A method for discriminating and alarming multi-source signals of a drop-type lightning arrester

Abstract

The application discloses a kind of drop lightning arrester multi-source signal discrimination and alarm method, it is related to the on-line monitoring and fault early warning technical field of power system equipment, the application can effectively separate high-frequency sub-band signal by wavelet packet decomposition technology, adapt to non-stationary vibration environment, ensure the robustness of feature extraction;The introduction of approximate entropy and peak factor and other characteristics, quantifies the randomness and impact of signal, can sensitively capture the subtle changes in the fatigue accumulation process of lead wire;Single-class classification model trained only by healthy sample is used, such as support vector data description, decision boundary is constructed, so that the model has high sensitivity to abnormal state, and can early warning before lead wire fracture occurs;This kind of evaluation mode based on healthy baseline avoids the false alarm problem of traditional threshold method in complex environment.

Description

Technical Field

[0001] This invention relates to the field of online monitoring and fault early warning technology for power system equipment, and in particular to a method for multi-source signal discrimination and alarm of drop-out surge arresters. Background Technology

[0002] In coastal power grid scenarios where typhoons are frequent, the connecting leads of line-type / drop-out surge arresters are subjected to non-steady vibrations and impact loads induced by strong winds for a long time, which can easily lead to metal fatigue accumulation. This process occurs when the electrical and dielectric parameters of the arrester are basically normal. Traditional online monitoring, which mainly relies on leakage current / harmonics, discharge count, and temperature and humidity, cannot directly reflect the mechanical degradation of the leads.

[0003] Field experience and research indicate that mechanical problems with the lead wire / detacher are common weaknesses in TLSA applications. Wind-induced swaying and vortex-induced vibration can both lead to fatigue and fracture, while current standards pay limited attention to lead wire mechanical testing. Recent information fusion solutions (including cross-phase comparison and UAV inspection) can improve the completeness of vessel condition assessment, but still lack real-time perception of lead wire vibration characteristics. Methods such as FRA can be used to identify early vessel degradation, but usually require controlled conditions and are difficult to use as a continuous online method during typhoon season. Drop alarms based on position switches are only triggered after fracture / drop, limiting their early warning capabilities. Therefore, a monitoring and early warning mechanism that can characterize lead wire fatigue risk online in strong wind and non-stationary vibration environments still needs improvement. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] This invention provides a multi-source signal discrimination and alarm method for drop-out surge arresters, which solves the problem that surge arrester leads are prone to fatigue fracture due to strong wind vibration in coastal areas during typhoons, and that traditional monitoring lacks real-time mechanical condition perception and early warning mechanisms.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a method for multi-source signal discrimination and alarm of a drop-out surge arrester, comprising: Step S1: Collect the leakage current signal of the surge arrester and perform preprocessing; Step S2: Extract frequency domain or time-frequency domain features related to the mechanical vibration of the connecting lead from the leakage current signal; Step S3: Based on the single-class classification model trained only on healthy samples, evaluate the fatigue state of the connection leads of the features and output the deviation. Step S4: When the deviation exceeds the threshold, a risk warning is triggered.

[0007] As a preferred embodiment of the multi-source signal discrimination and alarm method for a drop-out surge arrester described in this invention, step S2 includes: Wavelet packet decomposition is performed on the leakage current signal, and high-frequency subbands are separated using a multi-level decomposition tree structure; The subbands are sorted according to their energy entropy, and the target subband is selected as the input for feature extraction.

[0008] As a preferred embodiment of the multi-source signal discrimination and alarm method for drop-out surge arresters described in this invention, the features selected as the target sub-band for feature extraction include at least one of approximate entropy and peak factor. The approximate entropy is used to quantify non-stationary randomness, and the peak factor is used to characterize impulsiveness. The two are calculated within a fixed or sliding window to form a feature vector. The approximate entropy is calculated as follows: a1. Obtain the sequence within the sliding time window The window length is The dimension of the construction is The embedding vectors are used to perform similarity counting with the infinity norm. The calculation steps are as follows: , , , in, As an approximate entropy, For dimension The average neighborhood log probability, For the first The neighborhood ratio centered on the embedding vector The sample sequence is preprocessed within the window. The number of samples within the window. For the embedding dimension, For similarity tolerance, This is an indicator function; it returns 1 if the event is true and 0 otherwise. For embedding vector indexes, For component index, Indicates the first The embedding vector and the first The maximum absolute value of the component differences between two embedding vectors is used to calculate the infinite norm distance between them, and based on this, to determine whether the two embedding vectors fall within the similarity tolerance. Within the defined neighborhood; a2. Set tolerances within the same window using a robust metric, as follows: , , in, For similarity tolerance, This is the tolerance factor. The median absolute deviation among classmates For median operators; a3. Calculate within each sliding window according to steps a1-a2. And output the approximate entropy feature of the window; The peak factor is calculated as follows: b1. Apply amplitude clamping to the sequence within the same window to obtain... : , in, The sample after clamping. For symbolic functions, It is a binary minimal operator. This refers to the clamping ratio; b2. In Extract robust peak values ​​and effective values, and calculate robust peak factor: , in, As a robust peak factor, For a stable peak, Based on Robust effective value (RMS); b3. Output the window's output As a characteristic of peak factor.

[0009] As a preferred embodiment of the multi-source signal discrimination and alarm method for drop-out surge arresters described in this invention, the single-class classification model is implemented using Support Vector Data Description (SVDD), a hypersphere decision boundary is constructed in the feature space using a Gaussian kernel function, the training data only contains healthy state samples, and the training adopts sequential optimization or an equivalent incremental optimization algorithm.

[0010] As a preferred embodiment of the multi-source signal discrimination and alarm method for drop-out surge arresters described in this invention, the threshold is an adaptive threshold that is dynamically adjusted based on environmental wind speed data; the environmental wind speed data comes from a local wind speed sensor or an external meteorological service and is acquired through a wireless communication module, and is synchronized with the features by timestamp.

[0011] As a preferred embodiment of the multi-source signal discrimination and alarm method for a drop-out arrester described in this invention, the adaptive threshold adopts a recursive least squares method to update the threshold parameter by window, including parameter state update and error correction. The steps for updating the threshold parameters using the recursive least squares method include: On the timestamp series, a regression vector is constructed using wind speed characteristics, and a quasi-linear expression for the threshold is given: , in, Indicates time Adaptive threshold, Indicates wind speed The quasi-linear eigenvectors formed Represents the corresponding parameter vector, superscript Indicates the transpose operator; Update the parameter vector at each time step as follows: , in, and Let these represent the parameter vectors at the current and previous time steps, respectively. Represents the current gain vector. Indicates the reference threshold sample. Represents the current regression vector, with superscript. Indicates the transpose operator; calculate And perform the covariance matrix recursion as follows: , , in, and Let these represent the parameter covariance matrices at the current time step and the previous time step, respectively. Indicates the forgetting factor, Indicates and Identity matrices of the same order, in the denominator This indicates that a scalar is assigned to a single term.

[0012] As a preferred embodiment of the multi-source signal discrimination and alarm method for drop-out surge arresters described in this invention, the risk warning includes a warning level and an alarm level: Early warning levels trigger periodic review tasks, while alarm levels trigger immediate inspection tasks, and corresponding events and data fragments are reported to the operations and maintenance platform.

[0013] As a preferred embodiment of the multi-source signal discrimination and alarm method for drop-out surge arresters described in this invention, the method further includes: cross-validation based on leakage current data of multi-phase surge arresters; specifically, phases A, B, and C at the same tower location are used to calculate inter-phase correlation coefficients or cross-correlation indices under time alignment conditions, and common-mode components are identified and suppressed or thresholded based on correlation deviations.

[0014] As a preferred embodiment of the multi-source signal discrimination and alarm method for a drop-out arrester described in this invention, the time alignment includes a unified sampling frequency and delay compensation, and feature calculation is performed using a sliding time window of a unified length; common-mode suppression is achieved by constructing inter-phase common-mode reference components and subtracting them from the target phase signal or feature.

[0015] As a preferred embodiment of the multi-source signal discrimination and alarm method for drop-out surge arresters described in this invention, the preprocessing includes at least one of anti-aliasing filtering, removal of power frequency and fundamental frequency leakage components, and abnormal pulse elimination; and a healthy baseline library and drift calibration are introduced during the training phase, while incremental learning can be used to update the model or threshold during the operation phase.

[0016] The beneficial effects of this invention are as follows: The multi-source signal discrimination and alarm method for drop-out surge arresters provided by this invention achieves indirect real-time monitoring of lead fatigue state by extracting frequency domain or time-frequency domain features related to mechanical vibration of connecting leads from leakage current signals.

[0017] This invention utilizes wavelet packet decomposition technology to effectively separate high-frequency subband signals, adapting to non-stationary vibration environments and ensuring robustness of feature extraction. The introduction of features such as approximate entropy and peak factor quantifies the randomness and impulsiveness of the signal, enabling sensitive capture of subtle changes in the lead wire during fatigue accumulation. A single-class classification model trained solely on healthy samples, such as support vector data description, is used to construct the decision boundary, making the model highly sensitive to abnormal states and providing early warnings before lead wire breakage. This baseline-based assessment avoids false alarms common in traditional threshold methods under complex environments. The alarm threshold is dynamically adjusted based on environmental wind speed data, and parameters are updated online using recursive least squares, allowing the system to adapt to wind speed changes and maintain consistent warning accuracy under non-stationary conditions. A tiered alarm mechanism triggers different responses based on the degree of deviation; early warning levels trigger periodic reviews, while high alarm levels trigger immediate inspections, optimizing resource allocation and improving response efficiency. Multiphase data cross-validation eliminates common-mode interference by calculating inter-phase correlation coefficients, further enhancing the system's anti-interference capability and reliability.

[0018] This invention links electrical signals with mechanical status, making up for the shortcomings of traditional monitoring in sensing mechanical degradation. It provides an effective online early warning tool for surge arrester maintenance during typhoon season, extends equipment life, and ensures the safe and stable operation of the power grid. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0020] Figure 1 This is a flowchart illustrating the multi-source signal discrimination and alarm method for drop-out surge arresters in the embodiments. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0022] All terms used in this application (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0023] For example, the terms “first” and “second” used in this application are only used to distinguish and describe similar objects, to differentiate the first object from another object, and are not used to describe a specific order or sequence, nor should they be interpreted as indicating or implying relative importance.

[0024] This application proposes a multi-source signal discrimination and alarm method for drop-out surge arresters, combined with... Figure 1 As shown, the method includes: Step S1: Acquire the leakage current signal of the surge arrester and perform preprocessing. In this embodiment, the leakage current signal is the equivalent primary side leakage current obtained through current sampling and isolation. The acquisition link includes front-end anti-aliasing filtering and analog-to-digital conversion. The three-phase channels are time-aligned using unified timing. The sampling frequency is 2kHz by default and can be adjusted within the range of 1~10kHz according to line noise and target frequency band. The preprocessing window length is 2s by default and can be set within the range of 0.5~4s. The window overlap is 50% by default. If timing jitter causes the three-phase alignment error to exceed 2ms, the window is marked as low confidence and does not participate in model update. When the data quality is abnormal for three consecutive windows, a data link self-check is triggered, during which the most recent valid threshold remains unchanged.

[0025] Step S2 involves extracting frequency domain or time-frequency domain features related to the mechanical vibration of the connecting leads from the leakage current signal. Specifically, the mechanical vibration of the connecting leads is considered as modulation traces of high-frequency disturbances and intermittent impacts on the leakage current, with the target frequency band covering tens to hundreds of hertz and its envelope at even higher frequencies. The feature calculation window is consistent with step S1, using a default sliding step size of 25% to 50% to balance timeliness and smoothness. If necessary, a robust amplitude clamp is added before the features to suppress distortion caused by individual extreme samples. If the missing measurement ratio within the window does not exceed 5%, amplitude-limited interpolation and boundary smoothing can be performed; if it exceeds this ratio, the features of that window will not participate in the alarm determination.

[0026] Step S3: Based on the single-class classification model trained only on healthy samples, evaluate the fatigue state of the connection leads of the features and output the deviation. Step S4: When the deviation exceeds the threshold, a risk warning is triggered; In one embodiment, step S2 includes: Wavelet packet decomposition is performed on the leakage current signal, and high-frequency subbands are separated using a multi-level decomposition tree structure; Subbands are sorted based on energy entropy, and target subbands are selected as feature extraction inputs. For example, the wavelet basis is preferably a mother wavelet with tight support and good orthogonality, and the order is chosen to be in the middle range to balance time-frequency resolution. The default number of decomposition layers is 4 to 7, and one or more subbands are aligned with the tens to hundreds of hertz region based on the sampling frequency and the target frequency band. Boundary processing preferably uses symmetrical extension to reduce artifacts. When there are ties in the sorting, the subband with the lower center frequency and narrower bandwidth is preferred as the feature input. If they still cannot be distinguished, the two subbands are retained in parallel and concatenated at the model input. If the target subband is missing due to limited device sampling bandwidth, the selection degenerates to the suboptimal selection of energy entropy of adjacent subbands in the same layer while keeping the threshold unchanged.

[0027] In one embodiment, the features selected as the target subband for feature extraction input include at least one of approximate entropy and peak factor. Approximate entropy is used to quantify non-stationary randomness, and peak factor is used to characterize impact. The two are calculated within a fixed or sliding window to form a feature vector. The approximate entropy is calculated as follows: a1. Obtain the sequence within the sliding time window The window length is The dimension of the construction is The embedding vectors are used to perform similarity counting with the infinity norm. The calculation steps are as follows: , , , in, As an approximate entropy, For dimension The average neighborhood log probability, For the first The neighborhood ratio centered on the embedding vector The sample sequence is preprocessed within the window. The number of samples within the window. For the embedding dimension, For similarity tolerance, This is an indicator function; it returns 1 if the event is true and 0 otherwise. For embedding vector indexes, For component index, Indicates the first The embedding vector and the first The maximum absolute value of the component differences between two embedding vectors is used to calculate the infinite norm distance between them, and based on this, to determine whether the two embedding vectors fall within the similarity tolerance. Within the defined neighborhood; a2. To improve robustness, tolerances are set within the same window using a robustness metric, as follows: , , in, For similarity tolerance, This is the tolerance factor. The median absolute deviation among classmates This is the median operator; recommended values: (priority ), , Window step size ; a3. Calculate within each sliding window according to steps a1-a2. The approximate entropy feature of the window is output. Similarly, the approximate entropy parameters are set according to robust principles in engineering: the embedding dimension is set to 2 or 3 by default; the similarity tolerance is determined based on a fixed proportion of the absolute deviation of the median within the same window, with a default proportion of 0.15~0.35; the number of samples within the window is 2048 by default, and can be selected in the range of 1024~4096; the window step size is set to one-quarter to one-half of the window length by default. When there are continuous saturated or clipped segments within the same window, robust extremum removal is performed first before calculating the approximate entropy; if the effective samples are insufficient to support the embedding construction, the approximate entropy of the window is marked as missing and filled by the robust quantiles of adjacent windows, which is only used for trend display and does not participate in alarms.

[0028] The peak factor is calculated as follows: b1. Apply amplitude clamping to the sequence within the same window to obtain... : , in, The sample after clamping. For symbolic functions, It is a binary minimal operator. This refers to the clamping ratio; b2. In Extract robust peak values ​​and effective values, and calculate robust peak factor: , in, As a robust peak factor, For a stable peak, Based on Robust RMS value; Recommended value: With classmates Consistent estimation; optional preprocessing using 3-point Hampel denoising before clamping; b3. Output the window's output As a characteristic of the peak factor, unclamped peak values ​​are saved in parallel for comparison when necessary, but they are not included in the alarm. Optionally, the amplitude clamping factor of the robust peak factor is 3 to 5 times the median absolute deviation of the same window by default, and the robust RMS value is the root mean square of the same window. To avoid the zero value caused by open circuit or momentary interruption at the front end dragging down the RMS value, isolated zero value segments with a duration of less than 20ms are removed. If a step occurs due to range switching in the sampling link, the peak factor of the current window is only used for comparison and does not participate in the threshold determination. This restriction is automatically lifted after the subsequent two windows return to stability.

[0029] Specifically, two complementary features are established using a unified window. The approximate entropy focuses on the probability of occurrence of similar segments in the embedding space, reflects the degree of decay of the sequence regularity through the difference of the logarithmic mean, and sets the neighborhood radius based on the median scale to reduce the impact of occasional spikes and amplitude drift. The peak factor focuses on the impact measure, and uses the ratio of the clamped peak value to the effective value to represent the proportion of transient energy to the overall energy, avoiding the imbalance of the numerator and denominator scales caused by individual extreme samples. Both are completed within the same window, maintaining time alignment and dimensional consistency, which facilitates coupling with the input requirements of subsequent single-class models. The parameter range has engineering margins, which can adapt to different sampling rates and noise backgrounds, and maintain cross-scene reproducibility. In one embodiment, the single-class classification model is implemented using Support Vector Data Description (SVDD). A hypersphere decision boundary is constructed in the feature space using a Gaussian kernel function. The training data contains only healthy state samples, and training employs sequential optimization or an equivalent incremental optimization algorithm. Furthermore, healthy samples are derived from historical low-wind-speed and non-lightning-strike periods, covering day and night and various temperature and humidity environments. The samples are normalized using a uniform mean-standard deviation before being used in training. Deviation is defined in the implementation as the distance or equivalent score of a sample to the decision boundary, linearly mapped to 0-1 for threshold management. During the online phase, incremental learning is evaluated every 30 days by default. The triggering condition is that the data quality is acceptable and there are no major alarms in the most recent evaluation period, avoiding the inclusion of abnormal states within the healthy boundary.

[0030] In one embodiment, the threshold is an adaptive threshold that is dynamically adjusted based on environmental wind speed data. The environmental wind speed data comes from a local wind speed sensor or an external meteorological service and is acquired via a wireless communication module, synchronized with the features by timestamp. In this embodiment, the maximum allowable error for time alignment between wind speed and features is no more than 2 seconds by default, the wind speed sampling period is 1-10 seconds by default, and the threshold refresh period is consistent with the feature window. When both local and external wind speed data are available simultaneously, the local data is prioritized and verified using the external result; if either data source is interrupted, the remaining data source continues to operate; if both data sources are unavailable, the fixed threshold strategy is reverted and the source switching event is recorded.

[0031] In one embodiment, the adaptive threshold uses a recursive least squares method to update the threshold parameter by window, including parameter state update and error correction to adapt to wind speed changes. The steps for updating the threshold parameters using the recursive least squares method include: On the timestamp series, a regression vector is constructed using wind speed characteristics, and a quasi-linear expression for the threshold is given: , in, Indicates time Adaptive threshold, Indicates wind speed The resulting quasi-linear eigenvectors (constant term, linear term, and quadratic term). Represents the corresponding parameter vector, superscript This represents the transpose operator; if a linear model is required, it can be taken as... ; Update the parameter vector at each time step as follows: , in, and Let these represent the parameter vectors at the current and previous time steps, respectively. Represents the current gain vector. The reference threshold sample (from a healthy baseline library or a robust upper quantile statistic of a sliding window under the same working conditions, used as the expected threshold observation) is indicated. Represents the current regression vector, with superscript. Indicates the transpose operator; calculate And perform the covariance matrix recursion as follows: , , in, and Let these represent the parameter covariance matrices at the current time step and the previous time step, respectively. Indicates the forgetting factor, Indicates and Identity matrices of the same order, in the denominator This indicates that a scalar is assigned to a single term; Initial values ​​can be taken , , Forgetting factor When wind conditions change more rapidly, take the smaller value. Reference threshold sample The reference threshold samples are obtained by robustly aggregating healthy sample libraries according to wind speed neighborhoods. Specifically, the reference threshold samples are obtained by calculating percentiles within historical healthy windows of similar wind speed ranges, with the 90th percentile used by default. The wind speed neighborhood width is plus or minus one-tenth of the target wind speed by default. To ensure statistical stability, the number of windows participating in aggregation is no less than thirty; if this number is insufficient, it degenerates into statistics from the global healthy library for the same season and time period. In scenarios of sudden wind changes, the reference threshold samples are updated with priority given to the most recent hour to improve adaptive response speed.

[0032] Specifically, this paper presents a step-by-step application framework for recursive least squares, establishing the threshold and wind speed features on a simple quasi-linear structure, enabling the parameter recursion to stably track wind conditions during slow drift or abrupt changes. The backbone adopts a parameter update formula, where the parameters cover the closed recursion of gain and covariance, facilitating direct implementation as an online algorithm. The range setting of the forgetting factor balances tracking and noise suppression, and the larger initial covariance provides rapid convergence capability. The reference threshold samples are derived from robust statistics of a healthy baseline, which helps maintain the physical meaning and engineering interpretability of the threshold under environmental disturbances. This structure is friendly to the regression vector dimension, allowing the addition of quadratic terms or other derived wind speed features without changing the recursive framework, to adapt to differences in towers and terrain. In one embodiment, the risk warning includes a warning level and an alarm level: Warning levels trigger periodic review tasks, while alarm levels trigger immediate inspection tasks, and corresponding events and data segments are reported to the operations and maintenance platform. Furthermore, warning levels require three consecutive windows to deviate from the threshold; alarm levels require a duration of at least 10 seconds or at least two instances to simultaneously exceed the threshold. Reported data segments by default include the original signal and feature vector for 5 seconds before and after exceeding the threshold to support rapid verification. To avoid jitter, a 5% hysteresis strategy is used to remove the threshold, and a 60-second cooldown time is set to prevent repeated alarms.

[0033] In one embodiment, the method further includes: performing cross-validation based on leakage current data of multiphase surge arresters to eliminate common-mode interference; specifically, calculating the inter-phase correlation coefficient or cross-correlation index of phases A, B, and C at the same tower location under time alignment conditions, identifying common-mode components based on correlation deviations, and performing suppression or threshold correction; for example, the inter-phase correlation is calculated with zero time delay after time alignment. When the correlation between any two phases is higher than 0.8 and the third phase deviates significantly, it is determined that there is a common-mode disturbance, and a reference component is constructed using the median value of the three-phase signals, which is subtracted from the characteristics of the target phase to suppress the common-mode effect; if only two-phase data are available, the median value of the two phases is used to replace the reference component to maintain input-output consistency. If all three phases have missing measurements, the common-mode correction for that window is paused and the original characteristics are retained for trend observation.

[0034] In one embodiment, time alignment includes uniform sampling frequency and delay compensation, and feature calculation is performed using a sliding time window of uniform length; common-mode suppression is achieved by constructing inter-phase common-mode reference components and subtracting them from the target phase signal or feature; similarly, the weight of the common-mode reference components can be dynamically set according to the data quality index of each phase, with the median value as the default reference, and smaller weights are given to phases with higher noise; if a negative value or non-physical quantity appears after subtraction, it is truncated with zero as the lower bound and the number of truncations is recorded in the log for subsequent verification.

[0035] In one embodiment, preprocessing includes at least one of anti-aliasing filtering, removal of power frequency and fundamental frequency leakage components, and abnormal pulse removal; a healthy baseline library and drift calibration are introduced during the training phase, and incremental learning can be used to update the model or thresholds during the runtime phase; optionally, the cutoff frequency of anti-aliasing filtering is set to 40%~45% of the sampling frequency, adaptive notch filtering is used for power frequency removal to be compatible with 50 and 60 Hz and their low-order harmonics, and abnormal pulse removal uses a median-based three-point robust detection; drift calibration is performed during low-load periods each day, using the healthy baseline to perform zero-point and scale correction on the data from the most recent 24 hours. If storage or computing power is insufficient, the minimum executable approach is to retain only the window from the most recent 1 hour for alarm determination and suspend incremental learning until resources are restored.

[0036] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0037] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of this application and form different embodiments. For example, all the embodiments above can be used in any combination. The information disclosed in this background section is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art.

Claims

1. A method for discriminating and alarming multiple sources of signals of a surge arrester, characterized in that, include: Step S1: Collect the leakage current signal of the surge arrester and perform preprocessing; Step S2: Extract frequency domain or time-frequency domain features related to the mechanical vibration of the connecting lead from the leakage current signal; Step S2 includes: Wavelet packet decomposition is performed on the leakage current signal, and high-frequency subbands are separated using a multi-level decomposition tree structure; The subbands are sorted according to energy entropy, and the target subband is selected as the input for feature extraction. The features selected as the target sub-band for feature extraction include at least one of approximate entropy and peak factor. The approximate entropy is used to quantify non-stationary randomness, and the peak factor is used to characterize impact. The two are calculated within a fixed or sliding window and form a feature vector. Step S3: Based on the single-class classification model trained only on healthy samples, evaluate the fatigue state of the connection leads of the features and output the deviation. Step S4: When the deviation exceeds the threshold, a risk warning is triggered; The threshold is an adaptive threshold that is dynamically adjusted based on environmental wind speed data. The environmental wind speed data comes from a local wind speed sensor or an external meteorological service and is acquired through a wireless communication module, and is synchronized with the feature by timestamp. The adaptive threshold uses a recursive least squares method to update the threshold parameters by window, including parameter state updates and error correction.

2. The method for multi-source signal discrimination and alarm of a drop-out surge arrester as described in claim 1, characterized in that, The approximate entropy is calculated as follows: a1. Obtain the sequence within the sliding time window The window length is The dimension of the construction is The embedding vectors are used to perform similarity counting with the infinity norm. The calculation steps are as follows: ， ， ， in, As an approximate entropy, For dimension The average neighborhood log probability, For the first The neighborhood ratio centered on the embedding vector The sample sequence is preprocessed within the window. The number of samples within the window. For the embedding dimension, For similarity tolerance, This is an indicator function; it returns 1 if the event is true and 0 otherwise. For embedding vector indexes, For component index, Indicates the first The embedding vector and the first The maximum absolute value of the component differences between two embedding vectors is used to calculate the infinite norm distance between them, and based on this, to determine whether the two embedding vectors fall within the similarity tolerance. Within the defined neighborhood; a2. Set tolerances within the same window using a robust metric, as follows: ， ， in, For similarity tolerance, This is the tolerance factor. The median absolute deviation among classmates For median operators; a3. Calculate within each sliding window according to steps a1-a2. And output the approximate entropy feature of the window; The peak factor is calculated as follows: b1. Apply amplitude clamping to the sequence within the same window to obtain... : ， in, The sample after clamping. For symbolic functions, It is a binary minimal operator. This refers to the clamping ratio; b2. In Extract robust peak values ​​and effective values, and calculate robust peak factor: ， in, As a robust peak factor, For a stable peak, Based on Robust effective value (RMS); b3. Output the window's... As a characteristic of peak factor.

3. The method for multi-source signal discrimination and alarm of a drop-out surge arrester as described in claim 1, characterized in that, The single-class classification model is implemented using Support Vector Data Description (SVDD). A hypersphere decision boundary is constructed in the feature space using a Gaussian kernel function. The training data contains only healthy state samples, and training employs sequential optimization or an equivalent incremental optimization algorithm.

4. The method for multi-source signal discrimination and alarm of a drop-out surge arrester as described in claim 1, characterized in that, The steps for updating the threshold parameters using the recursive least squares method include: On the timestamp series, a regression vector is constructed using wind speed characteristics, and a quasi-linear expression for the threshold is given: ， in, Indicates time Adaptive threshold, Indicates wind speed The quasi-linear eigenvectors formed Represents the corresponding parameter vector, superscript Indicates the transpose operator; Update the parameter vector at each time step as follows: ， in, and Let these represent the parameter vectors at the current and previous time steps, respectively. Represents the current gain vector. Indicates the reference threshold sample. Represents the current regression vector, with superscript. Indicates the transpose operator; calculate The covariance matrix is ​​then recursively derived as follows: ， ， in, and Let these represent the parameter covariance matrices at the current time step and the previous time step, respectively. Indicates the forgetting factor, Indicates and Identity matrices of the same order, in the denominator This indicates that a scalar is assigned to a single term.

5. The method for multi-source signal discrimination and alarm of a drop-out surge arrester as described in claim 1, characterized in that, The risk warning includes a warning level and an alarm level: Early warning levels trigger regular review tasks, while alarm levels trigger immediate inspection tasks, and corresponding events and data fragments are reported to the operations and maintenance platform.

6. The method for multi-source signal discrimination and alarm of a drop-out surge arrester as described in claim 1, characterized in that, The method also includes: cross-validation based on leakage current data of multiphase surge arresters; specifically, phases A, B, and C at the same tower location are used to calculate the inter-phase correlation coefficient or cross-correlation index under time alignment conditions, and common-mode components are identified and suppressed or thresholded based on the correlation deviation.

7. The method for multi-source signal discrimination and alarm of a drop-out surge arrester as described in claim 6, characterized in that, The time alignment includes uniform sampling frequency and delay compensation, and feature calculation is performed using a sliding time window of uniform length; common mode suppression is achieved by constructing interphase common mode reference components and subtracting them from the target phase signal or feature.

8. The method for multi-source signal discrimination and alarm of a drop-out surge arrester as described in claim 1, characterized in that, The preprocessing includes at least one of anti-aliasing filtering, removal of power frequency and fundamental frequency leakage components, and abnormal pulse rejection; and a healthy baseline library and drift calibration are introduced during the training phase.

Images