the power equipment, all power equipment nodes identified as abnormal state are extracted as an abnormal node set;

[0135] Taking each abnormal node in the abnormal node set as a starting point, along the directed causal edges in the dynamic causal graph that have direct or indirect causal connection relationship with the current abnormal node, the reverse path traversal is performed according to the direction relationship of the causal edges;

[0136] In the traversal process, every time a node is reached, the path depth value is recorded as one, until it cannot continue to backtrack, the maximum depth value of all backtracking paths is obtained, the longest causal path layer from the abnormal node is obtained, and the maximum path depth value is output as the fault propagation depth;

[0137] It should be noted that the power equipment account data is derived from the factory data provided by the equipment manufacturer and the field detection record, which contains device technical parameters, manufacturing batch, installation location and maintenance history.

[0138] S6.2, based on the causal edge weight in the backtracking path and the real-time running state of the equipment, the contribution degree of each node to the fault propagation is evaluated, and the node contribution degree is obtained.

[0139] For the selected specific fault propagation path in the dynamic causal graph, the causal edge weight value between all adjacent power equipment in the specific fault propagation path is extracted; the real-time running state data of the power equipment is obtained to construct the current running state data of each power equipment, and the current running state data of each power equipment is normalized, so that the running state data of different power equipment is in the same numerical scale range, which is convenient for weighted processing in subsequent calculation;

[0140] The normalized running state values of the two power equipment connected by each causal edge are combined with the causal edge weight value, and the state response value on the causal edge is calculated to represent the strength of the propagation effect on the causal edge. The state response value is equal to the weighted product of the causal edge weight value and the normalized running state value of the two power equipment. For all causal edges contained in the entire backtracking path, the cumulative state response value of each power equipment is calculated. The specific way is: the causal edge state response values of all power equipment with the current calculation node contribution degree as the target node or source node are weighted and summed, and the result is normalized as the node contribution degree value of the power equipment with the current calculation node contribution degree.

[0141] S6.3, aggregate the fault propagation depth, node contribution degree, device account data and real-time running state of the device to generate a power device safety assessment report.

[0142] It should be noted that based on the fault propagation depth, the hierarchical range and propagation degree of fault propagation are reflected, and based on the node contribution degree, the importance and influence of each power device in the fault propagation process are embodied; then, the power device account data and the real-time running state of the device are aligned and logically integrated at the field level to form a complete device information set containing static attributes and dynamic indicators, ensuring that the safety assessment report can fully reflect the actual characteristics and running conditions of the power device; according to the above-mentioned fault propagation depth, node contribution degree, device account data and real-time running state of the device, the fault propagation depth, node contribution degree, device account data and real-time running state of the device are coded and organized in a unified data structure specification, ensuring that each index field is clear and hierarchical, facilitating information association and query, and generating a power device safety assessment report.

[0143] The embodiment also provides a power device safety state assessment system, comprising: a data acquisition module, configured to acquire power device communication interaction data flow, perform time synchronization and outlier elimination, and construct a multi-modal data matrix;

[0144] a feature extraction module, configured to input the multi-modal data matrix into a Transformer model, extract harmonic energy proportion of electrical quantity data and statistical features of non-electrical quantity data, and output an abnormal probability score and a preliminary abnormal type;

[0145] an abnormality determination module, configured to evaluate the abnormal probability score and an abnormal probability threshold, obtain an abnormal state flag, and generate a safety event data packet in combination with the preliminary abnormal type and the power device communication interaction data flow;

[0146] a relationship analysis module, configured to use a LiNGAM causal discovery algorithm to perform non-Gaussianity test and time series lag analysis on an abnormal event time series in the safety event data packet, and construct a directed acyclic causal graph;

[0147] a strength adjustment module, configured to dynamically adjust the causal strength of the directed acyclic causal graph according to real-time device load rate and communication quality indicators, and output a dynamic causal graph;

[0148] a fault assessment module, configured to combine the dynamic causal graph with device account data and real-time running state of the device, perform path backtracking analysis, assess fault propagation depth and node contribution degree, and generate a power device safety assessment report.

[0149] The embodiment also provides a computer device suitable for the power equipment safety state evaluation method, including a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions to realize the power equipment safety state evaluation method proposed in the above embodiment.

[0150] The computer device can be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with external terminals in a wired or wireless manner. The wireless manner can be achieved through WIFI, an operator network, NFC (Near Field Communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, a trackball or a touchpad arranged on the shell of the computer device, or an external keyboard, a touchpad or a mouse, etc.

[0151] The embodiment also provides a storage medium having a computer program stored thereon, the program being executed by a processor to realize the power equipment safety state evaluation method proposed in the above embodiment. The storage medium can be realized by any type of volatile or non-volatile storage device or a combination thereof, such as a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic memory, a flash memory, a magnetic disk or an optical disk.

[0152] To sum up, the application realizes modeling of the causal relationship between abnormal events by constructing a directed acyclic causal graph, can reveal the propagation path and the chronological relationship between the abnormalities, and enhances the logical clarity and analysis depth of the power equipment fault diagnosis; the dynamic causal graph is constructed by introducing the equipment load rate and the communication quality index to dynamically adjust the causal strength, so that the causal relationship can reflect the change of the running state in real time, the accuracy and adaptability of the abnormal propagation analysis are improved, and finally the real-time and reliability of the power equipment safety evaluation are enhanced.

[0153] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application but not limit the present application, although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application, and they should be covered in the scope of the claims of the present application.

Claims

1. A method for assessing the safety status of power equipment, characterized by: include, Collect communication interaction data streams of power equipment, perform time synchronization and outlier removal, and construct a multimodal data matrix; The multimodal data matrix is ​​input into the Transformer model to extract the harmonic energy ratio of the electrical quantity data and the statistical characteristics of the non-electricity data, and output the anomaly probability score and preliminary anomaly type; Evaluate the abnormal probability score and the abnormal probability threshold to obtain the abnormal status flag, and combine the preliminary abnormal type and the power equipment communication interaction data flow to generate a security event data packet; The LiNGAM causal discovery algorithm is used to perform non-Gaussianity tests and time series lag analysis on the time series of abnormal events in security event data packets, and to construct a directed acyclic causal graph. According to the real-time device load rate and communication quality indicators, the causal strength of the directed acyclic causal graph is dynamically adjusted to output a dynamic causal graph; Combine the dynamic causal graph with equipment ledger data and the real-time operating status of the equipment to conduct path backtracking analysis, evaluate the fault propagation depth and node contribution, and generate a power equipment safety assessment report.

2. The method for evaluating the safety status of electric power equipment according to claim 1, wherein: The time synchronization and outlier removal are performed to construct a multimodal data matrix. The specific steps are as follows: Resampling the communication interaction data stream of the power equipment to obtain synchronous communication data; the communication interaction data stream of the power equipment includes electrical quantity data and non-electrical quantity data; Calculate the mean and standard deviation of electrical quantity data and eliminate abnormal data that exceeds the normal range; Draw a box plot for non-electricity data and remove abnormal data that are beyond the upper and lower quartiles; The cleaned electrical quantity data, cleaned non-electrical quantity data and synchronous communication data are aligned and normalized by timestamp, and then spliced ​​into a multi-dimensional matrix by column to construct a multimodal data matrix.

3. The method for evaluating the safety status of electric power equipment according to claim 1, wherein: The specific steps of outputting abnormal probability scores and preliminary abnormality types are as follows: The multimodal data matrix is ​​standardized, and the electrical quantity data is subjected to fast Fourier transform to extract the harmonic energy ratio of the electrical quantity data; Using the mean and standard deviation of non-electricity data, the volatility and stability characteristics of non-electricity data are extracted to obtain statistical characteristics; The harmonic energy proportion and statistical features are concatenated into a comprehensive feature vector and input into the Transformer model. The vector is mapped to a high-dimensional feature space through an embedding layer to obtain a unified feature representation. The self-attention mechanism is used to extract the dependency of the unified feature representation, and the fully connected layer outputs the anomaly probability score. The Softmax layer classifies the anomaly probability score into anomaly types and outputs a preliminary anomaly type.

4. The method for evaluating the safety status of electric power equipment according to claim 1, wherein: The specific steps of generating a security event data packet are as follows: Based on the statistical distribution of historical abnormal probability scores, the optimal split point is selected through the ROC curve, and dynamic adjustments are made based on the real-time operating status of the equipment collected in real time to set the abnormal probability threshold; Compare the abnormal probability score with the abnormal probability threshold. When the abnormal probability score is higher than the abnormal probability threshold, it is judged as an abnormal state. Otherwise, it is judged as a normal state and a corresponding abnormal state flag is generated. Associate the abnormal state flag with the preliminary abnormality type, correspond each abnormal state to the preliminary abnormality type, and mark the time period to which the abnormal source data belongs; Extract all data fragments within the time period of the abnormal source data from the power equipment communication interaction data stream as the abnormal source data source; The abnormal status flag, preliminary abnormal type, abnormal probability score and abnormal data source are structurally encapsulated to generate a security event data packet.

5. The method for evaluating the safety status of electric power equipment according to claim 1, wherein: The specific steps of constructing a directed acyclic causal graph are as follows: Extract abnormal source data corresponding to the abnormal event time period from the security event data packet to form an abnormal event time series. Input the data into a multi-node parallel computing architecture to perform multi-dimensional non-Gaussianity testing, output non-Gaussian time series data, and perform feature extraction and dynamic time series lag identification to obtain optimized lag time series data. The LiNGAM causal discovery algorithm is used to perform linear non-Gaussian modeling and topological sorting on optimized lagged time series data, inferring direct causal paths between variables and outputting a set of candidate causal relationship edges. The set of candidate causal relationships is drawn into a directed acyclic causal graph in the order of lag.

6. The method for evaluating the safety status of electric power equipment according to claim 1, wherein: The specific steps of outputting the dynamic causal graph are as follows: The real-time load rate and communication quality indicators of power equipment are collected and normalized to form standardized performance data. The causal strength adjustment coefficient is calculated using the weighted linear combination method to correct the initial strength value of each causal edge in the directed acyclic causal graph. The corrected initial strength value is assigned to the corresponding causal edge of the directed acyclic causal graph, and a dynamic causal graph is output.

7. The method for evaluating the safety status of electric power equipment according to claim 1, wherein: The specific steps of generating the power equipment safety assessment report are as follows: The nodes and causal edges in the dynamic causal graph are associated and integrated with the equipment ledger data and the real-time operating status of the equipment to form a comprehensive data structure. The abnormal nodes are then extracted. Starting from the abnormal nodes, the causal path is traversed in reverse along the dynamic directed causal edges to perform path backtracking analysis and output the fault propagation depth. Based on the causal edge weights in the backtracking path and the real-time operating status of the equipment, the contribution of each node to the fault propagation is evaluated to obtain the node contribution; The fault propagation depth, node contribution, equipment inventory data and real-time equipment operating status are summarized to generate a power equipment safety assessment report.

8. A power equipment safety status assessment system, based on the power equipment safety status assessment method according to any one of claims 1 to 7, characterized in that: include, The data acquisition module is used to collect the communication interaction data stream of power equipment, perform time synchronization and outlier removal, and build a multimodal data matrix; The feature extraction module is used to input the multimodal data matrix into the Transformer model, extract the harmonic energy ratio of the electrical quantity data and the statistical features of the non-electricity data, and output the anomaly probability score and preliminary anomaly type; The anomaly determination module is used to evaluate the anomaly probability score and the anomaly probability threshold, obtain the anomaly status flag, and generate a security event data packet based on the preliminary anomaly type and the power equipment communication interaction data flow; The relationship analysis module uses the LiNGAM causal discovery algorithm to perform non-Gaussianity testing and time series lag analysis on the abnormal event time series in the security event data packet, and construct a directed acyclic causal graph; Strength adjustment module, used to dynamically adjust the causal strength of the directed acyclic causal graph according to the real-time device load rate and communication quality indicators, and output a dynamic causal graph; The fault assessment module is used to combine the dynamic causal graph with equipment ledger data and the real-time operating status of the equipment to perform path backtracking analysis, evaluate the fault propagation depth and node contribution, and generate a power equipment safety assessment report.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method for evaluating the safety status of electric power equipment according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for evaluating the safety status of electric power equipment according to any one of claims 1 to 7 are implemented.

Images