A method, system, device and medium for intelligently identifying abnormal behavior of a relay protection device

Abstract

This invention relates to a method, system, device, and medium for intelligent identification of abnormal behavior in relay protection devices. The method includes: performing time-frequency alignment on electrical quantity data and physical state quantity data to generate a multi-dimensional time series matrix; analyzing and extracting physical state quantity data and weighted fusing them to output a physical degradation index sequence; performing wavelet packet decomposition and reconstruction on the electrical quantity data to generate fault feature vectors and noise feature vectors; inputting the fault feature vectors and noise feature vectors into a twin network for feature dimensionality reduction and fusing them to generate a discrimination probability vector; and combining the original start signal input into a rule engine for logical decision-making to generate a final decision result. This invention effectively distinguishes between hardware aging noise and real fault signals by fusing physical layer contact state information and electrical quantity waveform features, avoiding protection malfunctions caused by equipment interface contact state degradation, and improving the operational reliability of relay protection devices under complex operating conditions.

Description

Technical Field

[0001] This invention relates to the field of power system automation technology, specifically to a method, system, device, and medium for intelligent identification of abnormal behavior of relay protection devices. Background Technology

[0002] With the continuous expansion of the power grid and the increasing proportion of new energy integration, relay protection devices, as the first line of defense for ensuring the safe and stable operation of the power system, directly affect the grid's fault isolation capability and power supply continuity. Driven by the construction of smart grids, relay protection anomaly behavior identification technology based on electrical quantity analysis has been widely applied. This type of method identifies potential fault characteristics or abnormal patterns by performing waveform analysis and feature extraction on electrical quantity data such as voltage and current collected by the protection device, providing decision support for grid dispatching and maintenance personnel.

[0003] However, in actual operating environments, especially in older substations with long service lives, existing relay protection anomaly identification methods based on pure electrical quantity analysis face severe challenges. After long-term operation, primary equipment interfaces may experience physical aging phenomena such as oxidation and loosening, leading to increased contact resistance. This, in turn, causes amplitude attenuation, waveform distortion, and noise superposition in the electrical quantity signals acquired on the secondary side. These signal distortions caused by the deterioration of physical layer contact conditions are highly similar in time-frequency domain characteristics to the transient signals generated by actual short-circuit faults. This makes it difficult for traditional identification methods to effectively distinguish between hardware aging noise and real fault signals. They are prone to misjudging anomalies caused by equipment interface aging as faults within the protection zone, leading to malfunctions of protection devices, or drowning out real fault signals in noise, causing protection to fail to operate, seriously affecting the reliability of power grid operation. Summary of the Invention

[0004] Based on this, the purpose of the present invention is to provide a method, system, device and medium for intelligent identification of abnormal behavior of relay protection devices that can integrate physical layer contact state information and electrical quantity waveform characteristics and accurately distinguish between hardware aging noise and real fault signals.

[0005] The objective of this invention is achieved through the following solution:

[0006] In a first aspect, the present invention provides a method for intelligent identification of abnormal behavior of a relay protection device, comprising the following steps:

[0007] S1: Perform time-frequency alignment on the collected electrical quantity data of the relay protection device access point and the physical state quantity data reflecting the contact status of the equipment interface to generate a multi-dimensional time series matrix containing electrical quantity information and physical state quantity information;

[0008] S2: Based on sliding window analysis, the contact resistance statistical characteristics, temperature-resistance correlation characteristics and vibration amplitude characteristics of physical state data in the multidimensional time series matrix are extracted and weighted and fused to output the physical degradation index of each window and expand it into a physical degradation index sequence aligned with the time of electrical quantity data.

[0009] S3: Based on the physical degradation index sequence, wavelet packet decomposition and reconstruction processing is performed on the electrical quantity data in the multidimensional time series matrix. The original electrical quantity signal is decoupled into fault characteristic path signal and noise characteristic path signal. The time-frequency domain statistical features of the fault characteristic path signal and noise characteristic path signal within the current analysis window are extracted to generate fault feature vector and noise feature vector.

[0010] S4: Input the fault feature vector into the preset Siamese network for feature dimensionality reduction to obtain the fault latent vector. Then, concatenate the noise feature vector and the mean of the physical degradation index corresponding to the current window and input them into the preset Siamese network for feature dimensionality reduction to obtain the noise latent vector. Finally, perform attention mechanism fusion processing on the fault latent vector and the noise latent vector to generate the discrimination probability vector of the current window.

[0011] S5: Obtain the original start signal of the relay protection device in the current window, and input the discrimination probability vector, the average physical degradation index and the original start signal into the preset rule engine for logical decision-making to generate the final decision result. The final decision result is used to indicate one of the following: allow protection action, block protection output, and issue maintenance warning.

[0012] In one embodiment, S2 of the intelligent identification method for abnormal behavior of a relay protection device provided by the present invention specifically includes the following steps:

[0013] S21: Based on the sliding window technique, the contact resistance data in the multidimensional time series matrix is ​​segmented and truncated. The arithmetic mean and standard deviation of the contact resistance in each window are calculated to generate contact resistance statistical features. The contact resistance statistical features are used to characterize the steady-state level of the interface contact state.

[0014] S22: Perform correlation analysis on temperature data and contact resistance data in the multidimensional time series matrix, calculate the Pearson correlation coefficient between temperature change and contact resistance change within the window, and generate temperature-resistance correlation features. Temperature-resistance correlation features are used to distinguish between physical loosening caused by thermal expansion and contraction and fault arc.

[0015] S23: Perform root mean square value calculation on the vibration data in the multidimensional time series matrix, calculate the square root of the mean of the sum of squares of vibration amplitudes within the window, and generate vibration amplitude features. The vibration amplitude features are used to quantify the severity of mechanical loosening.

[0016] S24: Based on preset weighting coefficients, the contact resistance statistical characteristics, temperature-resistance correlation characteristics and vibration amplitude characteristics are linearly weighted and fused to generate the physical degradation index for each window;

[0017] S25: A zero-order hold is used to extend the physical degradation index of each window along the time axis, extending the discrete window index values ​​into a continuous sequence with the same sampling rate as the electrical quantity data, thereby generating a physical degradation index sequence that is time-aligned with the electrical quantity data.

[0018] In one embodiment, the formula for calculating the physical degradation index of the intelligent identification method for abnormal behavior of a relay protection device provided by the present invention is as follows:

[0019]

[0020] in, Let be the physical degradation index of the w-th window. This represents the average contact resistance within the window. The standard deviation of contact resistance within the window. The Pearson correlation coefficient is the relationship between the temperature change and the contact resistance change within the window. This represents the root mean square value of the vibration amplitude within the window. This is the preset reference contact resistance. The preset reference vibration amplitude, , , , These are the preset weighting coefficients.

[0021] In one embodiment, S3 of the intelligent identification method for abnormal behavior of a relay protection device provided by the present invention specifically includes the following steps:

[0022] S31: Based on the electrical quantity data in the multi-dimensional time series matrix, wavelet packet decomposition is performed on the electrical quantity signals of each channel. The electrical quantity signals are then decomposed into four levels using the db4 wavelet basis to generate wavelet coefficients for each frequency band node.

[0023] S32: Dynamically adapt the wavelet packet denoising threshold of each channel's electrical quantity signal according to the physical degradation index sequence, and adaptively modulate the reference threshold through the physical degradation index to generate a dynamic denoising threshold that changes with the degree of equipment degradation.

[0024] S33: Based on the dynamic denoising threshold, the wavelet coefficients are subjected to threshold screening and reconstruction. The low-frequency nodes and high-frequency fault feature frequency band nodes in the fault feature set are reconstructed after soft threshold shrinkage to generate fault feature path signals. The high-frequency noise-dominant frequency band nodes are subjected to hard threshold truncation, and the part exceeding the dynamic threshold is retained for reconstruction to generate noise feature path signals.

[0025] S34: Using the preset power frequency cycle as the analysis window, perform time-frequency domain statistical feature extraction processing on the fault characteristic path signal and the noise characteristic path signal, calculate the kurtosis and waveform entropy of the fault characteristic path signal and the noise characteristic path signal respectively, and generate the fault characteristic parameters and noise characteristic parameters of each channel.

[0026] S35: Perform feature combination processing on the fault feature parameters and noise feature parameters of all channels, and concatenate the kurtosis and entropy of each channel into a feature vector in sequence to generate the global fault feature vector and global noise feature vector of the current window.

[0027] In one embodiment, the calculation formula for the dynamic denoising threshold of the intelligent identification method for abnormal behavior of a relay protection device provided by the present invention is as follows:

[0028]

[0029] in, Let be the dynamic denoising threshold for the m-th channel at time t. Let be the baseline threshold for the m-th channel. The preset adjustment coefficient, Let be the physical degradation index at time t.

[0030] In one embodiment, S4 of the intelligent identification method for abnormal behavior of a relay protection device provided by the present invention specifically includes the following steps:

[0031] S41: Extract the index value corresponding to the current analysis window from the physical degradation index sequence, average the physical degradation index of all sampling points in the window, and generate the average physical degradation index of the current window.

[0032] S42: Input the fault feature vector into the preset fault encoder for nonlinear dimensionality reduction processing, extract deep features through a three-layer fully connected network, and generate a fault latent vector;

[0033] S43: The noise feature vector is concatenated with the mean of the physical degradation index to form a fused feature, which is then input into a preset noise encoder for nonlinear dimensionality reduction. Deep features are extracted through a three-layer fully connected network to generate a noise latent vector.

[0034] S44: Perform multi-head attention fusion processing on the fault latent vector and the noise latent vector, calculate the attention weight between the two latent vectors, and concatenate the weighted features to generate a fused feature vector;

[0035] S45: The fused feature vector input and output layers are processed by softmax classification, and mapped to a two-dimensional probability space through a fully connected network to generate the discrimination probability vector of the current window. The discrimination probability vector contains the true fault probability and the hardware noise anomaly probability.

[0036] In one embodiment, S5 of the intelligent identification method for abnormal behavior of a relay protection device provided by the present invention specifically includes the following steps:

[0037] S51: Window the original start signal of the relay protection device within the current analysis window, aggregate the start status at any sampling time within the window into a Boolean value, and generate the start status identifier of the current window.

[0038] S52: Compare the true fault probability in the discrimination probability vector with the preset true fault judgment threshold, compare the hardware noise abnormal probability in the discrimination probability vector with the preset hardware noise judgment threshold, and compare the average physical degradation index with the preset degradation exceeding judgment threshold to generate a comparison result vector containing the three comparison results.

[0039] S53: Input the comparison result vector and the start status flag into the preset rule engine for logical matching. If the actual fault probability exceeds the actual fault judgment threshold and the start status flag is 1, then match the protection action rule and generate the protection action instruction. If the hardware noise abnormal probability exceeds the hardware noise judgment threshold and the average physical degradation index exceeds the degradation exceeding judgment threshold, then match the lockout protection and warning rule, generate the lockout protection output and issue the maintenance warning instruction. If the actual fault probability and the hardware noise abnormal probability do not exceed their respective thresholds and are within the preset fuzzy range, then match the edge state verification rule and generate the extended data window verification instruction.

[0040] Among them, the protection action instruction is used to instruct the relay protection device to perform a trip output operation, the maintenance early warning instruction is used to instruct the monitoring system to push a maintenance notification form containing noise characteristic path signals and contact resistance data, and the data window extension verification instruction is used to instruct the system to extend the current analysis window to two power frequency cycles and re-execute steps S3 and S4.

[0041] Secondly, the present invention provides an intelligent identification system for abnormal behavior of relay protection devices, which is configured with the following modules:

[0042] The multi-source time-frequency alignment module includes time-frequency alignment of the electrical quantity data collected from the relay protection device access point and the physical state quantity data reflecting the contact status of the equipment interface, generating a multi-dimensional time series matrix containing electrical quantity information and physical state quantity information;

[0043] The physical degradation index construction module includes extracting contact resistance statistical features, temperature-resistance correlation features, and vibration amplitude features from the physical state data in the multidimensional time series matrix based on sliding window analysis, performing weighted fusion, outputting the physical degradation index of each window, and expanding it into a physical degradation index sequence aligned with the time of the electrical quantity data;

[0044] The electrical quantity time-frequency feature extraction module includes wavelet packet decomposition and reconstruction processing of electrical quantity data in a multi-dimensional time-series matrix based on the physical degradation index sequence, decoupling the original electrical quantity signal into fault feature path signal and noise feature path signal, extracting the time-frequency domain statistical features of the fault feature path signal and noise feature path signal within the current analysis window, and generating fault feature vector and noise feature vector;

[0045] The Siamese network attention fusion module includes inputting the fault feature vector into a preset Siamese network for feature dimensionality reduction to obtain the fault latent vector, concatenating the noise feature vector and the mean of the physical degradation index corresponding to the current window and inputting them into a preset Siamese network for feature dimensionality reduction to obtain the noise latent vector, and performing attention mechanism fusion processing on the fault latent vector and the noise latent vector to generate the discrimination probability vector of the current window.

[0046] The rule engine logic decision module includes acquiring the original start signal of the relay protection device in the current window, and inputting the discrimination probability vector, the average physical degradation index and the original start signal into the preset rule engine for logic decision-making to generate the final decision result. The final decision result is used to indicate one of the following: allow protection action, block protection output, and issue maintenance warning.

[0047] Thirdly, this application provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement any of the above-mentioned intelligent identification methods for abnormal behavior of relay protection devices.

[0048] Fourthly, this application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the above-mentioned intelligent identification methods for abnormal behavior of relay protection devices.

[0049] In summary, the intelligent identification method for abnormal behavior of relay protection devices provided in this application achieves the goal of fusing multi-source data from the physical and information layers by aligning the collected electrical quantity data and physical state quantity data reflecting the contact state of the equipment interface in a time-frequency manner, thus providing a unified data foundation for subsequent cross-domain correlation analysis. By extracting contact resistance statistical features, temperature-resistance correlation features, and vibration amplitude features from the physical state quantity data through sliding window analysis and performing weighted fusion, a quantitative assessment of the degradation degree of the equipment interface contact state can be achieved, generating a physical degradation index sequence aligned with the electrical quantity data in time to accurately reflect the physical health status of the primary equipment interface. Based on the physical degradation index sequence, wavelet packet decomposition and reconstruction processing of the electrical quantity data allows for dynamic adjustment of the denoising threshold according to the actual degradation degree of the equipment interface, accurately decoupling the original electrical quantity signal into fault information. The characteristic path signal and noise characteristic path signal can effectively separate real fault features from hardware aging noise, avoiding excessive suppression of fault transient components by traditional fixed threshold denoising methods. By inputting the fault feature vector and the noise feature vector fused with the mean of the physical degradation index into the Siamese network for feature dimensionality reduction, and performing cross-domain fusion through an attention mechanism, deep interaction between physical domain features and information domain features can be achieved to construct a discrimination probability vector that can distinguish between hardware noise and real faults. Combined with the original start signal input rule engine for logical decision-making, the final decision results such as allowing protection actions, blocking protection outputs, or issuing maintenance warnings can be generated based on different discrimination results. This can effectively avoid protection malfunctions caused by the degradation of equipment interface contact conditions, and at the same time provide precise maintenance guidance for operation and maintenance personnel, thereby significantly improving the operational reliability of relay protection devices under complex operating conditions.

[0050] To better understand and implement this invention, the following detailed description is provided in conjunction with the accompanying drawings. Attached Figure Description

[0051] Figure 1 A flowchart illustrating an intelligent identification method for abnormal behavior of a relay protection device provided in this application embodiment;

[0052] Figure 2 A schematic diagram of the process for generating physical degradation index and physical degradation index sequence provided in the embodiments of this application;

[0053] Figure 3 This is a schematic diagram of the structure of an intelligent identification system for abnormal behavior of a relay protection device, provided as another embodiment of this application. Detailed Implementation

[0054] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.

[0055] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0056] In one embodiment, such as Figure 1 As shown, an intelligent identification method for abnormal behavior of relay protection devices is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, and further to a system including both a terminal and a server, and is implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0057] S1: Time-frequency alignment is performed on the electrical quantity data of the relay protection device access point and the physical state quantity data reflecting the contact status of the equipment interface to generate a multi-dimensional time series matrix containing electrical quantity information and physical state quantity information.

[0058] Specifically, the system acquires electrical quantity data from the relay protection device's connection point and physical state data reflecting the contact status of the equipment interface. The electrical quantity data consists of power system electrical-related monitoring data continuously collected by the relay protection device during operation, while the physical state data consists of physical characteristic data collected by monitoring units deployed at the interface of the primary equipment associated with the relay protection device. Both types of data carry their own time stamp information during acquisition. The system performs data preprocessing on the acquired electrical quantity data and physical state data. The system identifies invalid data in both types of data through data continuity verification. Invalid data refers to missing data and invalid identifier data generated during acquisition due to transmission interruption or abnormal acquisition unit. The system removes the identified invalid data and then uses data completion to fill the gaps left by the removal of invalid data, ensuring the continuity of both types of data.

[0059] Furthermore, the system determines the time identifier of the electrical quantity data as the reference time axis and performs matching calibration on the time identifier of the physical state quantity data. The system adjusts the acquisition nodes of the physical state quantity data through time interpolation to achieve a one-to-one correspondence between the acquisition nodes of the physical state quantity data and the acquisition nodes of the electrical quantity data in the time dimension. Based on this, the system completes the time-frequency alignment processing of the electrical quantity data and the physical state quantity data. After completing the time-frequency alignment, the system performs dimensional integration on the preprocessed and time-frequency aligned electrical quantity data and physical state quantity data. The system uses the time identifier as the core dimension for data arrangement and uses various types of electrical quantity data and various types of physical state quantity data as different feature dimensions. The system arranges all data in chronological order to generate a multi-dimensional time series matrix containing electrical quantity information and physical state quantity information. Each data unit in this matrix corresponds to a unique time identifier and feature dimension.

[0060] S2: Based on sliding window analysis, the contact resistance statistical characteristics, temperature-resistance correlation characteristics, and vibration amplitude characteristics of physical state data in the multidimensional time series matrix are extracted and weighted and fused to output the physical degradation index of each window and expand it into a physical degradation index sequence aligned with the time of electrical quantity data.

[0061] Specifically, the system invokes a preset sliding window analysis mechanism to extract features from the physical state data in the multi-dimensional time series matrix. The system continuously truncates the physical state data according to a preset window advancement method, maintaining the temporal continuity of the physical state data during truncation to ensure that the physical state data within each truncation window reflects the physical state changes of the device interface within the corresponding time period. The system then extracts contact resistance statistical features from the physical state data within each truncation window. A full statistical analysis of the contact resistance-related data within the window is performed, extracting relevant statistical features that reflect the overall state, dispersion, and trend of contact resistance based on the temporal and numerical distribution of the data. This ensures that the extracted features comprehensively characterize the operating status of the contact resistance.

[0062] Furthermore, the system simultaneously performs correlation analysis on the temperature and resistance data within the window, extracting temperature-resistance correlation features. During the analysis, the system eliminates interference from environmental factors on the temperature-resistance correlation, ensuring that the extracted features reflect the inherent temperature-resistance relationship of the device interface itself. The system also extracts features from the vibration amplitude data within the window. Based on the time-domain distribution characteristics of the vibration data, it extracts vibration amplitude features that reflect the overall vibration intensity, impact characteristics, and waveform distribution, ensuring that the extracted features comprehensively characterize the mechanical vibration state of the device interface. Finally, the system performs feature normalization on the extracted contact resistance statistical features, temperature-resistance correlation features, and vibration amplitude features, ensuring that all features are within a uniform numerical range and eliminating the impact of dimensional differences between different features on subsequent fusion analysis.

[0063] Preferably, the system performs weighted fusion processing on the three types of features after normalization. During the fusion process, the system adopts a preset weight allocation method, which is determined based on the contribution of each feature to the physical degradation degree of the equipment interface. The system calculates the physical degradation index corresponding to each intercept window through weighted fusion calculation. This index is a quantitative indicator that can comprehensively represent the physical degradation degree of the equipment interface within the corresponding time period. After the system completes the calculation of the physical degradation index for all windows, it performs time-dimensional extension processing on the physical degradation index. The system maps the physical degradation index of each window to each time acquisition node of the electrical quantity data through data interpolation, generating a physical degradation index sequence aligned with the time of the electrical quantity data. Each data unit in the index sequence corresponds one-to-one with the time identifier of the electrical quantity data, realizing a comprehensive match between physical degradation information and electrical quantity information in the time dimension.

[0064] S3: Based on the physical degradation index sequence, wavelet packet decomposition and reconstruction processing is performed on the electrical quantity data in the multidimensional time series matrix. The original electrical quantity signal is decoupled into fault characteristic path signal and noise characteristic path signal. The time-frequency domain statistical features of the fault characteristic path signal and noise characteristic path signal within the current analysis window are extracted to generate fault feature vector and noise feature vector.

[0065] Specifically, the system uses the physical degradation index sequence as the weighting basis for data processing, and performs wavelet packet decomposition and reconstruction on the electrical quantity data in the multi-dimensional time series matrix. The system first selects wavelet basis functions suitable for power system electrical quantity signal analysis, and then performs multi-level wavelet packet decomposition on the original electrical quantity signal based on the selected basis functions. This decomposes the original electrical quantity signal into sub-band signals of different frequency bands, each sub-band signal corresponding to the feature information of different frequency ranges in the original electrical quantity signal. The system combines the quantization information of the physical degradation index sequence to classify the decomposed sub-band signals. Based on the distribution patterns of different types of features in the frequency domain and the influence of the physical degradation index on signal features, the system classifies the sub-band signals corresponding to noise features caused by equipment interface aging as noise feature sub-bands, and the sub-band signals corresponding to features caused by power system faults as fault feature sub-bands.

[0066] Furthermore, the system reconstructs the noise characteristic subband signal to obtain the noise characteristic path signal. The system then performs weighted modulation on the fault characteristic subband signal using a physical degradation index sequence before reconstructing it to obtain the fault characteristic path signal. Based on this, the system decouples the original electrical quantity signal from the fault characteristic path signal and the noise characteristic path signal, allowing the two types of signals to respectively represent the fault and noise characteristic information in the original electrical quantity signal. The system performs time-frequency domain statistical feature extraction on the fault characteristic path signal and the noise characteristic path signal within the current analysis window. First, the system performs time-domain analysis on the fault characteristic path signal, extracting time-domain statistical features that reflect the signal amplitude characteristics, energy characteristics, and waveform distribution characteristics. Then, through frequency-domain transformation, the fault characteristic path signal is converted from the time domain to the frequency domain, extracting frequency-domain statistical features that reflect the signal frequency distribution, energy proportion, and spectral complexity. The system then combines the extracted time-domain and frequency-domain statistical features in an ordered manner to generate a fault feature vector. The system adopts the same method as the feature extraction of fault feature path signals to perform statistical feature extraction in the time and frequency domains of noise feature path signals. The extracted features are then combined in an orderly manner to generate noise feature vectors. During the feature extraction process, the system ensures that the feature dimensions of the fault feature vector and the noise feature vector are consistent, providing a unified feature dimension basis for subsequent feature dimensionality reduction processing and avoiding feature processing deviations caused by dimensionality differences.

[0067] S4: Input the fault feature vector into a preset Siamese network for feature dimensionality reduction to obtain the fault latent vector. Then, concatenate the noise feature vector and the mean of the physical degradation index corresponding to the current window and input them into a preset Siamese network for feature dimensionality reduction to obtain the noise latent vector. Finally, perform attention mechanism fusion processing on the fault latent vector and the noise latent vector to generate the discrimination probability vector of the current window.

[0068] Specifically, the system inputs the generated fault feature vector into a pre-defined Siamese network for feature dimensionality reduction. The Siamese network extracts features and performs nonlinear mapping on the input fault feature vector. Through the network's feature learning mechanism, it mines the core feature information in the fault feature vector, mapping the high-dimensional fault feature vector to a low-dimensional feature space. Based on this, the system obtains a fault latent vector that can characterize the core features of the fault. The system first calculates the mean of the physical degradation index sequence within the current analysis window. This mean is used as a new feature dimension and concatenated with the noise feature vector. The concatenated feature data contains both the feature information of the noise feature path signal and the overall physical degradation degree of the equipment interface within the corresponding time period. The system inputs the concatenated feature data into the pre-defined Siamese network for feature dimensionality reduction. The Siamese network uses the same feature learning and mapping mechanism as the fault feature vector processing to process the concatenated feature data and map it to the same low-dimensional feature space. Based on this, the system obtains the noise latent vector.

[0069] Furthermore, the system invokes a preset attention mechanism fusion module to fuse the obtained fault latent vectors and noise latent vectors. This module performs importance analysis on the features of each dimension of the fault and noise latent vectors, determining the weighting coefficients for each dimension. These coefficients are based on the distinguishing effect of each feature on the representation of real faults and aging noise. The system then weights the features of each dimension of the fault and noise latent vectors according to these coefficients, and fuses the two weighted latent vectors to obtain a fused latent vector. This fused latent vector comprehensively includes the core information of both fault and noise features, while also incorporating information related to the degree of physical degradation. The system inputs the fused latent vector into a preset fully connected layer for probability mapping. An activation function maps the fused latent vector into a vector data form of probability distribution, generating a discrimination probability vector for the current window. Each component in the discrimination probability vector represents the probability distribution of different types of electrical signals within the corresponding time period. The system ensures that the sum of the values ​​of each component is a constant, guaranteeing the integrity of the probability distribution and enabling the discrimination probability vector to clearly represent the probability proportion of different signal types.

[0070] S5: Obtain the original start signal of the relay protection device in the current window, and input the discrimination probability vector, the average physical degradation index and the original start signal into the preset rule engine for logical decision-making to generate the final decision result. The final decision result is used to indicate one of the following: allow protection action, block protection output, and issue maintenance warning.

[0071] Specifically, the system acquires the original start signal of the relay protection device within the current analysis window. This original start signal is the start judgment signal output by the relay protection device after analyzing the collected electrical quantity data based on traditional electrical quantity analysis criteria. The system integrates this original start signal with the previously generated discrimination probability vector and the average physical degradation index corresponding to the current window. These three types of data are used as input data for the preset rule engine, ensuring that the rule engine can obtain multi-dimensional judgment criteria. The rule engine internally has a set of logical decision rules built based on power system operation specifications and relay protection device operation and maintenance requirements. Each rule in the rule set includes preconditions and corresponding decision actions. Preconditions are logical combination judgment conditions of the input data, and decision actions are execution instructions corresponding to the preconditions. All rule preconditions have clear logical judgment criteria with no overlapping or ambiguous judgment boundaries.

[0072] The system inputs the integrated input data into the rule engine. The rule engine matches each rule in the input data according to a preset forward reasoning mechanism. Starting from the feature judgment of the input data, the forward reasoning mechanism sequentially verifies whether the preconditions in the rule set are met until a rule that meets all preconditions is found, ensuring the uniqueness and accuracy of the rule matching. After completing the rule matching, the rule engine generates a final decision result based on the decision action corresponding to the matched rule. The final decision result is divided into three types: allowing protection action, blocking protection output, and issuing maintenance warning. The decision result allowing protection action instructs the relay protection device to perform fault isolation action according to the preset protection logic. The decision result blocking protection output instructs the relay protection device to stop performing protection action. The decision result issuing maintenance warning instructs the operation and maintenance monitoring system to issue operation and maintenance prompts for the equipment interface.

[0073] The system transmits the final decision result to the corresponding execution unit according to the preset communication protocol. The execution unit is either a relay protection device or an operation and maintenance monitoring system. The execution unit performs the corresponding operation according to the received final decision result. Through this process, the system realizes intelligent identification and control of abnormal behavior of relay protection devices, and combines the action judgment of relay protection devices with the physical state of the equipment interface and the characteristics of electrical quantity signals to ensure the safe and stable operation of the power system.

[0074] In summary, the intelligent identification method for abnormal behavior of relay protection devices provided in this application achieves the goal of fusing multi-source data from the physical and information layers by aligning the collected electrical quantity data and physical state quantity data reflecting the contact state of the equipment interface in a time-frequency manner, thus providing a unified data foundation for subsequent cross-domain correlation analysis. By extracting contact resistance statistical features, temperature-resistance correlation features, and vibration amplitude features from the physical state quantity data through sliding window analysis and performing weighted fusion, a quantitative assessment of the degradation degree of the equipment interface contact state can be achieved, generating a physical degradation index sequence aligned with the electrical quantity data in time to accurately reflect the physical health status of the primary equipment interface. Based on the physical degradation index sequence, wavelet packet decomposition and reconstruction processing of the electrical quantity data allows for dynamic adjustment of the denoising threshold according to the actual degradation degree of the equipment interface, accurately decoupling the original electrical quantity signal into fault information. The characteristic path signal and noise characteristic path signal can effectively separate real fault features from hardware aging noise, avoiding excessive suppression of fault transient components by traditional fixed threshold denoising methods. By inputting the fault feature vector and the noise feature vector fused with the mean of the physical degradation index into the Siamese network for feature dimensionality reduction, and performing cross-domain fusion through an attention mechanism, deep interaction between physical domain features and information domain features can be achieved to construct a discrimination probability vector that can distinguish between hardware noise and real faults. Combined with the original start signal input rule engine for logical decision-making, the final decision results such as allowing protection actions, blocking protection outputs, or issuing maintenance warnings can be generated based on different discrimination results. This can effectively avoid protection malfunctions caused by the degradation of equipment interface contact conditions, and at the same time provide precise maintenance guidance for operation and maintenance personnel, thereby significantly improving the operational reliability of relay protection devices under complex operating conditions.

[0075] In one embodiment, such as Figure 2 As shown, S2 of the intelligent identification method for abnormal behavior of a relay protection device provided by the present invention specifically includes the following steps:

[0076] S21: Based on the sliding window technique, the contact resistance data in the multidimensional time series matrix is ​​segmented and truncated. The arithmetic mean and standard deviation of the contact resistance in each window are calculated to generate contact resistance statistical features. The contact resistance statistical features are used to characterize the steady-state level of the interface contact state.

[0077] Specifically, the system uses a sliding window technique to segment and extract contact resistance data from a multi-dimensional time-series matrix. The sliding window extracts continuous contact resistance data in an orderly manner according to a preset progression, covering all contact resistance data in the multi-dimensional time-series matrix to ensure no data omissions and no duplicate extractions. The system then extracts all contact resistance data within each segmented window and calculates its arithmetic mean using a specific formula:

[0078]

[0079] in, This represents the arithmetic mean of the contact resistance of the w-th window. This represents the total number of contact resistance data within this window. This represents the i-th contact resistance data within the w-th window. The system simultaneously calculates the standard deviation of the contact resistance within each window, using the following formula:

[0080]

[0081] in, This represents the standard deviation of the contact resistance for the w-th window. The system combines the arithmetic mean and standard deviation of the contact resistance for each window to generate a statistical feature of the contact resistance, which is used to characterize the steady-state level of the interface contact state.

[0082] S22: Perform correlation analysis on temperature data and contact resistance data in the multidimensional time series matrix, calculate the Pearson correlation coefficient between temperature change and contact resistance change within the window, and generate temperature-resistance correlation features. Temperature-resistance correlation features are used to distinguish between physical loosening caused by thermal expansion and contraction and fault arcing.

[0083] Specifically, the system performs correlation analysis on temperature and contact resistance data in the multidimensional time-series matrix. The analysis is conducted within the same window for each completed contact resistance data segment, ensuring complete consistency in the time range of the temperature and contact resistance data segments, thus achieving precise matching of the two types of data in the time dimension. The system calculates the temperature and contact resistance changes at each acquisition time within each window, and calculates the Pearson correlation coefficient based on these two types of changes. The calculation formula is as follows:

[0084]

[0085] in, The Pearson correlation coefficient represents the w-th window. This represents the covariance between the temperature change and the contact resistance change within the window. This represents the standard deviation of the temperature change within the window. This represents the standard deviation of the change in contact resistance within the window. The system uses the Pearson correlation coefficient calculated for each window as a temperature-resistance correlation feature, which is used to distinguish between physical loosening caused by thermal expansion and contraction and fault arcing.

[0086] S23: Perform root mean square value calculation on the vibration data in the multidimensional time series matrix, calculate the square root of the mean of the sum of squares of vibration amplitudes within the window, and generate vibration amplitude features. Vibration amplitude features are used to quantify the severity of mechanical loosening.

[0087] Specifically, the system performs root mean square (RMS) calculation on the vibration data in the multidimensional time-series matrix. The calculation is performed within a window whose intercept range matches that of the contact resistance and temperature data, ensuring the consistency of the vibration data's time range with other physical state data and guaranteeing the temporal matching of each feature analysis. The system extracts vibration amplitude data from all acquisition times within each window, squares each amplitude data point, sums all the squared results, divides the sum by the number of vibration amplitude data points within the window to obtain the mean, and finally takes the square root of the mean to obtain the RMS value. The calculation follows the formula:

[0088]

[0089] in, This represents the root mean square value of the vibration amplitude in the w-th window. This represents the total number of vibration amplitude data within the window. This represents the i-th vibration amplitude data within the w-th window. The system uses the root mean square value of the vibration amplitude in each window as a vibration amplitude feature, which is used to quantify the severity of mechanical loosening.

[0090] S24: Based on preset weighting coefficients, the contact resistance statistical characteristics, temperature-resistance correlation characteristics and vibration amplitude characteristics are linearly weighted and fused to generate the physical degradation index for each window.

[0091] Specifically, the system performs linear weighted fusion of the contact resistance statistical characteristics, temperature-resistance correlation characteristics, and vibration amplitude characteristics corresponding to each window based on preset weight coefficients to generate a physical degradation index for each window. The fusion calculation follows a preset formula, and the formula for calculating the physical degradation index is as follows:

[0092]

[0093] in, Let be the physical degradation index of the w-th window. This represents the average contact resistance within the window. The standard deviation of contact resistance within the window. The Pearson correlation coefficient is the relationship between the temperature change and the contact resistance change within the window. This represents the root mean square value of the vibration amplitude within the window. This is the preset reference contact resistance. The preset reference vibration amplitude, , , , These are the preset weighting coefficients. The system first normalizes each feature, and then performs linear weighted fusion according to this formula to obtain the physical degradation index of each window.

[0094] S25: A zero-order hold is used to extend the physical degradation index of each window along the time axis, extending the discrete window index values ​​into a continuous sequence with the same sampling rate as the electrical quantity data, thereby generating a physical degradation index sequence that is time-aligned with the electrical quantity data.

[0095] Specifically, the system uses a zero-order hold to extend the physical degradation index calculated for each window along the time axis. The system first obtains the sampling rate of electrical quantity data in the multi-dimensional time series matrix, clarifying the acquisition nodes and intervals of the electrical quantity data, using this as the benchmark for time axis extension. The system associates the physical degradation index of each window with its corresponding time range, determining the time interval for each discrete index value. The zero-order hold extends the physical degradation index of each window to all electrical quantity acquisition nodes within its corresponding time interval according to preset rules. The extension process follows the formula:

[0096]

[0097] in, This represents the physical degradation index value corresponding to time t. This represents the physical degradation index value of the w-th window. This represents the time interval corresponding to the w-th window. The system uses this method to extend the discrete window index values ​​into a continuous sequence with the same sampling rate as the electrical quantity data, ultimately generating a physical degradation index sequence that is time-aligned with the electrical quantity data.

[0098] In one embodiment, S3 of the intelligent identification method for abnormal behavior of a relay protection device provided by the present invention specifically includes the following steps:

[0099] S31: Based on the electrical quantity data in the multi-dimensional time series matrix, wavelet packet decomposition is performed on the electrical quantity signals of each channel. The electrical quantity signals are then decomposed into four levels using the db4 wavelet basis to generate wavelet coefficients for each frequency band node.

[0100] Specifically, the system performs wavelet packet decomposition on the electrical quantity signals of each channel based on the electrical quantity data in the multi-dimensional time-series matrix. The system first determines that the wavelet basis used for wavelet packet decomposition is the db4 wavelet basis, which is used to achieve frequency band decomposition of the electrical quantity signals. The system sets the wavelet packet decomposition level to 4 levels, performing 4-level wavelet packet decomposition operations on the electrical quantity signals of each channel, decomposing the electrical quantity signal of each channel into multiple sub-band signals of different frequency bands, and simultaneously generating wavelet coefficients corresponding to each frequency band node. The wavelet packet decomposition process follows the formula below:

[0101]

[0102] in, Represents the original electrical quantity signals of each channel. Represents the initial scale of the decomposition. Represents the total number of decomposition levels. Represents the low-frequency component coefficients. Represents the scaling function. Wavelet coefficients representing nodes in each frequency band, This represents the wavelet function. Through this decomposition process, the system separates the different frequency characteristics of the original electrical signal into nodes of each frequency band, and the generated wavelet coefficients are used for subsequent thresholding and signal reconstruction.

[0103] S32: Dynamically adapt the wavelet packet denoising threshold of each channel's electrical quantity signal according to the physical degradation index sequence, and adaptively modulate the reference threshold through the physical degradation index to generate a dynamic denoising threshold that changes with the degree of equipment degradation.

[0104] Specifically, the system dynamically adapts the wavelet packet denoising threshold of the electrical quantity signals in each channel based on the physical degradation index sequence. The system first obtains a preset baseline threshold for each channel, which serves as the basic threshold for denoising the electrical quantity signals in each channel. The system extracts the physical degradation index value corresponding to each moment in the physical degradation index sequence and uses this value as the basis for dynamic adjustment. It then adaptively modulates the baseline threshold for each channel, enabling the denoising threshold to dynamically adjust according to changes in the physical degradation degree of the device interface, generating dynamic denoising thresholds for each channel at different moments. Preferably, the calculation formula for the dynamic denoising threshold is:

[0105]

[0106] in, Let be the dynamic denoising threshold for the m-th channel at time t. Let be the baseline threshold for the m-th channel. The preset adjustment coefficient, Let t be the physical degradation index at time t. The system uses this formula to correlate the baseline threshold with the physical degradation index, enabling the dynamic denoising threshold to match the noise characteristics of the equipment under different degradation states, thus improving denoising accuracy.

[0107] S33: Based on the dynamic denoising threshold, the wavelet coefficients are subjected to threshold screening and reconstruction. The low-frequency nodes and high-frequency fault feature frequency band nodes in the fault feature set are reconstructed after soft threshold shrinkage to generate fault feature path signals. The high-frequency noise-dominant frequency band nodes are subjected to hard threshold truncation, and the part exceeding the dynamic threshold is retained for reconstruction to generate noise feature path signals.

[0108] Specifically, based on a dynamic denoising threshold, the system performs threshold filtering and reconstruction processing on the wavelet coefficients generated by decomposing the electrical quantity signals of each channel. The system first classifies the wavelet coefficients of each frequency band node, grouping low-frequency nodes and high-frequency fault characteristic frequency band nodes in the fault feature set into one category, and high-frequency noise-dominated frequency band nodes into another. The system then performs soft thresholding on the wavelet coefficients of the low-frequency nodes and high-frequency fault characteristic frequency band nodes in the fault feature set. The soft thresholding process follows the formula:

[0109]

[0110] in, Represents the processed wavelet coefficients. Represents the original wavelet coefficients. This represents the dynamic denoising threshold for the corresponding channel. This represents the positive operation. After processing, the wavelet coefficients of this type are reconstructed to generate fault feature path signals. The system performs hard threshold truncation on the wavelet coefficients of nodes in the high-frequency noise-dominated frequency band, retaining the wavelet coefficients exceeding the dynamic denoising threshold. The retained wavelet coefficients are then reconstructed to generate noise feature path signals, achieving effective separation of fault features and noise features.

[0111] S34: Using a preset power frequency cycle as the analysis window, perform time-frequency domain statistical feature extraction processing on the fault characteristic path signal and the noise characteristic path signal, calculate the kurtosis and waveform entropy of the fault characteristic path signal and the noise characteristic path signal respectively, and generate fault characteristic parameters and noise characteristic parameters for each channel.

[0112] Specifically, the system uses a preset power frequency cycle as the analysis window and performs time-frequency domain statistical feature extraction processing on the generated fault characteristic path signals and noise characteristic path signals respectively. The system first determines the analysis window as the preset power frequency cycle to ensure that each analysis window can completely cover the signal features within one power frequency cycle. The system calculates kurtosis and waveform entropy for the fault characteristic path signals within each analysis window, and simultaneously performs the same kurtosis and waveform entropy calculations for the noise characteristic path signals within the same window. Preferably, the kurtosis calculation formula is:

[0113]

[0114] in, Represents kurtosis, This represents the amount of signal data within the window. Represents signal data, This represents the signal mean. The formula for calculating waveform entropy is:

[0115]

[0116] in, Represents waveform entropy. Represents the number of signal segments. This represents the energy proportion of each signal segment. The system calculates the kurtosis and waveform entropy of the fault characteristic path signal, which are used as fault characteristic parameters for each channel; and calculates the kurtosis and waveform entropy of the noise characteristic path signal, which are used as noise characteristic parameters for each channel.

[0117] S35: Perform feature combination processing on the fault feature parameters and noise feature parameters of all channels, and concatenate the kurtosis and entropy of each channel into a feature vector in sequence to generate the global fault feature vector and global noise feature vector of the current window.

[0118] Specifically, the system performs feature combination processing on the fault characteristic parameters and noise characteristic parameters of all channels separately. First, the system acquires the fault characteristic parameters corresponding to each channel, including the kurtosis and waveform entropy of the fault characteristic path signal for that channel. Simultaneously, it acquires the noise characteristic parameters corresponding to each channel, including the kurtosis and waveform entropy of the noise characteristic path signal for that channel. The system sequentially concatenates the fault characteristic parameters of all channels according to a preset channel order, arranging the kurtosis and waveform entropy of each channel in sequence to form a feature vector. This feature vector is the global fault characteristic vector for the current analysis window, comprehensively reflecting the fault characteristic information of all channels. Similarly, the system sequentially concatenates the noise characteristic parameters of all channels according to a preset channel order, arranging the kurtosis and waveform entropy of each channel in sequence to form a feature vector. This feature vector is the global noise characteristic vector for the current analysis window, comprehensively reflecting the noise characteristic information of all channels.

[0119] In one embodiment, S4 of the intelligent identification method for abnormal behavior of a relay protection device provided by the present invention specifically includes the following steps:

[0120] S41: Extract the index value corresponding to the current analysis window from the physical degradation index sequence, average the physical degradation index of all sampling points in the window, and generate the average physical degradation index of the current window.

[0121] Specifically, the system extracts all physical degradation index values ​​corresponding to the current analysis window from the physical degradation index sequence. The extraction process strictly matches the time range of the current analysis window to ensure that all extracted index values ​​belong to the sampling points within the current window, with no index values ​​exceeding the window range and no index values ​​within the window being missed. The system then averages all extracted physical degradation index values, following the mean calculation formula:

[0122]

[0123] in, This represents the mean physical degradation index of the current w-th analysis window. This represents the total number of sampling points within the current analysis window. This represents the physical degradation index value corresponding to the t-th sampling point within the current analysis window. Through this averaging process, the system transforms the discrete physical degradation index values ​​within the current window into a single quantitative indicator, which comprehensively reflects the overall physical degradation degree of the device interface within the current window.

[0124] S42: Input the fault feature vector into the preset fault encoder for nonlinear dimensionality reduction processing, extract deep features through a three-layer fully connected network, and generate a fault latent vector.

[0125] Specifically, the system inputs the generated global fault feature vector into a preset fault encoder for nonlinear dimensionality reduction. The fault encoder consists of a three-layer fully connected network, with each layer sequentially performing linear transformations and nonlinear mappings of the features. A nonlinear activation function maps the high-dimensional fault feature vector to a low-dimensional feature space, extracting deep core information from the fault features, thus achieving dimensionality reduction and refinement of the fault features. The nonlinear dimensionality reduction process follows the formula:

[0126]

[0127] in, Represents the generated latent fault vector. This represents the input global fault feature vector. These represent the weight matrices of a three-layer fully connected network. These represent the bias terms of a three-layer fully connected network. This represents a nonlinear activation function. Through this process, the system generates a low-dimensional fault latent vector that retains the core information of the fault features while reducing the computational complexity of subsequent feature fusion and discrimination.

[0128] S43: The noise feature vector is concatenated with the mean of the physical degradation index to form a fused feature, which is then input into a preset noise encoder for nonlinear dimensionality reduction. Deep features are extracted through a three-layer fully connected network to generate a noise latent vector.

[0129] Specifically, the system first obtains the average physical degradation index of the current analysis window, uses this average as a new feature dimension, and concatenates it with the global noise feature vector. The concatenation process adds the average physical degradation index to the end of the global noise feature vector according to a preset order, forming a fused feature containing both noise features and physical degradation information. The system inputs this fused feature into a preset noise encoder for nonlinear dimensionality reduction. The noise encoder has the same structure as the fault encoder, both consisting of a three-layer fully connected network. Through linear transformations and nonlinear mappings of each layer, deep core information in the fused feature is extracted, achieving dimensionality reduction and purification of the fused feature. Preferably, the nonlinear dimensionality reduction process follows the formula:

[0130]

[0131] in, Represents the generated noise latent vector. This represents the fusion characteristics after splicing. These represent the weight matrices of the three fully connected layers of the noise encoder network. These represent the corresponding bias terms. This represents a nonlinear activation function. Through this process, the system generates a low-dimensional noise latent vector that integrates noise features with physical degradation information.

[0132] S44: Perform multi-head attention fusion processing on the fault latent vector and the noise latent vector, calculate the attention weight between the two latent vectors, and concatenate the weighted features to generate a fused feature vector.

[0133] Specifically, the system performs multi-head attention fusion processing on the generated fault latent vector and noise latent vector. First, it performs feature mapping on the two latent vectors through a preset attention head, mapping the two latent vectors to multiple attention subspaces respectively. Then, within each attention subspace, it calculates the attention weight between the fault latent vector and the noise latent vector. The weight calculation follows the formula:

[0134]

[0135] in, Let represent the attention weights between the j-th dimension of the fault latent vector and the k-th dimension of the noise latent vector under the i-th attention head. The mapping feature representing the fault latent vector in the i-th attention subspace. The mapping feature representing the noise latent vector in the i-th attention subspace. The system represents the dimension of the latent vector. Based on the calculated attention weights, the system weights the features of each dimension of the two latent vectors, and then concatenates the weighted features of all attention heads to generate a fused feature vector. This vector integrates the core information of the fault latent vector and the noise latent vector, improving the accuracy of subsequent discrimination probability calculations.

[0136] S45: The fused feature vector input and output layers are processed by softmax classification, and mapped to a two-dimensional probability space through a fully connected network to generate the discrimination probability vector of the current window. The discrimination probability vector contains the true fault probability and the hardware noise anomaly probability.

[0137] Specifically, the system inputs the fused feature vector generated by multi-head attention fusion into a preset output layer for softmax classification. The output layer consists of a fully connected network, which maps the fused feature vector to a two-dimensional probability space. This two-dimensional space corresponds to two types: real faults and hardware noise anomalies. The softmax function then transforms the mapping result into a probability distribution, generating the discrimination probability vector for the current window. Preferably, the softmax classification process follows the formula:

[0138]

[0139] in, Represents the discriminant probability vector. Represents the fused feature vector. The weight matrix represents the weights of the fully connected output layer. Represents the bias term of the output layer. These represent the row vectors corresponding to the two probability dimensions of the weight matrix. These represent the bias values ​​for the corresponding dimensions. The discrimination probability vector contains two components: the true failure probability and the hardware noise anomaly probability.

[0140] In one embodiment, S5 of the intelligent identification method for abnormal behavior of a relay protection device provided by the present invention specifically includes the following steps:

[0141] S51: Window the original start signal of the relay protection device within the current analysis window, aggregate the start status of any sampling time within the window into a Boolean value, and generate the start status identifier of the current window.

[0142] Specifically, the system acquires the original start-up signals of the relay protection devices within the current analysis window. These original start-up signals are binary start-up status data for each sampling point within the corresponding window, reflecting the start-up status of the relay protection devices at each sampling time. The system performs windowing processing on the acquired original start-up signals, iterates through the original start-up signals of all sampling points within the current analysis window, and performs aggregation calculations on the start-up status of all sampling points. The aggregation calculation follows the formula:

[0143]

[0144] in, This represents the startup status indicator of the current w-th analysis window. This represents the total number of sampling points within the current analysis window. This represents the original start signal value of the t-th sampling point within the current window. Through this aggregation operation, the system integrates the start states of multiple discrete sampling points within the window into a unified Boolean value. If the original start signal of any sampling point within the window is 1, the aggregated start state identifier is 1; if the original start signals of all sampling points within the window are 0, the aggregated start state identifier is 0. This process generates the start state identifier for the current window, achieving unified quantification of start signals within the window and providing a standardized basis for the logical matching of the subsequent rule engine.

[0145] S52: Compare the true fault probability in the discrimination probability vector with the preset true fault judgment threshold, compare the hardware noise abnormal probability in the discrimination probability vector with the preset hardware noise judgment threshold, and compare the mean physical degradation index with the preset degradation exceeding judgment threshold to generate a comparison result vector containing the three comparison results.

[0146] Specifically, the system acquires three preset judgment thresholds: a true fault judgment threshold, a hardware noise judgment threshold, and a degradation exceeding the standard judgment threshold. These three thresholds serve as the baseline parameters for the rule engine to perform logical judgments. The system extracts the discrimination probability vector of the current window, separating the true fault probability and the hardware noise anomaly probability from the vector. Simultaneously, it extracts the mean physical degradation index of the current window and performs threshold comparison calculations for each. The system compares the true fault probability with the true fault judgment threshold, the hardware noise anomaly probability with the hardware noise judgment threshold, and the mean physical degradation index with the degradation exceeding the standard judgment threshold. The comparison results are represented in binary form, and a comparison result vector is constructed. The construction process follows the formula below:

[0147]

[0148] in, Represents the comparison result vector. This is a comparison between the actual failure probability and the actual failure judgment threshold. This is a comparison of the probability of hardware noise anomalies with the hardware noise judgment threshold. The comparison result is the result of comparing the mean of the physical degradation index with the threshold for determining excessive degradation. A comparison result of 1 indicates that the corresponding probability or mean exceeds the threshold, while a result of 0 indicates that the threshold is not exceeded. Based on this, the system generates a comparison result vector containing the three comparison results.

[0149] S53: Input the comparison result vector and the start status flag into the preset rule engine for logical matching. If the actual fault probability exceeds the actual fault judgment threshold and the start status flag is 1, then match the protection action rule and generate the protection action instruction. If the hardware noise abnormal probability exceeds the hardware noise judgment threshold and the average physical degradation index exceeds the degradation exceeding the judgment threshold, then match the lockout protection and early warning rule, generate the lockout protection output and issue the maintenance early warning instruction. If neither the actual fault probability nor the hardware noise abnormal probability exceeds their respective thresholds and is within the preset fuzzy range, then match the edge state verification rule and generate the extended data window verification instruction.

[0150] Specifically, the system inputs the generated comparison result vector and the start status identifier into a preset rule engine. The rule engine has three preset core logic rules: rules for allowing protection actions, rules for locking protection and issuing warnings, and rules for edge state verification. The rule engine performs logical matching processing sequentially according to preset priorities. The system first extracts the... and startup status identifier ,like and If the actual fault probability exceeds the actual fault determination threshold and the activation status flag is 1, the rule engine matches the rules for allowing protection actions and generates a command to allow protection actions. This command is used to instruct the relay protection device to perform a trip output operation. If the above conditions are not met, the system extracts... and ,like and If the probability of hardware noise anomalies exceeds the hardware noise judgment threshold and the average physical degradation index exceeds the degradation exceedance judgment threshold, the rule engine matches the interlocking protection and early warning rules, generates an interlocking protection exit, and issues a maintenance early warning instruction. This maintenance early warning instruction is used to instruct the monitoring system to push a maintenance notification containing noise characteristic path signals and contact resistance data. If neither of the above two conditions is met, the system determines whether the actual fault probability and the probability of hardware noise anomalies do not exceed their respective thresholds and are within a preset fuzzy range. If they are met, the system matches the edge state verification rule and generates an extended data window verification instruction. This instruction is used to instruct the system to expand the current analysis window to two power frequency cycles and re-execute steps S3 and S4.

[0151] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0152] Based on the same inventive concept, this application also provides an intelligent identification system for abnormal behavior of relay protection devices, used to implement the intelligent identification method for abnormal behavior of relay protection devices described above. The solution provided by this system is similar to the implementation scheme described in the above method. Therefore, the specific limitations of one or more embodiments of the intelligent identification system for abnormal behavior of relay protection devices provided below can be found in the limitations of the intelligent identification method for abnormal behavior of relay protection devices described above, and will not be repeated here.

[0153] Preferably, such as Figure 3 As shown, the present invention provides an intelligent identification system 600 for abnormal behavior of relay protection devices, which is configured with the following modules:

[0154] The multi-source time-frequency alignment module 610 includes time-frequency alignment of the collected electrical quantity data of the relay protection device access point and the physical state quantity data reflecting the contact status of the equipment interface, generating a multi-dimensional time series matrix containing electrical quantity information and physical state quantity information;

[0155] The physical degradation index construction module 620 includes extracting contact resistance statistical features, temperature-resistance correlation features and vibration amplitude features from the physical state quantity data in the multidimensional time series matrix based on sliding window analysis, performing weighted fusion, outputting the physical degradation index of each window and expanding it into a physical degradation index sequence aligned with the time of the electrical quantity data;

[0156] The electrical quantity time-frequency feature extraction module 630 includes wavelet packet decomposition and reconstruction processing of electrical quantity data in a multi-dimensional time-series matrix based on the physical degradation index sequence, decoupling the original electrical quantity signal into fault feature path signal and noise feature path signal, extracting the time-frequency domain statistical features of the fault feature path signal and noise feature path signal within the current analysis window, and generating fault feature vector and noise feature vector;

[0157] The Siamese network attention fusion module 640 includes inputting the fault feature vector into a preset Siamese network for feature dimensionality reduction to obtain the fault latent vector, concatenating the noise feature vector and the mean of the physical degradation index corresponding to the current window and inputting them into a preset Siamese network for feature dimensionality reduction to obtain the noise latent vector, and performing attention mechanism fusion processing on the fault latent vector and the noise latent vector to generate the discrimination probability vector of the current window.

[0158] The rule engine logic decision module 650 includes acquiring the original start signal of the relay protection device in the current window, and inputting the discrimination probability vector, the average physical degradation index and the original start signal into the preset rule engine for logic decision-making to generate the final decision result. The final decision result is used to indicate one of the following: allow protection action, block protection output, and issue maintenance warning.

[0159] Preferably, the physical degradation index construction module 620 provided in this application is configured with the following units:

[0160] The contact resistance statistics unit is used to segment and extract contact resistance data in a multidimensional time series matrix based on the sliding window technique, calculate the arithmetic mean and standard deviation of the contact resistance in each window, and generate contact resistance statistical features. The contact resistance statistical features are used to characterize the steady-state level of the interface contact state.

[0161] The temperature-resistance correlation calculation unit is used to perform correlation analysis on temperature data and contact resistance data in a multidimensional time series matrix. It calculates the Pearson correlation coefficient between temperature change and contact resistance change within the window and generates temperature-resistance correlation features. These features are used to distinguish between physical loosening caused by thermal expansion and contraction and fault arcing.

[0162] The vibration amplitude feature unit is used to calculate the root mean square value of vibration data in a multidimensional time series matrix. It calculates the square root of the mean of the sum of squares of vibration amplitudes within the window and generates vibration amplitude features. Vibration amplitude features are used to quantify the severity of mechanical loosening.

[0163] The physical degradation index fusion unit is used to linearly weight and fuse contact resistance statistical characteristics, temperature-resistance correlation characteristics and vibration amplitude characteristics based on preset weight coefficients to generate physical degradation indices for each window.

[0164] The time axis extension alignment unit is used to extend the physical degradation index of each window using a zero-order hold, extending the discrete window index values ​​into a continuous sequence with the same sampling rate as the electrical quantity data, and generating a physical degradation index sequence that is time-aligned with the electrical quantity data.

[0165] Preferably, the electrical quantity time-frequency feature extraction module 630 provided in this application is configured with the following units:

[0166] The wavelet packet decomposition unit is used to perform wavelet packet decomposition on the electrical quantity signals of each channel based on the electrical quantity data in the multi-dimensional time-series matrix. The electrical quantity signals are decomposed into 4 levels through the db4 wavelet basis to generate wavelet coefficients for each frequency band node.

[0167] The dynamic threshold adaptation unit is used to dynamically adapt the wavelet packet denoising threshold of each channel's electrical quantity signal according to the physical degradation index sequence. It adaptively modulates the reference threshold through the physical degradation index to generate a dynamic denoising threshold that changes with the degree of equipment degradation.

[0168] The dual-path signal reconstruction unit is used to perform threshold screening and reconstruction processing on wavelet coefficients based on dynamic denoising threshold. It performs soft threshold shrinkage processing on low-frequency nodes and high-frequency fault feature frequency band nodes in the fault feature set and then reconstructs them to generate fault feature path signals. It performs hard threshold truncation processing on high-frequency noise-dominant frequency band nodes, retains the part exceeding the dynamic threshold for reconstruction, and generates noise feature path signals.

[0169] The time-frequency feature extraction unit is used to perform time-frequency domain statistical feature extraction processing on fault feature path signals and noise feature path signals with a preset power frequency period as the analysis window, calculate the kurtosis and waveform entropy of fault feature path signals and noise feature path signals respectively, and generate fault feature parameters and noise feature parameters for each channel.

[0170] The global feature combination unit is used to perform feature combination processing on the fault feature parameters and noise feature parameters of all channels, and concatenates the kurtosis and entropy of each channel into a feature vector in sequence to generate the global fault feature vector and global noise feature vector of the current window.

[0171] Preferably, the twin network attention fusion module 640 provided in this application is configured with the following units:

[0172] The degradation index mean calculation unit is used to extract the index value corresponding to the current analysis window from the physical degradation index sequence, average the physical degradation index of all sampling points in the window, and generate the physical degradation index mean of the current window.

[0173] The fault feature encoding dimensionality reduction unit is used to input the fault feature vector into the preset fault encoder for nonlinear dimensionality reduction processing, and extract deep features through a three-layer fully connected network to generate the fault latent vector.

[0174] The noise degradation fusion coding unit is used to concatenate the noise feature vector with the mean of the physical degradation index to form a fusion feature, which is then input into a preset noise encoder for nonlinear dimensionality reduction. Deep features are extracted through a three-layer fully connected network to generate a noise latent vector.

[0175] The latent vector attention fusion unit is used to perform multi-head attention fusion processing on fault latent vectors and noise latent vectors, calculate the attention weights between the two latent vectors and concatenate the weighted features to generate a fused feature vector;

[0176] The probability discrimination output unit is used to process the fused feature vector into the input and output layers using softmax classification. It is then mapped to a two-dimensional probability space through a fully connected network to generate the discrimination probability vector for the current window. The discrimination probability vector contains the true fault probability and the hardware noise anomaly probability.

[0177] Preferably, the rule engine logic decision module 650 provided in this application is configured with the following units:

[0178] The startup status aggregation unit is used to perform windowing processing on the original startup signals of the relay protection device within the current analysis window, aggregate the startup status at any sampling time within the window into a Boolean value, and generate the startup status identifier of the current window.

[0179] The multi-threshold comparison unit is used to compare the true fault probability in the discrimination probability vector with the preset true fault judgment threshold, compare the hardware noise abnormal probability in the discrimination probability vector with the preset hardware noise judgment threshold, and compare the average physical degradation index with the preset degradation exceeding judgment threshold, generating a comparison result vector containing the three comparison results.

[0180] The rule engine decision unit is used to input the comparison result vector and the start status flag into the preset rule engine for logical matching processing. If the actual fault probability exceeds the actual fault judgment threshold and the start status flag is 1, then the rule that allows protection action is matched and the instruction that allows protection action is generated. If the hardware noise abnormal probability exceeds the hardware noise judgment threshold and the average physical degradation index exceeds the degradation exceeding the judgment threshold, then the rule that locks out protection and issues an early warning is matched, the lockout protection exit is generated and the maintenance early warning instruction is issued. If neither the actual fault probability nor the hardware noise abnormal probability exceeds its respective threshold and is within the preset fuzzy range, then the rule that verifies the edge state is matched and the instruction that extends the data window is generated.

[0181] Among them, the protection action instruction is used to instruct the relay protection device to perform a trip output operation, the maintenance early warning instruction is used to instruct the monitoring system to push a maintenance notification form containing noise characteristic path signals and contact resistance data, and the data window extension verification instruction is used to instruct the system to extend the current analysis window to two power frequency cycles and re-execute the feature extraction and probability discrimination steps.

[0182] In one embodiment, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described intelligent identification method for abnormal behavior of relay protection devices.

[0183] In one embodiment, this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described intelligent identification method for abnormal behavior of relay protection devices.

[0184] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0185] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0186] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in this application, and these should all be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for intelligent identification of abnormal behavior of a relay protection device, characterized in that, Includes the following steps: S1: Perform time-frequency alignment on the collected electrical quantity data of the relay protection device access point and the physical state quantity data reflecting the contact status of the equipment interface to generate a multi-dimensional time series matrix containing electrical quantity information and physical state quantity information; S2: Based on sliding window analysis, extract the contact resistance statistical characteristics, temperature-resistance correlation characteristics and vibration amplitude characteristics of the physical state data in the multidimensional time series matrix, perform weighted fusion, output the physical degradation index of each window and expand it into a physical degradation index sequence aligned with the time of the electrical quantity data; S3: Based on the physical degradation index sequence, wavelet packet decomposition and reconstruction processing is performed on the electrical quantity data in the multidimensional time series matrix to decouple the original electrical quantity signal into fault feature path signal and noise feature path signal. The time-frequency domain statistical features of the fault feature path signal and the noise feature path signal in the current analysis window are extracted to generate fault feature vector and noise feature vector. S4: Input the fault feature vector into a preset Siamese network for feature dimensionality reduction to obtain the fault latent vector, and concatenate the noise feature vector and the mean of the physical degradation index corresponding to the current window into the preset Siamese network for feature dimensionality reduction to obtain the noise latent vector. Then, perform attention mechanism fusion processing on the fault latent vector and the noise latent vector to generate the discrimination probability vector of the current window. S5: Obtain the original start signal of the relay protection device in the current window, and input the discrimination probability vector, the average physical degradation index and the original start signal into the preset rule engine for logical decision-making to generate the final decision result. The final decision result is used to indicate one of the following: allow protection action, block protection output, and issue maintenance warning.

2. The method according to claim 1, characterized in that, S2 includes: S21: Based on the sliding window technique, the contact resistance data in the multidimensional time series matrix is ​​segmented and truncated. The arithmetic mean and standard deviation of the contact resistance in each window are calculated to generate contact resistance statistical features. The contact resistance statistical features are used to characterize the steady-state level of the interface contact state. S22: Perform correlation analysis on the temperature data and contact resistance data in the multidimensional time series matrix, calculate the Pearson correlation coefficient between the temperature change and the contact resistance change within the window, and generate temperature-resistance correlation features. The temperature-resistance correlation features are used to distinguish between physical loosening caused by thermal expansion and contraction and fault arc. S23: Perform root mean square value calculation on the vibration data in the multidimensional time series matrix, calculate the square root of the mean of the sum of squares of vibration amplitudes within the window, and generate vibration amplitude features. The vibration amplitude features are used to quantify the severity of mechanical loosening. S24: Based on preset weighting coefficients, the contact resistance statistical characteristics, the temperature-resistance correlation characteristics, and the vibration amplitude characteristics are linearly weighted and fused to generate the physical degradation index for each window; S25: The physical degradation index of each window is extended along the time axis using a zero-order hold, and the discrete window index values ​​are extended into a continuous sequence with the same sampling rate as the electrical quantity data, thereby generating a physical degradation index sequence that is time-aligned with the electrical quantity data.

3. The method according to claim 2, characterized in that, The formula for calculating the physical degradation index is as follows: in, Let be the physical degradation index of the w-th window. This represents the average contact resistance within the window. The standard deviation of contact resistance within the window. The Pearson correlation coefficient is the relationship between the temperature change and the contact resistance change within the window. This represents the root mean square value of the vibration amplitude within the window. This is the preset reference contact resistance. The preset reference vibration amplitude, , , , These are the preset weighting coefficients.

4. The method according to claim 1, characterized in that, S3 includes: S31: Based on the electrical quantity data in the multidimensional time series matrix, perform wavelet packet decomposition on the electrical quantity signals of each channel, and decompose the electrical quantity signals into 4 levels using the db4 wavelet basis to generate wavelet coefficients for each frequency band node. S32: Dynamically adapt the wavelet packet denoising threshold of each channel's electrical quantity signal according to the physical degradation index sequence, and adaptively modulate the reference threshold through the physical degradation index to generate a dynamic denoising threshold that changes with the degree of equipment degradation. S33: Based on the dynamic denoising threshold, the wavelet coefficients are subjected to threshold screening and reconstruction processing. The low-frequency nodes and high-frequency fault feature frequency band nodes in the fault feature set are reconstructed after soft threshold shrinkage processing to generate fault feature path signals. The high-frequency noise-dominant frequency band nodes are subjected to hard threshold truncation processing, and the part exceeding the dynamic threshold is retained for reconstruction to generate noise feature path signals. S34: Using a preset power frequency cycle as the analysis window, perform time-frequency domain statistical feature extraction processing on the fault characteristic path signal and the noise characteristic path signal, calculate the kurtosis and waveform entropy of the fault characteristic path signal and the noise characteristic path signal respectively, and generate fault characteristic parameters and noise characteristic parameters for each channel. S35: Perform feature combination processing on the fault feature parameters and noise feature parameters of all channels, and concatenate the kurtosis and entropy of each channel into a feature vector in sequence to generate the global fault feature vector and global noise feature vector of the current window.

5. The method according to claim 4, characterized in that, The formula for calculating the dynamic denoising threshold is: in, Let be the dynamic denoising threshold for the m-th channel at time t. Let be the baseline threshold for the m-th channel. The preset adjustment coefficient, Let be the physical degradation index at time t.

6. The method according to claim 1, characterized in that, S4 includes: S41: Extract the index value corresponding to the current analysis window from the physical degradation index sequence, average the physical degradation index of all sampling points in the window, and generate the average physical degradation index of the current window. S42: Input the fault feature vector into a preset fault encoder for nonlinear dimensionality reduction processing, extract deep features through a three-layer fully connected network, and generate a fault latent vector; S43: The noise feature vector is concatenated with the mean of the physical degradation index to form a fused feature, which is then input into a preset noise encoder for nonlinear dimensionality reduction. Deep features are extracted through a three-layer fully connected network to generate a noise latent vector. S44: Perform multi-head attention fusion processing on the fault latent vector and the noise latent vector, calculate the attention weight between the two latent vectors, and concatenate the weighted features to generate a fused feature vector; S45: The fused feature vector is input to the output layer and subjected to softmax classification. It is then mapped to a two-dimensional probability space through a fully connected network to generate a discrimination probability vector for the current window. The discrimination probability vector includes the true fault probability and the hardware noise anomaly probability.

7. The method according to any one of claims 1-6, characterized in that, The final decision result includes one of the following: a protection action command, a maintenance early warning command, and an extended data window verification command. S5 includes: S51: Window the original start signal of the relay protection device within the current analysis window, aggregate the start status at any sampling time within the window into a Boolean value, and generate the start status identifier of the current window. S52: Compare the true fault probability in the discrimination probability vector with the preset true fault judgment threshold, compare the hardware noise abnormal probability in the discrimination probability vector with the preset hardware noise judgment threshold, and compare the average physical degradation index with the preset degradation exceeding judgment threshold to generate a comparison result vector containing three comparison results. S53: Input the comparison result vector and the start-up status identifier into a preset rule engine for logical matching. If the actual fault probability exceeds the actual fault judgment threshold and the start-up status identifier is 1, then match the allow protection action rule and generate an allow protection action instruction. If the hardware noise abnormal probability exceeds the hardware noise judgment threshold and the average physical degradation index exceeds the degradation exceeding judgment threshold, then match the lockout protection and early warning rule, generate the lockout protection exit and issue a maintenance early warning instruction. If neither the actual fault probability nor the hardware noise abnormal probability exceeds their respective thresholds and is within a preset fuzzy range, then match the edge state verification rule and generate an extended data window verification instruction. The instruction to allow protection action is used to instruct the relay protection device to perform a trip output operation; the instruction to perform maintenance early warning is used to instruct the monitoring system to push a maintenance notification form containing noise characteristic path signals and contact resistance data; and the instruction to extend the data window for verification is used to instruct the system to extend the current analysis window to two power frequency cycles and re-execute steps S3 and S4.

8. An intelligent identification system for abnormal behavior of a relay protection device, characterized in that, The system includes: The multi-source time-frequency alignment module includes time-frequency alignment of the electrical quantity data collected from the relay protection device access point and the physical state quantity data reflecting the contact status of the equipment interface, generating a multi-dimensional time series matrix containing electrical quantity information and physical state quantity information; The physical degradation index construction module includes extracting contact resistance statistical features, temperature-resistance correlation features and vibration amplitude features from the physical state data in the multidimensional time series matrix based on sliding window analysis, performing weighted fusion, outputting the physical degradation index of each window and expanding it into a physical degradation index sequence aligned with the time of the electrical quantity data; The electrical quantity time-frequency feature extraction module includes wavelet packet decomposition and reconstruction processing of electrical quantity data in the multi-dimensional time series matrix based on the physical degradation index sequence, decoupling the original electrical quantity signal into fault feature path signal and noise feature path signal, extracting the time-frequency domain statistical features of the fault feature path signal and the noise feature path signal within the current analysis window, and generating fault feature vector and noise feature vector; The Siamese network attention fusion module includes inputting the fault feature vector into a preset Siamese network for feature dimensionality reduction to obtain a fault latent vector, concatenating the noise feature vector and the mean of the physical degradation index corresponding to the current window and inputting them into a preset Siamese network for feature dimensionality reduction to obtain a noise latent vector, and performing attention mechanism fusion processing on the fault latent vector and the noise latent vector to generate a discrimination probability vector for the current window. The rule engine logic decision module includes acquiring the original start signal of the relay protection device in the current window, and inputting the discrimination probability vector, the average physical degradation index and the original start signal into a preset rule engine for logic decision-making to generate a final decision result. The final decision result is used to indicate one of the following: allow protection action, block protection output, and issue maintenance warning.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images