A power distribution line arrester on-line monitoring method and system

Abstract

The application provides a power distribution line lightning arrester online monitoring method and system, and relates to the technical field of data processing.The method comprises the following steps: distributing each sampling point into a corresponding convex region unit; calculating the surface area of each region unit based on the distribution density and spatial structure characteristics of the sampling points in each unit; generating a region density index representing local distribution characteristics; weighting and optimizing the acoustic characteristic parameters by using the region density index; performing multi-modal feature fusion on the optimized acoustic characteristic parameters and current characteristic indexes to form a high-dimensional joint feature matrix; inputting the feature matrix into a one-dimensional convolutional neural network state recognition model which has been pre-trained to perform intelligent diagnosis; and outputting the classification result of the lightning arrester operating state.The application improves the accuracy and reliability of early fault identification of the lightning arrester.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for online monitoring of surge arresters in power distribution lines. Background Technology

[0002] Existing online monitoring technologies for surge arresters mainly rely on the acquisition and analysis of single physical quantities. For example, leakage current monitoring is used to determine the degree of insulation degradation, or infrared thermal imaging technology is used to detect local overheating defects. Although some systems have achieved multi-parameter acquisition (such as leakage current, temperature, number of operations, etc.), there are still significant limitations in feature processing and state recognition. In practical applications, early faults of surge arresters often manifest as subtle changes in acoustic signals (such as abnormal ultrasonic spectra caused by partial discharge) and weak fluctuations in leakage current. Traditional methods have failed to effectively integrate multi-modal features, resulting in insufficient fault identification sensitivity.

[0003] Specifically, the key flaws of existing technology are:

[0004] The processing of acoustic features lacks the ability to analyze high-dimensional spatial distribution, making it difficult to effectively distinguish between normal operation fluctuations and early fault signals. Traditional methods typically employ linear dimensionality reduction techniques such as Principal Component Analysis (PCA) to process acoustic spectral data. However, these methods are prone to losing subtle local features when dealing with high-dimensional nonlinear acoustic signals. Furthermore, the feature fusion process fails to consider the differences in the discriminative value of features in different regions, treating all sampling points with equal weight, resulting in fault features being submerged in the fluctuations of normal data. For example, in early water ingress defects caused by poor sealing of surge arresters, the acoustic signal energy generated by partial discharge is weak, and its characteristics are easily masked by environmental noise or normal equipment noise. Existing technologies struggle to accurately identify such early defects, only issuing an alarm when the leakage current significantly increases, thus missing the optimal maintenance opportunity. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide an online monitoring method and system for surge arresters in power distribution lines, which improves the accuracy and reliability of early fault identification of surge arresters.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0007] A method and system for online monitoring of surge arresters in power distribution lines, the method comprising:

[0008] Acquire acoustic spectrum signals and leakage current waveform data generated by the surge arrester during operation;

[0009] Synchronous signal preprocessing is performed on the acoustic spectrum signal and leakage current waveform data to extract acoustic feature parameters and current feature indices, and the acoustic feature parameters are converted into corresponding spatial scatter datasets through high-dimensional feature mapping.

[0010] A two-dimensional feature mapping coordinate system is constructed based on the spatial scattered point dataset, and a minimum boundary range that can enclose all sampling points is determined in this coordinate system;

[0011] Based on the spatial distribution characteristics of the sampling points within the minimum boundary range, the boundary range is divided into regions to generate a series of non-overlapping convex region units.

[0012] Each sampling point is assigned to a corresponding convex region unit. Based on the distribution density and spatial structure characteristics of the sampling points in each unit, the surface area of ​​each region unit is calculated to generate a region density index that characterizes the local distribution characteristics.

[0013] The acoustic feature parameters are weighted and optimized using the region density index. The optimized acoustic feature parameters are then fused with the current feature index in a multimodal manner to form a high-dimensional joint feature matrix. This feature matrix is ​​then input into a pre-trained one-dimensional convolutional neural network state recognition model for intelligent diagnosis, and the classification result of the arrester's operating state is output.

[0014] Based on the classification results, it is determined whether the surge arrester is in an abnormal working state. If it is determined to be abnormal, the alarm information generation program is started and an abnormal state warning signal is output.

[0015] An online monitoring system for surge arresters in power distribution lines includes:

[0016] The acquisition module is used to acquire acoustic spectrum signals and leakage current waveform data generated by the surge arrester during operation;

[0017] The conversion module is used to perform synchronous signal preprocessing on the acoustic spectrum signal and leakage current waveform data to extract acoustic feature parameters and current feature indicators, and convert the acoustic feature parameters into corresponding spatial scatter datasets through high-dimensional feature mapping.

[0018] The construction module is used to construct a two-dimensional feature mapping coordinate system based on the spatial scattered point dataset, and to determine a minimum boundary range in the coordinate system that can enclose all sampling points;

[0019] The partitioning module is used to divide the boundary range into regions based on the spatial distribution characteristics of the sampling points within the minimum boundary range, generating a series of non-overlapping convex region units.

[0020] The calculation module is used to assign each sampling point to the corresponding convex region unit, calculate the surface area of ​​each region unit based on the distribution density and spatial structure characteristics of the sampling points in each unit, and generate a region density index that characterizes the local distribution characteristics.

[0021] The prediction module is used to perform weighted optimization of the acoustic feature parameters using the region density index, fuse the optimized acoustic feature parameters with the current feature index for multimodal features to form a high-dimensional joint feature matrix, and input the feature matrix into a pre-trained one-dimensional convolutional neural network state recognition model for intelligent diagnosis, outputting the classification result of the arrester's operating state; based on the classification result, it determines whether the arrester is in an abnormal operating state, and if it is determined to be abnormal, it starts the alarm information generation program and outputs an abnormal state warning signal.

[0022] The above-described solution of the present invention has at least the following beneficial effects:

[0023] Referring to the acquisition and processing logic of acoustic signature features in the online monitoring system for power transformer acoustic signatures, this method simultaneously acquires the acoustic spectrum signal and leakage current waveform data of the surge arrester. Combined with high-dimensional feature mapping and regional density analysis, it can capture early subtle fault signals that are difficult to detect by traditional single-parameter monitoring, thus avoiding delays in fault investigation.

[0024] Based on the reliability characteristics of power frequency and pulse magnetic field immunity tests, this method reduces the impact of environmental noise and electromagnetic interference on single features by using multimodal feature fusion and a one-dimensional convolutional neural network model, thereby reducing false alarms and false negatives and improving the accuracy of status diagnosis. Continuing the system's advantages in data storage and feature display, this method can automatically activate the alarm program after an anomaly is detected, and output warning signals in a timely manner, facilitating rapid response by operation and maintenance personnel, avoiding power outages and other safety accidents caused by surge arrester failures, and ensuring the stable operation of the power grid. Attached Figure Description

[0025] Figure 1 This is a schematic flowchart of an online monitoring method for surge arresters in power distribution lines, provided by an embodiment of the present invention.

[0026] Figure 2 This is a schematic flowchart of an online monitoring system for surge arresters in power distribution lines, provided by an embodiment of the present invention. Detailed Implementation

[0027] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0028] like Figure 1 As shown in the figure, an embodiment of the present invention proposes an online monitoring method and system for surge arresters in power distribution lines. The method includes the following steps:

[0029] Step S100: Acquire the acoustic spectrum signal and leakage current waveform data generated by the surge arrester during operation;

[0030] Step S200: Perform synchronous signal preprocessing on the acoustic spectrum signal and leakage current waveform data to extract acoustic feature parameters and current feature indices, and convert the acoustic feature parameters into corresponding spatial scatter datasets through high-dimensional feature mapping.

[0031] Step S300: Construct a two-dimensional feature mapping coordinate system based on the spatial scattered point dataset, and determine a minimum boundary range in the coordinate system that can encompass all sampling points;

[0032] Step S400: Based on the spatial distribution characteristics of the sampling points within the minimum boundary range, the boundary range is divided into regions to generate a series of non-overlapping convex region units.

[0033] Step S500: Assign each sampling point to the corresponding convex region unit, calculate the surface area of ​​each region unit based on the distribution density and spatial structure characteristics of the sampling points in each unit, and generate a region density index that characterizes the local distribution characteristics.

[0034] Step S600: The acoustic feature parameters are weighted and optimized using the region density index. The optimized acoustic feature parameters are then fused with the current feature index in a multimodal manner to form a high-dimensional joint feature matrix. The feature matrix is ​​then input into a pre-trained one-dimensional convolutional neural network state recognition model for intelligent diagnosis, and the classification result of the arrester's operating state is output.

[0035] Step S700: Based on the classification result, determine whether the surge arrester is in an abnormal working state. If it is determined to be abnormal, start the alarm information generation program and output the abnormal state warning signal.

[0036] In this embodiment of the invention, the acoustic spectrum signal and leakage current waveform data of the surge arrester can be collected simultaneously. By simultaneously preprocessing to extract the two types of features and performing multimodal fusion, the limitations of single data monitoring can be overcome, and the comprehensiveness and accuracy of feature extraction can be improved. Through high-dimensional feature mapping, convex region division and region density index calculation, abstract signal features can be transformed into concrete spatial distribution features, allowing for a more detailed characterization of the local characteristics of the surge arrester's operating state. State recognition is performed through a pre-trained one-dimensional convolutional neural network, which can replace the traditional manual judgment method, reduce subjective errors, and improve the automation level and reliability of anomaly diagnosis. After identifying an abnormal state, an alarm program can be quickly activated and a warning signal can be output, forming a closed loop. This promptly reminds maintenance personnel to troubleshoot the fault, preventing serious accidents such as power outages and equipment damage caused by surge arrester failure, and effectively ensuring the safe operation of the power distribution line.

[0037] In a preferred embodiment of the present invention, in order to effectively acquire acoustic spectrum signals and leakage current waveform data of the surge arrester in the operating state in step S100, specific operations need to be carried out in terms of acquisition equipment selection, installation and deployment, parameter configuration, synchronization control and anti-interference processing, etc. The specific implementation process is as follows:

[0038] First, the acquisition equipment is selected. For acoustic spectrum signal acquisition, an acoustic sensor with ultrasonic frequency band detection capability is chosen. This sensor needs to cover the ultrasonic frequency spectrum range that may be generated by partial discharge of the surge arrester to ensure that it can capture the weak ultrasonic signals at the early stage of the fault, while also having high sensitivity to distinguish between environmental noise and the equipment's own signals. For leakage current waveform data acquisition, a high-precision current sensor is selected. This sensor needs to have microampere-level detection accuracy to accurately capture the weak fluctuations of the surge arrester's leakage current, and it needs to be compatible with the current specifications of the surge arrester's grounding circuit to avoid data acquisition distortion due to range mismatch.

[0039] Next, the sensor installation and deployment are carried out. The selected acoustic sensors are installed in key parts of the surge arrester body, with priority given to the upper middle area of ​​the porcelain bushing or metal casing of the surge arrester. This area is close to the core insulating components inside the surge arrester and can more directly receive the acoustic signals generated by partial discharge. During installation, ensure that the sensor is in close contact with the surface of the surge arrester to reduce signal transmission loss. At the same time, a special shielding cover is added to the sensor to reduce the interference of external environmental noise (such as wind and rain noise, equipment operating noise) on the acquired signal. The current sensor is connected in series in the grounding circuit of the surge arrester, ensuring that the primary side of the current sensor is reliably connected to the grounding wire, and the secondary side is connected to the data acquisition unit through a shielded cable. The installation location should be far away from strong electromagnetic interference sources (such as high-voltage busbars and transformers) to avoid the influence of electromagnetic coupling on the leakage current acquisition data.

[0040] Then, the acquisition parameters are configured. For the acoustic spectrum signal, a suitable sampling frequency is set. The sampling frequency must be higher than twice the highest frequency of the ultrasonic spectrum to satisfy the sampling theorem. It is usually set to the range of tens of kilohertz to several megahertz. At the same time, a reasonable sampling duration is set. Each sampling duration must be able to cover at least one complete partial discharge signal cycle to avoid missing signal features due to insufficient sampling duration. In addition, the amplitude range of the signal acquisition must be set to ensure that weak early fault signals can be captured without data distortion due to signal overload. For the leakage current waveform data, a sampling frequency that matches the acoustic spectrum signal is set to ensure that the two types of signals are synchronized in the time dimension. The sampling accuracy is set to the microampere level to accurately reflect the slight changes in leakage current. At the same time, the sampling period is set. The sampling period needs to be determined according to the operating characteristics of the surge arrester to ensure continuous monitoring of leakage current fluctuations in different operating periods.

[0041] Then, synchronous acquisition control is implemented. A synchronous triggering module is integrated into the data acquisition unit. This module establishes signal connections with the acoustic sensor and the current sensor respectively. When the acquisition program is started, the synchronous triggering module sends acquisition start commands to both types of sensors simultaneously to ensure that the acoustic spectrum signal and leakage current waveform data have consistent timestamps from the moment the acquisition begins. This avoids deviations in subsequent feature correlation analysis due to differences in acquisition start time. At the same time, during the acquisition process, the synchronous triggering module monitors the acquisition status of the two types of sensors in real time. If an acquisition interruption or delay occurs, it immediately issues an abnormal prompt and restarts the acquisition process to ensure the continuity and synchronization of the acquired data.

[0042] Finally, anti-interference processing is carried out during the data acquisition process. In addition to installing shielding devices on the sensors, shielded cables are used on the data acquisition lines, and the cable joints are sealed to reduce the impact of external electromagnetic interference on the acquired signals. At the same time, a filtering circuit is set up inside the acquisition unit to specifically filter the acoustic spectrum signal, remove power frequency interference and high-frequency noise, and perform power frequency harmonic filtering on the leakage current waveform data to eliminate the interference of the power grid power frequency signal on the detection of weak fluctuations in leakage current. The acquired raw data is transmitted to the local storage module in real time for temporary storage. A data verification mechanism is used during the storage process to ensure that the data is not lost or tampered with during transmission and storage.

[0043] In a preferred embodiment of the present invention, step S200 involves performing synchronous signal preprocessing on the acoustic spectrum signal and leakage current waveform data to extract acoustic feature parameters and current feature indices, and converting the acoustic feature parameters into a corresponding spatial scatter dataset through high-dimensional feature mapping, including:

[0044] Step S210 involves performing time synchronization calibration on the synchronously acquired acoustic signal and leakage current signal to obtain time-aligned acoustic signal sequences and leakage current signal sequences. Specifically, this includes: determining the time reference for time synchronization, using the high-precision system clock of the data acquisition unit as a unified time reference. This clock needs to be pre-calibrated with the power grid standard time to ensure that the time accuracy reaches the microsecond level, providing an accurate basis for subsequent signal alignment; then extracting feature anchor points from the acquired leakage current signal sequence, selecting the zero-crossing point within each cycle of the leakage current signal as the anchor point, because the zero-crossing point of the leakage current has obvious signal change characteristics, is easy to accurately identify, and each zero-crossing point corresponds to a fixed moment in the power grid cycle, providing a stable time marker.

[0045] Next, a signal segment synchronous with the zero-crossing point of the leakage current is searched in the acoustic signal sequence. By comparing the signal change trends of the two in the same time interval, the time deviation between the acoustic signal sequence and the leakage current signal sequence is calculated. If the deviation exceeds a preset threshold, the timestamp of the acoustic signal sequence is adjusted point by point, and each data point of the acoustic signal sequence is shifted according to the deviation value so that each time point in the acoustic signal sequence can accurately match the corresponding time point in the leakage current signal sequence. Finally, the two types of signal sequences after calibration are synchronously verified. Multiple different time periods are selected to compare whether the signal characteristics of the two types of signals at the corresponding time maintain time consistency. For example, observe whether the acoustic signal mutation caused by partial discharge and the weak fluctuation of leakage current appear in the same time interval. If the time alignment requirement is met in multiple consecutive verifications, the time-aligned acoustic signal sequence and leakage current signal sequence are confirmed.

[0046] Step S220 involves windowing and framing the time-aligned acoustic signal sequence and performing sliding window averaging filtering on the leakage current signal sequence to obtain preprocessed acoustic and leakage current signals. Specifically, this includes: windowing and framing the time-aligned acoustic signal sequence. First, the frame length and frame shift parameters are determined. The frame length needs to cover a complete cycle of the partial discharge signal of the surge arrester to ensure that each frame contains complete partial discharge acoustic features. Typically, the frame length is set to a few milliseconds to tens of milliseconds, and the frame shift is set to one-third to one-half of the frame length to avoid information loss or excessive overlap between adjacent frames. Then, a Hamming window is selected as the windowing function. The Hamming window is multiplied with each frame of the acoustic signal sequence. Through the smooth transition characteristics of the Hamming window, the spectral leakage phenomenon generated after signal framing is reduced, making the spectral characteristics of each frame of acoustic signal closer to the real signal characteristics.

[0047] After windowing and framing, multiple independent frame acoustic signals are obtained, forming the preprocessed acoustic signal. For the time-aligned leakage current signal sequence, sliding window averaging filtering is performed. First, the window size of the sliding window is determined, and the window size is set to an integer multiple of the power grid cycle, usually 1 to 3 power grid cycles, to ensure that random noise in the leakage current signal can be effectively smoothed, while retaining the weak fluctuation characteristics of the leakage current and avoiding over-filtering that would cause the weak fault-related signals to be filtered out. Then, the sliding step size is set to one-quarter to one-third of the window size, so that the sliding window moves point by point on the leakage current signal sequence. Each time it moves, the average value of all leakage current data points in the current window is calculated, and this average value is used as the filtered data point at the current sliding position. After completing the calculation of all sliding windows in sequence, a continuous smooth leakage current signal is obtained, which constitutes the preprocessed leakage current signal.

[0048] Step S230 involves performing wavelet threshold denoising on the preprocessed acoustic signal to suppress environmental noise interference, and simultaneously performing adaptive notch filtering on the preprocessed leakage current signal to suppress power grid frequency interference, thereby obtaining the denoised acoustic signal and leakage current signal. Specifically, this includes:

[0049] To address the pulse characteristics of partial discharge acoustic signals from surge arresters—namely, short duration and abrupt amplitude changes—a dedicated atomic library is constructed. This library contains various pulse signal templates, with parameters for each template (such as pulse width, rise time, and attenuation coefficient) covering typical characteristics of partial discharge signals, such as 1-10ms width and microsecond-level rise time. It also incorporates common forms of environmental noise, such as random broadband noise and low-frequency stable noise from equipment vibration, ensuring the ability to distinguish between valid signals and noise.

[0050] The preprocessed acoustic signal is input into the matching pursuit algorithm. The algorithm sequentially selects the atom most similar to the current signal from the atom library (by calculating the correlation between the signal and the atom, the atom with the highest correlation is selected), and separates the energy of the atom from the original signal. After each separation, the remaining residual signal (the original signal minus the superimposed signal of the matched atoms) is calculated, and the above selection and separation process is repeated until the energy of the residual signal drops to a preset threshold (this threshold is set based on the average energy of normal environmental noise to ensure that the residual noise is less than 10% of the effective signal). All atoms that match the partial discharge signal are collected, and these atoms are superimposed in the original separation order to reconstruct the acoustic signal after removing the environmental noise. After reconstruction, it needs to be verified by comparing the partial discharge pulse characteristics (such as pulse peak value and duration) of the reconstructed signal with those of the original signal. If the characteristic deviation is less than 5%, the denoising is confirmed to be effective.

[0051] The pure power frequency signal of the power grid is synchronously acquired through a voltage transformer. This signal contains only a 50Hz power frequency component and has no weak fluctuation characteristics of leakage current, serving as an interference reference signal. During acquisition, it is necessary to ensure that the reference signal and the leakage current signal are synchronized in time, using the zero-crossing point of the power grid voltage as the time reference, aligning the sampling times of the two to avoid cancellation failure due to time deviation. The leakage current signal is used as the main input signal, and the error between the main input signal and the reference signal is calculated in real time, which is the weighted sum of the main input signal and the reference signal. The weight of the reference signal is dynamically adjusted according to the magnitude of the error. If the error is large, it indicates that the current weight has not effectively canceled the power frequency interference, so the weight of the reference signal is increased. If the error is small (indicating that the interference has been partially canceled), the weight is finely adjusted until the error stabilizes in a very small range, usually less than 1% of the amplitude of the original leakage current signal.

[0052] Using the final determined weights, the reference signal is weighted and then subtracted from the main input signal to obtain the leakage current signal after canceling the power frequency interference. During verification, the signal spectrum needs to be analyzed and compared with the spectrum before and after cancellation. If the energy ratio of the power frequency component drops from more than 60% before cancellation to less than 5%, and the weak fluctuations of the leakage current are still completely preserved, then the interference suppression is confirmed to be effective.

[0053] Step S240 involves performing a short-time Fourier transform on the denoised acoustic signal to extract the Mel-frequency cepstral coefficient feature vector as acoustic feature parameters, specifically including:

[0054] First, perform signal activity analysis on the denoised acoustic signal, and statistically analyze the amplitude change rate of the signal in different time periods. If the signal amplitude changes frequently in a certain time period, it may contain partial discharge pulses. In this case, shorten the frame length of that segment (e.g., set to 5ms) and reduce the frame shift (e.g., set to 2ms) to improve the time-frequency resolution. If the signal is stable in a certain time period (mostly due to residual environmental noise), extend the frame length (e.g., set to 20ms) and increase the frame shift (e.g., set to 10ms) to reduce redundant calculations. Finally, ensure that each frame of signal can completely cover at least one partial discharge pulse (if it exists) and that there are no information gaps between adjacent frames.

[0055] Each frame of signal is smoothed using a Hanning window. The edge attenuation of the Hanning window is more gradual, which can further reduce signal distortion caused by framing. Then, a Fourier transform is performed on each frame of signal after windowing. The number of transform points is adjusted according to the frame length. For example, a 5ms frame length corresponds to 1024 transform points to ensure that the frequency resolution reaches 1Hz. After the transform, a three-dimensional time-frequency diagram of time, frequency and amplitude of each frame of signal is obtained, with the frequency range covering 0-20kHz, which is the main distribution frequency band of partial discharge ultrasonic signals.

[0056] The time-frequency graph is segmented by energy thresholds, and the average energy of the entire frequency band is calculated. Frequency points with energy below 1 / 3 of the average energy are set to zero, as these points are mostly residual noise. Frequency points with concentrated energy are retained, corresponding to the characteristic frequency bands of the partial discharge signal, forming a time-frequency graph for focusing the effective signal. Based on the characteristics of human hearing (more sensitive to low-frequency signals), the 0-20kHz frequency band is divided into 24 non-uniform frequency bands according to the Mel scale. The low-frequency band (20Hz-5kHz) has a small interval (e.g., 200Hz / band) to capture the low-frequency characteristics of the partial discharge signal; the high-frequency band (5kHz-20kHz) has a large interval (e.g., 1000Hz / band) to reduce redundant information. Each frequency band corresponds to a characteristic energy range, and only the energy of these ranges is extracted subsequently.

[0057] From the focused time-frequency graph, the total energy of all frequency points within each Mel band is calculated, i.e., the cumulative energy of that band, resulting in 24 band energy values. These energy values ​​are then normalized by dividing each energy value by the sum of all band energies to eliminate the influence of overall signal amplitude fluctuations. For example, changes in the amplitude of partial discharge signals do not affect the relative proportion of band energies.

[0058] Discrete cosine transform was performed on the energy values ​​of the 24 normalized frequency bands to extract the first 12 transform coefficients. These 12 coefficients can reflect the core characteristics of the energy distribution. High-frequency redundancy was removed and used as the basic Mel frequency cepstral coefficients. In order to capture the temporal dynamic changes of the signal, the inter-frame difference of the basic coefficients of 10 consecutive frames was calculated, that is, the coefficient of the next frame minus the coefficient of the previous frame, to obtain the first-order dynamic feature. Then, the inter-frame difference of the first-order dynamic feature was calculated to obtain the second-order dynamic feature. Finally, the basic coefficients, the first-order dynamic feature, and the second-order dynamic feature were merged to form a 36-dimensional feature vector, which serves as the core representation parameter of the acoustic signal.

[0059] Step S250 involves performing full-cycle waveform analysis on the denoised leakage current signal, extracting the peak amplitude, third harmonic component, and phase difference parameters of the leakage current signal as current characteristic indicators. Specifically, this includes: using the grid voltage zero-crossing point as the synchronization reference (collecting voltage signals through voltage transformers and detecting the zero-crossing time), the interference-free leakage current signal is truncated into continuous periodic waveform segments according to the grid cycle (20ms for a 50Hz grid); each waveform segment is sampled with 2048 data points to ensure that the subtle fluctuations of the waveform (such as μA-level current changes) can be completely preserved, avoiding peak loss due to insufficient sampling points.

[0060] For each waveform segment of a cycle, a triple sliding window is used to detect peak values. The first window (5 sampling points) initially locates possible peak points. The second window (10 sampling points) verifies whether the point is the maximum value within the window. The third window (15 sampling points) confirms that the signal on both sides of the point exhibits a pulse pattern of rising and falling, eliminating false peaks caused by noise from a single sampling point. After triple verification, the positive and negative peak values ​​for each cycle are recorded. The mean and standard deviation of the positive and negative peak values ​​for 10 consecutive cycles are calculated. Abnormal peak values ​​exceeding the mean ± 2 times the standard deviation are removed, as these are mostly caused by instantaneous fluctuations in the power grid. The mean is then recalculated using the peak values ​​of the remaining 8 cycles to obtain stable average positive and average negative peak values, which serve as amplitude peak characteristics.

[0061] Fourier decomposition is performed on each waveform segment of the cycle, decomposing the waveform into multiple frequency components such as the fundamental wave (50Hz), second harmonic (100Hz), and third harmonic (150Hz). During the decomposition, windowing processing (Hamming window) is used to reduce spectral leakage and ensure that the amplitude calculation error of each harmonic is less than 3%. The fundamental wave amplitude and the third harmonic amplitude are extracted from the decomposition results. The fundamental wave amplitude reflects the main component of the leakage current, and the third harmonic amplitude reflects the early characteristics of insulation degradation. The proportion of the third harmonic is low in normal surge arresters and increases when the insulation is deteriorated. The ratio of the third harmonic amplitude to the fundamental wave amplitude is calculated, and the average of the ratio over 10 consecutive cycles is taken to obtain the stable third harmonic component characteristics. The original grid voltage signal (containing the fundamental wave, harmonics, and a small amount of noise) is collected from the voltage transformer and firstly, the signal is purified: by calculating the amplitude distribution of the signal, abnormal values ​​exceeding ±10% of the normal voltage range (mostly transient interference) are removed. Then, the signal is smoothed (using a moving average of 5 sampling points to reduce the influence of high-frequency noise) to obtain a stable voltage signal sequence.

[0062] The preprocessed voltage signal is subjected to Fourier decomposition, covering a range of 0-500Hz (including major harmonic components). The component corresponding to the 50Hz frequency is selected from the decomposition results, and its amplitude and waveform information are extracted to obtain a pure voltage fundamental signal (only the 50Hz sine wave component is retained). A threshold point of 0 for voltage amplitude is set. When the signal crosses 0 from negative to positive, the moment is recorded as a positive zero-crossing point; when it crosses 0 from positive to negative, it is recorded as a negative zero-crossing point. To avoid false zero-crossings caused by noise, a continuous verification method is adopted. If a point is determined to be a zero-crossing, the signal trend of the three sampling points before and after it needs to be checked (e.g., the three points before a positive zero-crossing need to be continuously negative, and the three points after it need to be continuously positive). After verification, the valid zero-crossing is confirmed. Taking the moment of the positive zero-crossing as the phase starting point (defined as 0°), the phase increases by 1° every 1 / 360 of the power grid cycle (approximately 55.56μs for a 50Hz power grid) along the time axis. After a complete cycle (20ms), the phase returns to 360° (i.e., 0°), forming a continuous voltage fundamental phase sequence, which serves as the subsequent reference phase.

[0063] The interference-free leakage current signal (with power frequency interference suppressed and fundamental and weak fluctuations retained) is subjected to Fourier decomposition, with the decomposition range consistent with the voltage signal (0-500Hz). The component corresponding to the 50Hz frequency is selected, and its amplitude and waveform information are extracted to obtain the leakage current fundamental signal. A standard sine wave signal with a frequency of 50Hz is generated, whose phase can be precisely adjusted and continuously changed from 0° to 360° with a step size of 0.1°. It is also synchronized with the grid voltage signal in time, with the voltage zero-crossing point as the time reference, to ensure that the time axis of the standard signal is consistent with the leakage current fundamental signal.

[0064] The fundamental frequency signal of the leakage current is cross-correlated with a standard sinusoidal signal.

[0065] The sliding window is set to one grid cycle (20ms), and the window step size is one sampling point (approximately 10μs, determined according to the sampling rate). For each window, the cross-correlation coefficient between the leakage current fundamental signal and the standard sine wave signal within the current window is calculated (the closer the coefficient is to 1, the higher the waveform matching degree between the two). The sliding window moves along the time axis, and the cross-correlation coefficients for each window are repeatedly calculated, recording all coefficient values ​​and the corresponding standard sine wave signal phase. The maximum value is selected from all cross-correlation coefficients, and the window position corresponding to this maximum value is the moment when the leakage current fundamental signal and the standard sine wave waveform are most matched. At this time, the phase of the standard sine wave signal is the phase of the leakage current fundamental wave. The phase results of 10 consecutive grid cycles are averaged. If the average fluctuation is less than 0.5°, the average value is confirmed as the stable phase of the leakage current fundamental wave. The absolute difference between the reference voltage phase and the leakage current fundamental wave phase is calculated as the phase difference parameter. The phase difference of 10 consecutive cycles is averaged. If the average fluctuation is less than 2°, the phase difference characteristic is confirmed to be stable.

[0066] Step S260 involves mapping each feature component of the extracted acoustic feature parameters to a three-dimensional feature space using a nonlinear dimensionality reduction method, forming a set of spatial coordinate points and constituting a spatial scattered dataset. Specifically, this includes: calculating the mean and standard deviation of the 36-dimensional Mel frequency cepstral coefficient feature vector dimension by dimension, converting the feature value of each dimension into the form of (original value - mean) / standard deviation, so that the mean of each dimension is 0 and the standard deviation is 1; this process can eliminate the dimensional differences of features in different dimensions (such as the different amplitude ranges of the basic coefficient and the dynamic coefficient), and avoid the excessive influence of a certain dimension feature on subsequent mapping; for the standardized feature vector, the isolated forest algorithm is used to initially screen outliers. This algorithm constructs multiple decision trees to determine feature vectors that differ greatly from the distribution of most samples as outliers (such as abnormal features caused by sudden noise); after removing the abnormal feature vectors, more than 95% of the normal feature vectors are retained to ensure the accuracy of subsequent mapping.

[0067] Input the pre-screened high-dimensional feature vectors, set the core parameters of the UMAP algorithm, set the number of nearest neighbors to 15 (determined by calculating the local density of the feature vectors to ensure that the correlation between adjacent features can be captured), set the minimum distance to 0.1 (to avoid excessive overlap of sample points after mapping), and set the target dimension to 3 (to facilitate the subsequent construction of a two-dimensional coordinate system and division of regions). The algorithm will automatically calculate the similarity matrix between feature vectors, mapping feature vectors with high similarity to adjacent points in three-dimensional space, and mapping feature vectors with low similarity to distant points.

[0068] After mapping, the structure of the point cloud in 3D space is verified, and the local clustering coefficient (measuring the degree of clustering of adjacent points) and global distance deviation (measuring the change in the relative distance between samples before and after mapping) are calculated. If the local clustering coefficient is greater than 0.6 (indicating that the clustering of partial discharge features is obvious) and the global distance deviation is less than 10% (indicating that the global structure is well preserved), the mapping is confirmed to be effective. Using a 1-minute time window (covering multiple partial discharge cycles to ensure sample representativeness), 1000 frames of acoustic signals are continuously collected, corresponding to 3D mapping points (each frame of signal corresponds to 1 3D point), forming an initial scatter set. For the initial scatter set, the mean and standard deviation of all points in three dimensions are calculated, and outliers exceeding the mean ± 3 times the standard deviation in any dimension are removed. These are mostly calculation errors in the mapping process or noise features that have not been pre-screened. If the number of remaining scatter points is still greater than 900 after removing outliers, the dataset is confirmed to be effective.

[0069] In this embodiment of the invention, microsecond-level high-precision clock calibration is used to align the timing of acoustic and leakage current signals. Combined with zero-crossing anchor point verification, feature correlation deviations caused by signal timing misalignment are avoided, ensuring the accuracy of subsequent multimodal feature fusion and solving the problem of fault misjudgment caused by signal asynchrony in traditional monitoring. For acoustic signals, adaptive frame length windowing (adjusting frame length / frame shift as needed) is used to cover the complete partial discharge cycle while avoiding information loss or excessive overlap. For leakage current, sliding window averaging filtering (a window that is an integer multiple of the power grid cycle) smooths random noise while retaining weak μA-level fluctuations, preserving key fault information for subsequent feature extraction. Precisely separates partial discharge pulses from environmental noise (residual noise is less than 10% of the effective signal); adaptive noise cancellation is used for leakage current, reducing the proportion of power frequency interference from over 60% to below 5%, while fully preserving fault-related weak fluctuations, solving the defect of traditional filtering that easily filters out useful signals and improving the identifiability of early fault signals; improves time-frequency resolution through adaptive frame length, and constructs a 36-dimensional acoustic feature vector by combining non-uniform division of the Mel band and dynamic features (first-order / second-order difference) to comprehensively characterize subtle changes in partial discharge; extracts stable amplitude, third harmonic (sensitive indicator of early degradation), and phase difference features by using triple peak detection, Fourier harmonic decomposition, and cross-correlation phase calculation, solving the problem of insufficient sensitivity of traditional methods for identifying early faults (such as poor sealing and water ingress).

[0070] In a preferred embodiment of the present invention, step S300, constructing a two-dimensional feature mapping coordinate system based on the spatial scattered point dataset, and determining a minimum boundary range that can encompass all sampling points in the coordinate system, includes:

[0071] Step S310: Extract the three-dimensional coordinate information of all spatial points in the spatial scatter dataset, and obtain the X-axis coordinate value, Y-axis coordinate value and Z-axis coordinate value of each spatial point in the Cartesian coordinate system. Specifically, this includes: retrieving the spatial scatter dataset generated in step S260, which contains no less than nine hundred valid three-dimensional scatter points. These scatter points are all valid data that can truly reflect the acoustic characteristics of the surge arrester after outlier removal; and extracting the complete coordinate information of each valid scatter point in the dataset in the Cartesian coordinate system, that is, the X-axis coordinate value, Y-axis coordinate value and Z-axis coordinate value corresponding to each spatial point. Referring to the outlier judgment criteria set in step S260, that is, the coordinate values ​​must be within the range of the mean plus or minus three standard deviations, the X-axis coordinate values, Y-axis coordinate values, and Z-axis coordinate values ​​of each extracted spatial point are checked a second time. During the check, invalid points with missing coordinates, empty coordinates, or coordinate values ​​exceeding the reasonable range are excluded to ensure that all three-dimensional scattered points participating in the subsequent two-dimensional projection operation are valid data that can accurately reflect the acoustic characteristics of the surge arrester, and to avoid invalid data interfering with the subsequent determination of the boundary range.

[0072] Step S320 involves using the X-axis and Y-axis coordinates of each spatial point as projection references to map all spatial points onto a two-dimensional feature projection plane formed by the X-axis and Y-axis, thus forming a two-dimensional feature distribution map. Specifically, this includes: determining the projection dimension based on the effective three-dimensional scatter point coordinates extracted in step S310, combined with the UMAP mapping results from step S260; analysis shows that the combination of dimensions formed by the X-axis and Y-axis can best preserve the local clustering information and global distribution trend of acoustic features, which is more conducive to subsequent boundary delineation and region density analysis. Therefore, the X-axis and Y-axis coordinates of each spatial point are selected as the reference for two-dimensional projection; a projection operation is performed on each three-dimensional scatter point, directly mapping the X-axis of the three-dimensional scatter point... The coordinate values ​​are used as the x-coordinates of the corresponding 2D projection points, and the Y-coordinates of the 3D scattered points are used as the y-coordinates of the corresponding 2D projection points. The Z-coordinates are temporarily ignored during this projection process because they are only used for the initial 3D feature space construction. The 2D mapping stage focuses on the distribution features reflected by the X-axis and Y-axis dimensions. After all the 3D scattered points have been projected, the positions of each projection point are marked in the preset 2D plane coordinate system according to the correspondence between the x-coordinates and y-coordinates of each 2D projection point, forming an intuitive 2D feature distribution map. At the same time, to facilitate quick matching when it is necessary to backtrack the original 3D scattered point information later, a unique association identifier between each 2D projection point and the original 3D scattered points is recorded.

[0073] Step S330: On the two-dimensional feature projection plane, traverse the X-axis coordinate values ​​of all projection points to determine the minimum and maximum projection coordinate values ​​in the X-axis direction. Specifically, this includes: for all two-dimensional projection points in the two-dimensional feature distribution map generated in step S320, read the abscissa of each projection point one by one. This abscissa is the X-axis coordinate value corresponding to the original three-dimensional scatter point. During the reading process, the read abscissa values ​​are compared and updated in real time to determine the minimum and maximum projection coordinate values ​​in the X-axis direction. At the beginning of the traversal operation, the abscissa of the first read two-dimensional projection point is temporarily set. The initial minimum and maximum projected coordinates in the X-axis direction are set. Then, for each new 2D projection point, the x-coordinate is read and compared with the currently recorded minimum and maximum projected coordinates in the X-axis direction. If the x-coordinate is less than the currently recorded minimum projected coordinate, the current minimum projected coordinate is updated to this x-coordinate; if the x-coordinate is greater than the currently recorded maximum projected coordinate, the current maximum projected coordinate is updated to this x-coordinate; if the x-coordinate is between the current minimum and maximum projected coordinates, the current extreme value remains unchanged. After traversing all 2D projection points, the final determined minimum and maximum projected coordinates in the X-axis direction are recorded. These two values ​​will serve as key parameters for the X-axis direction when constructing the rectangular boundary. Simultaneously, the number of 2D projection points actually read during the traversal is checked against the number of valid 2D projection points generated in step S320 to ensure no data is missed during the traversal and that the determined extreme values ​​in the X-axis direction cover all valid projection points.

[0074] Step S340 involves traversing the Y-axis coordinates of all projection points on the two-dimensional feature projection plane to determine the minimum and maximum projection coordinates in the Y-axis direction. Specifically, this includes: referring to the operation logic of step S330, traversing all two-dimensional projection points in the two-dimensional feature distribution map generated in step S320, focusing on the ordinate of each two-dimensional projection point, which is the Y-axis coordinate corresponding to the original three-dimensional scatter point. The minimum and maximum projection coordinates in the Y-axis direction are determined through this traversal. At the start of the traversal, the ordinate of the first two-dimensional projection point is set as the initial minimum and initial maximum projection coordinates in the Y-axis direction. Subsequently, the ordinates of new two-dimensional projection points are read point by point, and these ordinates are compared with the currently recorded minimum and maximum projection coordinates in the Y-axis direction. If the ordinate value is less than the currently recorded minimum projected coordinate value, then update the current minimum projected coordinate value; if the ordinate value is greater than the currently recorded maximum projected coordinate value, then update the current maximum projected coordinate value; if the ordinate value is between the current minimum and maximum projected coordinate values, then do not adjust the current recorded extreme value. After traversing all 2D projection points, record the final determined minimum and maximum projected coordinate values ​​in the Y-axis direction. This completes the extraction of extreme values ​​for all projection points in the 2D plane in both the X and Y axes, ensuring that the initially constructed rectangular boundary can completely cover the distribution range of all projection points in the Y-axis direction, avoiding the situation where some projection points cannot be enveloped by the initial boundary due to incomplete extraction of extreme values ​​in the Y-axis direction.

[0075] Step S350: Based on the obtained minimum and maximum projected coordinates along the X-axis and Y-axis, determine an initial rectangular boundary that can enclose all projection points. Specifically, this includes: determining the four vertices of the initial rectangular boundary on the two-dimensional feature projection plane based on the minimum and maximum projected coordinates along the X-axis obtained in step S330 and the minimum and maximum projected coordinates along the Y-axis obtained in step S340. Specifically, the coordinates of the lower left vertex of the initial rectangular boundary are a combination of the minimum and minimum projected coordinates along the X-axis and Y-axis; the coordinates of the upper right vertex are a combination of the maximum and maximum projected coordinates along the X-axis and Y-axis; the coordinates of the upper left vertex are a combination of the minimum and maximum projected coordinates along the X-axis and Y-axis; and the coordinates of the lower right vertex are a combination of the maximum and minimum projected coordinates along the X-axis and Y-axis. Connect these four vertices sequentially to form a closed initial rectangular boundary. To verify the validity of the initial rectangular boundary, at least fifty valid projection points are randomly selected from all 2D projection points. These selected projection points must cover points distributed in both the edge and center regions of the 2D feature projection plane to ensure representativeness. The x-coordinate of each selected projection point is checked to ensure it lies between the minimum and maximum projection coordinate values ​​in the X-axis direction, and the y-coordinate is also checked to ensure it lies between the minimum and maximum projection coordinate values ​​in the Y-axis direction. If all selected projection points are within the initial rectangular boundary, and it is confirmed that no 2D projection point falls outside the initial rectangular boundary, then the initial rectangular boundary is confirmed to completely enclose all 2D projection points. If any projection point exceeds the initial rectangular boundary, the calculation results of the extreme coordinates in the X-axis and Y-axis directions are rechecked, calculation errors are identified and corrected, and the verification within the projection point boundary is performed again until the initial rectangular boundary completely encloses all 2D projection points.

[0076] Step S360: Extend the initial rectangular boundary in the X-axis and Y-axis directions according to the preset boundary extension coefficients to obtain the minimum bounding rectangular boundary range after extension. Specifically, this includes: determining the preset boundary extension coefficient according to different operating states of the surge arrester. The judgment criteria and boundary extension coefficients corresponding to different operating states are set as follows.

[0077] When the surge arrester is in normal operation, the partial discharge signal of the surge arrester fluctuates little and has no obvious abnormal characteristics. In step S260, the number of effective scattered points within a one-minute window is no less than 1,100, and the scattered points are evenly distributed in three-dimensional space with no obvious sparse edge areas. In this scenario, the boundary expansion coefficient is taken as 1.05. This value can completely enclose all scattered points while minimizing the blank area within the boundary, thus avoiding redundant calculations when dividing the convex area later.

[0078] When the surge arrester is in the routine monitoring period, the partial discharge signal fluctuates slightly. In step S260, the number of effective scattered points within a one-minute window is between 950 and 1100. The scattered points are relatively evenly distributed, but there are a few sparse points at the local edges. In this scenario, the boundary expansion coefficient is set to 1.07, which can cover the weak feature points at the sparse edges without causing too many blank areas due to an excessively large coefficient, thus balancing coverage integrity and computational efficiency. When the surge arrester is in the fault warning period, historical data shows that the fluctuation of the partial discharge signal has increased significantly, or there have been minor abnormal alarms. In step S260, the number of effective scattered points within a one-minute window is between 900 and 950. The scattered points are unevenly distributed, and there are many discrete potential fault feature points at the edges. In this scenario, the boundary expansion coefficient is set to 1.09. By using a slightly larger coefficient, more space is reserved to avoid missing discrete points at the edges, while reducing the interference of irrelevant blank areas on the subsequent area density calculation. In special circumstances, such as a sudden increase in partial discharge signals due to transient fluctuations in grid voltage, even if the number of effective scattered points within a one-minute window in step S260 is not less than one thousand, the number of edge points may suddenly increase, or the surge arrester may have just undergone maintenance and restarted, resulting in poor signal stability and chaotic scattered point distribution. In this scenario, the boundary expansion coefficient is taken as 1.08, which is slightly larger than that during normal operation to cover temporary edge points while avoiding excessive expansion.

[0079] Only when the surge arrester is in a high-risk warning period, and a slight increase in the third harmonic component and fluctuations in phase difference have been detected, and the number of effective scattered points within a one-minute window in step S260 does not exceed 950 and the proportion of sparse edge points exceeds 15%, the boundary expansion coefficient is set to 1.1, which is the maximum value, ensuring that all potential fault feature points are enveloped. After determining the boundary expansion coefficient, the initial rectangular boundary is extended in both the X-axis and Y-axis directions. In the X-axis direction, the minimum projected coordinate value of the initial rectangular boundary is multiplied by one and the difference between the minimum and maximum projected coordinate values ​​of the initial rectangular boundary is subtracted from the boundary expansion coefficient to obtain the minimum X-axis coordinate value after expansion; the maximum projected coordinate value of the initial rectangular boundary is multiplied by one and added to the boundary expansion coefficient to obtain the maximum X-axis coordinate value after expansion.

[0080] In the Y-axis direction, the same expansion method as in the X-axis direction is adopted. The minimum projected Y-axis coordinate value of the initial rectangular boundary is multiplied by one and the difference between the minimum and the boundary expansion coefficient to obtain the expanded minimum Y-axis coordinate value. The maximum projected Y-axis coordinate value of the initial rectangular boundary is multiplied by one and the sum of the boundary expansion coefficient to obtain the expanded maximum Y-axis coordinate value. Based on the expanded minimum, maximum, minimum, and maximum X-axis coordinate values, a new rectangular boundary is constructed, which is the range of the expanded minimum bounding rectangle. After the expansion is completed, it is verified again to confirm that all the original two-dimensional projection points are within the expanded boundary range and that there is no excessively large blank area within the boundary. This ensures that the boundary can completely enclose all effective feature points and provide a reasonable spatial range for the convex region division in the subsequent step S400, thereby improving the accuracy and efficiency of region division.

[0081] In a preferred embodiment of the present invention, step S400 involves dividing the boundary range into regions based on the spatial distribution characteristics of the sampling points within the minimum boundary range, generating a series of non-overlapping convex region units, including:

[0082] Step S410: Based on the expanded minimum bounding rectangle boundary range, calculate the overall distribution density of spatial scattered points within the boundary range, and set an adaptive density threshold parameter based on the distribution density. Specifically, this includes: based on the expanded minimum bounding rectangle boundary range obtained in step S360, counting the total number of all spatial scattered points within the boundary range; calculating the area of ​​the rectangular boundary range, i.e., obtaining the X-axis length by subtracting the minimum X-axis coordinate value from the expanded maximum X-axis coordinate value, obtaining the Y-axis length by subtracting the minimum Y-axis coordinate value from the expanded maximum Y-axis coordinate value, and multiplying the two to obtain the total area of ​​the boundary range; dividing the total number of spatial scattered points by the total area of ​​the boundary range to obtain the overall distribution density of the spatial scattered points, which reflects the overall density of scattered points within the boundary range.

[0083] When using an adaptive density threshold parameter, the scattered distribution characteristics of the surge arrester under normal and abnormal conditions are considered. When the surge arrester is in normal operation, the scattered distribution is relatively uniform, and the overall density threshold is 1.2 times the overall distribution density to distinguish dense areas in normal fluctuations. When the surge arrester has early fault characteristics, local dense areas will appear in the scattered distribution. At this time, the density threshold is adjusted to 1.5 times the overall distribution density to accurately capture high-density feature areas related to faults. At the same time, the threshold setting needs to refer to the density difference between normal and fault states in historical data to ensure that the threshold can effectively distinguish different distribution states without causing missed or misjudged feature areas due to the threshold being too high or too low.

[0084] Step S420: Based on the density threshold parameter, the standard deviation calculation method is used to analyze the dispersion of the spatial scatter points in the X-axis direction and determine the number of partitions required in the X-axis direction. Specifically, this includes: based on the density threshold parameter determined in step S410, focusing on the coordinate values ​​of the spatial scatter points in the X-axis direction, and calculating the standard deviation of the X-axis coordinates of all scatter points; the larger the standard deviation, the higher the dispersion of the scatter points in the X-axis direction, and the more partitions are needed to refine and capture local features; the smaller the standard deviation, the more concentrated the distribution, and the fewer partitions are needed accordingly.

[0085] Set a standard deviation grading standard. When the standard deviation is less than or equal to the preset low dispersion threshold, divide the X-axis into three partitions, which is sufficient to balance computational efficiency and feature capture requirements. When the standard deviation is greater than the low dispersion threshold but less than or equal to the medium dispersion threshold, divide the X-axis into five partitions to accommodate feature distributions with medium dispersion. When the standard deviation is greater than the medium dispersion threshold, divide the X-axis into seven partitions to ensure that subtle distribution differences in highly discrete states can be effectively captured. After determining the number of partitions, it is necessary to verify whether the partitioning results can cover the X-axis distribution range of all scattered points, and whether the number of scattered points in each partition matches the density threshold, so as to avoid feature aliasing due to too few partitions or computational redundancy due to too many partitions.

[0086] Step S430: Based on the density threshold parameter, use the variance analysis method to evaluate the dispersion of the spatial scatter points in the Y-axis direction and determine the number of partitions required in the Y-axis direction. Specifically, this includes: based on the density threshold parameter obtained in step S410, calculating the variance of the Y-axis coordinates of all scatter points for the coordinate values ​​in the Y-axis direction; the variance value reflects the dispersion of the scatter points in the Y-axis direction. The larger the variance, the more dispersed the distribution of scatter points in the Y-axis direction, and more partitions are needed for refined analysis; the smaller the variance, the more concentrated the distribution, and the number of partitions can be appropriately reduced.

[0087] A variance grading rule was established: when the variance is less than or equal to a preset low dispersion standard, the Y-axis is divided into three partitions to meet basic feature capture requirements; when the variance is greater than the low dispersion standard but less than or equal to the medium dispersion standard, five partitions are divided to accommodate distribution characteristics under medium-medium dispersion conditions; when the variance is greater than the medium dispersion standard, seven partitions are divided to ensure the capture of local subtle features under high dispersion conditions. After determining the number of partitions, it was checked whether the scatter point distribution within each partition matched the density threshold to ensure that the partitioning results could cover all scatter points in the Y-axis direction while avoiding feature loss or redundancy due to unreasonable partitioning.

[0088] Step S440: Based on the calculated number of partitions in the X-axis and Y-axis directions, the boundary range of the minimum bounding rectangle is divided into several rectangular partition units of equal area. Specifically, this includes: calculating the partition interval in the X-axis and Y-axis directions based on the number of partitions in the X-axis direction determined in step S420 and the number of partitions in the Y-axis direction determined in step S430; in the X-axis direction, dividing the difference between the maximum and minimum X-axis coordinate values ​​by the expanded maximum X-axis coordinate value by the number of X-axis partitions to obtain the width of each X-axis partition; in the Y-axis direction, dividing the difference between the maximum and minimum Y-axis coordinate values ​​by the expanded maximum Y-axis coordinate value by the number of Y-axis partitions to obtain the height of each Y-axis partition.

[0089] Based on the calculated partition width and height, starting from the lower left corner of the expanded minimum bounding rectangle boundary, vertical partition lines corresponding to the number of X-axis partitions are sequentially drawn along the X-axis, and horizontal partition lines corresponding to the number of Y-axis partitions are sequentially drawn along the Y-axis. The vertical and horizontal partition lines intersect each other, dividing the minimum bounding rectangle boundary range into several rectangular partition units, and each rectangular partition unit has the same area, that is, the product of the width and height of each unit is the same. After the division is completed, it is checked whether all rectangular partition units completely cover the minimum bounding rectangle boundary range, and whether there is no overlap or gap between adjacent units, ensuring that each spatial point can be uniquely assigned to a rectangular partition unit.

[0090] Step S450 involves triangulating each rectangular partition unit along its diagonal to generate two triangular partition units. Based on the spatial relationship between adjacent rectangular partition units, an alternating optimization strategy is used to determine the diagonal partitioning direction. Specifically, this includes triangulating each rectangular partition unit generated in step S440, i.e., dividing each rectangle into two triangular partition units by connecting one diagonal line. To ensure that the triangulation better reflects the local distribution characteristics of the scattered points, an alternating optimization strategy is used to determine the diagonal partitioning direction based on the spatial relationship between adjacent rectangular partition units.

[0091] Specifically, for two horizontally adjacent rectangular partition units, if the left rectangle is partitioned diagonally from the top left corner to the bottom right corner, then the right rectangle is partitioned diagonally from the top right corner to the bottom left corner; for two vertically adjacent rectangular partition units, if the upper rectangle is partitioned diagonally from the top left corner to the bottom right corner, then the lower rectangle is partitioned diagonally from the top right corner to the bottom left corner. This alternating partitioning method avoids the formation of elongated shapes in adjacent triangular units, ensuring that each triangular unit can better fit the local distribution density of the scattered points, while reducing the characteristic directional deviation caused by the consistent partitioning direction, thus providing a more reasonable basic unit for subsequent region density calculation.

[0092] Step S460: Perform Delaunay triangulation optimization on all triangular partitioning units. Based on the optimization results, generate multiple non-overlapping convex polygon partitioning units using a polygon fusion algorithm. Specifically, this includes:

[0093] Collect all triangular partition units generated in step S450, extract the coordinate information of the three vertices of each triangular unit on the two-dimensional feature projection plane, and assign a unique identifier index to each vertex to ensure that the spatial position of each vertex can be accurately traced when adjusting the vertex connection relationship in the future, so as to avoid the triangular mesh structure disorder caused by vertex confusion and lay an accurate data foundation for optimization processing.

[0094] Based on the extracted vertex coordinates and identifier indexes, an initial triangular mesh structure is constructed, linking all triangular partition units according to the relative positions of their vertex coordinates on the two-dimensional plane to form a complete initial triangular mesh. Then, a Delaunay criterion verification process is initiated, checking each triangular unit in the initial triangular mesh one by one. Specifically, for a single triangular unit, the coordinates of the center and radius of its circumcircle are first calculated based on the coordinates of its three vertices. The center is determined by finding the intersection of the perpendicular bisectors of the three sides of the triangle, and the radius is calculated by the distance from the center to any vertex. Next, the two-dimensional plane covered by the circumcircle is searched to see if it contains vertices of other triangular units. If a vertex is found to fall within the circumcircle of the current triangle, the vertex connection relationship of the triangle is determined to not conform to the Delaunay criterion and needs adjustment.

[0095] For triangular elements that do not conform to the Delaunay criterion, an edge flipping operation is used to optimize vertex connections. When two adjacent triangular elements share an edge, and the circumcircle of one triangle contains a vertex of the other triangle, the shared edge is first deleted. Then, the two non-shared vertices of the two triangles are connected to form two new edges, thus reconstructing the original two triangles into two new triangular elements. After edge flipping, the circumcircles of these two new triangular elements are recalculated, and it is checked again whether the circumcircles contain other vertices. If there are still cases that do not conform to the criterion, the edge flipping operation is repeated until both new triangles satisfy the Delaunay criterion that there are no other triangle vertices within their circumcircles.

[0096] Throughout the triangulation optimization process, optimization of elongated triangles is carried out simultaneously. When one of the three interior angles of a triangular element contains an acute angle less than 30 degrees, the triangle is identified as an elongated triangle. Such triangles tend to distort the local distribution characteristics of scattered points, which is not conducive to subsequent fault feature identification. For elongated triangles, edge flipping is prioritized on the shared edges with adjacent triangles to adjust vertex connections, making the distribution of the three interior angles of the triangle more uniform, and gradually reducing the number of elongated triangles. The steps of criterion verification, edge flipping adjustment, and elongated triangle optimization are continuously iterated, traversing all triangular elements in the initial triangulation and their adjacent relationships until all triangular elements in the entire triangulation meet the Delaunay criterion, and the proportion of elongated triangles in all triangular elements is less than 5%. At this point, the optimized Delaunay triangulation can more realistically reflect the local aggregation state of spatial scattered points, avoid the distortion of scattered point distribution characteristics caused by unreasonable triangulation structure, solve the problem of easy loss of local subtle features when processing high-dimensional nonlinear acoustic signals using traditional methods, and ensure that the weak acoustic signal features generated by early faults of surge arresters can be completely preserved in the triangulation.

[0097] After optimizing the Delaunay triangulation, the polygon fusion stage begins. First, the density threshold parameter set in step S410 is retrieved, and a similarity criterion for the distribution characteristics of scattered points is established. Similarity in density change trends between adjacent triangular units means the difference in density variation from the edge to the center of the scattered points within the two units is less than 20%. Similarity in coordinate distribution range means the overlap of the coordinate ranges of the two units in the X and Y axes is greater than 60%. Then, each adjacent triangular unit in the optimized triangulation is analyzed individually. First, the number of spatial scattered points within each adjacent triangular unit is counted, and the scattered point distribution density is calculated based on the unit area. It is then checked whether the densities of the two units are both higher than the density threshold set in step S410. Finally, by comparing the density change curves and coordinate ranges of the two units, it is determined whether their scattered point distribution characteristics meet the similarity criteria.

[0098] If two adjacent triangular units simultaneously satisfy the conditions that their densities are both above the threshold and their distribution characteristics are similar, then the polygon fusion operation is initiated, merging the two triangular units into a single convex polygon unit. Based on all the vertices of the two triangular units, duplicate vertices are removed, and the vertices are connected in a clockwise order on the two-dimensional plane to form a closed convex polygon, ensuring that the convex polygon can completely enclose all the scattered points within the original two triangular units. If two adjacent triangular units have densities below the threshold, density differences greater than 30%, or distribution characteristic similarity that does not meet the standard, then the two units remain independent and are not merged.

[0099] The analysis and fusion operation of adjacent triangular units is repeated. After each fusion, the adjacency relationship between the newly formed convex polygonal unit and the surrounding triangular units is re-examined, and the existence of fusionable units is checked again until no new fusionable units appear after traversing all units. Finally, multiple convex polygonal partition units are generated. The coordinates are checked to confirm that these units do not overlap and that all spatial scattered points are contained within these units. These convex polygonal units can accurately enclose groups of scattered points with similar distribution characteristics. The unit boundaries are used to determine the regions that distinguish different distribution characteristics, which fully considers the difference in the discriminative value of different regional features. This avoids the problem of fault feature submersion caused by equal weighting of all sampling points in traditional methods. It enables the weak acoustic features generated by early faults of surge arresters (such as the ultrasonic spectrum anomaly of partial discharge) to stand out in the corresponding convex polygonal units, providing a reliable partitioning basis for subsequent calculation of regional density indicators and feature weighting optimization, and improving the sensitivity of early fault identification.

[0100] In a preferred embodiment of the present invention, step S500 involves assigning each sampling point to a corresponding convex region unit, calculating the surface area of ​​each region unit based on the distribution density and spatial structure characteristics of the sampling points within each unit, and generating a region density index characterizing local distribution characteristics, including:

[0101] Step S510 involves establishing a spatial index mapping relationship based on the spatial coordinate information of the convex polygon partitioning units, assigning each spatial point to the corresponding convex polygon partitioning unit according to its projected coordinate position. Specifically, this includes: determining the boundary range of each convex polygon based on the vertex coordinate information of each convex polygon partitioning unit generated in step S460; constructing the boundary contour of the polygon by recording the coordinates of all vertices of each convex polygon on the two-dimensional feature projection plane, and determining the coordinate range it covers; establishing a spatial index mapping relationship, assigning a unique identifier to each convex polygon partitioning unit, and storing the coordinate range corresponding to its boundary vertex coordinates; and for the two-dimensional projected coordinates of all spatial points generated in step S320, determining one by one which convex polygon partitioning unit's boundary range the coordinate position of each projected point belongs to.

[0102] The determination method involves checking whether the x-coordinate and y-coordinate of the projected point both fall within the closed region formed by the edges of the convex polygon, ensuring that each spatial point is assigned to only one corresponding convex polygon partition unit. If a projected point happens to fall on the boundary line of two convex polygons, it is assigned to a unit with more similar characteristics based on the similarity of the distribution characteristics between that point and the scattered points within the two polygons. After all the scattered points have been assigned, the set of scattered points corresponding to each convex polygon partition unit is recorded.

[0103] Step S520: Based on the spatial scatter point allocation results, count the number of spatial scatter points contained in each convex polygon partition unit, and calculate the spatial scatter point distribution density index based on the area of ​​the partition unit. Specifically, this includes: based on the spatial scatter point allocation results of step S510, counting the number of spatial scatter points contained in each convex polygon partition unit to ensure that the statistical results are consistent with the actual number of scatter points within the boundary range of the unit, avoiding omissions or duplicates; retrieving the surface area of ​​each convex polygon partition unit calculated in step S550, dividing the number of spatial scatter points in the unit by its surface area to obtain the spatial scatter point distribution density index of the convex polygon partition unit. This index reflects the density of scatter points per unit area; the higher the value, the denser the scatter point distribution in the area. For convex polygon partition units with zero scatter points, their distribution density index is marked as zero. After calculation, compare the distribution density index of each unit with the density threshold set in step S410, and mark the units with a density higher than the threshold. These units may contain key characteristic information of the arrester's operating status, especially the weak acoustic signal characteristics generated by early faults.

[0104] Step S530: Based on the spatial scatter point distribution density index, the spatial autocorrelation analysis method is used to evaluate the distribution uniformity of spatial scatter points within each partition unit, and the distribution uniformity index is calculated. Specifically, this includes: based on the spatial scatter point distribution density index obtained in step S520, the spatial autocorrelation analysis method is used to evaluate the distribution uniformity of scatter points within each convex polygon partition unit. First, the scatter points within each unit are divided into several smaller grid sub-regions according to their projected coordinates, and the scatter point density of each grid sub-region is calculated. By analyzing the density differences between adjacent grid sub-regions, the degree of aggregation of scatter points within the unit is evaluated. If the density differences between grid sub-regions are small, it indicates that the scatter point distribution is relatively uniform; if some grid sub-regions have significantly higher densities than other regions, it indicates that the scatter point distribution is uneven and there is a local aggregation phenomenon.

[0105] Based on the statistical results of density differences, a distribution uniformity index is calculated. The higher the index value, the more uniform the distribution of scattered points, and the lower the value, the more uneven the distribution. This index can effectively reflect the spatial structure characteristics of scattered points within a unit and is of great significance for identifying the area of ​​partial discharge signal accumulation caused by early faults in surge arresters. It avoids the traditional method from masking the characteristics of local faults by ignoring the distribution uniformity.

[0106] Step S540: For each convex polygon partition unit, the Delaunay triangulation algorithm is used to divide it into several triangular sub-units based on its vertex coordinates to obtain the triangular mesh partitioning result. Specifically, for each convex polygon partition unit generated in step S460, the coordinate information of all its vertices is extracted to ensure that the vertex coordinates are arranged in a clockwise or counterclockwise order to form a closed polygonal outline.

[0107] The Delaunay triangulation algorithm is used to process each convex polygon. The vertices of the polygon are used as the triangulation vertices to construct a triangular mesh inside the polygon. During the triangulation process, it is ensured that the vertices of each triangular sub-unit are vertices of the original convex polygon, and that all triangular sub-units do not overlap, completely covering the entire convex polygon region. Simultaneously, the triangulation results are checked to ensure they meet the Delaunay criterion, i.e., the circumcircle of each triangular sub-unit does not contain vertices of other triangular sub-units, ensuring the reasonable shape of the triangular sub-units and avoiding elongated triangles that could affect the accuracy of subsequent area calculations. After triangulation, the coordinates of all triangular sub-units and their vertices corresponding to each convex polygon partition are recorded, forming the triangular mesh triangulation result.

[0108] Step S550: Based on the triangular mesh partitioning results, calculate the area of ​​each triangular sub-unit, and obtain the surface area of ​​the corresponding convex polygon partition unit by summing the results. Specifically, this includes: based on the triangular mesh partitioning results obtained in step S540, for each triangular sub-unit, calculate the area of ​​the triangle according to the coordinates of its three vertices. During the calculation, the side length and height of the triangle are determined by the vertex coordinates, or a coordinate-based geometric calculation method is used to ensure the accuracy of the area calculation; sum the areas of all triangular sub-units contained in each convex polygon partition unit, and the sum is the surface area of ​​the convex polygon partition unit. After the calculation is completed, the surface area is verified by comparing it with the area directly calculated from the boundary coordinates of the convex polygon to ensure that the error is within the allowable range, and to avoid the accuracy of subsequent density indicators being affected by triangular partitioning or area calculation errors; finally, record the surface area of ​​each convex polygon partition unit as the basic parameter for calculating the distribution density index in step S520.

[0109] Step S560 involves standardizing the calculated spatial scatter distribution density index, distribution uniformity index, and partition unit surface area. Based on the standardized parameters, a weighted fusion algorithm is used to generate a comprehensive regional attribute index for each convex polygon partition unit. Specifically, this includes standardizing the spatial scatter distribution density index obtained in step S520, the distribution uniformity index obtained in step S530, and the partition unit surface area obtained in step S550. During the standardization process, the values ​​of each index are converted to the range of zero to one to eliminate the influence of differences in dimensions and numerical ranges among different indices, making each index comparable.

[0110] Based on the requirements for surge arrester condition identification, weights are assigned to the standardized parameters. Among them, the spatial scatter point distribution density index has the highest weight because it directly reflects the density of local features and has the strongest correlation with early fault signals. The distribution uniformity index is second, used to help judge the clustering of features. The surface area of ​​the partition unit has a relatively low weight, mainly used to correct the influence of region size on the density index. A weighted fusion algorithm is adopted to multiply the standardized parameters with their corresponding weights and then sum them to obtain the comprehensive regional attribute index of each convex polygon partition unit. This index comprehensively reflects the density, uniformity, and spatial size characteristics of the region, which can highlight high-value fault feature areas and avoid the problem of fault features being submerged by normal data due to equal weighting of all sampling points in traditional methods.

[0111] In a preferred embodiment of the present invention, step S600 involves using the region density index to perform weighted optimization on the acoustic feature parameters, fusing the optimized acoustic feature parameters with the current feature index through multimodal feature fusion to form a high-dimensional joint feature matrix, and inputting the feature matrix into a pre-trained one-dimensional convolutional neural network state recognition model for intelligent diagnosis, outputting a classification result of the arrester's operating state, including:

[0112] Step S610 involves extracting the initially acquired acoustic feature parameters based on the comprehensive regional attribute index of the convex polygon partition unit, and establishing a mapping relationship between the attribute index and the feature parameters. Specifically, this includes retrieving the initial acoustic feature parameters extracted in step S240. These parameters are 36-dimensional Mel frequency cepstral coefficient feature vectors, with each spatial scatter point corresponding to a complete set of acoustic feature parameters. Simultaneously, it retrieves the allocation results of the spatial scatter points and convex polygon partition units from step S510, as well as the comprehensive regional attribute index of each convex polygon partition unit calculated in step S560.

[0113] Based on the scatter point allocation results, the acoustic feature parameters corresponding to each spatial scatter point are associated with the convex polygon partition unit to which the scatter point belongs, ensuring that each set of acoustic feature parameters can correspond to a unique convex polygon partition unit. On this basis, a mapping relationship between comprehensive regional attribute indicators and acoustic feature parameters is established, that is, the comprehensive regional attribute indicators of each convex polygon partition unit correspond to the set of acoustic feature parameters of all spatial scatter points in the unit. After establishing the mapping relationship, the set of acoustic feature parameters corresponding to each convex polygon partition unit is classified and organized, and the number of acoustic feature parameters and the specific feature component values ​​in each unit are recorded. This ensures that the regional attribute basis corresponding to each acoustic feature parameter can be determined during subsequent weighted optimization, avoiding the problem of disconnect between acoustic feature parameters and regional features in traditional methods.

[0114] Step S620: Based on the comprehensive regional attribute index values ​​of each partition unit, determine the weighting coefficients of the corresponding acoustic feature parameters. Specifically, this includes: sorting the comprehensive regional attribute index values ​​of all convex polygon partition units and classifying the index levels according to the sorting results; classifying the units with index values ​​in the top 30% as high-value regions, which typically contain richer fault feature information; classifying the units with index values ​​in the middle 40% as medium-value regions, corresponding to normal operating characteristics; and classifying the units with index values ​​in the bottom 30% as low-value regions, which mostly contain noise or normal fluctuation characteristics.

[0115] Weighting coefficient ranges were set for different levels of regional units: the weighting coefficient for high-value regions was set to 0.8 to 1.0 to highlight the discriminative value of acoustic feature parameters within the region; the weighting coefficient for medium-value regions was set to 0.5 to 0.7 to balance feature contribution and noise interference; and the weighting coefficient for low-value regions was set to 0.2 to 0.4 to reduce the impact of noise features on subsequent analysis. Within the set coefficient range, fine-tuning was performed based on the specific values ​​of the comprehensive regional attribute indicators for each unit. For example, in high-value regions, the higher the indicator value of a unit, the closer the weighting coefficient was to 1.0; in low-value regions, the lower the indicator value of a unit, the closer the weighting coefficient was to 0.2. At the same time, the regional attribute indicators under normal and fault conditions in historical monitoring data were referenced to verify the rationality of the weighting coefficients, ensuring that the coefficients could effectively distinguish regions with different discriminative values ​​and solving the problem of fault features being overwhelmed by equal weighting in traditional methods.

[0116] Step S630 involves dynamically adjusting the weights of each feature component in the acoustic feature parameters based on the weighting coefficients to obtain optimized acoustic feature parameters. Specifically, this includes retrieving the weighting coefficients corresponding to each convex polygon partition unit determined in step S620, as well as the values ​​of each feature component of all acoustic feature parameters within that unit. For each of the 36 feature components of an acoustic feature parameter, dynamic weight adjustment is performed using the weighting coefficients of its respective unit. Specifically, the adjustment method involves multiplying the original value of each feature component in the acoustic feature parameter by its corresponding weighting coefficient to obtain the adjusted value of that feature component. For example, if the weighting coefficient of a high-value region unit is 0.9, and the original value of a certain feature component of an acoustic feature parameter within that unit is 2.5, then the adjusted value is 2.5 multiplied by 0.9. After adjusting the weights of all feature components, the optimized acoustic feature parameters corresponding to each acoustic feature parameter are obtained. After adjustment, the optimization results need to be checked to ensure that the feature component values ​​of the acoustic feature parameters in high-value regions are reasonably amplified, and the feature component values ​​in low-value regions are appropriately suppressed, thereby highlighting the weak acoustic features corresponding to early faults and avoiding interference from normal fluctuations or noise features on fault identification.

[0117] Step S640 involves concatenating the optimized acoustic feature parameters and current feature indices using multimodal features to construct a multidimensional feature vector. Specifically, this includes retrieving the optimized acoustic feature parameters obtained in step S630, which are 36-dimensional feature vectors; and simultaneously retrieving the current feature indices extracted in step S250, which contain three feature parameters: peak amplitude, third harmonic component, and phase difference, forming a 3-dimensional feature vector.

[0118] For each spatial scatter point, the optimized acoustic feature parameters are paired with the current feature index corresponding to the synchronously acquired leakage current signal to ensure that each set of optimized acoustic feature parameters can be matched with the current feature index at the same monitoring time. This pairing is based on the time synchronization calibration in step S210, ensuring the consistency of acoustic and current features in the time dimension. The paired 36-dimensional optimized acoustic feature parameters and 3-dimensional current feature index are feature-stitched together and combined in the order of acoustic features first and current features last to form a 39-dimensional multi-dimensional feature vector. Each multi-dimensional feature vector contains the acoustic and current multi-modal information of the surge arrester at the same monitoring time, solving the problem of the traditional method not effectively integrating single physical quantities or multiple parameters.

[0119] Step S650 involves standardizing and normalizing the multidimensional feature vectors to eliminate differences in feature dimensions and obtain a standardized feature matrix. Specifically, this includes: statistically analyzing the numerical sets of each feature dimension in all multidimensional feature vectors; and for each feature dimension, calculating the mean and standard deviation of all values ​​within that dimension. The mean reflects the overall level of that feature dimension, and the standard deviation reflects the degree of dispersion of the values.

[0120] Based on the calculated mean and standard deviation, the values ​​of each dimension in each multidimensional feature vector are standardized and normalized. The process involves subtracting the mean of that dimension from its original value and then dividing by its standard deviation, resulting in a mean of zero and a standard deviation of one for each feature dimension. After standardization of all multidimensional feature vectors, these vectors are arranged in chronological order of monitoring time to form a standardized feature matrix. After standardization, the numerical range of each feature dimension in the matrix needs to be verified to ensure that the values ​​of all dimensions are within similar orders of magnitude. This eliminates the influence of differences in dimensions between acoustic and current features (e.g., acoustic features are coefficients, while current features are microampere amplitudes or angles), and prevents a single feature from dominating the model learning process due to its excessively large dimension. This ensures that the subsequent one-dimensional convolutional neural network can learn the effective information from the multimodal features in a balanced manner, thereby improving the accuracy of fault identification.

[0121] Step S660: Input the standardized feature matrix into the pre-trained one-dimensional convolutional neural network model, and extract deep feature representations through alternating processing of multiple one-dimensional convolutional layers and pooling layers;

[0122] Step S670: Based on the extracted deep feature representation, state recognition is performed through a fully connected layer and a softmax classifier, and the classification probability distribution of the arrester's operating state is output.

[0123] In this embodiment of the invention, by establishing a mapping between comprehensive regional attribute indicators and acoustic features, the acoustic features of high-priority regions (including high-density, low-uniformity regions with potential fault features) are given higher weights, effectively amplifying early weak fault signals, such as the partial discharge features of water ingress due to poor sealing, preventing them from being submerged by normal operation fluctuations, and solving the problem that fault features are not highlighted due to traditional equal weighting. By combining optimized acoustic features (reflecting details of partial discharge) with current features (reflecting core indicators of insulation degradation), the system covers the multi-physical quantities of early-stage surge arrester faults. This overcomes the shortcomings of traditional single-feature monitoring (such as measuring only leakage current) in comprehensively characterizing complex faults, thus improving the completeness of feature representation. Standardization unifies the dimensions of each feature dimension, such as acoustic MFCC, current microampere value, and phase angle, preventing the model from being overly biased towards a certain type of feature due to differences in feature numerical ranges. This ensures that all features participate fairly in model calculations and reduces diagnostic bias caused by dimensional interference. Through alternating convolutional pooling of 1D-CNN, deep-level fault patterns in features can be automatically captured, which is more accurate than traditional manual extraction and avoids the loss of details due to linear dimensionality reduction. The softmax output probability distribution not only determines the normal / abnormal classification results but also provides the confidence level of each state, making it easier for maintenance personnel to intuitively judge the fault risk level. At the same time, the pre-trained model replaces manual judgment, reducing subjective errors and improving the level of diagnostic automation and the reliability of results.

[0124] In a preferred embodiment of the present invention, step S660 involves inputting the standardized feature matrix into a pre-trained one-dimensional convolutional neural network model, and extracting deep-level feature representations through alternating processing of multiple one-dimensional convolutional layers and pooling layers, including:

[0125] Step S661 involves inputting the standardized feature matrix into the input layer of a pre-trained one-dimensional convolutional neural network for data format conversion to generate an initial feature tensor. Specifically, this includes retrieving the standardized feature matrix generated in step S650. This matrix integrates the multimodal feature information of the surge arrester during operation, including the 36-dimensional acoustic feature parameters optimized in step S630 (corresponding to the ultrasonic spectrum features generated by partial discharge, which can capture the weak acoustic changes of early faults such as poor sealing and water ingress) and the 3-dimensional current feature indicators extracted in step S250 (including peak amplitude, third harmonic component, and phase difference, which can reflect the weak fluctuations of leakage current). Each sample at each monitoring time corresponds to a complete set of multimodal feature data, overcoming the limitation of traditional methods that rely solely on the acquisition of a single physical quantity.

[0126] The standardized feature matrix is ​​input into the input layer of the pre-trained one-dimensional convolutional neural network. The input layer first parses the dimensional structure of the matrix and confirms that the first dimension of the matrix is ​​the number of monitoring samples (covering the arrester status data of different operating stages), the second dimension is the 39-dimensional feature dimension of each sample (36-dimensional acoustic + 3-dimensional current), and the third dimension is the number of feature channels (fixed to 1, because the multimodal features have been spliced ​​and integrated, so there is no need for multi-channel input, but the correlation of multi-features under a single channel needs to be retained).

[0127] Subsequently, data format conversion is performed. According to the preset receiving specifications of the network input layer, the matrix is ​​adjusted to a three-dimensional initial feature tensor to ensure that the number of samples in the feature tensor is consistent with the original matrix, the feature dimension accurately corresponds to the 39-dimensional multimodal features, and the number of channels remains unchanged at 1. At the same time, the data type of the feature tensor is uniformly converted to the floating-point type supported by the network. Each sample's feature data is checked one by one for missing or outlier values. If an anomaly is found, it is filled in based on the feature trend of adjacent samples in the same time period to avoid the loss of key information in subsequent feature extraction due to data format incompatibility or outliers. This process is specifically designed to address the defect of traditional methods in processing high-dimensional nonlinear acoustic signals, which is prone to losing local subtle features. Through precise format conversion and data verification, the weak multimodal features of early faults are completely preserved.

[0128] Step S662 involves performing a one-dimensional convolution operation on the initial feature tensor through the first convolutional layer to extract the primary feature map, and then implementing a nonlinear feature transformation using the ReLU activation function. Specifically, the first convolutional layer pre-configures 32 one-dimensional convolutional kernels, each with a size of 5. This size is chosen based on the spectral distribution characteristics of the ultrasonic signal from partial discharge of the surge arrester, ensuring coverage of the correlation information across five consecutive feature dimensions. This avoids missing weak spectral anomalies due to excessively small kernel size, or feature aliasing due to excessively large kernel size. Simultaneously, the parameters of the convolutional kernels are pre-trained and optimized, learning the feature difference patterns between normal operation and early fault states. This enhances the ability to capture low-amplitude features, particularly addressing the weak energy of the ultrasonic signal during water ingress due to poor sealing.

[0129] The initial feature tensor generated in step S661 is input into the first convolutional layer. The convolutional kernel slides along the feature dimension direction with a stride of 1. Each time it slides, the convolutional kernel is multiplied element-wise with the five consecutive feature dimensions of the region corresponding to the feature tensor and then summed to generate a feature mapping value. By fully sliding, all 39 feature dimensions are covered. Each convolutional kernel outputs a primary feature mapping containing 35 feature values. A total of 32 primary feature mappings are output by 32 convolutional kernels. Each mapping focuses on the correlation information of different dimensions in the multimodal features. For example, some mappings capture the low-frequency changes of the ultrasonic spectrum, and some mappings capture the synchronous fluctuations of the third harmonic of the leakage current and the acoustic features.

[0130] The ReLU activation function is applied to each of the 32 generated primary feature maps, converting all negative feature map values ​​to 0 and retaining positive feature map values. This nonlinear transformation is specifically designed to address the shortcomings of traditional linear dimensionality reduction methods (such as PCA) in characterizing high-dimensional nonlinear acoustic signals. It can effectively enhance the network's ability to express nonlinear features. For example, when water enters due to poor sealing, the difference between the ultrasonic signal generated by partial discharge and the normal operating noise exhibits nonlinear characteristics. ReLU activation can amplify this difference, suppress meaningless negative noise signals, and clearly retain the weak feature traces of early faults in the primary feature maps.

[0131] Step S663: Input the nonlinearly transformed feature map into the first pooling layer for spatial downsampling to obtain the dimensionality-reduced feature representation. Specifically, the first pooling layer adopts max pooling, with the pooling window size set to 2 and the step size set to 2. This parameter configuration is based on the fluctuation pattern of the surge arrester feature data: the feature fluctuation is smooth during normal operation, and local peak features will appear during faults (such as the instantaneous amplitude increase of the ultrasonic signal and the pulse fluctuation of the leakage current). Max pooling can accurately retain these peak features while removing redundant smooth fluctuation data.

[0132] The 32 primary feature maps obtained from the nonlinear transformation in step S662 are input into the first pooling layer. The pooling window slides along the feature dimension with a step size of 2. Each time it slides, the feature value with the largest value among the two consecutive feature values ​​covered by the window is selected as the output value of the window. For example, if the two feature values ​​in a certain window correspond to normal noise fluctuations and weak peaks of partial discharge, the peak value is selected as the output to ensure that no key features related to the fault are lost. After spatial downsampling, the dimension of each primary feature map is compressed from 35 to 17 (35 / 2 rounded down). The number of 32 feature maps remains unchanged, and the overall feature data volume is reduced by half. This process not only solves the problem of data redundancy causing fault features to be submerged in traditional methods, but also strengthens the early fault discrimination criteria by retaining peak features. For example, in the early stage of water ingress due to poor sealing of the surge arrester, the peak of the ultrasonic signal of partial discharge is weak and sparse. Max pooling can filter out these sparse peaks to avoid being covered by a large amount of normal data, while reducing the computational complexity of subsequent deep feature extraction.

[0133] Step S664 involves inputting the dimensionality-reduced feature representation into the second convolutional layer for deep feature extraction, generating a high-level feature map, and then applying the ReLU activation function for nonlinear activation. Specifically, the second convolutional layer is configured with 64 one-dimensional convolutional kernels, each with a size of 3, which is smaller than that of the first convolutional layer. The purpose is to capture deep correlation features more precisely in the feature map after the first dimensionality reduction, such as the synchronous correlation between subtle changes in the high-frequency band of the ultrasonic spectrum (corresponding to the pulse width change of partial discharge) and the phase difference of the leakage current. These features cannot be identified in traditional methods because they are not multimodal deep fusion. The number of 64 convolutional kernels is set to comprehensively explore the complex correlation patterns between multimodal features through more channels, especially for the weak correlation of multimodal features in early faults, thereby enhancing the comprehensiveness of feature extraction.

[0134] The 32 feature representations after dimensionality reduction in step S663 are input into the second convolutional layer. The convolutional kernel slides along the feature dimension direction with a stride of 1. During each slide, the convolutional kernel is element-wise multiplied and summed with the three consecutive feature dimensions of the region corresponding to the input feature, generating a high-level feature map value. By fully sliding and covering all 17 feature dimensions, each convolutional kernel outputs a high-level feature map containing 15 feature values ​​(17-3+1). A total of 64 convolutional kernels output 64 high-level feature maps. These maps are no longer single-dimensional features, but rather integrate deep correlation information between acoustic and current features. For example, a certain high-level feature map can reflect the synchronous trend of the change in the Mel frequency cepstral coefficient of the ultrasonic signal and the third harmonic component of the leakage current. This is precisely the effective fusion of multimodal features that traditional methods have failed to achieve, solving the problem of neglecting the correlation of multiple features in early faults.

[0135] The ReLU activation function was applied again to the 64 high-level feature maps generated, and nonlinear activation processing was performed. This activation further enhanced the effective information in the deep correlation features and suppressed the interference of irrelevant noise. For example, when the surge arrester is operating normally, the correlation between acoustic and current features is weak and exhibits random fluctuations. ReLU activation can filter out the negative signals of these random fluctuations, while the weak correlation features during early faults (such as the synchronous occurrence of ultrasonic anomalies and leakage current phase difference shifts) will be retained and amplified in the form of positive values, so that the correlation pattern of fault features is clearly presented in the high-level feature maps.

[0136] Step S665 involves inputting the activated high-level feature map into the second pooling layer for further spatial downsampling to obtain a compressed deep feature representation. Specifically, the second pooling layer also employs max pooling, with a pooling window size of 2 and a stride of 2, maintaining consistency with the parameters of the first pooling layer. This ensures the continuity and stability of the feature reduction process and avoids the loss of key features due to parameter mutations. This pooling configuration is specifically designed for the dispersed and weak characteristics of early fault features. By further filtering peak features, redundant information in normal operation data is compressed, increasing the proportion of fault features in the overall features.

[0137] The 64 high-level feature maps activated in step S664 are input into the second pooling layer. The pooling window slides along the feature dimension with a step size of 2, selecting the maximum value from the two consecutive feature values ​​covered by the window as the output value each time. For example, if two feature values ​​within a window correspond to the peak values ​​of normal feature association and weak association of early faults, respectively, only the peak value is retained after pooling, ensuring that the fault features with deep associations are not masked by normal association information. After this spatial downsampling process, the dimension of the 64 high-level feature maps is compressed from 15 to 7 (rounded down from 15 / 2), reducing the feature data volume by half again, while further improving the discriminative value of the features. At this point, the feature representation has highly condensed the deep association information of multimodal features, including key feature traces of early faults such as poor sealing and water ingress, such as the synchronous association between abnormal ultrasonic spectrum and weak fluctuations in leakage current, and the correspondence between the increase of the third harmonic component and changes in acoustic features. This solves the problem of insufficient feature fusion leading to insufficient fault identification sensitivity in traditional methods.

[0138] Step S666 involves vectorizing the feature tensor that has undergone two convolution and pooling processes to generate a one-dimensional feature vector. Specifically, this includes performing dimensionality analysis on the feature tensor that has undergone two convolution and pooling processes to confirm its current three-dimensional structure: the first dimension is the number of monitoring samples (consistent with the number of samples in the initial input to ensure the feature integrity of each sample), the second dimension is the feature dimension of each sample after two dimensionality reductions (7 feature values), and the third dimension is the number of feature maps (64, corresponding to 64 deep-association feature channels).

[0139] The 3D feature tensor is then vectorized, following the order of the feature maps: for each sample, the first 7 feature values ​​from the first feature map are extracted and arranged into a continuous numerical sequence according to the feature dimension; then the second 7 feature values ​​from the second feature map are extracted and concatenated after the first sequence; and so on, until all feature values ​​from all 64 feature maps are concatenated to form a continuous one-dimensional array. The length of the one-dimensional array corresponding to each sample is the product of the feature dimension and the number of feature maps (7 × 64 = 448), meaning each sample corresponds to a 448-dimensional one-dimensional feature vector.

[0140] During the transformation process, each sample's one-dimensional feature vector is checked one by one for missing data or disordered order. This ensures that each vector fully integrates all the deep-level feature information extracted by the two convolutions and pooling processes. This information includes multimodal key correlation features of the arrester's operating state, especially the weak feature traces of early faults (such as poor sealing and water ingress). This avoids the defects of isolated parameters and insufficient feature fusion in traditional methods. The generated one-dimensional feature vector can be directly input into the subsequent fully connected layer for state classification, providing a comprehensive and condensed feature basis for accurate identification of early faults. This solves the problem that fault features are submerged by normal data fluctuations due to insufficient feature expression in traditional methods.

[0141] In a preferred embodiment of the present invention, step S670, based on the extracted deep feature representation, performs state recognition through a fully connected layer and a softmax classifier, and outputs the classification probability distribution of the arrester's operating state, including:

[0142] Step S671 involves inputting the vectorized feature representation into the first fully connected layer for feature dimensionality reduction and feature recombination to generate a recombined feature vector. Specifically, this includes retrieving the vectorized feature representation generated in step S666. This feature vector is a one-dimensional structure, with each sample corresponding to a vector of length 448. It integrates the deep multimodal features extracted by two convolutions and pooling in a one-dimensional convolutional neural network. It includes subtle changes in the ultrasonic spectrum of partial discharge of the surge arrester (corresponding to weak acoustic signals of early faults such as poor sealing and water ingress) and subtle fluctuations in leakage current (such as slight increases in the third harmonic component and slight shifts in phase difference). This avoids the feature bias caused by traditional methods that rely solely on the acquisition of a single physical quantity. At the same time, it preserves the deep correlation information between multimodal features and solves the defect of losing local subtle features when traditional linear dimensionality reduction methods (such as PCA) process high-dimensional nonlinear acoustic signals.

[0143] The one-dimensional feature vector is input into the pre-trained first fully connected layer. This layer is pre-configured with suitable weight parameters, which are optimized through training on a large number of historical samples of normal operation of surge arresters, early anomalies (such as the initial stage of water ingress due to poor sealing), and severe anomalies (such as a significant increase in leakage current). This allows for accurate identification of feature association patterns valuable for state classification. The core function of the first fully connected layer is feature dimensionality reduction and feature recombination: on the one hand, it compresses the 448-dimensional high-dimensional feature vector to a medium-dimensional 128-dimensional vector, reducing subsequent computational complexity while avoiding feature redundancy caused by excessive dimensionality, ensuring that the retained features are all key information strongly related to state discrimination; on the other hand, it recombines the features through linear transformation of the weight matrix, integrating the multimodal correlation features originally scattered across different dimensions (such as the synchronous features of ultrasonic spectrum anomalies and leakage current phase difference shifts) into a more concentrated feature expression, strengthening the proportion of weak early fault signals in the feature vector, and preventing these signals from being masked by the stable features of normal operation.

[0144] During feature dimensionality reduction and recombination, the numerical distribution of the feature vector is monitored in real time to ensure that the feature values ​​of each dimension are within a reasonable range (avoiding feature failure caused by numerical overflow or excessively small values). At the same time, the overall feature distribution is adjusted by pre-trained bias parameters to make the recombined feature vector more in line with the subsequent classification requirements. Finally, a 128-dimensional recombined feature vector is generated, with each dimension carrying key correlation information of multimodal features.

[0145] Step S672: Input the recombined feature vector into the second fully connected layer to extract classification and discrimination features and obtain the final feature representation. Specifically, this includes: inputting the 128-dimensional recombined feature vector generated in step S671 into the second fully connected layer. The weight parameters of this layer have also been optimized through training with historical fault samples, and the weight configuration is more focused on the discrimination difference between normal and abnormal states. In particular, for early abnormal states that are difficult to identify by traditional methods (such as weak ultrasonic signal energy and small leakage current fluctuations caused by poor sealing and water ingress), this layer will assign higher weights to the feature dimensions corresponding to such states to ensure that early fault features are not overwhelmed by a large number of stable features of normal operation.

[0146] The core task of the second fully connected layer is to extract classification and discrimination features. Through the operation of the weight matrix and the recombined feature vector, it filters and amplifies the feature information directly related to the operation status of the surge arrester. For example, the feature dimensions under normal operation (such as stable ultra-high frequency spectrum, high proportion of leakage current fundamental wave, and stable phase difference) will form a specific numerical pattern. The feature dimensions under early abnormal conditions (such as sporadic high-frequency anomalies in ultra-high frequency spectrum, slight increase in the proportion of third harmonic leakage current, and small fluctuations in phase difference) will form another type of numerical pattern. The feature dimensions under severe abnormal conditions (such as dense high-frequency anomalies in ultra-high frequency spectrum, significant increase in leakage current amplitude, and drastic fluctuations in phase difference) will form a third type of numerical pattern.

[0147] During feature extraction, activation functions (such as ReLU) are used to further enhance the discriminative power of the discriminative features. Low-value features that are meaningless for state classification are suppressed to zero, while features with discriminative value (especially weak features of early anomalies) are retained and amplified to avoid meaningless features interfering with the classification results. Finally, a 32-dimensional final feature representation is generated, where each dimension of the vector corresponds to a specific state discriminative feature. The features of early abnormal states have been effectively highlighted through weight adjustment and activation processing, solving the problems of insufficient feature fusion and neglect of the discriminative value of fault features in traditional methods.

[0148] Step S673: Input the final feature representation into the softmax classifier and calculate the posterior probability distribution of each state category. Specifically, this includes: inputting the 32-dimensional final feature representation generated in step S672 into the softmax classifier. The classifier is pre-set with three output categories, corresponding to the normal operation state, early abnormal state, and severe abnormal state of the surge arrester, respectively. The parameters of the classifier have been trained with historical samples to adapt to the discrimination rules of multimodal features. In particular, for the weak features of early abnormal states (such as the initial stage of water ingress due to poor sealing), the classifier can improve its sensitivity to such features by adjusting the parameters, avoiding the defect of traditional classification methods that do not respond to early faults.

[0149] The classifier first performs a linear transformation on the final feature representation, mapping the 32-dimensional features to three category dimensions. Each category dimension corresponds to a raw score, which reflects the degree of matching between the current sample and the features of that category. Then, the softmax function converts the three raw scores into probability values ​​between 0 and 1, with the sum of the three probability values ​​being 1, forming a complete posterior probability distribution. The probability of a normal operating state reflects the degree to which the sample conforms to the characteristics of stable operation; the probability of an early abnormal state reflects the degree to which the sample has weak fault signals (such as abnormal ultrasonic spectrum or small fluctuations in leakage current); and the probability of a severe abnormal state reflects the degree to which the sample has obvious fault signals.

[0150] During probability calculation, the contribution of weak features is specifically preserved for the probability calculation of early abnormal states. Even if the ultrasonic signal energy is weak and the leakage current fluctuation is small, as long as these features match the feature patterns of historical early fault samples, the classifier will still output an early abnormal probability with a certain confidence level, avoiding the problem of traditional methods directly judging normal states due to low fault signal energy. At the same time, the probability distribution is checked for reasonableness to ensure that there are no abnormal cases with negative probability values ​​or probabilities exceeding 1. If a certain type of probability is significantly abnormal (such as being much higher than the other two types and exceeding the reasonable range), the final input feature representation is re-checked for any deviation to ensure that the probability distribution can truly reflect the state characteristics of the sample.

[0151] Step S674: Based on the calculated posterior probability distribution, the maximum a posteriori probability criterion is used to determine the classification result of the surge arrester's operating status, including three diagnostic results: normal operating status, early abnormal status, and severe abnormal status. Specifically, this includes: determining the probability judgment rules for the three types of operating status. These rules are set based on the probability distribution of historical fault samples, and the probability judgment threshold for early abnormal status is appropriately lowered to meet the identification needs of early abnormal status, avoiding being ignored because the probability of early faults is lower than that of severe abnormalities. The judgment criterion for normal operating status is that its probability value is greater than 0.6 and significantly higher than the other two types, the difference between it and the probability of early abnormality is greater than 0.2, and the difference between it and the probability of severe abnormality is greater than 0. 3. For stable operation scenarios where the surge arrester has no partial discharge abnormalities and the leakage current is stable; the criteria for judging early abnormal states is that its probability value is greater than 0.3 and higher than the probability of normal operation (difference greater than 0.05) or higher than the probability of severe abnormal states, with a difference greater than 0.1, corresponding to early fault scenarios such as poor sealing and water ingress. At this time, there are slight abnormalities in the ultrasonic partial discharge signal and slight fluctuations in the leakage current, but it does not increase significantly; the criteria for judging severe abnormal states is that its probability value is greater than 0.5 and significantly higher than the other two categories, with a difference of greater than 0.4 from the probability of normal operation and a difference of greater than 0.2 from the probability of early abnormal states, corresponding to fault scenarios with significantly increased leakage current and obvious partial discharge abnormalities.

[0152] Based on the above rules, the maximum a posteriori probability criterion is applied to the posterior probability distribution generated in step S673: the probability values ​​of the three categories are compared, and the category with the largest probability value is taken as the preliminary classification result. If the category corresponding to the maximum probability value meets the judgment criteria of that category, it is directly determined as the classification result of the arrester's operating status; if special cases occur, such as the maximum probability value being an early abnormal state but the probability is only 0.35, and the difference between the probability value and the probability value of the normal operating state is only 0.04, then a secondary verification is initiated, retrieving the original multimodal features (optimized acoustic feature parameters and current feature indicators) corresponding to the sample, and manually checking for any abnormal ultrasonic spectrum or leakage current fluctuations to avoid misjudgment caused by a single probability criterion and ensure that early abnormal states are not missed.

[0153] After determining the classification result, record the corresponding probability distribution information (e.g., early abnormal state probability 0.38, normal operation 0.33, severe abnormality 0.29). If it is an early abnormal state, even if the probability does not reach a high value, timely inspection should be arranged (e.g., checking the sealing of the surge arrester) to avoid maintenance delays caused by traditional methods that only alarm when the leakage current increases significantly. If it is a normal operation state, monitor according to the regular cycle. If it is a severe abnormal state, immediately initiate the emergency maintenance procedure. At the same time, compare the classification results with historical data of the same period to analyze the trend of state change (e.g., whether it gradually transitions from normal to early abnormality), further improve the foresight of fault early warning, and solve the core problems of insufficient fault identification sensitivity and alarm lag in traditional methods.

[0154] In a preferred embodiment of the present invention, step S700 involves determining whether the surge arrester is in an abnormal operating state based on the classification result. If it is determined to be abnormal, an alarm information generation program is initiated, and an abnormal state warning signal is output, including:

[0155] Step S710 involves receiving the operational status classification results and extracting the probability distribution data of the status categories. Specifically, this includes receiving the arrester operational status classification results output in step S674. These results include three diagnostic conclusions: normal operation status, early abnormal status, and severe abnormal status, along with the posterior probability distribution data corresponding to each status category. These classification results and probability data are generated based on multimodal feature fusion analysis, integrating the subtle changes in the ultrasonic spectrum of partial discharge of the arrester and the weak fluctuations in leakage current. Unlike traditional methods that rely solely on a single physical quantity (such as leakage current amplitude) to arrive at partial conclusions, this approach can completely preserve the weak signal traces of early faults, avoiding misjudgments of the status due to a single feature.

[0156] Extract the probability distribution data of the state categories from the classification results, and record the specific values ​​of the probability of normal operation, early abnormal state, and severe abnormal state one by one to ensure the completeness and accuracy of the probability data. For example, if the classification result is an early abnormal state, the probability value corresponding to this state, as well as the probability values ​​of normal operation and severe abnormal state, need to be extracted simultaneously to provide a complete basis for subsequent threshold analysis. During the extraction process, check whether the probability data meets the basic rule that the sum is one. If the probability sum deviates from the allowable range, the posterior probability calculation process in step S673 is called back to check for abnormal feature vector processing or classifier parameter deviation, ensuring that the extracted probability distribution data can truly reflect the actual operating state of the surge arrester and solve the problems of traditional methods lacking probability support and having strong subjectivity in state judgment.

[0157] The extracted state categories are associated with and stored with probability distribution data, establishing a correspondence between state categories, probability values, and monitoring time. This facilitates subsequent tracing of the time node when the state occurs and analysis of the state's changing trend over time, providing a data foundation for early fault prediction and avoiding the limitations of traditional methods that only record the current state and cannot trace historical changes.

[0158] Step S720: Analyze the probability distribution of state categories based on preset state judgment thresholds to determine the working state level of the surge arrester. Specifically, this includes: setting judgment thresholds for three types of operating states in advance. These thresholds are optimized and determined based on a large amount of historical operating data and fault cases of surge arresters. The threshold standards are adjusted with a focus on the identification needs of early abnormal states, solving the problem of missed detection of early faults due to excessively high thresholds in traditional methods. The specific thresholds are set as follows: The threshold for determining normal operation is a probability value greater than 0.6, with a difference of greater than 0.2 from the probability of early abnormal operation and a difference of greater than 0.3 from the probability of severe abnormal operation, ensuring the reliability of normal operation determination; the threshold for determining early abnormal operation is a probability value greater than 0.3, which is at least 0.05 higher than the probability of normal operation or at least 0.1 higher than the probability of severe abnormal operation. This threshold is lower than the abnormality determination standard of traditional methods and can capture weak signals of early faults such as poor sealing and water ingress; the threshold for determining severe abnormal operation is a probability value greater than 0.5, with a difference of greater than 0.4 from the probability of normal operation and a difference of greater than 0.2 from the probability of early abnormal operation, corresponding to fault scenarios with significantly increased leakage current and obvious abnormal partial discharge.

[0159] Based on the extracted probability distribution data of the state categories, each category is analyzed against a preset threshold. If the probability of normal operation meets the judgment threshold and no other state probabilities exceed the standard, the arrester's operating state level is determined to be normal operation. At this time, the partial discharge ultrasonic spectrum of the arrester is stable, the leakage current does not fluctuate significantly, and there is no need to activate the alarm. If the probability of early abnormal state meets the judgment threshold, regardless of whether the probabilities of other states are close to the threshold, the operating state level is determined to be early abnormal level. The arrester may have early defects such as poor sealing and water ingress. The ultrasonic signal energy generated by partial discharge is weak, and the leakage current fluctuates slightly but does not increase significantly. If the probability of severe abnormal state meets the judgment threshold, the operating state level is determined to be severe abnormal level. At this time, the arrester insulation deteriorates significantly, the leakage current increases significantly, and the partial discharge ultrasonic signal is abnormally dense.

[0160] If the probability distribution data does not clearly meet any state threshold, for example, the probability of an early abnormal state is 0.32 and the probability of a normal operating state is 0.30, with a difference of only 0.02, a secondary verification process is initiated. The original multimodal feature data (optimized acoustic feature parameters and current feature indices) corresponding to the sample is retrieved, and manual verification is performed to check for sporadic high-frequency anomalies in the ultrasonic spectrum and whether the third harmonic component of the leakage current is slightly elevated. The state level is then re-determined based on the actual feature details, avoiding missed early anomalies due to a single probability criterion, ensuring the accuracy of state level determination, and overcoming the shortcomings of traditional methods with fixed thresholds that cannot handle ambiguous states.

[0161] Step S730: When an abnormal working state is determined, a multi-level alarm triggering mechanism is activated. Corresponding alarm commands are generated based on the abnormality level. Specifically, this includes establishing a multi-level alarm triggering mechanism that corresponds one-to-one with the working state levels determined in step S720. Differentiated alarm response strategies are set for different abnormality levels to avoid the problem of traditional methods only triggering alarms when the fault is severe, thus missing early maintenance opportunities. The specific mechanism is as follows: When the working state level is normal operation, no alarm mechanism is triggered; only status information is recorded and the next monitoring cycle begins. When the working state level is early abnormality, a mild alarm mechanism is triggered. This mechanism focuses on reminding maintenance personnel to pay attention and conduct inspections, without initiating an emergency response process. When the working state level is severe abnormality, a severe alarm mechanism is triggered. This mechanism focuses on rapid response and emergency handling to ensure the fault does not escalate further.

[0162] When an early abnormal state is identified, a mild alarm trigger command is initiated. This command includes information such as the abnormality level (early stage), possible fault type (e.g., poor sealing leading to water ingress, slight local insulation degradation), and recommended inspection time (e.g., within 24 hours). This information is derived based on the multimodal characteristics corresponding to the early abnormal state (e.g., sporadic abnormalities in the high-frequency band of the ultrasonic spectrum, slight fluctuations in the leakage current phase difference), providing maintenance personnel with targeted inspection directions and avoiding the problem of traditional alarms only indicating abnormalities without fault location information.

[0163] When a severe abnormal condition is determined, a severe alarm trigger command is initiated. This command includes the abnormality level (severe), possible fault types (such as severe insulation degradation, internal discharge), and emergency handling suggestions, such as immediate shutdown for maintenance and replacement of the surge arrester. It also indicates the urgency level of the fault (e.g., Level 1 Emergency) to ensure maintenance personnel can quickly determine the severity of the fault and activate the appropriate emergency plan. When the severe alarm command is triggered, key parameters such as the current leakage current amplitude and the intensity of the partial discharge ultrasonic signal are recorded simultaneously, solving the problems of lacking key parameters and difficulty in fault analysis in traditional severe fault alarms.

[0164] Step S740: Based on the alarm command, the alarm information generation module is invoked to construct a multi-dimensional alarm data packet containing timestamp, device identifier, anomaly type, and risk level. Specifically, based on the alarm command generated in step S730, the built-in alarm information generation module of the system is invoked. This module pre-stores basic information of the surge arrester, such as device model, installation location, commissioning time, and the line number to which it belongs, as well as the standard format of alarm information, to ensure that the generated alarm data packet is complete and uniform in format, avoiding the problems of fragmented alarm information and missing key information in traditional alarms.

[0165] Construct a multi-dimensional alarm data packet, which should contain at least the following core information: timestamp, accurate to milliseconds, recording the specific time the alarm was triggered, facilitating the accurate tracing of the moment the anomaly occurred; device identifier, including the surge arrester's unique number, installation location, and associated circuit information, ensuring accurate location of abnormal devices in the multi-surge arrester monitoring system; anomaly type, clearly indicating the anomaly level (early or severe) and possible causes of the fault, derived based on multimodal characteristics, such as poor sealing leading to water ingress or localized insulation degradation; and risk level, labeled according to the anomaly level, such as early anomalies corresponding to low risk levels and severe anomalies corresponding to high risk levels, providing a basis for maintenance personnel to determine the priority of handling.

[0166] During the construction process, the accuracy of information in each dimension is verified. For example, the timestamp must be synchronized with the system's high-precision clock to avoid misjudging the time of anomalies due to time deviations; the device identifier must be consistent with the surge arrester ledger information to avoid device confusion; the anomaly type must match the anomaly level and feature analysis results in the alarm command to ensure that the fault cause inference is reasonable; if any dimension of information is found to be missing or incorrect, it must be supplemented or corrected from the system database in a timely manner to ensure that the alarm data packet can comprehensively reflect the anomaly situation and solve the problems of traditional alarm information being singular and unable to support accurate maintenance.

[0167] Step S750: The alarm data packet is transmitted to the monitoring center via the communication interface. Simultaneously, the local audible and visual alarm device is activated, outputting an abnormal status warning signal. Specifically, this includes: transmitting the constructed alarm data packet to the remote monitoring center via a preset communication interface; the communication interface supports multiple transmission methods such as Ethernet, 4G, or 5G, automatically selecting the optimal transmission path based on the on-site network environment to ensure stable and unlossable data packet transmission, avoiding the problems of traditional alarm signal transmission methods being singular and susceptible to interference. During transmission, a data encryption and verification mechanism is employed to encrypt the alarm data packet, preventing data tampering during transmission. Simultaneously, a checksum verifies the integrity of the data packet. If the receiving end detects a damaged or incomplete data packet, it automatically requests retransmission, ensuring the monitoring center accurately receives the alarm information.

[0168] Simultaneously with the initiation of remote transmission, the local audible and visual alarm device of the surge arrester is triggered. The local alarm device includes an indicator light and a buzzer. Different abnormal levels correspond to different audible and visual signals: In the early abnormal state, the indicator light flashes yellow light at a frequency of once every 3 seconds, and the buzzer emits a low-frequency (e.g., 1 time / second) warning sound, which alerts on-site personnel without causing excessive interference; In the severe abnormal state, the indicator light flashes red light at a frequency of once every 1 second, and the buzzer emits a high-frequency (e.g., 3 times / second) urgent alarm sound, ensuring that on-site personnel can immediately detect the emergency fault.

[0169] An online monitoring system for surge arresters in power distribution lines includes:

[0170] The acquisition module is used to acquire acoustic spectrum signals and leakage current waveform data generated by the surge arrester during operation;

[0171] The conversion module is used to perform synchronous signal preprocessing on the acoustic spectrum signal and leakage current waveform data to extract acoustic feature parameters and current feature indicators, and convert the acoustic feature parameters into corresponding spatial scatter datasets through high-dimensional feature mapping.

[0172] The construction module is used to construct a two-dimensional feature mapping coordinate system based on the spatial scattered point dataset, and to determine a minimum boundary range in the coordinate system that can enclose all sampling points;

[0173] The partitioning module is used to divide the boundary range into regions based on the spatial distribution characteristics of the sampling points within the minimum boundary range, generating a series of non-overlapping convex region units.

[0174] The calculation module is used to assign each sampling point to the corresponding convex region unit, calculate the surface area of ​​each region unit based on the distribution density and spatial structure characteristics of the sampling points in each unit, and generate a region density index that characterizes the local distribution characteristics.

[0175] The prediction module is used to perform weighted optimization of the acoustic feature parameters using the region density index, fuse the optimized acoustic feature parameters with the current feature index for multimodal features to form a high-dimensional joint feature matrix, and input the feature matrix into a pre-trained one-dimensional convolutional neural network state recognition model for intelligent diagnosis, outputting the classification result of the arrester's operating state; based on the classification result, it determines whether the arrester is in an abnormal operating state, and if it is determined to be abnormal, it starts the alarm information generation program and outputs an abnormal state warning signal.

[0176] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for online monitoring of surge arresters in power distribution lines, characterized in that, The method includes: Acquire acoustic spectrum signals and leakage current waveform data generated by the surge arrester during operation; Synchronous signal preprocessing is performed on the acoustic spectrum signal and leakage current waveform data to extract acoustic feature parameters and current feature indices, and the acoustic feature parameters are converted into corresponding spatial scatter datasets through high-dimensional feature mapping. A two-dimensional feature mapping coordinate system is constructed based on the spatial scattered point dataset, and a minimum boundary range that can enclose all sampling points is determined in this coordinate system; Based on the spatial distribution characteristics of the sampling points within the minimum boundary range, the boundary range is divided into regions to generate a series of non-overlapping convex region units. Each sampling point is assigned to a corresponding convex region unit. Based on the distribution density and spatial structure characteristics of the sampling points in each unit, the surface area of ​​each region unit is calculated to generate a region density index that characterizes the local distribution characteristics. The acoustic feature parameters are weighted and optimized using the region density index. The optimized acoustic feature parameters are then fused with the current feature index in a multimodal manner to form a high-dimensional joint feature matrix. This feature matrix is ​​then input into a pre-trained one-dimensional convolutional neural network state recognition model for intelligent diagnosis, and the classification result of the arrester's operating state is output. Based on the classification results, it is determined whether the surge arrester is in an abnormal working state. If it is determined to be abnormal, the alarm information generation program is started and an abnormal state warning signal is output.

2. The online monitoring method for surge arresters in power distribution lines according to claim 1, characterized in that, Synchronous signal preprocessing is performed on the acoustic spectrum signal and leakage current waveform data to extract acoustic feature parameters and current feature indices. The acoustic feature parameters are then converted into a corresponding spatial scatter dataset through high-dimensional feature mapping, including: The synchronously acquired acoustic signal and leakage current signal are time-synchronized and calibrated to obtain time-aligned acoustic signal sequence and leakage current signal sequence; The time-aligned acoustic signal sequence is windowed and framed, and the leakage current signal sequence is subjected to sliding window averaging filtering to obtain the preprocessed acoustic signal and leakage current signal. Wavelet threshold denoising is performed on the preprocessed acoustic signal to suppress environmental noise interference, and adaptive notch filtering is performed on the preprocessed leakage current signal to suppress power grid frequency interference, thus obtaining the denoised acoustic signal and leakage current signal. Short-time Fourier transform is performed on the denoised acoustic signal to extract the Mel frequency cepstral coefficient feature vector as the acoustic feature parameter; Full-cycle waveform analysis was performed on the denoised leakage current signal to extract the peak amplitude, third harmonic component, and phase difference parameters of the leakage current signal as current characteristic indicators. Each feature component in the extracted acoustic feature parameters is mapped to a three-dimensional feature space through a nonlinear dimensionality reduction method, forming a set of spatial coordinate points, which constitutes a spatial scattered dataset.

3. The online monitoring method for surge arresters in power distribution lines according to claim 2, characterized in that, A two-dimensional feature mapping coordinate system is constructed based on the spatial scattered point dataset, and a minimum boundary range that can encompass all sampling points is determined in this coordinate system, including: Extract the three-dimensional coordinate information of all spatial points in the spatial scatter dataset, and obtain the X-axis coordinate value, Y-axis coordinate value and Z-axis coordinate value of each spatial point in the Cartesian coordinate system; Using the X-axis and Y-axis coordinates of each spatial point as projection references, all spatial points are mapped onto a two-dimensional feature projection plane composed of the X-axis and Y-axis, forming a two-dimensional feature distribution map; On the two-dimensional feature projection plane, traverse the X-axis coordinate values ​​of all projection points to determine the minimum and maximum projection coordinate values ​​in the X-axis direction; On the two-dimensional feature projection plane, traverse the Y-axis coordinate values ​​of all projection points to determine the minimum and maximum projection coordinate values ​​in the Y-axis direction; Based on the obtained minimum and maximum projected coordinate values ​​in the X-axis direction and the minimum and maximum projected coordinate values ​​in the Y-axis direction, determine an initial rectangular boundary that can enclose all projected points. The initial rectangular boundary is extended along the X-axis and Y-axis according to a preset boundary expansion coefficient to obtain the range of the extended minimum bounding rectangle boundary.

4. The online monitoring method for surge arresters in power distribution lines according to claim 3, characterized in that, Based on the spatial distribution characteristics of the sampling points within the minimum boundary range, the boundary range is divided into regions to generate a series of non-overlapping convex region units, including: Based on the extended minimum bounding rectangle boundary range, calculate the overall distribution density of spatial scattered points within the boundary range, and set an adaptive density threshold parameter based on the distribution density. Based on the density threshold parameter, the standard deviation calculation method is used to analyze the dispersion of spatial scatter points in the X-axis direction, and to determine the number of partitions required in the X-axis direction. Based on the density threshold parameter, the variance analysis method is used to evaluate the dispersion of spatial scatter points in the Y-axis direction and determine the number of partitions required in the Y-axis direction. Based on the calculated number of partitions in the X and Y axes, the boundary range of the minimum bounding rectangle is divided into several rectangular partition units of equal area. Each rectangular partition unit is triangulated along its diagonal to generate two triangular partition units. Based on the spatial relationship between adjacent rectangular partition units, an alternating optimization strategy is used to determine the diagonal partitioning direction. All triangular partitioning units are optimized using Delaunay triangulation. Based on the optimization results, multiple non-overlapping convex polygon partitioning units are generated using a polygon fusion algorithm.

5. The online monitoring method for surge arresters in power distribution lines according to claim 4, characterized in that, Each sampling point is assigned to a corresponding convex region cell. Based on the distribution density and spatial structure characteristics of the sampling points within each cell, the surface area of ​​each region cell is calculated, generating a region density index characterizing local distribution characteristics, including: Based on the spatial coordinate information of the convex polygon partitioning unit, a spatial index mapping relationship is established, and each spatial point is assigned to the corresponding convex polygon partitioning unit according to its projected coordinate position. Based on the spatial scatter point allocation results, the number of spatial scatter points contained in each convex polygon partition unit is counted, and the spatial scatter point distribution density index is calculated based on the area of ​​the partition unit. Based on the spatial scatter point distribution density index, the spatial autocorrelation analysis method is used to evaluate the distribution uniformity of spatial scatter points within each partition unit, and the distribution uniformity index is calculated. For each convex polygon partitioning unit, the Delaunay triangulation algorithm is used to divide it into several triangular sub-units based on its vertex coordinates to obtain the triangular mesh partitioning result; Based on the triangular mesh partitioning results, the area of ​​each triangular sub-unit is calculated, and the surface area of ​​the corresponding convex polygon partition unit is obtained by summing the results. The calculated spatial scatter distribution density index, distribution uniformity index, and partition unit surface area are standardized. Based on the standardized parameters, a weighted fusion algorithm is used to generate a comprehensive regional attribute index for each convex polygon partition unit.

6. The online monitoring method for surge arresters in power distribution lines according to claim 5, characterized in that, The acoustic feature parameters are weighted and optimized using the region density index. The optimized acoustic feature parameters are then fused with the current feature index using multimodal features to form a high-dimensional joint feature matrix. This feature matrix is ​​input into a pre-trained one-dimensional convolutional neural network state recognition model for intelligent diagnosis, outputting a classification result of the arrester's operating state, including: Based on the comprehensive regional attribute index of the convex polygon partition unit, the initially acquired acoustic feature parameters are extracted, and the mapping relationship between the attribute index and the feature parameters is established. Based on the comprehensive regional attribute index values ​​of each partition unit, determine the weighting coefficients of the corresponding acoustic feature parameters; Based on the weighting coefficients, the weights of each feature component in the acoustic feature parameters are dynamically adjusted to obtain the optimized acoustic feature parameters. The optimized acoustic feature parameters and current feature indices are combined using multimodal feature concatenation to construct a multidimensional feature vector; The multidimensional feature vectors are standardized and normalized to eliminate differences in feature dimensions and obtain a standardized feature matrix. The standardized feature matrix is ​​input into a pre-trained one-dimensional convolutional neural network model, and deep feature representations are extracted through alternating processing of multiple one-dimensional convolutional layers and pooling layers. Based on the extracted deep feature representation, state recognition is performed through a fully connected layer and a softmax classifier, and the classification probability distribution of the arrester's operating state is output.

7. The online monitoring method for surge arresters in power distribution lines according to claim 6, characterized in that, The standardized feature matrix is ​​input into a pre-trained one-dimensional convolutional neural network model. Through alternating processing of multiple one-dimensional convolutional layers and pooling layers, deep feature representations are extracted, including: The standardized feature matrix is ​​input into the input layer of a pre-trained one-dimensional convolutional neural network to perform data format conversion and generate an initial feature tensor. The initial feature tensor is subjected to a one-dimensional convolution operation by the first convolutional layer to extract the primary feature map, and a nonlinear feature transformation is implemented by the ReLU activation function. The feature map after nonlinear transformation is input into the first pooling layer for spatial downsampling to obtain the dimensionality-reduced feature representation. The dimensionality-reduced feature representation is input into the second convolutional layer for deep feature extraction, generating a high-level feature map, and then the ReLU activation function is applied again for non-linear activation; The activated high-level feature map is input into the second pooling layer for further spatial downsampling to obtain the compressed deep feature representation. The feature tensor, which has undergone two convolution and pooling processes, is vectorized to generate a one-dimensional feature vector.

8. The online monitoring method for surge arresters in power distribution lines according to claim 7, characterized in that, Based on the extracted deep feature representation, state recognition is performed through a fully connected layer and a softmax classifier, outputting the classification probability distribution of the surge arrester's operating state, including: The vectorized feature representation is input into the first fully connected layer for feature dimensionality reduction and feature recombination to generate the recombined feature vector. The recombined feature vector is input into the second fully connected layer to extract classification and discrimination features, and obtain the final feature representation. The final feature representation is input into the softmax classifier to calculate the posterior probability distribution of each state category; Based on the calculated posterior probability distribution, the maximum a posteriori probability criterion is used to determine the classification of the arrester's operating status, including three diagnostic results: normal operating status, early abnormal status, and severe abnormal status.

9. The online monitoring method for surge arresters in power distribution lines according to claim 8, characterized in that, Based on the classification results, it is determined whether the surge arrester is in an abnormal operating state. If it is determined to be abnormal, an alarm information generation program is initiated, and an abnormal state warning signal is output, including: Receive the operation status classification results and extract the status category probability distribution data; Based on the preset state determination threshold, the probability distribution of state categories is analyzed to determine the working state level of the surge arrester. When an abnormal working state is determined, a multi-level alarm triggering mechanism is activated to generate corresponding alarm commands according to the level of abnormality. Based on the alarm command, the alarm information generation module is invoked to construct a multi-dimensional alarm data packet containing timestamp, device identifier, anomaly type and risk level; The alarm data packet is transmitted to the monitoring center through the communication interface, and the local audible and visual alarm device is activated at the same time to output an abnormal status warning signal.

10. An online monitoring system for surge arresters in power distribution lines, characterized in that, The system is used to perform the method as described in any one of claims 1 to 9, comprising: The acquisition module is used to acquire acoustic spectrum signals and leakage current waveform data generated by the surge arrester during operation; The conversion module is used to perform synchronous signal preprocessing on the acoustic spectrum signal and leakage current waveform data to extract acoustic feature parameters and current feature indicators, and convert the acoustic feature parameters into corresponding spatial scatter datasets through high-dimensional feature mapping. The construction module is used to construct a two-dimensional feature mapping coordinate system based on the spatial scattered point dataset, and to determine a minimum boundary range in the coordinate system that can enclose all sampling points; The partitioning module is used to divide the boundary range into regions based on the spatial distribution characteristics of the sampling points within the minimum boundary range, generating a series of non-overlapping convex region units. The calculation module is used to assign each sampling point to the corresponding convex region unit, calculate the surface area of ​​each region unit based on the distribution density and spatial structure characteristics of the sampling points in each unit, and generate a region density index that characterizes the local distribution characteristics. The prediction module is used to perform weighted optimization of the acoustic feature parameters using the region density index, fuse the optimized acoustic feature parameters with the current feature index for multimodal features to form a high-dimensional joint feature matrix, and input the feature matrix into a pre-trained one-dimensional convolutional neural network state recognition model for intelligent diagnosis, outputting the classification result of the arrester's operating state; based on the classification result, it determines whether the arrester is in an abnormal operating state, and if it is determined to be abnormal, it starts the alarm information generation program and outputs an abnormal state warning signal.

Images