A diagnostic and early warning method and system for surge arresters with dual time-scale current and quality gating

Abstract

This invention discloses a method and system for diagnosing and warning surge arresters using dual-timescale current and quality gating. Electrical current signals are acquired through a current sensing module electromagnetically coupled to a pointer-type leakage current meter measuring branch, a leakage current sensor, or an online monitoring terminal. Integer power frequency periodic waveforms are resampled, and a trend sequence across acquisition times is constructed. Waveform quality and trend quality are calculated separately. Periodic waveform representations are extracted using a multi-sensory-field one-dimensional convolutional encoder, and long-term trend representations are extracted using a gated cyclic encoder. Reliability and gating weights are calculated based on branch quality, normalized uncertainty, and available branch markers. In the offline phase, a model is trained using classification loss, degradation monotonicity loss, state transition loss, and three-phase consistency loss. In the online phase, the model parameters are fixed for diagnosis, and the state level, fault risk, and diagnostic evidence are output through a persistence window and hysteresis threshold.

Description

Technical Field

[0001] This invention relates to the fields of online monitoring, signal processing and intelligent fault diagnosis of power equipment, specifically to the construction of periodic waveform and long-term trend dual-timescale data of surge arrester leakage current signal, data quality assessment, quality gating fusion, status diagnosis and early warning output, and especially to a surge arrester diagnosis and early warning method and system with dual-timescale current and quality gating. Background Technology

[0002] Metal oxide surge arresters are critical protective devices in substations for limiting lightning overvoltages and switching overvoltages. During long-term operation, surge arresters are affected by factors such as varistor aging, internal moisture absorption, surface contamination, seal failure, operating voltage fluctuations, and changes in ambient temperature and humidity. These factors alter the amplitude, waveform, harmonic components, and long-term trend of their leakage current. Continuous monitoring and diagnosis of leakage current provides a basis for condition-based maintenance and fault early warning of surge arresters.

[0003] Pointer-type leakage current meters are commonly installed in substations, connected in series with the grounding branch of surge arresters. Traditional maintenance relies mainly on manual inspections and manual recording of meter readings, which suffers from long inspection cycles, insufficient continuity, and delayed anomaly detection. Some digital transformation solutions use cameras to capture dial images and read pointer positions using visual algorithms; however, cameras and lighting devices are susceptible to the effects of light, condensation, dirt, obstructions, and lens aging, and image acquisition, transmission, and recognition increase power consumption and communication burden.

[0004] Existing technologies have focused on surge arrester current monitoring and intelligent diagnosis. For example, patent CN114660387A discloses a surge arrester monitoring method based on a leakage current sensor and a BP neural network algorithm, combining a leakage current sensor, harmonic analysis, and a BP neural network for aging monitoring of metal oxide surge arresters; patent CN118094402A discloses a fault diagnosis method for medium and low voltage surge arresters based on neural networks, using defect data mining, automatic encoders, and one-dimensional convolutional networks for state classification based on operating voltage, leakage current, resistive current, and number of operations; and patent CN105242128A discloses an online surge arrester monitoring system based on a fuzzy neural network, which simultaneously collects bus voltage and surge arrester total current, extracts the fundamental resistance current, and uses a weighted K-nearest neighbor algorithm for fault judgment. The above scheme has promoted the development of surge arrester monitoring from manual criteria to data-driven diagnosis. However, its data structure, feature extraction and model fusion methods are different from those of this invention. It has not yet established dual-timescale modeling for integer power frequency period waveforms and long-term trends across acquisition times, nor has it utilized the fusion contribution of two types of characterizations: waveform quality, trend quality and branch uncertainty control.

[0005] Furthermore, existing intelligent diagnostic solutions typically extract features from leakage current, resistive current, harmonics, or environmental quantities, and then use support vector machines, K-nearest neighbors, artificial neural networks, autoencoders, or convolutional neural networks for state classification. These methods generally have the following shortcomings: First, they fail to distinguish between waveform information within the power frequency cycle and long-term trend information across time periods, and sampling scales and feature calculation conditions are easily confused; second, data quality is often used only as a preprocessing screening condition, without utilizing waveform quality, trend quality, and model uncertainty to control the contribution of different diagnostic branches; third, purely data-driven models may produce state transitions or risk reversals inconsistent with the arrester's degradation patterns; fourth, early warning outputs lack corresponding key waveforms, trend intervals, quality assessments, and constraint triggering evidence, making it difficult to support on-site verification.

[0006] Electromagnetically coupling a through-hole current sensor to the measuring branch of a pointer-type leakage current meter allows for the direct acquisition of electrical current signals without altering the original metering circuit. This signal is not a dial image, but rather a sampling sequence reflecting the time-varying leakage current of the surge arrester. If the sampling frequency and duration meet waveform analysis conditions, information such as RMS values, waveform factors, and harmonic ratios can be extracted from one or more power frequency cycles. Furthermore, continuously arranging waveform packets from different times can reveal long-term trends reflecting slow changes such as moisture absorption and aging.

[0007] However, the strong electromagnetic environment and low-power communication conditions in substations can cause sampling clock jitter, waveform clipping, local missing data, non-target interference, inconsistent reporting intervals, and data packet loss. Periodic waveforms and long-term trends exhibit different sensitivities to these quality issues: local waveforms are significantly affected by sampling integrity and interference, while long-term trends can still provide state evolution information even with a few missing waveform points, but are affected by continuous reporting gaps and baseline drift. Simply concatenating the two types of data may result in low-quality branches making unreasonable contributions to the diagnostic results.

[0008] Therefore, a method is needed that does not rely on camera images, can correspond to the actual sampling conditions of current signals, can collaboratively model periodic waveforms and long-term trends, can adaptively fuse based on on-site quality, and can incorporate arrester degradation knowledge into model training and early warning judgment. Summary of the Invention

[0009] The technical problem to be solved by this invention is: to address the problems of mixed waveform and trend scales, low-quality data interference diagnosis, unstable identification of a few fault samples, state prediction not conforming to the degradation law, and lack of evidence chain in the early warning results of the electrical signals of the measuring branch of the pointer leakage current meter under strong electromagnetic interference, sampling quality fluctuations and low power communication conditions, and to provide a method and system for the diagnosis and early warning of surge arresters with dual time scale current and quality gating.

[0010] A surge arrester diagnostic and early warning method with dual time-scale current and quality gating includes the following steps: S1: Acquire the electrical current signal of the surge arrester leakage current meter measurement branch, organize the current sampling values ​​obtained in one acquisition that cover an integer number of power frequency cycles into a periodic waveform packet, and associate the periodic waveform packet at different acquisition times with the equipment identifier, phase and timestamp; the electrical current signal is output by a current sensing module, leakage current sensor or online monitoring terminal that is electromagnetically coupled to the pointer-type leakage current meter measurement branch; S2: Perform power frequency cycle positioning, integer period segmentation, equal-point resampling, DC bias removal, abnormal sampling point marking, and amplitude normalization on the periodic waveform packet. Calculate the waveform quality based on period integrity, saturation distortion ratio, sampling clock jitter, interference band energy ratio, and missing sampling ratio. And generate waveform branches that can be marked. ; S3: Constructing a periodic waveform tensor from the preprocessed periodic waveform. Furthermore, packet-level current summary indicators are extracted from each periodic waveform packet, and these packet-level current summary indicators within a preset trend window are used to form a long-term trend tensor. Trend quality is calculated based on packet missing rate, reporting interval dispersion, abnormal summary rate, and baseline stability. And generate available markers for trend branches based on the number of valid acquisition times and trend quality. ; S4: Transfer the periodic waveform tensor Input a multi-receptive-field one-dimensional convolutional encoder to obtain waveform representations. and the branch uncertainty of the original waveform and the original waveform branch uncertainty Normalization ; the long-term trend tensor Input a gated cyclic trend encoder to obtain trend representation. and the uncertainty of the original trend branch and the uncertainty of the original trend branch Normalization ; S5: Based on the waveform quality Trend Quality Normalized waveform branch uncertainty Normalized trend branch uncertainty and branch available markers and The reliability, evidence quantity, and gating weights of the two branches are calculated. The waveform representation and trend representation are then weighted and fused using the gating weights to obtain the fused representation. ; S6: During the offline training phase, a joint loss function including classification loss, degradation monotonicity loss, state transition loss, and three-phase consistency loss is used to iteratively update the model parameters of the dual-branch encoder, quality gating module, and diagnostic output layer based on historical labeled samples, resulting in a trained diagnostic model. During the online diagnostic phase, the model parameters of the trained diagnostic model are fixed, and the fused representation of the current diagnostic window is input into the diagnostic model to perform forward computation, obtaining the arrester state probability and fault risk probability. The degradation monotonicity loss is constructed based on the difference between the degradation representation increment and the risk probability, and is used to penalize the prediction that the fault risk probability decreases in the opposite direction when there is no repair event and the current degradation representation continues to increase. The state transition loss constrains the state changes at adjacent diagnostic times to conform to a preset state transition relationship. S7: Perform persistence and hysteresis determination on the state probability and fault risk probability within the continuous diagnostic window to generate normal, attention, abnormal or severe warning status; S8: Outputs early warning status, health score, key period waveforms, key trend intervals, triggered degradation characteristics, data quality information, and operation and maintenance suggestions.

[0011] Preferably, the current sensing module in step S1 includes a through-hole zero-flux current sensor, a signal conditioning circuit, and an analog-to-digital converter circuit; the through-hole zero-flux current sensor is sleeved outside the conductor of the measuring branch of the pointer-type leakage current meter; the signal conditioning circuit isolates, filters, adjusts the gain, and performs zero-bias correction on the sensor output; and the analog-to-digital converter circuit adjusts the sampling frequency... Obtain the current sampling sequence ,in For the conditioned sampling voltage, To calibrate the gain, The bias is zero; and through bypass protection circuits, surge suppressors, or clamping protection circuits, the induced voltage or induced current on the acquisition side exceeds the protection threshold. or It can conduct, discharge, or clamp at any time to suppress the overvoltage and overcurrent induced by the lightning strike current on the acquisition side, protect the signal conditioning circuit and analog-to-digital conversion circuit, and ensure that the current sensing module does not change the original metering circuit and indicating function of the pointer-type leakage current meter.

[0012] Preferably, the waveform quality described in step S2 Calculate using the following formula: (1) In the formula, Indicates waveform quality. This indicates the proportion of incomplete power frequency cycles. Indicates the proportion of saturated or clipped sampling points. This represents the normalized amount of jitter during the sampling interval. This indicates the proportion of non-target interference frequency band energy. Indicates the proportion of missing samples. to The weighting coefficients are non-negative; the equal-point resampling maps each power frequency cycle to the same number of phase sampling points; the... This is a dimensionless quality score, ranging from 0 to 1; when waveform branches are available, Greater than 0; when waveform branches are unavailable, The value is 0; the larger the value, the higher the waveform quality.

[0013] Preferably, in step S3, the first... Resampled current sequence of periodic waveform packets Calculate the packet-level current summary vector ;in, For valid values, Peak value, Peak-to-peak value For waveform factor, As the peak factor, This represents the ratio of the third harmonic amplitude to the fundamental frequency amplitude. This represents the ratio of the fifth harmonic amplitude to the fundamental frequency amplitude. Represents the total harmonic distortion rate. Represents the offset relative to the device health baseline; the long-term trend tensor consists of continuous... The packet-level current summary vector at each acquisition time and its multi-window derivation constitute the multi-window derivation, wherein the multi-window derivation includes at least the increment. robust slope volatility Historical quantile position and duration of continuous over-limit ; Let the preset trend window length be The actual number of effective data acquisition times is The preset minimum number of valid data acquisitions is... ;when Less than and not less than At that time, the input is padded according to the original timeline, and the missing mask and actual time interval are set synchronously for each time step. The derived quantity is calculated only based on the valid acquisition time. Not less than and When the preset trend availability threshold is reached, let =1; when Less than or When the value is below the preset trend availability threshold, let =0, where when Less than Time still order It is 0.

[0014] Preferably, the multi-receptive-field one-dimensional convolutional encoder in step S4 includes The nth one-dimensional convolutional branch, the th Each one-dimensional convolutional branch has a convolutional kernel size. or expansion rate Each branch with respect to the periodic waveform tensor Branch features are obtained after convolution, nonlinear transformation, channel mapping, and pooling. The waveform representation is obtained by splicing the features of each branch and then projecting them. The gated cyclic trend encoder processes the long-term trend tensor according to the order of acquisition time. And introduce a missing mask and quality mask Calculate attention weights This reduces the contribution of missing or low-quality moments to trend convergence, ultimately leading to a trend representation. The original waveform branch uncertainty and the uncertainty of the original trend branch Calculated from at least one of output entropy, repeated inference variance, or model ensemble variance; and then processed through at least one of the following methods: maximum entropy scaling normalization, pruning normalization based on robust quantiles of training or validation data, or monotonic compression. and Normalization and ,in and All values ​​are dimensionless uncertainties ranging from 0 to 1, with larger values ​​indicating higher uncertainties.

[0015] Preferably, the waveform reliability in step S5 Trend reliability Waveform evidence quantity Trend Evidence Quantity Gating weights and and fusion characterization satisfy:

[0016] In the formula, Indicates waveform reliability. Indicates the reliability of the trend. Indicates the amount of waveform evidence. Indicates the amount of evidence for a trend. Indicates the waveform branch gating weight. Indicates the trend branch gating weight. and For a dimensionless quality score with values ​​ranging from 0 to 1, and For the normalized uncertainty ranging from 0 to 1, and A flag can be used to indicate branches that take the value 0 or 1. It has a positive temperature coefficient; and The value range is from 0 to 1. The preset evidence threshold is non-negative; when the total amount of available evidence is no greater than... At this time, no deterministic device status diagnosis is output; only a message indicating insufficient data or retesting of the acquisition link is output. Indicates fusion representation, and These represent waveform representation and trend representation, respectively. and These represent projection functions that map two types of representations to the same dimension.

[0017] Preferably, the joint loss function described in step S6 satisfies:

[0018] In the formula, Weighted classification loss based on sample quality. To mitigate monotonicity loss, For state transition loss, For three-phase consistency loss, The weighting coefficients are non-negative; for two samples of the same surge arrester, the increment of the degradation characteristic at a later time step relative to the previous time step is not less than the preset increment. Furthermore, if no repair or repositioning event occurs during this period, the degradation monotonicity loss results in a lower probability of failure at a later time than at a previous time, exceeding the allowable margin. Punishment will be imposed in such circumstances.

[0019] Preferably, the state transition loss is based on the permissible state transition matrix between normal, attention, abnormal, and severe warning states. and penalty matrix Calculation, if based on state to state If the transfer is allowed, then If the abnormal state or severe warning state directly changes to the normal state in the absence of a repair or restoration event, then A positive penalty coefficient is applied, and if a confirmed repair and repositioning event exists, the corresponding penalty is lifted. The three-phase consistency loss applies to the same group of three-phase surge arresters under the same voltage level, the same interval, or the same busbar operating conditions. During the health tag period, the distance between the common changes in the three phases is minimized, and during the single-phase abnormal tag period, the deviation of the abnormal phase relative to the other two phases is retained through the phase-to-phase residual or interval margin.

[0020] Preferably, in step S7, when continuously At least one diagnostic window Failure risk probability of each diagnostic window Reaching the upgrade threshold The warning status is raised in a timely manner; when it is continuous Failure risk probability of each diagnostic window Below the downgrade threshold The warning status was lowered at that time, among which Less than or equal to , Higher than When the data quality is lower than the preset quality threshold, the equipment fault level is not directly increased, but a retest prompt for the data acquisition link is generated instead.

[0021] Preferably, waveform branches can be set using markers. Trend branches can be marked When periodic waveform packets are unavailable, sampling frequency... Less than the frequency threshold required for harmonic analysis, or period integrity Below the preset integrity threshold season 0, Set to 0 and retain missing markers; when the actual number of valid acquisition times is... Less than the preset minimum number of valid samples season 0, =0; when Not less than but When the value is below the preset trend availability threshold, let =0; when =0 and When it is 1, directly set 0, Setting it to 1 degrades the gating fusion to a diagnosis based on the long-term trend tensor input trend branch; when =0 and When it is 1, directly set For 1, The value is 0; when both branches are unavailable or the total amount of available evidence is not greater than the preset evidence threshold. At this time, no deterministic device status diagnosis is output; only a data shortage or acquisition link retest prompt is output, and the diagnostic confidence level is reduced.

[0022] A substation surge arrester diagnosis and early warning system includes a current signal acquisition module, a periodic waveform processing module, a trend sample construction module, a quality assessment module, a dual-branch coding module, a quality gating fusion module, a knowledge constraint diagnosis module, and an early warning output module. The current signal acquisition module acquires the electrical current signal of the surge arrester's leakage current meter measurement branch and generates periodic waveform packets or low-frequency trend data. The quality assessment module calculates waveform quality, trend quality, and missing data markers. The quality gating fusion module calculates reliability, evidence quantity, and gating weights based on quality scores, normalized branch uncertainty, and branch availability markers. The knowledge constraint diagnosis module performs classification, degradation monotonicity, state transition, and three-phase consistency constraints. The early warning output module performs persistence window and hysteresis threshold determination. Each module is configured to perform a surge arrester diagnosis and early warning method with dual-timescale current and quality gating.

[0023] An electronic device includes a processor, a memory, and a communication interface. The memory stores a computer program, model parameters, quality thresholds, and warning thresholds. When the computer program is executed by the processor, the electronic device receives electrical current signals or trend data from the leakage current meter measurement branch of a surge arrester and implements a surge arrester diagnostic and warning method with dual time-scale current and quality gating.

[0024] A computer-readable storage medium storing a computer program thereon, the computer program including data preprocessing instructions, dual-timescale feature construction instructions, quality gating fusion instructions, knowledge constraint diagnosis instructions, and continuous window warning instructions, wherein when the computer program is executed by a processor, it implements a surge arrester diagnosis and warning method with dual-timescale current and quality gating.

[0025] Compared with existing technologies, the present invention has at least the following advantages: The present invention directly processes the electrical current signal obtained by electromagnetic coupling with the measuring branch of the pointer-type leakage current meter, without relying on cameras, supplementary lighting, dial images, or visual readings. This avoids the influence of light, condensation, obstruction, and lens contamination on the data source, and reduces the burden of on-site data acquisition and transmission. The present invention clearly separates the integer power frequency period waveform and the long-term trend across acquisition times, ensuring that harmonics, waveform morphology, and trend characteristics are calculated separately on data that meets their sampling conditions, avoiding the mixing of low-frequency scalars and power frequency waveform characteristics. This invention constructs waveform quality based on periodic integrity, saturation distortion, clock jitter, interference energy, and missing conditions; constructs trend quality based on continuous reporting and baseline stability; and combines quality scores with normalized uncertainty to form branch reliability. By combining branch availability labels to determine the amount of evidence and fusion weight, it can reduce the unreasonable impact of low-quality branches on diagnostic results. This invention incorporates the operating rules of surge arresters into the model training by degrading monotonicity, state transition, and three-phase consistency constraints. Compared with using a fixed threshold only at the output, this can reduce risk reversal and state jumps without physical basis. This invention employs a continuous window and hysteresis threshold to output early warnings, while simultaneously providing waveform, trend, quality, and constraint evidence. This facilitates maintenance personnel in distinguishing between equipment malfunctions and data acquisition link malfunctions, thereby improving the verifiability of diagnostic results. Attached Figure Description

[0026] Figure 1 This is a flowchart of the overall process of the present invention; Figure 2 This is a flowchart of the dual-time-scale current data construction process of the present invention; Figure 3 This is a structural diagram of the quality-gated dual-branch diagnostic model of the present invention; Figure 4 This is a flowchart of the knowledge constraint training and interpretable early warning process of this invention. Detailed Implementation

[0027] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are all within the protection scope of the present invention.

[0028] The surge arrester diagnostic and early warning method with dual time-scale current and quality gating includes the following in step S1: Step S1-1: At the operation site of the surge arrester in the substation, establish an equipment file for the surge arrester to be monitored, and record the equipment identification, the bay to which it belongs, the voltage level, the phase, the model, the commissioning time, historical maintenance records, live-line testing records, and confirmed abnormal handling records, forming the basic equipment status information corresponding to the subsequent current diagnosis results.

[0029] Step S1-2: Install a current sensing module outside the measuring branch of the pointer-type leakage current meter for the surge arrester. The current sensing module includes a through-core zero-flux current sensor, a signal conditioning circuit, and an analog-to-digital converter circuit. The current sensing module is electromagnetically coupled to the measuring branch of the pointer-type leakage current meter without cutting off the original conductor, and without changing the series connection, metering loop, or indicating function of the pointer-type leakage current meter. The signal conditioning circuit is used for isolation, filtering, gain adjustment, and zero-bias correction. The analog-to-digital converter circuit operates according to the sampling frequency. Forming a current sampling sequence ,in For the conditioned sampling voltage, To calibrate the gain, Zero bias. Through bypass protection circuits, surge suppressors, or clamping protection circuits, if the induced voltage or current on the acquisition side exceeds the protection threshold... or It can switch on, discharge, or clamp at any time to protect the signal conditioning circuit and analog-to-digital conversion circuit.

[0030] Steps S1-3: Acquire the electrical current signal of the surge arrester leakage current meter measurement branch through a current sensing module, leakage current sensor, or online monitoring terminal electromagnetically coupled to the pointer-type leakage current meter measurement branch, forming current sampling values ​​covering an integer number of power frequency cycles, and organizing the current sampling values ​​obtained in one acquisition into a periodic waveform packet. The diagnostic input is an electrical current signal or numerical current data formed therefrom.

[0031] Steps S1-4: Associate the periodic waveform packet with the device identifier, phase, timestamp, sampling frequency, sampling start and end time, and acquisition terminal status. The acquisition terminal status includes one or more of the following: communication signal strength, power supply status, sampling gain, range status, and impulse protection status.

[0032] Steps S1-5: For low-power remote acquisition scenarios, the acquisition terminal uploads data in a timed acquisition, event-triggered acquisition, or a combination of timed and event-triggered acquisition methods; the station server, edge gateway, or remote monitoring platform caches, deduplicates, verifies integrity, and arranges the time sequence of the uploaded data to obtain the original set of periodic waveform packets for subsequent steps.

[0033] Combination Figure 2 Preferably, step S2 includes: Step S2-1: Power Frequency Cycle Location. The original current sequence in the periodic waveform packet is de-DC biased, and the power frequency cycle boundary is determined by zero-crossing location, phase-locked loop estimation, or fundamental spectrum peak estimation, thereby obtaining one or more complete power frequency cycle segments.

[0034] Step S2-2: Integer period segmentation. The current sampling sequence is segmented according to the power frequency period boundary. If the period integrity of a certain segment is lower than the preset integrity threshold, the segment is marked as an incomplete period and is not directly used for harmonic feature calculation, but its missing information is used in waveform quality calculation.

[0035] Step S2-3: Equal-point resampling. For each complete power frequency cycle, phase interpolation is used to map it to the same number of phase sampling points, ensuring that the periodic waveforms under different acquisition times, different sampling frequencies, or different small frequency offsets have uniform phase coordinates. Typically, each power frequency cycle can be resampled to 64, 128, or 256 phase sampling points.

[0036] Step S2-4: Marking Abnormal Sampling Points. Saturation points, clipping peaks, consecutive freeze points, missing points, and transition points where adjacent rates of change exceed the physical upper limit are marked in the periodic waveform. For short missing segments, they can be filled in using in-phase sampling points from adjacent periods or interpolation between adjacent phases; for long missing segments, no fictitious filling is performed, but the missing mask is retained.

[0037] Step S2-5: Amplitude normalization and anti-interference processing. The amplitude of the resampled periodic waveform is normalized, and one or more of the following methods can be used to suppress non-target frequency band interference and local impulse noise: bandpass filtering, wavelet thresholding denoising, variational mode decomposition, notch filtering, or median filtering.

[0038] Step S2-6: Waveform quality calculation. Calculate waveform quality based on period integrity, saturation distortion ratio, sampling clock jitter, interference band energy ratio, and missing sampling ratio. The formula for its calculation is: (1) In the formula, Indicates waveform quality. This indicates the proportion of incomplete power frequency cycles. Indicates the proportion of saturated or clipped sampling points. This represents the normalized amount of jitter during the sampling interval. This indicates the proportion of non-target interference frequency band energy. Indicates the proportion of missing samples. to The non-negative weighting coefficients; This is a dimensionless quality score, ranging from 0 to 1; when waveform branches are available, Greater than 0; when waveform branches are unavailable, The value is 0; the larger the value, the higher the waveform quality.

[0039] Preferably, step S3 includes: Step S3-1: Construction of Periodic Waveform Tensor. Multiple power frequency periodic waveforms, after period localization, integer period segmentation, equal-point resampling, and normalization, are organized according to the acquisition time, period number, and phase sampling point order to construct a periodic waveform tensor. It is used to preserve the local waveform morphology and harmonic modes within the power frequency cycle.

[0040] Step S3-2: Packet-level current summary index extraction. For the first... Resampled current sequence of periodic waveform packets Calculate the packet-level current summary vector ,in For valid values, Peak value, Peak-to-peak value For waveform factor, As the peak factor, The ratio of the third harmonic amplitude to the fundamental frequency amplitude. The ratio of the fifth harmonic amplitude to the fundamental frequency amplitude. The total harmonic distortion (THD) is the harmonic distortion rate. This represents the offset relative to the equipment health baseline. The harmonic index, as a proxy for resistive current variation, characterizes the nonlinear variation of the total leakage current and is not directly equivalent to the precise resistive current in the absence of voltage phase information.

[0041] Step S3-3: Constructing the long-term trend tensor. This involves... The packet-level current summary vectors at each acquisition time are arranged in chronological order to form a long-term trend tensor. Further, multi-window derivations with different window lengths are added to this long-term trend tensor. These multi-window derivations include increments. Robust linear slope volatility Continuous overtime Historical quantile position and drift relative to the healthy baseline; assuming the actual number of effective data collection times is... .when Less than and not less than When the time frame is reached, the long-term trend tensor is completed according to the original time axis. A time-by-time missing mask is set for the missing positions, and the actual time interval is recorded. The increment, robust slope, and volatility are calculated only based on the valid time frame. Less than At that time, the actual observation is not replaced with a filler value, and the trend branch is marked as unavailable.

[0042] Step S3-4: Trend Quality Calculation. Normalize the package missing ratio, reporting interval dispersion, abnormal summary ratio, and healthy baseline instability within the long-term trend window. Calculate the trend quality based on the normalized quality defect quantity. The This is a dimensionless quality score ranging from 0 to 1, with higher values ​​indicating higher quality trend data. Not less than and When the preset trend availability threshold is reached, let Set to 1; otherwise, set to 1. It is 0.

[0043] Step S3-5: Low-frequency data compatibility processing. Set waveform branch available markers. Trend branches can be marked When periodic waveform packets are unavailable, sampling frequency... Less than the frequency threshold required for harmonic analysis, or period integrity Below the preset integrity threshold At that time, without calculating the periodic waveform tensor and harmonic waveform characteristics, let 0, Set to 0 and retain the missing marker; when Less than season 0, =0; when Not less than But trend quality When the value is below the preset trend availability threshold, let The value is 0. When only the trend branch is available, the diagnosis is performed by the trend branch; when only the waveform branch is available, the diagnosis is performed by the waveform branch; when neither branch is available, only an insufficient data or acquisition link retest prompt is output.

[0044] Combination Figure 3 Preferably, step S4 includes: Step S4-1: Construct a multi-receptive-field one-dimensional convolutional encoder. The multi-receptive-field one-dimensional convolutional encoder includes... Each with different kernel size Or different expansion rates One-dimensional convolution branches; for example, parallel convolution branches with kernel sizes of 3, 7 and 15 can be set, or dilated convolution branches with dilation rates of 1, 2 and 4 can be set to extract short-neighborhood waveform details and waveform morphology across a longer phase range, respectively.

[0045] Step S4-2: Waveform representation extraction. Extract the periodic waveform tensor. Input the multi-receptive-field one-dimensional convolutional encoder, the first Branch pairs Branch features are obtained by performing convolution, nonlinear transformation, channel mapping, and pooling. The features of each branch are spliced ​​together and then projected using a function. Waveform characterization The waveform characterization is used to describe the local morphology, distortion mode, phase-dependent fluctuations, and harmonic modes of periodic waveforms.

[0046] Step S4-3: Calculate the waveform branch uncertainty. Estimate the original waveform branch uncertainty through random inactivation repeated inference, deep integration, output entropy, or prediction variance. The method employs at least one of the following: maximum entropy proportional normalization, pruning normalization based on robust quantiles of training or validation data, or monotonic compression, to... Normalization ,make The value range is 0 to 1; when the periodic waveform has clipping, strong interference, or missing period, Increase.

[0047] Step S4-4: Construct a gated cyclic trend encoder. Convert the long-term trend tensor... The input is a gated cyclic trend encoder, which may employ a gated cyclic unit, a long short-term memory network, or a bidirectional cyclic structure, and processes the long-term trend tensor according to the order of acquisition time. .

[0048] Step S4-5: Trend representation extraction. The gated cyclic trend encoder introduces a time-by-time missing mask during the attention convergence process. and quality mask Attention weights are calculated based on the hidden state at each time step, the missing mask, and the quality mask. This ensures that missing moments, anomalous summary moments, or low-quality moments do not participate in or reduce their contribution to trend convergence, ultimately yielding a trend representation. and the uncertainty of the original trend branch The same normalization principle as the waveform branching is used to... Normalization ,make The value range is from 0 to 1.

[0049] Preferably, step S5 includes: Step S5-1: Characterize the waveform and trend representation respectively through projection function and Mapping to the same dimension yields the waveform projection representation. and trend projection representation .

[0050] Step S5-2: Based on waveform quality Trend Quality Normalized waveform branch uncertainty Normalized trend branch uncertainty and branch available markers and Calculate the waveform reliability sequentially. Trend reliability Waveform evidence quantity Trend Evidence Quantity and gating weights and The formula for its calculation is:

[0051] In the formula, and These represent waveform reliability and trend reliability, respectively. and These represent the waveform evidence quantity and the trend evidence quantity, respectively. and These represent the waveform branch gating weights and the trend branch gating weights, respectively. and For a dimensionless quality score with values ​​ranging from 0 to 1, and For the normalized uncertainty ranging from 0 to 1, and A flag can be used to indicate branches that take the value 0 or 1. It has a positive temperature coefficient.

[0052] Step S5-3: When the total amount of available evidence exceeds the preset evidence threshold At that time, based on the gating weight and Waveform characterization and trend representation We perform weighted fusion to obtain the fusion representation. The formula for its calculation is:

[0053] In the formula, Indicates fusion representation, Indicates waveform representation, Indicates trend representation, and These represent projection functions that map waveform representations and trend representations to the same dimension, respectively.

[0054] when =0 and When it is 1, directly set 0, =1; when =0 and When it is 1, directly set For 1, =0; when both branches are unavailable or and The sum is not greater than the preset evidence threshold. At this time, deterministic device status diagnosis is not performed; only a message indicating insufficient data or retesting of the acquisition link is output.

[0055] Step S5-4: The quality-gated fusion process retains the dual directionality of quality and uncertainty: reduced quality will reduce the reliability of branches, and increased uncertainty will also reduce the reliability of branches; the amount of evidence for unusable branches is strictly 0, thereby suppressing the unreasonable contribution of low-quality, high-uncertainty, or unusable branches to the diagnostic results.

[0056] Preferably, step S6 includes: Step S6-1: Construct the diagnostic output layer and divide the model operation phase. The diagnostic output layer is used to output the arrester state probability and fault risk probability based on the fused characterization. The state probabilities include at least four types of state probabilities: normal, alert, abnormal, and severe warning. The model operation phase includes an offline training phase and an online diagnostic phase.

[0057] Step S6-2: Construct a sample quality-weighted classification loss. The classification loss is calculated based on sample labels, class weights, and sample quality weights, ensuring that outlier samples, high-quality samples, and high-confidence labeled samples confirmed through experimentation or verification make a more reasonable contribution to model training.

[0058] Step S6-3: Construct the degradation monotonicity loss. For samples of the same surge arrester at two different times, the degradation characteristic at the later time relative to the earlier time is not less than a preset increment. Furthermore, if there are no repair or repositioning events during this period, and the probability of failure at a later time is lower than the probability of failure at a previous time and exceeds the allowable margin... If so, a penalty is imposed on the model. The degradation characteristics include one or more of the following: RMS leakage current, baseline offset, third harmonic to fundamental amplitude ratio, total harmonic distortion rate, duration of continuous over-limit, and trend slope.

[0059] Step S6-4: Construct the state transition loss. Based on the permissible state transition matrix between normal, attention, abnormal, and severe warning states. and penalty matrix Calculate the state transition loss; if from state to state If the transfer is allowed, then If the abnormal state or severe warning state directly transitions to the normal state without any confirmed repair or relocation event, then A positive penalty coefficient is applied, and the corresponding penalty is lifted when a confirmed maintenance, replacement, or test reset event occurs.

[0060] Step S6-5: Construct three-phase consistency loss. The three-phase consistency loss applies to three-phase surge arresters in the same group operating under the same voltage level, same interval, or same busbar conditions. During the health label period, the distance between the common changes in the three phases is constrained to not exceed a preset consistency threshold; during the confirmed single-phase abnormal period, the abnormal phase is not forced to be consistent with the other two phases, but rather the deviation of the abnormal phase relative to other phases is retained as diagnostic evidence through inter-phase residuals or interval margins.

[0061] Step S6-6: Train the diagnostic model using a joint loss function during the offline training phase. Iteratively update the model parameters of the dual-branch encoder, quality gating module, and diagnostic output layer using historical labeled samples; the joint loss function includes a sample quality-weighted classification loss. Deterioration of monotonicity loss State transition loss and three-phase consistency loss The formula for its calculation is:

[0062] In the formula, Weighted classification loss based on sample quality. To mitigate monotonicity loss, For state transition loss, For three-phase consistency loss, These are non-negative weighting coefficients.

[0063] Steps S6-7: Deploy the trained diagnostic model and fix the model parameters during the online diagnosis phase, and fuse the representation of the current diagnostic window. The input diagnostic output layer performs forward computation to obtain the state probability and fault risk probability; the online diagnostic phase does not calculate the joint loss and does not perform backpropagation or model parameter updates.

[0064] Combination Figure 4 Preferably, step S7 includes: Step S7-1: Continuously determine the state probability and fault risk probability within the continuous diagnostic window. When continuous... At least one diagnostic window Failure risk probability of each diagnostic window Reaching the upgrade threshold At that time, the early warning status will be upgraded, including Less than or equal to .

[0065] Step S7-2: Determine the hysteresis of the early warning status. When continuous Failure risk probability of each diagnostic window Below the downgrade threshold At that time, the warning status will be lowered; among them, Higher than , and These are the number of upgrade decision windows and the number of downgrade decision windows, respectively. This is the minimum number of windows required to reach the upgrade threshold in the upgrade determination process, in order to suppress alarm jitter caused by isolated strong interference, short-term communication anomalies, or transient noise.

[0066] Step S7-3: When both branches are unavailable, the waveform quality and trend quality are both below the preset quality threshold, or the total amount of available evidence is not greater than the preset evidence threshold. Instead of directly raising the equipment fault level, it generates a prompt for insufficient data or retesting of the data acquisition link; when at least one branch is available, there is sufficient available evidence, and the fault risk continues to rise, it outputs a normal, warning, abnormal, or severe warning status.

[0067] Step S7-4: Review the continuous diagnostic results according to the state transition rule, degradation monotonicity rule, and three-phase consistency rule. If the model output conflicts with the domain knowledge constraints, it is corrected or marked as requiring manual review based on the model confidence, quality score, and rule triggering strength.

[0068] Preferably, step S8 includes: Step S8-1: Output the warning status and health score. Health Score The value ranges from 0 to 100, with lower values ​​indicating a higher risk of arrester degradation. One possible calculation formula is:

[0069] In the formula, Indicates health score, Indicates the probability of failure. The value ranges from 0 to 100, and the lower the value, the higher the risk of the surge arrester deteriorating.

[0070] In other implementations, a health score can be calculated by weighting the probability of a state based on the risk coefficient corresponding to that state.

[0071] Step S8-2: Output the key period waveform and key trend interval. The key period waveform includes one or more power frequency cycles that contribute most to the diagnosis and their waveform quality. The key trend interval includes the long-term trend window, trend slope, and duration of continuous exceedance that contribute most to the diagnosis.

[0072] Step S8-3: Output the triggered degradation characteristics and data quality information. The degradation characteristics include one or more of the following: RMS value, third harmonic to fundamental amplitude ratio, total harmonic distortion rate, baseline offset, phase residual, duration of continuous out-of-limit, and trend slope; the data quality information includes waveform quality, trend quality, missing proportion, interference proportion, normalized branch uncertainty, branch reliability, and evidence quantity.

[0073] Step S8-4: Output maintenance recommendations. If the effective value of leakage current, health baseline deviation, and trend slope continue to increase, it is recommended to shorten the inspection cycle and carry out live-line testing; if the third harmonic ratio and total harmonic distortion rate continue to increase, it is recommended to check the risk of valve plate deterioration, moisture absorption, or sealing failure in conjunction with operating voltage and temperature and humidity; if the three phases change significantly, prioritize checking the common operating conditions; if only a single phase deviates significantly, prioritize checking the surge arrester of that phase and its acquisition branch; if the data quality remains low, prioritize checking the sensors, signal conditioning, sampling clock, and communication link.

[0074] Step S8-5: Write each diagnostic result, warning status, diagnostic evidence, manual review conclusion, and repair / replacement event into the equipment status file. Confirmed samples can be used for periodic model retraining after meeting data volume and quality requirements; the retrained model should be verified through an independent equipment test set, version rollback mechanism, and on-site review process before redeployment.

[0075] Preferably, after completing step S8, a verification step is also included: Step 1: Divide the training set, validation set, and test set according to the device number, so that adjacent time samples from the same device do not enter the training set and test set at the same time, in order to avoid information leakage caused by adjacent time samples.

[0076] Step 2: Set up a comparison scheme that includes at least the following: a traditional classification model that uses only package-level statistical features, a one-dimensional convolutional model that uses only periodic waveforms, a recurrent model that uses only long-term trends, a direct splicing model without quality gating, and a bi-branch model without knowledge constraints.

[0077] Step 3: Accuracy, outlier recall, macro average F1 score, false alarm rate, false negative rate, probability calibration error, and early warning time are used as evaluation indicators. The role of each technical feature is verified by ablation tests that remove waveform quality, trend quality, normalized branch uncertainty, branch reliability, branch availability mask, degradation monotonicity loss, state transition loss, and three-phase consistency loss.

[0078] The above are merely embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for diagnostic warning of a surge arrester with double time scale current and mass gating, characterized in that, Includes the following steps: S1: Acquire the electrical current signal of the surge arrester leakage current meter measurement branch, organize the current sampling values ​​obtained in one acquisition that cover an integer number of power frequency cycles into a periodic waveform packet, and associate the periodic waveform packet at different acquisition times with the equipment identifier, phase and timestamp; the electrical current signal is output by a current sensing module, leakage current sensor or online monitoring terminal that is electromagnetically coupled to the pointer-type leakage current meter measurement branch; S2: performing power frequency cycle positioning, integer cycle segmentation, equal point number resampling, DC offset removal, abnormal sampling point marking and amplitude normalization on the periodic waveform package, calculating waveform quality according to cycle integrity, saturation distortion ratio, sampling clock jitter, interference frequency band energy ratio and missing sampling ratio , and generating a waveform branch available label ; S3: constructing a periodic waveform tensor from the preprocessed periodic waveforms and extracting packet-level current summary indicators from each periodic waveform packet, and grouping the packet-level current summary indicators within a preset trend window into a long-term trend tensor , calculating trend quality according to packet loss ratio, reporting interval dispersion, abnormal summary ratio, and baseline stability and generating a trend branch availability label according to the number of valid collection time points and the trend quality ; S4: Transfer the periodic waveform tensor Input a multi-receptive-field one-dimensional convolutional encoder to obtain waveform representations. and the branch uncertainty of the original waveform and the original waveform branch uncertainty Normalization ; the long-term trend tensor Input a gated cyclic trend encoder to obtain trend representation. and the uncertainty of the original trend branch and the uncertainty of the original trend branch Normalization ; S5: calculating the reliability, evidence quantity and gating weight of the two branches according to the waveform quality , the trend quality , the normalized waveform branch uncertainty , the normalized trend branch uncertainty and the branch available flag and , and fusing the waveform representation and the trend representation by using the gating weight to obtain a fused representation ; S6: During the offline training phase, a joint loss function including classification loss, degradation monotonicity loss, state transition loss, and three-phase consistency loss is used to iteratively update the model parameters of the dual-branch encoder, quality gating module, and diagnostic output layer based on historical labeled samples, resulting in a trained diagnostic model. During the online diagnostic phase, the model parameters of the trained diagnostic model are fixed, and the fused representation of the current diagnostic window is input into the diagnostic model to perform forward computation, obtaining the arrester state probability and fault risk probability. The degradation monotonicity loss is constructed based on the difference between the degradation representation increment and the risk probability, and is used to penalize the prediction that the fault risk probability decreases in the opposite direction when there is no repair event and the current degradation representation continues to increase. The state transition loss constrains the state changes at adjacent diagnostic times to conform to a preset state transition relationship. S7: Perform persistence and hysteresis determination on the state probability and fault risk probability within the continuous diagnostic window to generate normal, attention, abnormal or severe warning status; S8: Outputs early warning status, health score, key period waveforms, key trend intervals, triggered degradation characteristics, data quality information, and operation and maintenance suggestions.

2. The method according to claim 1, characterized in that, The current sensing module in step S1 includes a through-hole zero-flux current sensor, a signal conditioning circuit, and an analog-to-digital converter circuit. The through-hole zero-flux current sensor is sleeved outside the conductor of the measuring branch of the pointer-type leakage current meter. The signal conditioning circuit isolates, filters, adjusts the gain, and performs zero-bias correction on the sensor output. The analog-to-digital converter circuit adjusts the output according to the sampling frequency. Obtain the current sampling sequence , in For the conditioned sampling voltage, To calibrate the gain, The bias is zero; and through bypass protection circuits, surge suppressors, or clamping protection circuits, the induced voltage or induced current on the acquisition side exceeds the protection threshold. or It can conduct, discharge, or clamp at any time to suppress the overvoltage and overcurrent induced by the lightning strike current on the acquisition side, protect the signal conditioning circuit and analog-to-digital conversion circuit, and ensure that the current sensing module does not change the original metering circuit and indicating function of the pointer-type leakage current meter.

3. The method according to claim 1, characterized in that, The waveform quality described in step S2 Calculate using the following formula: (1) In the formula, Indicates waveform quality. This indicates the proportion of incomplete power frequency cycles. Indicates the proportion of saturated or clipped sampling points. This represents the normalized amount of jitter during the sampling interval. This indicates the proportion of non-target interference frequency band energy. Indicates the proportion of missing samples. to These are non-negative weighting coefficients; The equal-point resampling maps each power frequency cycle to the same number of phase sampling points; This is a dimensionless quality score, with a value ranging from 0 to 1. When waveform branches are available Greater than 0; when waveform branches are unavailable, The value is 0; the larger the value, the higher the waveform quality.

4. The method according to claim 1, characterized in that, In step S3, the first Resampled current sequence of periodic waveform packets Calculate the packet-level current summary vector ;in, For valid values, Peak value, Peak-to-peak value For waveform factor, As the peak factor, This represents the ratio of the third harmonic amplitude to the fundamental frequency amplitude. This represents the ratio of the fifth harmonic amplitude to the fundamental frequency amplitude. Represents the total harmonic distortion rate. Represents the offset relative to the device health baseline; the long-term trend tensor consists of continuous... The packet-level current summary vector at each acquisition time and its multi-window derivation constitute the multi-window derivation, wherein the multi-window derivation includes at least the increment. robust slope volatility Historical quantile position and duration of continuous over-limit ; Let the preset trend window length be The actual number of effective data acquisition times is The preset minimum number of valid data acquisitions is... ;when Less than and not less than At that time, the input is padded to the original timeline. Each time point is defined with a time-by-time missing mask and an actual time interval. The multi-window derived quantity is based solely on... Calculate the number of valid acquisition times; when Not less than and When the preset trend availability threshold is reached, mark the trend branches as available. =1; when Less than or When the value is below the preset trend availability threshold, let =0, where when Less than Time still order The value is 0, which lowers the confidence level of the diagnostic results.

5. The method according to claim 1, characterized in that, The multi-receptive-field one-dimensional convolutional encoder mentioned in step S4 includes The nth one-dimensional convolutional branch, the th Each one-dimensional convolutional branch has a convolutional kernel size. or expansion rate Each branch with respect to the periodic waveform tensor Branch features are obtained after convolution, nonlinear transformation, channel mapping, and pooling. The waveform representation is obtained by splicing the features of each branch and then projecting them. The gated cyclic trend encoder processes the long-term trend tensor according to the order of acquisition time. And introduce a missing mask and quality mask Calculate attention weights This reduces the contribution of missing or low-quality moments to trend convergence, ultimately leading to a trend representation. ; The original waveform branch uncertainty and the uncertainty of the original trend branch It is calculated from at least one of output entropy, repeated inference variance, or model integration variance; By using at least one of the following methods: maximum entropy proportional normalization, pruning normalization based on robust quantiles of training or validation data, or monotonic compression, respectively... and Normalization and ,in and All values ​​are dimensionless uncertainties ranging from 0 to 1, with larger values ​​indicating higher uncertainties.

6. The method according to claim 1, characterized in that, The waveform reliability described in step S5 Trend reliability Waveform evidence quantity Trend Evidence Quantity Gating weights and and fusion characterization satisfy: In the formula, Indicates waveform reliability. Indicates the reliability of the trend. Indicates the amount of waveform evidence. Indicates the amount of evidence for a trend. Indicates the waveform branch gating weight. Indicates the trend branch gating weight. and For a dimensionless quality score with values ​​ranging from 0 to 1, and For the normalized uncertainty ranging from 0 to 1, and A flag can be used to indicate branches that take the value 0 or 1. It has a positive temperature coefficient; and The value range is from 0 to 1. The preset evidence threshold is non-negative; when the total amount of available evidence is no greater than... At this time, no deterministic device status diagnosis is output; only a message indicating insufficient data or retesting of the acquisition link is output. Indicates fusion representation, and These represent waveform representation and trend representation, respectively. and These represent projection functions that map two types of representations to the same dimension.

7. The method according to claim 1, characterized in that, The joint loss function described in step S6 satisfies: (9) In the formula, Weighted classification loss based on sample quality. To mitigate monotonicity loss, For state transition loss, For three-phase consistency loss, The weighting coefficients are non-negative; for two samples of the same surge arrester, the increment of the degradation characteristic at a later time step relative to the previous time step is not less than the preset increment. Furthermore, if no repair or repositioning event occurs during this period, the degradation monotonicity loss results in a lower probability of failure at a later time than at a previous time, exceeding the allowable margin. Punishment will be imposed in such circumstances.

8. The method according to claim 7, characterized in that, The state transition loss is based on the permissible state transition matrix between normal, alert, abnormal, and severe warning states. and penalty matrix Calculation, if based on state to state If the transfer is allowed, then If the abnormal state or severe warning state directly changes to the normal state in the absence of a repair or restoration event, then A positive penalty coefficient is applied, and if a confirmed repair and repositioning event exists, the corresponding penalty is lifted. The three-phase consistency loss applies to the same group of three-phase surge arresters under the same voltage level, the same interval, or the same busbar operating conditions. During the health tag period, the distance between the common changes in the three phases is minimized, and during the single-phase abnormal tag period, the deviation of the abnormal phase relative to the other two phases is retained through the phase-to-phase residual or interval margin.

9. The method according to claim 1, characterized in that, In step S7, when consecutive At least one diagnostic window Failure risk probability of each diagnostic window Reaching the upgrade threshold The warning status is raised in a timely manner; when it is continuous Failure risk probability of each diagnostic window Below the downgrade threshold The warning status was lowered at that time, among which Less than or equal to , Higher than When the data quality is lower than the preset quality threshold, the equipment fault level is not directly increased, but a retest prompt for the data acquisition link is generated instead.

10. The method according to claim 1, characterized in that, Set waveform branch available markers Trend branches can be marked ; When periodic waveform packets are unavailable, sampling frequency Less than the frequency threshold required for harmonic analysis, or period integrity Below the preset integrity threshold season 0, Set to 0 and retain the missing marker; When the actual effective data collection time Less than the preset minimum number of valid samples season 0, =0; when Not less than but When the value is below the preset trend availability threshold, let =0; when =0 and When it is 1, directly set 0, Setting it to 1 degrades the gated fusion to a diagnostic process based on the long-term trend tensor input trend branch; when =0 and When it is 1, directly set For 1, The value is 0; when both branches are unavailable or the total amount of available evidence is not greater than the preset evidence threshold. At this time, no deterministic device status diagnosis is output; only a data shortage or acquisition link retest prompt is output, and the diagnostic confidence level is reduced.

11. A substation surge arrester diagnostic and early warning system, characterized in that, The system includes a current signal acquisition module, a periodic waveform processing module, a trend sample construction module, a quality assessment module, a dual-branch coding module, a quality gating fusion module, a knowledge constraint diagnosis module, and an early warning output module. The current signal acquisition module acquires the electrical current signal of the surge arrester leakage current meter measurement branch and generates periodic waveform packets or low-frequency trend data. The quality assessment module calculates waveform quality, trend quality, and missing data markers. The quality gating fusion module calculates reliability, evidence quantity, and gating weights based on quality scores, normalized branch uncertainty, and branch availability markers. The knowledge constraint diagnosis module performs classification, degradation monotonicity, state transition, and three-phase consistency constraints. The early warning output module performs persistence window and hysteresis threshold determination. Each module is configured to perform the method described in any one of claims 1 to 10.

12. An electronic device, comprising a processor, a memory, and a communication interface, characterized in that, The memory stores computer programs, model parameters, quality thresholds, and warning thresholds. When the computer program is executed by the processor, it causes the electronic device to receive electrical current signals or trend data from the leakage current meter measuring branch of the surge arrester and to implement the method described in any one of claims 1 to 10.

13. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program includes data preprocessing instructions, dual-timescale feature construction instructions, quality gating fusion instructions, knowledge constraint diagnosis instructions, and continuous window warning instructions. When the computer program is executed by the processor, it implements the method described in any one of claims 1 to 10.

Images