Arc fault diagnosis and analysis method based on artificial intelligence algorithm

Abstract

The invention relates to the technical field of power system power distribution network fault detection, in particular to an arc fault diagnosis and analysis method based on an artificial intelligence algorithm, and the method comprises the four steps: synchronous data collection and topological excitation, deployment of terminals at transformer area nodes, injection of characteristic current, and synchronous collection of response waveforms; signal preprocessing: separating a key frequency band through double-digital band-pass filtering, and calculating energy ratio and other enhanced fault features; performing multi-source feature fusion and intelligent diagnosis, constructing a vector containing statistical features, time domain distortion features and topology identification, and inputting a gradient boosting decision tree and deep neural network hybrid model to obtain fault confidence and type; and based on fault positioning and verification of topology, scheduling multi-node cooperative monitoring, and determining a fault point in combination with a topological relation. The method improves the data reliability and diagnosis precision, achieves the precise positioning of a fault, and guarantees the safe operation of a power distribution network.

Description

Technical Field

[0001] This invention relates to the field of power system distribution network fault detection technology, specifically to an arc fault diagnosis and analysis method based on artificial intelligence algorithms. Background Technology

[0002] The current low-voltage line arc fault diagnosis technology in the front area has many shortcomings, making it difficult to meet the actual operation and maintenance needs. Firstly, in the data acquisition stage, existing monitoring solutions mostly use distributed sampling terminals. The lack of a unified clock reference for each terminal leads to time deviations in the current waveform data collected from different nodes, making it impossible to achieve spatiotemporal synchronous analysis. At the same time, existing terminals mostly rely on the natural current signal of the line for monitoring without actively injecting characteristic reference signals. However, the natural current is mixed with interference components such as load switching transient harmonics and motor starting impacts, making it difficult to distinguish fault signals from normal interference signals, and significantly reducing the effectiveness of the collected data. Secondly, in the signal preprocessing stage, traditional technologies often use a single-band filtering method, which only filters signals for power frequency or fixed high-frequency bands, without designing targeted filtering strategies based on the typical high-frequency oscillation characteristics of arc faults. In addition, the feature extraction dimension is singular, focusing on current amplitude changes or simple time-domain parameters, ignoring features that are strongly correlated with arc faults, such as the frequency domain distortion energy ratio. This results in low fault feature recognition, easy masking by noise interference, and inability to provide accurate feature input for subsequent diagnosis. Thirdly, in the intelligent diagnosis stage, existing diagnostic models mostly adopt a single algorithm architecture; threshold-based models are difficult to adapt to the differences in line parameters in different transformer areas, and are prone to false alarms or missed alarms due to unreasonable threshold settings; models based on simple machine learning, such as single decision tree support vector machine, have limited ability to distinguish features under complex operating conditions, especially unable to effectively identify the differences between series arc faults, parallel arc faults and load switching transients, resulting in insufficient diagnostic accuracy and generalization ability. Fourthly, in the fault location stage, existing technologies mostly rely on data from a single monitoring node to determine faults, without combining the topology of the distribution area to conduct multi-node collaborative analysis; they can only roughly locate the faulty branch, and cannot accurately determine the specific section where the fault is located, which requires maintenance personnel to check the entire branch one by one, resulting in low fault finding efficiency, prolonging fault handling time, and increasing the loss of power distribution network outages.

[0003] In summary, current arc fault diagnosis technologies for low-voltage lines in distribution areas have significant shortcomings in terms of data acquisition synchronization, signal preprocessing accuracy, diagnostic model generalization, and fault location accuracy. These shortcomings make it difficult to meet the high-efficiency and precise requirements of distribution network safety operation and maintenance. There is an urgent need for a new arc fault diagnosis and analysis method that can integrate multi-node synchronous acquisition, multi-dimensional feature extraction, intelligent model diagnosis, and topology collaborative location. Summary of the Invention

[0004] The purpose of this invention is to provide an arc fault diagnosis and analysis method based on artificial intelligence algorithms to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: An arc fault diagnosis and analysis method based on artificial intelligence algorithms includes the following steps: S1. Synchronous Data Acquisition and Topology Excitation: Deploy intelligent monitoring terminals at all levels of nodes in the low-voltage lines of the distribution area; when the system performs periodic topology identification or receives a fault diagnosis command, the master station sends a control command to the target line node, triggering the injection of a characteristic current signal of a preset frequency into the node; each intelligent monitoring terminal synchronously collects the response current waveform data of the line within this time period; S2. Targeted signal preprocessing: Preprocessing the acquired response current waveform data, including: Feature band extraction: Using a digital bandpass filter, the first frequency band signal centered on a preset feature frequency and the second frequency band signal containing the potential high-frequency oscillation component of the electric arc are separated. Fault Feature Enhancement: Calculate the energy ratio, distortion rate, or correlation coefficient between the second frequency band signal and the first frequency band signal in the time or frequency domain to generate a fault-sensitive feature sequence; S3. Multi-source feature fusion and intelligent diagnosis: Construct a fusion feature vector, which includes at least: statistical features of fault-sensitive feature sequences, time-domain distortion features of response current waveform data in the power frequency cycle, and topological identification information of injection nodes; input the fusion feature vector into a pre-trained arc fault diagnosis model to obtain diagnosis results, including fault confidence and preliminary fault type; S4. Precise Fault Location and Verification Based on Topology: The tested branch is determined based on the topology identifier of the injected node and its position in the transformer area topology map. When the fault confidence exceeds the threshold, the tested branch is judged to have a suspected arc fault. The monitoring terminals of the upstream and downstream adjacent nodes of the branch are coordinated for monitoring and diagnosis by the main station. Based on the spatiotemporal relationship of the multi-node diagnosis results, the fault point is finally confirmed and a diagnosis report containing the fault location, type and confidence level is reported.

[0006] As a preferred approach, the synchronous data acquisition and topology excitation in step S1 specifically includes the following steps: A topology map of the low-voltage lines in the transformer area is pre-established, and the topology map records the connection relationships and electrical parameters of nodes at all levels in the lines; At each node marked on the topology map, deploy intelligent monitoring terminals equipped with synchronous clocks, high-frequency sampling, and communication capabilities. When the system performs periodic topology identification or receives a fault diagnosis command, the master station selects the target line node based on the topology map and sends an injection control command to the intelligent monitoring terminal of that node. The intelligent monitoring terminal of the target node responds to the injection control command and injects a constant current characteristic signal with a preset characteristic frequency into the line it is on. During the injection of constant current characteristic signal, all intelligent monitoring terminals in the distribution area synchronously collect high-frequency waveform data of response current at their respective nodes based on a unified clock reference, and upload the collected high-frequency waveform data along with the corresponding node identifier and timestamp to the main station.

[0007] As a preferred approach, feature frequency band extraction includes: Configure a first digital bandpass filter, take the preset characteristic frequency as the center frequency, and set the first bandwidth according to the line impedance characteristics to filter the response current waveform data to obtain a first frequency band signal centered on the preset characteristic frequency. Configure a second digital bandpass filter, take the frequency band that is higher than the power frequency and covers the typical oscillation frequency of the electric arc as the passband, and set the second bandwidth according to the spectral characteristics of the electric arc noise, and filter the response current waveform data to obtain a second frequency band signal containing the potential high-frequency oscillation component of the electric arc. The first and second frequency band signals are time-domain aligned and amplitude-calibrated respectively to ensure that the two signals are synchronized on the time axis and have a unified amplitude reference, thus providing standardized input for subsequent feature calculations.

[0008] As a preferred embodiment, fault characteristic enhancement includes: Based on the first and second frequency band signals after time-domain alignment and amplitude calibration, the ratio of the energy of the second frequency band signal to the energy of the first frequency band signal is calculated in the time domain to obtain the energy ratio feature sequence. Alternatively, frequency domain transformation can be performed on the first frequency band signal and the second frequency band signal respectively to extract their spectral profiles in the preset frequency band, and the degree of distortion of the spectral profile of the second frequency band signal relative to the spectral profile of the first frequency band signal can be calculated to obtain the distortion rate feature sequence. Alternatively, the correlation coefficient between the first frequency band signal and the second frequency band signal can be calculated in the time domain to obtain the correlation feature sequence; The energy ratio feature sequence, distortion rate feature sequence, or correlation feature sequence are processed by time window moving average to generate a smoothed fault-sensitive feature sequence.

[0009] As a preferred solution, multi-source feature fusion and intelligent diagnosis include: Extract statistical features from the fault-sensitive feature sequence. The statistical features include at least the maximum value, minimum value, mean, standard deviation, and skewness. Calculate the time-domain distortion features of the response current waveform data within a complete power frequency cycle. The time-domain distortion features include at least the harmonic distortion rate, crest factor, and number of waveform abrupt change points. Concatenate and normalize the statistical features, time-domain distortion features, and topological identification information of the injected nodes to construct a unified dimension fused feature vector. The fused feature vector is input into a pre-trained arc fault diagnosis model, which is a hybrid model composed of gradient boosting decision tree and deep neural network. The hybrid model outputs the fault confidence and preliminary fault type of the tested branch. The preliminary fault type includes series arc fault, parallel arc fault or load switching transient. The arc fault diagnosis model calculates the fault confidence value representing the possibility of the fault based on the mapping relationship of the fused feature vectors, and outputs the corresponding preliminary fault type label, which together constitute the diagnosis result.

[0010] As a preferred approach, step S4, fault location and verification based on topological relationships, includes the following steps: Branch initial identification and suspected fault judgment: The master station queries the pre-established transformer area topology map based on the topology identifier of the injected node to determine the electrical branch to which the node belongs as the branch under test; when the fault confidence output by the arc fault diagnosis model exceeds the preset first threshold, the master station determines that there is a suspected arc fault in the branch under test and generates a suspected fault alarm containing the identifier of the branch under test and the preliminary fault type. Collaborative monitoring command issuance and data acquisition: Based on the suspected fault alarm and the transformer area topology map, the main station dispatches the intelligent monitoring terminals of the upstream and downstream adjacent nodes along the tested branch to issue collaborative monitoring commands to these adjacent nodes; upon receiving the commands, the intelligent monitoring terminals of the adjacent nodes synchronously collect the current waveform data of their respective nodes and upload them to the main station in the next power frequency cycle or multiple consecutive cycles. Multi-node feature comparison and propagation path analysis: The master station performs steps S2 and S3 on the current waveform data of the upstream and downstream adjacent nodes to obtain the fault-sensitive feature sequence, fused feature vector and preliminary diagnosis results corresponding to each adjacent node; The master station compares the amplitude attenuation characteristics of the fault-sensitive feature sequence of the upstream node and the downstream node and / or the trend of fault confidence change in the preliminary diagnosis results, and analyzes the source path of the arc fault characteristics in combination with the propagation direction of the signal on the tested branch; Fault point confirmation under topology constraints: The master station compares the results of multi-node feature comparison and propagation path analysis with the physical connection relationship and electrical distance of the tested branch in the transformer area topology map. If the analysis results show that the fault features originate from a certain segment between the upstream and downstream nodes of the tested branch, and the segment conforms to the topology connection logic, then the master station determines the segment as the precise fault point. If the diagnostic results of both upstream and downstream nodes show high-confidence faults, then the master station confirms the fault point at the common upstream node or the nearest line connection point of the tested branch. Comprehensive diagnostic report generation: The main station integrates the results of suspected fault determination, multi-node feature comparison and propagation path analysis, and fault point confirmation under topological constraints to generate the final diagnostic report; the diagnostic report includes at least the precise fault point location, the finally determined fault type, the comprehensive confidence level, and a summary description of the collaborative verification process.

[0011] As can be seen from the technical solution provided by the present invention above, the arc fault diagnosis and analysis method based on artificial intelligence algorithm provided by the present invention has the following beneficial effects: This invention deploys intelligent monitoring terminals with synchronous clock and high-frequency sampling functions at each level of low-voltage line nodes in the distribution area. It achieves synchronous data acquisition from multiple nodes based on a unified clock reference, and constructs a diagnostic benchmark by injecting characteristic current signals at a preset frequency. This design effectively avoids the data distortion problems caused by clock deviation and lack of signal reference in traditional acquisition methods. The acquired data is also accurately associated with node topology identifiers and timestamps to form a standardized raw dataset, which lays a solid data foundation for subsequent signal processing and fault diagnosis, and significantly reduces diagnostic errors caused by data quality defects. Step S2 uses a dual digital bandpass filter to selectively separate two key frequency band signals. The first frequency band signal serves as a reference, while the second frequency band signal focuses on the high-frequency characteristics of the arc. Time-domain alignment and amplitude calibration are then performed to ensure signal consistency. Subsequently, the fault characteristics are amplified by calculating the energy ratio, distortion rate, and correlation coefficient. Combined with time-window moving average to smooth random noise, irrelevant signals such as power frequency harmonics and line interference are successfully filtered out, making the high-frequency oscillation characteristics of the arc fault more prominent. Compared with traditional single signal processing methods, this invention can accurately capture weak fault signals even in complex interference environments, significantly improving the identifiability of fault characteristics. Step S3 integrates the statistical features of the fault-sensitive feature sequence, the time-domain distortion features of the current waveform, and the node topology identification information to construct a multi-dimensional fusion feature vector, avoiding the limitations of a single feature. A hybrid model consisting of gradient boosting decision trees and deep neural networks is adopted, which not only leverages the advantages of gradient boosting decision trees in extracting key features, but also utilizes the classification ability of deep neural networks for complex features. This not only accurately determines whether there is a fault in the line, but also precisely distinguishes between series arc faults, parallel arc faults, and load switching transients. This effectively reduces the situation in traditional diagnosis where normal transients are mistakenly identified as faults or minor faults are missed. The diagnostic accuracy and type differentiation ability far exceed those of a single algorithm model. Step S4 relies on the topology map of the distribution area and multi-node collaborative monitoring. It first identifies the branch under test by judging suspected faults, then schedules upstream and downstream nodes to collect data synchronously. Combining the amplitude attenuation characteristics and confidence change trends of fault sensitivity features, it analyzes the fault propagation path and finally determines the precise fault point through topology constraint matching. This can narrow the positioning range to a specific line section rather than just the branch. Compared with the traditional method that relies on manual inspection and has an ambiguous positioning range, this invention can significantly shorten the time for operation and maintenance personnel to find faults and reduce the downtime and operation and maintenance costs of the distribution network. This invention enables a fully automated process from data acquisition, signal preprocessing, and intelligent diagnosis to fault location and comprehensive report generation, all without human intervention. The main station can automatically complete data verification, model calculation, and result integration, generating a diagnostic report that includes fault location, type, comprehensive confidence level, and maintenance recommendations, forming a closed-loop automated process of "acquisition-analysis-diagnosis-location-reporting". This mode not only reduces the cost of manual operation and human judgment errors, but also responds to faults within minutes, promptly detects and handles arc faults, and avoids serious accidents such as line burnout and large-scale power outages caused by fault escalation, effectively ensuring the safe and stable operation of the low-voltage distribution network in the area. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the steps of an arc fault diagnosis and analysis method based on artificial intelligence algorithms according to the present invention. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0014] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific embodiments.

[0015] like Figure 1As shown, this embodiment of the invention provides an arc fault diagnosis and analysis method based on artificial intelligence algorithms, including the following steps: S1. Synchronous Data Acquisition and Topology Excitation: Deploy intelligent monitoring terminals at all levels of nodes in the low-voltage lines of the distribution area; when the system performs periodic topology identification or receives a fault diagnosis command, the master station sends a control command to the target line node, triggering the injection of a characteristic current signal of a preset frequency into the node; each intelligent monitoring terminal synchronously collects the response current waveform data of the line within this time period; S2. Targeted signal preprocessing: Preprocessing the acquired response current waveform data, including: Feature band extraction: Using a digital bandpass filter, the first frequency band signal centered on a preset feature frequency and the second frequency band signal containing the potential high-frequency oscillation component of the electric arc are separated. Fault Feature Enhancement: Calculate the energy ratio, distortion rate, or correlation coefficient between the second frequency band signal and the first frequency band signal in the time or frequency domain to generate a fault-sensitive feature sequence; S3. Multi-source feature fusion and intelligent diagnosis: Construct a fusion feature vector, which includes at least: statistical features of fault-sensitive feature sequences, time-domain distortion features of response current waveform data in the power frequency cycle, and topological identification information of injection nodes; input the fusion feature vector into a pre-trained arc fault diagnosis model to obtain diagnosis results, including fault confidence and preliminary fault type; S4. Precise Fault Location and Verification Based on Topology: The tested branch is determined based on the topology identifier of the injected node and its position in the transformer area topology map. When the fault confidence exceeds the threshold, the tested branch is judged to have a suspected arc fault. The monitoring terminals of the upstream and downstream adjacent nodes of the branch are coordinated for monitoring and diagnosis by the main station. Based on the spatiotemporal relationship of the multi-node diagnosis results, the fault point is finally confirmed and a diagnosis report containing the fault location, type and confidence level is reported.

[0016] In this embodiment, step S1 serves to construct the low-voltage line topology of the transformer area, deploy intelligent monitoring terminals, inject characteristic current signals, and synchronously collect response data. This provides standardized, spatiotemporally aligned raw data support for subsequent signal preprocessing and fault diagnosis, laying the foundation for fault feature extraction. The detailed steps are as follows: Step S1-1: Construction and parameter entry of the transformer area topology map: Based on the design drawings, field survey records, and historical operation and maintenance data of the low-voltage lines in the transformer area, a complete topology map of the low-voltage lines in the transformer area is constructed. The core content of the topology map includes the physical connection relationship and key electrical parameters of the nodes at all levels in the line. The physical connection relationship needs to clearly define the node level of the main line and branch lines, and mark the unique identification code of each node. The key electrical parameters cover the conductor type, conductor cross-sectional area, physical distance between nodes, resistance per unit length of the line, reactance per unit length of the line, and capacitance to ground of each level of the line. After data entry is completed, the system automatically performs topology logic verification and parameter consistency check: it checks whether there are loop conflicts or missing circuits in the node connection relationship to ensure that the topology structure conforms to the actual line layout; it checks whether the electrical parameters are within the typical range of similar lines, such as whether the resistance per unit length of copper core conductors meets the standard value of the corresponding cross-sectional specification; if there are topology logic conflicts or abnormal parameters, the system triggers a manual review process, and the data is re-entered after correction to ensure the accuracy and completeness of the topology map. Step S1-2, Deployment and Function Configuration of Intelligent Monitoring Terminal: The physical deployment of intelligent monitoring terminals should be completed at each level of nodes marked on the topology map. Deployment locations include main line segment nodes, branch line access nodes, and important user terminal access nodes. The deployed intelligent monitoring terminals must have three core functions: high-frequency sampling of synchronous clock and remote communication. The hardware configuration must meet the following requirements: the synchronous clock module supports satellite synchronization or network clock synchronization, and the clock accuracy is controlled at the microsecond level; the high-frequency sampling module has a sampling frequency of not less than 50kHz and a sampling bit depth of not less than 16 bits to ensure that it can capture the details of high-frequency oscillating current signals; the communication module supports wireless or wired communication methods and has the ability to upload data and receive commands in real time. After the terminal deployment is completed, the functions are configured and calibrated: the clock synchronization command is remotely issued by the master station to align the clocks of all terminals with the unified clock reference of the master station, ensuring that the clock deviation is less than 1 microsecond; the sampling parameters of the terminal are configured, including sampling frequency, sampling duration and data storage format; the current sampling channel of the terminal is calibrated by inputting a standard current signal to correct the amplitude error of the sampling data and ensure that the sampling amplitude reference of different terminals is consistent. Step S1-3: Target node selection and injection control command issuance: When the system triggers the data acquisition process, the target line node is determined according to the triggering scenario: if it is a periodic topology identification scenario, the main station selects the target node from the topology map in sequence according to the preset node polling order to ensure that all key nodes are covered in each period; if it is a fault diagnosis command scenario, the main station selects the core node in the section associated with the fault alarm information from the topology map as the target node. The master station sends an injection control command to the intelligent monitoring terminal of the target node. The command includes the preset characteristic frequency injection signal amplitude, injection duration, and synchronous acquisition trigger timestamp. A verification mechanism is used during the transmission of the control command. After the master station sends the command, it waits for the terminal to receive the confirmation feedback. If no feedback is received within the preset time, the master station automatically resends the command. The number of resends shall not exceed 3. If no feedback is received, the terminal is marked as having a communication abnormality and an alarm is triggered. Step S1-4, characteristic current signal injection execution: After receiving the injection control command, the intelligent monitoring terminal of the target node parses the parameter information in the command, starts the internal signal generation module, and generates a constant current characteristic signal with a preset characteristic frequency. The signal must meet the following technical indicators: the constant current accuracy error does not exceed ±1% to ensure the signal amplitude is stable; the frequency stability error does not exceed ±0.1% to avoid frequency drift affecting subsequent frequency band extraction; the signal injection duration is consistent with the synchronous acquisition duration to ensure that the acquired data completely covers the response process of the injected signal. During the injection process, the terminal monitors the parameters of the output signal in real time. If the signal amplitude or frequency deviates from the preset range, the closed-loop adjustment mechanism is immediately activated to correct the signal deviation by adjusting the internal circuit parameters, thereby ensuring the stability and accuracy of the injected signal. Steps S1-5: Multi-node synchronous data collection and uploading: While the characteristic current signal is injected, the main station sends a synchronous acquisition command to all intelligent monitoring terminals in the distribution area. All terminals start acquiring current waveform data based on a unified clock reference at the preset trigger timestamp. The acquisition parameters strictly follow the pre-configured standards: the sampling frequency is kept above 50kHz, and the sampling duration is not shorter than the duration of the injected signal, to ensure complete capture of the dynamic change process of the response current. After acquisition, the terminal preprocesses the raw current waveform data: the data is formatted according to a preset format, and the corresponding node identifier code and acquisition timestamp are added. The timestamp is accurate to the microsecond level and is consistent with the main station clock reference. The data integrity is checked by verifying whether there are any transmission or storage errors through the check code. If errors are found, the data is re-acquired. After preprocessing, the terminal uploads the data to the main station through the communication module. The upload process adopts a block transmission method, where large data volumes are split into multiple data blocks and transmitted sequentially. Each data block is verified and confirmed after transmission. After receiving the data, the main station checks the integrity and timestamp consistency of the data uploaded by each terminal to ensure that the data collected by all nodes are synchronized and aligned on the timeline. Finally, all data is classified and stored to form a response current waveform database that corresponds one-to-one with the topology nodes.

[0017] In this embodiment, step S2 purifies and enhances the response current waveform data synchronously acquired in step S1. By separating key frequency band signals and calculating fault-sensitive features, the original current data is transformed into an effective feature sequence that can be used for intelligent diagnosis, providing high-quality data support for subsequent multi-source feature fusion and intelligent diagnosis. The detailed steps are as follows: Step S2-1, Feature frequency band extraction: The core of feature frequency band extraction is to separate two types of signals that are strongly correlated with fault diagnosis from the response current waveform data using a dual digital bandpass filter, while ensuring the spatiotemporal consistency of the two types of signals. This process includes the following sub-steps: Step S2-1-1: Configuration and filtering of the first digital bandpass filter: First, configure the first digital bandpass filter, whose center frequency strictly matches the preset characteristic frequency injected in step S1; the setting of the first bandwidth needs to be adjusted in combination with the impedance characteristics of the low-voltage line in the distribution area: if the line is a short-distance copper core line (low impedance), the bandwidth is set to 5% to 10% of the preset characteristic frequency to reduce the mixing of irrelevant frequency signals; if the line is a long-distance aluminum core line (high impedance), the bandwidth needs to be expanded to 10% to 15% of the preset characteristic frequency to ensure that the signal corresponding to the preset characteristic frequency can be completely preserved. The response current waveform data collected in step S1 is input into the configured first digital bandpass filter. The frequency components are filtered by the filtering algorithm to remove all signals below the first bandwidth lower limit and above the first bandwidth upper limit, and finally the first frequency band signal centered on the preset characteristic frequency is obtained. This signal can be used as a reference for subsequent characteristic calculations and directly reflects the normal response law of the line to the injected characteristic current. Step S2-1-2, Configuration and Filtering of the Second Digital Bandpass Filter: When configuring the second digital bandpass filter, the passband needs to be set to a frequency band higher than the power frequency and covering the typical oscillation frequency of the electric arc: the power frequency is usually 50Hz or 60Hz, while the typical oscillation frequency of the electric arc is mostly concentrated in the range of 1kHz to 10kHz. Therefore, the lower limit of the passband is set to 200Hz (to avoid interference from power frequency harmonics) and the upper limit is set to 15kHz (to ensure coverage of most high-frequency oscillation components of the electric arc). The setting of the second bandwidth needs to be optimized in combination with the spectral characteristics of arc noise: if historical operation and maintenance data show that the arc noise of the transformer area is mainly concentrated in the range of 2kHz to 5kHz, this range can be set as the core bandwidth with a bandwidth value of 1kHz; the passband area outside the core range (200Hz to 2kHz, 5kHz to 15kHz) is set as the non-core bandwidth with a bandwidth value of 2kHz. This variable bandwidth design balances the noise suppression effect and the requirement to retain fault characteristics. The response current waveform data is input into a second digital bandpass filter to filter out low-frequency components at or below the power frequency and high-frequency noise above 15kHz, ultimately obtaining a second frequency band signal containing potential high-frequency arc oscillation components; this signal is directly related to the high-frequency characteristics of arc faults and is a key signal source for subsequent fault identification. Step S2-1-3, Time Domain Alignment and Amplitude Calibration: Based on the unified clock reference during data acquisition in step S1, the timestamp information of the first frequency band signal and the second frequency band signal is extracted. All sampling points of the two types of signals are matched one-to-one according to the time coordinate to ensure that the amplitude of the corresponding first frequency band signal and the amplitude of the second frequency band signal can be obtained at any time, thus achieving time domain synchronization. If there is a small time offset between the two types of signals due to filter delay (the offset is usually less than 1 sampling period), the sampling point position is adjusted by linear interpolation to control the offset within 1 sampling period, ensuring that the time axis is completely aligned. Select a stable segment from the two signal types (usually the initial stage of injecting characteristic current in step S1, when the line does not exhibit abnormal fluctuations and the signal is free from fault interference). Calculate the average amplitude of the first and second frequency band signals within this stable segment. Using the average amplitude of the first frequency band signal within its stable segment as a reference, calculate an amplitude calibration coefficient (the calibration coefficient equals the reference amplitude divided by the average amplitude of the second frequency band signal within its stable segment). Multiply the amplitude of all sampling points of the second frequency band signal by this calibration coefficient to ensure that the amplitude references of the two signal types are consistent. After calibration, the amplitude ranges of the two signal types under the same intensity excitation are consistent, avoiding interference to subsequent feature calculations due to amplitude differences. Step S2-2, Fault Feature Enhancement: The core of fault feature enhancement is to amplify the arc fault features in the second frequency band signal through time-domain or frequency-domain calculations, generating a smooth and stable fault-sensitive feature sequence. This specifically includes the following sub-steps: Step S2-2-1, Calculation of fault-sensitive feature sequence: Based on the needs of actual diagnostic scenarios, one or more features are selected from three categories—energy ratio, distortion rate, and correlation coefficient—for calculation. The specific method is as follows: Sub-step S2-2-1-1: Calculation of energy ratio characteristic sequence: Within the time domain, the first frequency band signal and the second frequency band signal are divided into multiple calculation windows at fixed time intervals (usually set to 1ms, which can ensure time resolution and avoid excessive computation). For each calculation window, the energy of the first frequency band signal and the energy of the second frequency band signal within the window are counted respectively (signal energy is measured by the sum of the squares of the amplitudes of all sampling points within the window, and the number of sampling points is determined by the window duration and sampling frequency). Calculate the ratio of the energy of the second frequency band signal to the energy of the first frequency band signal within each window, and arrange the energy ratios of all calculated windows in chronological order to form an energy ratio characteristic sequence. When there is an arc fault in the line, the energy of the second frequency band signal will increase significantly due to the high-frequency oscillation of the arc, and the corresponding energy ratio will show a clear peak value, which can directly reflect the existence and intensity of the fault. Sub-step S2-2-1-2: Calculation of distortion rate feature sequence: Perform Fast Fourier Transform on the first frequency band signal and the second frequency band signal respectively to convert the time domain signal into the frequency domain signal, and obtain the spectrum of the first frequency band signal and the spectrum of the second frequency band signal (the spectrum can intuitively reflect the amplitude distribution of the signal at different frequencies). Select a preset frequency band (this frequency band is consistent with the passband of the second digital bandpass filter, i.e., 200Hz to 15kHz), and count the total amplitude of the first frequency band signal spectrum and the total amplitude of the second frequency band signal spectrum within this frequency band. At the same time, calculate the sum of the amplitude differences between the second frequency band signal spectrum and the first frequency band signal spectrum (the sum of the differences is the sum of the absolute values ​​of the amplitude differences between the two types of signal spectra corresponding to each frequency point within this frequency band). The calculation window is divided into 1ms intervals. For the preset frequency band data in each window, the distortion rate is calculated (the distortion rate is equal to the sum of the differences in spectral amplitude between the second frequency band signal and the first frequency band signal divided by the total spectral amplitude of the first frequency band signal). The distortion rates of all windows are arranged in chronological order to form a distortion rate characteristic sequence. When an arc fault occurs in the line, the spectrum of the second frequency band signal will deviate from the normal pattern (the spectrum of the first frequency band signal represents the normal pattern), resulting in a significant increase in the distortion rate, which can effectively distinguish between normal signals and fault signals. Sub-step S2-2-1-3: Calculation of correlation coefficient feature sequence: Within the time domain, the calculation window is divided into 1ms intervals. For the first frequency band signal and the second frequency band signal in each window, the mean of the signal is calculated (the mean is the arithmetic mean of the amplitude of all sampling points in the window). The deviations between the amplitude and mean of the first frequency band signal and the amplitude and mean of the second frequency band signal corresponding to each sampling point within the statistical window are used to calculate the correlation between the two types of signals (the higher the correlation, the more consistent the changing trends of the two types of signals; the lower the correlation, the greater the difference in the changing trends of the two types of signals). The correlation results of each window are arranged in chronological order to form a correlation coefficient feature sequence. In normal lines, the first frequency band signal and the second frequency band signal have a high correlation (consistent trend). When an arc fault occurs in the line, the second frequency band signal will introduce random high-frequency components, which will increase the difference in the trend of the two types of signals and significantly reduce the correlation, which can be used as an important basis for fault judgment. Step S2-2-2: Time window moving average processing: The sliding window length is determined based on the power frequency cycle of the low-voltage lines in the transformer area: if the power frequency is 50Hz (corresponding to a cycle of 20ms), the sliding window length is set to 20 calculation windows (each window is 1ms, for a total duration of 20ms); if the power frequency is 60Hz (corresponding to a cycle of approximately 16.7ms), the sliding window length is set to 17 calculation windows (for a total duration of 17ms), ensuring that the sliding window can completely cover one power frequency cycle and smooth out the interference of power frequency fluctuations on the characteristic sequence; The energy ratio feature sequence, distortion rate feature sequence, or correlation feature sequence generated in step S2-2-1 are smoothed point by point using a sliding window: taking the current feature point as the center, the arithmetic mean of all feature points within the sliding window is taken as the smoothed feature value of the current feature point; all feature points are processed sequentially in time order to finally generate a smoothed fault-sensitive feature sequence; this sequence can effectively filter out random noise interference, highlight the changing trend of fault features, and provide stable and reliable feature input for multi-source feature fusion in step S3.

[0018] In this embodiment, step S3 integrates the fault-sensitive feature sequence and line-related features generated in step S2 to construct a standardized fusion feature vector. This vector is then used to complete fault identification through a pre-trained intelligent model, outputting the fault confidence level and preliminary type, providing core diagnostic basis for fault location and verification in step S4. The detailed steps are as follows: Step S3-1, Multi-source feature extraction: The core of multi-source feature extraction is to mine feature information related to arc faults from different dimensions, covering the statistical features of fault-sensitive feature sequences and the time-domain distortion features of response current waveforms, ensuring the comprehensiveness of features and their correlation with the fault. Specifically, it includes the following sub-steps: Step S3-1-1: Statistical feature extraction of fault-sensitive feature sequences: For the fault-sensitive feature sequence generated in step S2, statistical features that reflect the overall pattern of the sequence are extracted, including five core indicators. The calculation method and meaning of each indicator are as follows: Maximum value: Traverse all feature values ​​in the fault-sensitive feature sequence and filter out the feature point with the largest value. This value can reflect the strongest performance of the fault feature. If there is a peak in the sequence (such as the peak energy ratio during a sudden arc), the maximum value can directly capture the anomaly. Minimum value: Traverse all feature values ​​in the fault-sensitive feature sequence and filter out the feature point with the smallest value. This value can be used as a benchmark for normal fluctuations in the sequence. By comparing the difference between the maximum and minimum values, it is possible to preliminarily determine whether the fluctuation amplitude of the sequence is abnormal. Mean: Calculate the arithmetic mean of all feature values ​​in the fault-sensitive feature sequence. Add all feature values ​​in the sequence and divide by the total number of feature values. The mean can reflect the overall level of the sequence. If the mean is significantly higher than the historical normal data, it indicates that the line may have persistent fault characteristics. Standard deviation: First, calculate the difference between each characteristic value and the mean. Then, square all the differences and take the average. Finally, take the square root of the average. The standard deviation reflects the degree of dispersion of the characteristic values ​​around the mean. The greater the dispersion, the worse the stability of the line current signal, and there may be intermittent faults. Skewness: First, calculate the cube of the difference between each feature value and the mean. Then, sum all the cubed results and divide by the total number of feature values ​​to obtain the third central moment. Finally, divide the third central moment by the cube of the standard deviation. Skewness can measure the degree of asymmetry in the distribution of feature sequences. If the skewness is positive and large, it indicates that there are many abnormal feature points in the sequence that are higher than the mean, which may correspond to a sudden failure scenario. The calculation results of the above five types of statistical features are organized into a statistical feature set. Each feature value is retained to four decimal places to ensure that the data accuracy meets the requirements of subsequent model input. Step S3-1-2: Extraction of time-domain distortion features of the response current waveform: Based on the response current waveform data collected in step S1, focusing on a complete power frequency cycle (20ms when the power frequency is 50Hz, and approximately 16.7ms when the power frequency is 60Hz), time-domain distortion features that reflect the deviation of the current waveform from a normal sine wave are extracted. These features include three core indicators, and the calculation methods and meanings of each indicator are as follows: Harmonic distortion rate: First, decompose the response current waveform into the fundamental component and each harmonic component (such as the 3rd, 5th, and 7th harmonics). Calculate the effective value of the fundamental component and the sum of the squares of the effective values ​​of all harmonic components. Then, take the square root of the sum of the squares of the effective values ​​of the harmonic components and divide it by the effective value of the fundamental component. The higher the harmonic distortion rate, the more severe the nonlinear interference on the current waveform. Arc faults will cause a significant increase in harmonic components. This indicator can be directly related to the degree of fault. Crest factor: Calculate the peak value (maximum amplitude) and effective value of the response current waveform within one power frequency cycle, and divide the peak value by the effective value to obtain the crest factor. The crest factor of a normal sine wave is about 1.414. If the crest factor deviates significantly from this value (e.g., greater than 2.0), it indicates that the waveform has sharp peaks, which may be caused by the instantaneous large current generated by arc discharge. Number of abrupt waveform change points: Set an amplitude change threshold (usually 5% of the average current amplitude within the power frequency cycle), traverse all adjacent sampling points within the cycle, calculate the amplitude difference between adjacent sampling points, and if the absolute value of the difference exceeds the set threshold, then the position is determined to be a waveform abrupt change point. Count the total number of abrupt change points within the cycle. Arc faults can cause the current amplitude to change drastically in a short period of time, and the number of abrupt change points will be significantly more than that of normal lines (normal lines usually have fewer than 3 abrupt change points, while faulty lines may have more than 10). The calculation results of the above three types of time-domain distortion features are organized into a time-domain distortion feature set, which is consistent with the precision of the statistical feature set and retains four decimal places. Step S3-2, Constructing the fused feature vector: The core of feature vector fusion construction is to integrate multi-source features and topological information into a vector of unified dimension, and to eliminate differences in magnitude through normalization to ensure that the features have a fair impact on the model. The specific operation is as follows: Feature splicing: First, obtain the topology identification information of the injected nodes in step S1 (this information is a unique code, such as "Area A-Branch 1-Node 3", which needs to be converted into numerical form, such as mapping to the integer 103 through encoding). Then, splice the statistical feature set generated in step S3-1-1, the temporal distortion feature set generated in step S3-1-2, and the topology identification value in a fixed order to form the original feature combination; for example, the splicing order is: maximum value, minimum value, mean, standard deviation, skewness, harmonic distortion rate, peak factor, number of waveform steep change points, and topology identification value, for a total of 9 feature dimensions. Feature normalization: The Min-Max normalization method is used to process the original feature combination, scaling the value of each feature to the range of 0 to 1. During processing, the minimum and maximum values ​​of each feature in the historical training dataset are first calculated. Then, the current value of the feature is subtracted from the historical minimum value, and the difference is divided by the difference between the historical maximum and the historical minimum value to complete the normalization. The purpose of normalization is to eliminate the influence of differences in the magnitude of different features. For example, the topological identifier value may be a three-digit number, while the harmonic distortion rate may be a decimal. After normalization, all features are on the same magnitude, ensuring that the model can learn the importance of each feature fairly. Vector format conversion: Convert the normalized feature combination into a vector format that the model can recognize, such as a one-dimensional vector arranged in a row and nine columns according to the concatenation order, with each element corresponding to a normalized feature value, and finally forming a fused feature vector of a unified dimension. Step S3-3, Intelligent Diagnosis: The core of intelligent diagnosis is to input the fused feature vector into a pre-trained arc fault diagnosis model, and then, through the model's feature analysis and classification capabilities, output the fault confidence and preliminary fault type. The specific operation is as follows: Model Loading and Feature Input: The pre-trained arc fault diagnosis model is launched. This model is a hybrid model based on gradient boosting decision trees and deep neural networks. The gradient boosting decision tree part is composed of multiple decision trees and is responsible for extracting key distinguishing features (such as harmonic distortion rate, maximum energy ratio, and other features strongly correlated with the fault) from the fused feature vector. The deep neural network part consists of an input layer, hidden layers (usually 3 to 5 layers), and an output layer, and is responsible for fine-classifying the key features extracted by the gradient boosting decision tree. The fused feature vector generated in step S3-2 is input into the input layer of the model to trigger the model calculation process. Internal feature processing: The fused feature vectors first enter the gradient boosting decision tree part. Each decision tree splits the vector according to the feature threshold. Through voting and weighting of multiple trees, the 3 to 5 key features that contribute the most to fault judgment are selected. The key features are then fed into the hidden layer of the deep neural network. The features are non-linearly transformed by the activation function (such as the ReLU function) to gradually learn the mapping relationship between features and fault types. Finally, the feature vector output by the hidden layer is fed into the output layer. The output layer transforms the result into a probability distribution through the Softmax function. Diagnostic results output: The model outputs two core types of results: Fault confidence: The probability value corresponding to the "fault" category in the output layer probability distribution. The value ranges from 0 to 1. The closer the value is to 1, the higher the probability that the tested branch has an arc fault. The closer the value is to 0, the higher the probability that the branch is normal. This value is calculated by the model to determine the degree of correlation between features and faults, and reflects the reliability of the judgment. Preliminary fault type: Determined based on the category with the highest probability in the output layer probability distribution, it is divided into three categories: series arc fault, parallel arc fault, and load switching transient. Among them, series arc fault corresponds to the arc generated at the moment of line breakage, parallel arc fault corresponds to the arc between phase lines or between a phase line and the neutral line, and load switching transient is a non-fault scenario (such as instantaneous current fluctuations when a motor starts or a switch is closed). The model determines the final type by the degree of matching between features and historical data of each category. The fault confidence level is combined with the preliminary fault type to form a complete diagnostic result, providing a basis for judgment in the precise fault location and verification in step S4.

[0019] In this embodiment, step S4 combines the diagnostic results output from step S3 with the transformer area topology map information. Through multi-node collaborative monitoring, feature comparison, and topology constraint analysis, it further locates the suspected faulty branch to a specific segment, verifies the accuracy of the diagnostic results, and finally generates a diagnostic report containing complete fault information, providing maintenance personnel with accurate fault handling basis. The detailed steps are as follows: Step S4-1, Initial Branch Determination and Suspected Fault Judgment: The master station first obtains the topology identifier of the target node injected with the characteristic current signal in step S1. This topology identifier is a unique code of the node in the transformer area topology map, which corresponds one-to-one with the node information of each level recorded in the map. The master station queries the pre-established transformer area topology map through the topology identifier, determines the electrical branch to which the target node belongs from the line connection relationship recorded in the map, and defines the electrical branch as the branch under test. The main station then retrieves the fault confidence score of the tested branch output in step S3 and compares it with a preset first threshold. The first threshold is set based on historical fault data and maintenance experience of the transformer area, and is usually between 0.7 and 0.8 (the closer the value is to 1, the stricter the standard for judging suspected faults, which can reduce false alarms; the closer the value is to 0.7, the lower the risk of missed alarms. The specific value needs to be adjusted according to the actual fault occurrence rate of the transformer area). If the fault confidence score is greater than the first threshold, the main station determines that the tested branch has a suspected arc fault; if the fault confidence score is less than or equal to the first threshold, the tested branch is determined to be fault-free, the subsequent location process is terminated, and only the diagnosis result is recorded. When a suspected fault is detected, the main station automatically generates a suspected fault alarm. The alarm content includes the unique identifier of the tested branch (consistent with the topology map code), the preliminary fault type output in step S3 (series arc fault, parallel arc fault or load switching transient), and the current fault confidence value. At the same time, the alarm generation time is marked (accurate to the second) to provide a time reference for subsequent collaborative monitoring. Step S4-2: Issuance of collaborative monitoring commands and acquisition of data: Based on the identifier of the tested branch in the suspected fault alarm, the main station queries the transformer area topology map again and filters the upstream and downstream adjacent nodes along the tested branch from the node connection relationships recorded on the map. The upstream adjacent node refers to the nearest node to the power supply side of the injected node in the tested branch, and the downstream adjacent node refers to the nearest node to the load side of the injected node in the tested branch (if the tested branch has branches, the upstream and downstream adjacent nodes of all branches need to be filtered to ensure coverage of the entire branch range). Based on the screening results, the main station sends collaborative monitoring instructions to the intelligent monitoring terminals of these upstream and downstream adjacent nodes. The instructions include the monitoring start time, data acquisition duration, and sampling parameters: the monitoring start time is set to the start time of the next power frequency cycle (to ensure synchronization with the line power frequency and avoid data deviation caused by cycle misalignment); the data acquisition duration is set to 2 to 3 consecutive power frequency cycles (to capture the complete current fluctuation pattern while avoiding processing delays caused by excessive data volume); the sampling parameters are consistent with those in step S1 to ensure uniform format and accuracy of the collected data. Upon receiving the collaborative monitoring instruction, the intelligent monitoring terminal initiates current waveform data acquisition at the specified monitoring start time, based on the unified clock reference calibrated in step S1. During the acquisition process, the terminal stores the data in real time and adds node identifiers and timestamps (timestamps are accurate to microseconds and are fully synchronized with the master station clock). After acquisition, the terminal uploads the data to the master station through the communication module. After receiving the data, the master station first verifies the integrity of the data (confirming whether the number of sampling points meets the acquisition duration requirements) and the consistency of the timestamps (confirming whether the time axes of all node data are aligned). If there are missing data or time deviations, the master station issues a supplementary acquisition instruction to the corresponding terminal until complete and synchronized multi-node data is obtained. Step S4-3 Multi-node feature comparison and propagation path analysis The main station processes the current waveform data of the upstream and downstream adjacent nodes in sequence using steps S2 and S3: First, it separates the first frequency band signal and the second frequency band signal using a digital bandpass filter, and completes time-domain alignment, amplitude calibration, and fault-sensitive feature sequence generation; then, it extracts the statistical features of the fault-sensitive feature sequence and the time-domain distortion features of the current waveform, constructs a fused feature vector, and inputs it into the arc fault diagnosis model to obtain the fault-sensitive feature sequence, fused feature vector, and preliminary diagnosis results (including the fault confidence and preliminary fault type of the branch where each node is located) for each adjacent node. The main station then compares the amplitude attenuation characteristics of the fault-sensitive feature sequences of upstream and downstream nodes. Since current signals experience energy loss due to line impedance during propagation in low-voltage lines, under normal circumstances, the amplitude of the fault-sensitive feature sequence collected by the upstream node will be greater than that of the downstream node (if the fault feature originates from the upstream direction, the upstream amplitude is higher; if it originates from the downstream direction, the downstream amplitude is higher). The main station calculates the amplitude attenuation rate (the attenuation rate is the ratio of the downstream amplitude to the upstream amplitude; the smaller the ratio, the more significant the attenuation) by comparing the amplitude differences between the upstream and downstream sequences point by point. Simultaneously, it observes the trend of the fault confidence level of each node output in step S3: if the fault confidence level of the upstream node is higher than that of the downstream node, and the amplitude attenuation rate is less than 0.8 (indicating significant signal loss from upstream to downstream), the fault feature originates from the upstream direction; if the fault confidence level of the downstream node is higher than that of the upstream node, and the amplitude attenuation rate is greater than 1.2 (indicating less signal loss from downstream to upstream, and the fault is closer to the downstream), the fault feature originates from the downstream direction. By combining the signal propagation direction of the transformer area lines (default from the power supply side to the load side, i.e., from upstream to downstream), the main station comprehensively considers the amplitude attenuation characteristics and the confidence level change trend to determine the specific source path of the arc fault characteristics, thereby narrowing down the scope for subsequent fault location. Step S4-4: Fault point confirmation under topology constraints: The main station performs constraint matching between the fault feature source path analysis results obtained in step S4-3 and the physical connection relationship and electrical distance of the tested branch in the transformer area topology map. First, it extracts the physical connection details between the upstream and downstream adjacent nodes of the tested branch from the topology map, including the direction, length, conductor type (affecting the signal attenuation) of the line between the nodes and whether there are branch nodes. At the same time, it extracts the electrical distance between the two nodes (converted to the actual line length, in meters) as a reference for judging the location of the fault point. If the analysis results show that the fault characteristics originate from a certain segment between upstream and downstream adjacent nodes, and the physical connection relationship of this segment conforms to the signal propagation logic (such as no open circuits or abnormal branches within the segment, and matching the amplitude attenuation law), then the master station determines this segment as the precise fault point. For example, if the fault confidence of the upstream node is 0.92, the fault confidence of the downstream node is 0.65, the amplitude attenuation rate is 0.71, and the line length between the two nodes is 200 meters, and the signal attenuation law conforms to the normal loss of this conductor type, then the precise fault point is located within a 100-meter segment between the two nodes closer to the upstream node (because the upstream confidence is higher, the fault is closer to the signal source). If the preliminary diagnostic results of adjacent upstream and downstream nodes both show that the fault confidence exceeds the first threshold (i.e., both are high-confidence faults), it indicates that the fault characteristics do not originate from the segment between the two nodes, but from a more upstream common node. The main station queries the topology map for the common upstream node of the two nodes (i.e., the superior node that provides power to both nodes), or the nearest line connection point between the two nodes (such as a branch junction point), and determines the common upstream node or the nearest line connection point as the precise fault point. This situation usually corresponds to the fault occurring on the trunk line, causing the fault characteristics to be detected in both upstream and downstream branch nodes. After confirming the exact fault location, the main station re-verifies the topology logic of the fault location: checks whether the electrical parameters (such as resistance and reactance) of the section where the fault location is located match the signal attenuation characteristics, confirms that there are no special structures with historical maintenance markings in the section (such as transformers and capacitors, to avoid misjudging the normal operation signal of the equipment as a fault), and ensures that there is no topology conflict in the fault location. Step S4-5: Generation of comprehensive diagnostic report: The main site integrates all results from steps S4-1 to S4-4 to generate a final comprehensive diagnostic report; the report must include five core components: Precise fault location: Described by combining node codes and physical addresses in the transformer area topology map, such as the 100-meter section between upstream node 3 and downstream node 4 in transformer area A-main line 1, while also marking the latitude and longitude of the fault point (if the intelligent monitoring terminal has positioning function), to ensure that maintenance personnel can quickly find the on-site location. The final fault type is determined as follows: Based on the preliminary type in step S3 and the multi-node verification results in step S4, the fault type is corrected and confirmed. For example, if step S3 initially determines it to be a load switching transient, but both upstream and downstream nodes detect continuous high-frequency oscillation characteristics, it is corrected to a series arc fault. If the preliminary type is consistent with the multi-node results, it is directly used. Overall confidence level: Calculate the weighted average of the fault confidence level of the tested branch and the fault confidence levels of the upstream and downstream nodes in step S3 (the weight of the tested branch is 0.5, and the weight of each upstream and downstream node is 0.25). The weighted average result is the overall confidence level. This value can more comprehensively reflect the reliability of fault judgment. It usually needs to be greater than 0.85 to be used as a basis for operation and maintenance. Summary of Collaborative Verification Process: Briefly describe the number of nodes in collaborative monitoring, data acquisition duration, key conclusions of feature comparison (such as amplitude attenuation rate and confidence change trend), explain the core basis for fault location, and facilitate the rationality of subsequent traceability and diagnosis process; Maintenance recommendations: Provide targeted recommendations based on the type and location of the fault. For example, for series arc faults, check whether the line joints are loose, and for parallel arc faults, check whether the line insulation is damaged. Provide maintenance personnel with specific handling directions. After the report is generated, the main station pushes it to the operation and maintenance management platform through the communication system, and at the same time triggers alarm notifications (such as SMS, platform pop-ups) to ensure that operation and maintenance personnel receive fault information in a timely manner and start subsequent fault handling procedures.

[0020] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for diagnosing and analyzing electric arc faults based on artificial intelligence algorithms, characterized in that: Includes the following steps: S1. Synchronous data acquisition and topology excitation: Deploy intelligent monitoring terminals at all levels of nodes of the low-voltage lines in the transformer area; When the system performs periodic topology identification or receives a fault diagnosis command, the master station sends a control command to the target line node, triggering the node to inject a characteristic current signal of a preset frequency. Each intelligent monitoring terminal synchronously collects the response current waveform data of the line during this time period; S2. Targeted signal preprocessing: Preprocessing the acquired response current waveform data, including: Feature band extraction: Using a digital bandpass filter, the first frequency band signal centered on a preset feature frequency and the second frequency band signal containing the potential high-frequency oscillation component of the electric arc are separated. Fault Feature Enhancement: Calculate the energy ratio, distortion rate, or correlation coefficient between the second frequency band signal and the first frequency band signal in the time or frequency domain to generate a fault-sensitive feature sequence; S3. Multi-source feature fusion and intelligent diagnosis: Construct a fusion feature vector, which includes at least: statistical features of fault-sensitive feature sequences, time-domain distortion features of response current waveform data in the power frequency cycle, and topological identification information of injection nodes; input the fusion feature vector into a pre-trained arc fault diagnosis model to obtain diagnosis results, including fault confidence and preliminary fault type; S4. Precise Fault Location and Verification Based on Topology: The tested branch is determined based on the topology identifier of the injected node and its position in the transformer area topology map. When the fault confidence exceeds the threshold, the tested branch is judged to have a suspected arc fault. The monitoring terminals of the upstream and downstream adjacent nodes of the branch are coordinated for monitoring and diagnosis by the main station. Based on the spatiotemporal relationship of the multi-node diagnosis results, the fault point is finally confirmed and a diagnosis report containing the fault location, type and confidence level is reported.

2. The arc fault diagnosis and analysis method based on artificial intelligence algorithm according to claim 1, characterized in that: The synchronous data acquisition and topology excitation in step S1 specifically include the following steps: A topology map of the low-voltage lines in the transformer area is pre-established, and the topology map records the connection relationships and electrical parameters of nodes at all levels in the lines; At each node marked on the topology map, deploy intelligent monitoring terminals equipped with synchronous clocks, high-frequency sampling, and communication capabilities. When the system performs periodic topology identification or receives a fault diagnosis command, the master station selects the target line node based on the topology map and sends an injection control command to the intelligent monitoring terminal of that node. The intelligent monitoring terminal of the target node responds to the injection control command and injects a constant current characteristic signal with a preset characteristic frequency into the line it is on. During the injection of constant current characteristic signal, all intelligent monitoring terminals in the distribution area synchronously collect high-frequency waveform data of response current at their respective nodes based on a unified clock reference, and upload the collected high-frequency waveform data along with the corresponding node identifier and timestamp to the main station.

3. The arc fault diagnosis and analysis method based on artificial intelligence algorithm according to claim 1, characterized in that: The feature frequency band extraction includes: Configure a first digital bandpass filter, take the preset characteristic frequency as the center frequency, and set the first bandwidth according to the line impedance characteristics to filter the response current waveform data to obtain a first frequency band signal centered on the preset characteristic frequency. Configure a second digital bandpass filter, take the frequency band that is higher than the power frequency and covers the typical oscillation frequency of the electric arc as the passband, and set the second bandwidth according to the spectral characteristics of the electric arc noise, and filter the response current waveform data to obtain a second frequency band signal containing the potential high-frequency oscillation component of the electric arc. The first and second frequency band signals are time-domain aligned and amplitude-calibrated respectively to ensure that the two signals are synchronized on the time axis and have a unified amplitude reference, thus providing standardized input for subsequent feature calculations.

4. The arc fault diagnosis and analysis method based on artificial intelligence algorithm according to claim 3, characterized in that: The enhanced fault characteristics include: Based on the first and second frequency band signals after time-domain alignment and amplitude calibration, the ratio of the energy of the second frequency band signal to the energy of the first frequency band signal is calculated in the time domain to obtain the energy ratio feature sequence. Alternatively, frequency domain transformation can be performed on the first frequency band signal and the second frequency band signal respectively to extract their spectral profiles in the preset frequency band, and the degree of distortion of the spectral profile of the second frequency band signal relative to the spectral profile of the first frequency band signal can be calculated to obtain the distortion rate feature sequence. Alternatively, the correlation coefficient between the first frequency band signal and the second frequency band signal can be calculated in the time domain to obtain the correlation feature sequence; The energy ratio feature sequence, distortion rate feature sequence, or correlation feature sequence are processed by time window moving average to generate a smoothed fault-sensitive feature sequence.

5. The arc fault diagnosis and analysis method based on artificial intelligence algorithm according to claim 1, characterized in that: The multi-source feature fusion and intelligent diagnosis include: Extract statistical features from the fault-sensitive feature sequence. The statistical features include at least the maximum value, minimum value, mean, standard deviation, and skewness. Calculate the time-domain distortion features of the response current waveform data within a complete power frequency cycle. The time-domain distortion features include at least the harmonic distortion rate, crest factor, and number of waveform abrupt change points. Concatenate and normalize the statistical features, time-domain distortion features, and topological identification information of the injected nodes to construct a unified dimension fused feature vector. The fused feature vector is input into a pre-trained arc fault diagnosis model, which is a hybrid model composed of gradient boosting decision tree and deep neural network. The hybrid model outputs the fault confidence and preliminary fault type of the tested branch. The preliminary fault type includes series arc fault, parallel arc fault or load switching transient. The arc fault diagnosis model calculates the fault confidence value representing the possibility of the fault based on the mapping relationship of the fused feature vectors, and outputs the corresponding preliminary fault type label, which together constitute the diagnosis result.

6. The arc fault diagnosis and analysis method based on artificial intelligence algorithm according to claim 1, characterized in that: Step S4, fault location and verification based on topological relationships, includes the following steps: Branch initial identification and suspected fault judgment: The master station queries the pre-established transformer area topology map based on the topology identifier of the injected node to determine the electrical branch to which the node belongs as the branch under test; when the fault confidence output by the arc fault diagnosis model exceeds the preset first threshold, the master station determines that there is a suspected arc fault in the branch under test and generates a suspected fault alarm containing the identifier of the branch under test and the preliminary fault type. Collaborative monitoring command issuance and data acquisition: Based on the suspected fault alarm and the transformer area topology map, the main station dispatches the intelligent monitoring terminals of the upstream and downstream adjacent nodes along the tested branch to issue collaborative monitoring commands to these adjacent nodes; upon receiving the commands, the intelligent monitoring terminals of the adjacent nodes synchronously collect the current waveform data of their respective nodes and upload them to the main station in the next power frequency cycle or multiple consecutive cycles. Multi-node feature comparison and propagation path analysis: The master station performs steps S2 and S3 on the current waveform data of the upstream and downstream adjacent nodes to obtain the fault-sensitive feature sequence, fused feature vector and preliminary diagnosis results corresponding to each adjacent node; The master station compares the amplitude attenuation characteristics of the fault-sensitive feature sequence of the upstream node and the downstream node and / or the trend of fault confidence change in the preliminary diagnosis results, and analyzes the source path of the arc fault characteristics in combination with the propagation direction of the signal on the tested branch; Fault point confirmation under topology constraints: The master station compares the results of multi-node feature comparison and propagation path analysis with the physical connection relationship and electrical distance of the tested branch in the transformer area topology map. If the analysis results show that the fault features originate from a certain segment between the upstream and downstream nodes of the tested branch, and the segment conforms to the topology connection logic, then the master station determines the segment as the precise fault point. If the diagnostic results of both upstream and downstream nodes show high-confidence faults, then the master station confirms the fault point at the common upstream node or the nearest line connection point of the tested branch. Comprehensive diagnostic report generation: The main station integrates the results of suspected fault determination, multi-node feature comparison and propagation path analysis, and fault point confirmation under topological constraints to generate the final diagnostic report; the diagnostic report includes at least the precise fault point location, the finally determined fault type, the comprehensive confidence level, and a summary description of the collaborative verification process.

Images