An edge-computing-based local fault self-healing method and system for pole-mounted circuit breakers

Abstract

The application discloses a pole-mounted circuit breaker local fault self-healing method and system based on edge computing, which comprises the following steps: collecting line three-phase current, voltage and circuit breaker state data in real time, and extracting fault characteristic quantities by using a sliding time window; generating a fault judgment result containing fault type, fault distance and fault phase based on the fault characteristic quantities; inputting the fault judgment result into a locally deployed adversarial verification network to output an effective fault result verified and confirmed; generating fault isolation and recovery control instructions containing tripping instructions, reclosing instructions or blocking instructions according to the effective fault result; executing the fault isolation and recovery control instructions to generate a self-healing success or a treatment report reported to a master station. By using the embodiment of the application, the fault treatment delay can be shortened, the dependence on the master station communication can be reduced, and the self-healing accuracy and reliability in the high-resistance fault and complex fault scenarios can be improved.

Description

Technical Field

[0001] This invention belongs to the field of pole-mounted circuit breaker technology, and in particular, it is a method and system for local fault self-healing of pole-mounted circuit breakers based on edge computing. Background Technology

[0002] With the increasing complexity of distributed power generation and distribution network structures, traditional pole-mounted circuit breakers, relying on a master station system for fault detection and isolation, have revealed significant problems such as high communication latency, heavy processing load on the master station, and susceptibility to network fluctuations. While existing local feeder automation solutions can achieve local protection, they typically rely on a single overcurrent threshold or simple timing logic, making it difficult to accurately identify high-resistance faults, transient faults, and permanent faults, easily leading to false tripping or protection failure. In recent years, although some research has attempted to introduce deep learning into distribution network fault diagnosis, the models are often too complex to be deployed on pole-mounted circuit breakers with limited computing power. Furthermore, existing methods generally lack a mechanism to verify the reliability of fault determination results, making it impossible to effectively verify fault type identification and location errors, and compromising the reliability of self-healing actions. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for local fault self-healing of pole-mounted circuit breakers based on edge computing, so as to overcome the shortcomings of the prior art, shorten the fault handling delay, reduce the dependence on master station communication, and enhance the accuracy and reliability of self-healing in high-resistance fault and complex fault scenarios.

[0004] One embodiment of this application provides a local fault self-healing method for pole-mounted circuit breakers based on edge computing, the method comprising: The three-phase current, voltage and circuit breaker status data of the line are collected in real time by deploying computing nodes on the side edge of the pole-mounted circuit breaker, and fault characteristic quantities are extracted by using a sliding time window. Based on the fault features, a lightweight convolutional neural network model is used to identify the fault type and locate the fault segment, generating a fault determination result that includes the fault type, fault distance and fault phase. The fault determination result is input into the locally deployed adversarial verification network. The credibility of the fault determination result is evaluated through the game between the generator and the discriminator, and the valid fault result after verification is output. Based on the valid fault results, the locally stored fault self-healing strategy library is invoked to match the action logic corresponding to the current fault type and generate fault isolation and recovery control instructions containing tripping instructions, reclosing instructions, or blocking instructions. Execute the fault isolation and recovery control command, drive the operating mechanism of the pole-mounted circuit breaker to complete the opening and closing operation, and collect the line status for self-check after the operation, and generate a self-healing success or report to the main station.

[0005] Another embodiment of this application provides a local fault self-healing system for pole-mounted circuit breakers based on edge computing, the system comprising: The data acquisition module is used to collect three-phase current, voltage and circuit breaker status data in real time through the edge computing nodes deployed on the side of the pole-mounted circuit breaker, and to extract fault feature quantities using a sliding time window. The identification module is used to identify the fault type and locate the fault segment based on the fault feature quantity using a lightweight convolutional neural network model, and generate a fault determination result including fault type, fault distance and fault phase. The evaluation module is used to input the fault determination result into the locally deployed adversarial verification network, evaluate the credibility of the fault determination result through the game between the generator and the discriminator, and output the valid fault result that has been verified. The matching module is used to call the locally stored fault self-healing strategy library according to the valid fault result, match the action logic corresponding to the current fault type, and generate fault isolation and recovery control instructions including tripping instructions, reclosing instructions, or blocking instructions. The self-healing module is used to execute the fault isolation and recovery control commands, drive the operating mechanism of the pole-mounted circuit breaker to complete the opening and closing operations, and collect the line status for self-checking after the operation, and generate a self-healing success or reporting to the main station.

[0006] Another embodiment of this application provides a storage medium storing a computer program, wherein the computer program is configured to execute the method described in any of the preceding claims when running.

[0007] Another embodiment of this application provides an electronic device including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the method described in any of the preceding claims.

[0008] Compared with existing technologies, the present invention provides a local fault self-healing method for pole-mounted circuit breakers based on edge computing, which can shorten the fault handling delay, reduce the dependence on master station communication, and enhance the self-healing accuracy and reliability in high-resistance fault and complex fault scenarios. Attached Figure Description

[0009] Figure 1 Hardware structure block diagram of a computer terminal for a local fault self-healing method for pole-mounted circuit breakers based on edge computing, provided in an embodiment of the present invention; Figure 2 A flowchart illustrating a local fault self-healing method for pole-mounted circuit breakers based on edge computing, provided in an embodiment of the present invention. Figure 3This is a schematic diagram of a local fault self-healing system for pole-mounted circuit breakers based on edge computing, provided in an embodiment of the present invention. Detailed Implementation

[0010] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0011] This invention first provides a method for local fault self-healing of pole-mounted circuit breakers based on edge computing. This method can be applied to electronic devices, such as computer terminals, specifically ordinary computers.

[0012] The following detailed explanation uses a computer terminal as an example. Figure 1 This is a hardware structure block diagram of a computer terminal for a local fault self-healing method for pole-mounted circuit breakers based on edge computing, provided in an embodiment of the present invention. Figure 1 As shown, the computer device includes a processor, memory, and network interface connected via a system bus, wherein the memory may include non-volatile storage media and internal memory.

[0013] See Figure 2 The present invention provides a method for local fault self-healing of pole-mounted circuit breakers based on edge computing, which may include the following steps: S201 collects real-time data on three-phase current, voltage, and circuit breaker status by deploying edge computing nodes on the side of the pole-mounted circuit breaker, and extracts fault characteristic quantities using a sliding time window. Specifically, edge computing nodes can be deployed on the side of the pole-mounted circuit breaker to collect data on the three-phase current, three-phase voltage, circuit breaker opening and closing position, and energy storage status in real time through the current transformer at a sampling rate of 256 points per cycle, generating the original high-frequency time-series data stream. The core of this step is to rely on locally deployed edge computing devices and high-precision sampling transformers to complete high-frequency synchronous acquisition of electrical quantities of power distribution lines and status quantities of circuit breaker equipment. This transforms the physical operating status into continuous digital time-series data, providing complete raw data support for subsequent fault feature mining. The specific implementation method is as follows: The edge computing node adopts an on-site deployment mode, directly installed and fixed on the integrated mounting base on the side of the pole-mounted circuit breaker equipment. It achieves close hardware interface with the circuit breaker body and line sampling transformer, abandoning the traditional centralized cloud data acquisition mode and completely avoiding the latency, packet loss and interference problems caused by long-distance data transmission, realizing millisecond-level on-site data acquisition and processing. This edge computing node has independent data acquisition, caching and preprocessing computing power, and can adapt to the complex operating environment of outdoor power distribution lines, such as high temperature, low temperature, humidity and electromagnetic interference, ensuring uninterrupted data acquisition around the clock.

[0014] Data acquisition for the power line relies on high-precision power sampling transformers. The transformers are precisely connected to the three-phase busbars of the distribution line and are specifically adapted to the operating conditions of power distribution lines with a power frequency of 50Hz. A fixed sampling rate of 256 points per cycle is set. The duration of a single complete operating cycle of a power frequency of 50Hz line is 20ms. The 256-point sampling per cycle means that the equipment will uniformly complete 256 data samplings within a single 20ms cycle. The time interval between each sampling is 0.078125ms. The ultra-high sampling density can completely capture the subtle waveform changes under all operating conditions, including normal operation, fault transients, and fault steady states, without missing the instantaneous distortion characteristics at the moment of a fault.

[0015] The collected data includes four core categories: three-phase line current, three-phase line voltage, circuit breaker opening / closing position data, and circuit breaker energy storage status data. Each type of data has a clear collection standard and physical meaning. Three-phase current specifically refers to the real-time load current and fault current of phases A, B, and C of the distribution line, reflecting the current-carrying operating status of the line, measured in amperes. Three-phase voltage specifically refers to the real-time operating voltage of phases A, B, and C relative to ground, reflecting the line's power supply voltage level, measured in kilovolts. Circuit breaker opening / closing position data is equipment status switch data, containing only two status indicators: closed and open, accurately reflecting the current operating status of the circuit breaker. Circuit breaker energy storage status data is also switch data, including two states: energy storage complete and energy storage incomplete, used to determine whether the circuit breaker operating mechanism has the conditions for opening / closing operations.

[0016] The data acquisition process is a continuous, real-time, synchronous acquisition. After the edge computing node starts, it continuously sends sampling commands to the instrument transformers. The instrument transformers capture various types of data in real time according to a fixed sampling frequency. Each set of sampled data is bound with a high-precision millisecond-level timestamp, marking the precise acquisition time corresponding to the data. All timestamped discrete sampled data are automatically sorted and continuously spliced ​​in chronological order to form an uninterrupted, time-aligned original high-frequency time-series data stream. The data stream completely records the real-time operating status of the line and circuit breaker. Under normal operating conditions, the data stream waveform is stable and regular. When a fault occurs, the data stream will show instantaneous amplitude changes and waveform distortion, etc., leaving the most original and authentic operating data for subsequent fault analysis.

[0017] The original high-frequency time-series data stream is denoised and normalized. Median filtering is used to remove spike pulse interference, and the current and voltage amplitudes are normalized to the range of 0 to 1 to generate preprocessed standard data frames. The core of this step is to eliminate invalid noise interference from the on-site data acquisition environment, unify the dimensions and numerical range of electrical data, solve the problems of excessive noise and inconsistent magnitudes in the raw data, and transform the coarse raw data stream into standardized, high-precision usable data. This eliminates data errors for subsequent feature extraction. The specific implementation method is as follows: The raw high-frequency time-series data stream originates from real-time outdoor field acquisition. The power distribution site is subject to various interference factors, including electromagnetic radiation, line vibrations from light winds, instantaneous induced current from equipment switches, and environmental clutter. These factors result in a large amount of irregular spike pulse interference data mixed into the raw data stream. This type of interference data is invalid and abnormal, lacking practical operational reference value. Directly using it for fault feature extraction would lead to misjudgments of fault features and deviations in feature parameters. Therefore, denoising preprocessing of the raw data stream is essential.

[0018] The median filtering used in this step is a nonlinear filtering algorithm adapted for denoising power time-series waveforms. Its core working principle involves selecting a fixed-length continuous sampling data window, sorting all sampling points within the window from smallest to largest, and extracting the value at the middle position of the sorted value to replace the original sampling data at the center of the window. The core advantage of this algorithm is its ability to accurately remove randomly generated instantaneous spike noise without smoothing out genuine waveform changes and amplitude distortions during fault transients, perfectly meeting the denoising requirements of power fault waveforms. For the 256 sampling points per cycle, this scheme sets the filtering window to include 5 continuous sampling points. This window size can fully cover common short-term pulse interference in the field, balancing denoising effectiveness with the accuracy of retaining effective features.

[0019] After the data denoising process is completed, the amplitude data of three-phase current and three-phase voltage need to be normalized and standardized. The core purpose of the standardization process is to convert electrical parameters such as current and voltage, which have dimensions and large differences in numerical magnitude, into dimensionless standardized values ​​in the range of 0 to 1, so as to eliminate the influence of the magnitude difference of different electrical parameters on the subsequent algorithm calculation and adapt to the input data requirements of lightweight convolutional neural network models.

[0020] Per-unit calculations use the rated operating parameters of the distribution line as a fixed benchmark. The rated phase voltage benchmark for a conventional distribution line is set to 220V, and the rated steady-state load current benchmark is set to 500A. The real-time collected, denoised voltage and current values ​​are compared with the corresponding rated benchmark values ​​to obtain per-unit values ​​between 0 and 1. Under normal and stable operation, the per-unit values ​​of current and voltage remain stable between 0.80 and 0.95. When a short circuit or grounding fault occurs, the voltage amplitude drops significantly, and the per-unit voltage value decreases to below 0.6, while the current amplitude rises sharply, and the per-unit current value increases to above 1.0. The numerical changes are clearly defined and highly identifiable.

[0021] After completing the dual preprocessing of denoising and per-unit conversion, all data are regularized and filtered to remove empty data, garbled data, and abnormal out-of-limit data in the data stream. The preprocessed per-unit values ​​of three-phase current, three-phase voltage, circuit breaker opening and closing position data, and circuit breaker energy storage status data are encapsulated and integrated in a fixed order and format to form a preprocessed standard data frame with time alignment, unified format, and standardized values. Each frame of data corresponds to the preprocessing result of a complete power frequency cycle, ensuring the accuracy of subsequent time window truncation and feature extraction.

[0022] Based on the preprocessed standard data frame, a fixed-length sliding time window is set, which covers the complete transient process from the first half of the fault occurrence to the last five cycles after the fault occurrence. Data segments are truncated by sliding half a cycle in steps to generate a sliding window data block sequence. The core of this step is to segment standardized data into time series segments using a customized sliding time window, fully covering the entire lifecycle of a power fault from occurrence and transient oscillations to steady-state continuation. Overlapping sliding segmentation ensures that no fault features are missed, generating a continuous and complete set of time series data blocks. This provides a segmented data foundation for multi-dimensional fault feature extraction. The specific implementation method is as follows: This solution defines window parameters based on the operating characteristics of 50Hz power distribution lines. The duration of a single complete operating cycle of a power frequency line is 20ms, and the duration of a single half-cycle is 10ms. Based on this, the length and sliding step of the sliding time window are configured. The coverage of the fixed-length sliding time window is the half-cycle before the fault occurs and the five cycles after the fault occurs, corresponding to a total window duration of 10ms plus 100ms, totaling 110ms. This duration can accurately cover the entire transient process of all common faults in power distribution lines, including the normal steady-state operation stage before the fault occurs, the instantaneous change stage of the fault, the transient oscillation stage of the fault, and the steady-state continuation stage of the fault. It avoids losing key features in the middle and late stages of the fault due to an excessively short window duration, and also avoids including a large amount of invalid and redundant normal operation data due to an excessively long window duration, accurately adapting to the waveform change patterns of power distribution faults.

[0023] The sliding time window uses preprocessed standard data frames arranged in temporal order as its data source. It starts filling the initial window with data from the first data frame in the time series, constructing the first complete window data block. The window sliding step size is fixed at half a cycle of 10ms. That is, after each data block is truncated and stored, the window shifts backward along the time axis by 10ms, automatically discarding the temporal data in the first 10ms of the window, and simultaneously incorporating the newly added 10ms of preprocessed data from the back end, completing a new round of window data update and fragment truncation.

[0024] This half-cycle step sliding method allows adjacent window data blocks to overlap for half the duration of time sequence data, effectively solving the problem of fault instantaneous features falling on the window boundary and being truncated or missed. For minor faults with extremely short durations, such as single-phase instantaneous grounding and short-term phase-to-phase short circuits, it can capture weak fault waveform features in all aspects, greatly improving the completeness of fault data acquisition.

[0025] By iteratively sliding and trunculating with a fixed step size, the preprocessed standard data frame of the full time series is segmented continuously. Each sliding truncation generates a set of independent and complete window data blocks. All data blocks are arranged and integrated strictly according to the chronological order of their occurrence, ultimately forming a continuous and ordered sequence of sliding window data blocks. This sequence completely replicates the operating status of the line throughout the entire time period, covering all operating conditions such as normal operation, fault initiation, fault occurrence, fault persistence, and fault elimination. It provides a refined and segmented standard data carrier for subsequent segment-by-segment extraction of time-domain and frequency-domain fault features.

[0026] For each data block in the sliding window data block sequence, time-domain and frequency-domain features are extracted, and the amplitude of the sudden change in three-phase current, harmonic distortion rate, voltage drop depth, and effective value of zero-sequence current are calculated to generate a fault feature vector.

[0027] The core of this step is to mine key fault feature parameters in both the time and frequency domains from segmented window data, quantify the differences in electrical waveforms for different fault types, and transform the raw waveform data into highly recognizable digital feature vectors. This provides core feature basis for subsequent intelligent fault identification and localization. The specific implementation method is as follows: Fault feature extraction is divided into two dimensions: time domain feature extraction and frequency domain feature extraction. Time domain features focus on the instantaneous change characteristics of electrical quantities over time, which can intuitively reflect the changes in the line operating status at the moment the fault occurs. Frequency domain features focus on the spectral distortion characteristics of the fault waveform, which can accurately capture the nonlinear distortion problem of the waveform caused by the fault. The two types of features complement each other and can comprehensively characterize the core characteristics of the fault.

[0028] The time-domain characteristics mainly include two core parameters: the amplitude of the three-phase current surge and the voltage sag depth. The amplitude of the three-phase current surge is the absolute value of the difference between the real-time sampled values ​​of the currents in phases A, B, and C at the instant of the fault and the steady-state operating current values ​​before the fault, expressed in amperes. This parameter directly reflects the impact intensity and severity of the fault. In the example, the current in phase A during normal steady-state operation is 120A, and at the instant of the fault, the current in phase A surges to 960A, resulting in a calculated amplitude of 840A for the phase A current surge. The larger the amplitude, the more severe the impact of the line fault. The voltage sag depth is the percentage of the difference between the steady-state voltage of each phase after the fault and the rated phase voltage of the line. It is a dimensionless parameter used to quantify the degree of voltage attenuation caused by the fault. If the rated phase voltage of the line is 220V, and the voltage in phase B drops to 77V after the fault, the corresponding voltage sag depth is 65%. The voltage sag depth has a fixed range for different fault types and is a key indicator for distinguishing fault types.

[0029] Frequency domain characteristics are based on the three-phase current harmonic distortion rate as the core parameter. This parameter is calculated by performing spectral decomposition of the current waveform in the window data block using Fourier transform. It is the ratio of the total effective value of all higher harmonics to the effective value of the fundamental wave, and is a dimensionless parameter used to characterize the degree of distortion of the fault waveform. When the line is operating normally, the waveform has good sinusoidal properties, and the harmonic distortion rate is usually below 5%. When a short circuit or ground fault occurs, the waveform is severely distorted, the harmonic content increases significantly, and the harmonic distortion rate can rise to over 15%. Moreover, the harmonic distortion rate values ​​of different faults such as phase-to-phase short circuits and single-phase ground faults differ significantly, which can serve as a basis for refined fault classification.

[0030] The effective value of zero-sequence current is a core parameter for distinguishing ground faults. Zero-sequence current is calculated by superimposing the vectors of the three-phase currents. When the three phases of the line are symmetrical and operating normally, the vector sum of the three-phase currents approaches zero, and the effective value of zero-sequence current is basically 0A. When a single-phase ground fault or a two-phase ground fault occurs, the three-phase currents become unbalanced, resulting in a continuous zero-sequence current. By performing an integral calculation of the effective value of the zero-sequence current waveform within the window data block over the entire time period, an accurate effective value of zero-sequence current can be obtained, in amperes. This parameter can directly distinguish between ground faults and pure phase-to-phase short-circuit faults and is an indispensable key parameter in the fault characteristic system.

[0031] After calculating all individual feature parameters, the three-phase current surge amplitude, three-phase current harmonic distortion rate, three-phase voltage drop depth, and zero-sequence current effective value corresponding to each sliding window data block are integrated and arranged according to a fixed dimensional order and data format. This completes the dimensional alignment and numerical normalization of multi-dimensional features, ultimately generating a fault feature vector with a unified structure, fixed dimensions, and complete features. Each vector corresponds to all fault feature information of a time series window, which can be directly input into the subsequent lightweight convolutional neural network model to complete fault identification and segment location calculations.

[0032] S202, Based on the fault feature quantity, a lightweight convolutional neural network model is used to identify the fault type and locate the fault segment, and generate a fault determination result including fault type, fault distance and fault phase. Specifically, the fault feature vector can be input into the input layer of a lightweight convolutional neural network model. This model uses a depthwise separable convolutional structure to reduce the number of parameters. Through convolution and pooling operations, deep fault feature maps are extracted to generate a deep fault feature tensor. The core of this step is to complete the deep mining and dimensional transformation of shallow fault features. Leveraging the unique structural advantages of lightweight convolutional neural networks, while reducing the model's computational load and adapting to low-computing-power scenarios at edge nodes, the one-dimensional fault feature vector is transformed into a high-dimensional, highly discriminative deep feature tensor. This provides standardized feature data for subsequent accurate fault classification and ranging calculations. The specific implementation method is as follows: The fault feature vector is a standardized one-dimensional data vector obtained after data preprocessing and feature extraction. It integrates the core anomaly features in the time and frequency domains after a line fault. The vector dimension is fixed at 16 dimensions according to the types of extracted features. Each dimension corresponds to specific feature parameters such as the amplitude of three-phase current surge, the harmonic distortion rate, the three-phase voltage drop depth, and the effective value of zero-sequence current. All parameter values ​​have been normalized and the value range is uniformly locked in the interval between 0 and 1. The numerical precision retains four decimal places, which can accurately quantify the minor, moderate, and severe anomalies of line faults. It is the basic input data for model fault identification.

[0033] The lightweight convolutional neural network model is a dedicated deep learning model adapted for deployment on edge computing nodes of pole-mounted circuit breakers. Unlike traditional standard convolutional neural networks, its core adopts a depthwise separable convolutional structure to reduce the number of parameters and computational cost. Traditional standard convolution requires simultaneous channel convolution and spatial convolution operations, resulting in a large number of parameters and long computation time. Depthwise separable convolution, on the other hand, splits the convolution operation into two independent stages: depthwise convolution and pointwise convolution. Depthwise convolution performs spatial convolution operations on each feature channel separately, while pointwise convolution uses a 1×1 convolution kernel to complete the fusion and splicing of channel dimensions. This two-step operation splitting can reduce the overall number of model parameters by more than 75% and floating-point operation by more than 80%, perfectly adapting to the operating scenarios of edge computing nodes with limited computing power and high real-time requirements, avoiding the problems of computational delay and device overload that occur with large models.

[0034] The model input layer is a dedicated data receiving layer, adapted to the input specification of 16-dimensional fault feature vectors. It has data verification and dimensionality normalization functions, which can automatically identify the dimensional integrity and numerical compliance of the input vector, filter out abnormal values ​​and missing data, and ensure that the feature data input to the model is uniform and standardized. After the standardized fault feature vector is input into the input layer, the model performs convolution and pooling operations in sequence to complete deep feature extraction. The convolution operation uses a 3×3 kernel, with a stride of 1 and a padding value of 1. The kernel size is used to define the spatial range of a single feature extraction, the stride of 1 means that the kernel slides through all feature data pixel by pixel, and the padding value of 1 is used to ensure that the feature map size does not shrink before and after the convolution operation, thus fully preserving the edge fault feature information.

[0035] Max pooling is used for pooling operations, with a kernel size of 2×2 and a stride of 2. The core function of max pooling is to compress the feature map dimension, filter core fault features, eliminate redundant and invalid feature data, and retain the maximum response value of fault features, thereby enhancing the recognizability of fault features. The combination of kernel size and stride can minimize the amount of data while retaining core features. After multiple rounds of convolution and pooling iterations, the model discards redundant information from the original shallow features, uncovers fault correlation patterns and feature coupling relationships hidden in the shallow data, and generates a multi-channel, high-dimensional deep fault feature map. This feature map integrates multi-dimensional fault characteristics such as temporal fluctuations, frequency domain distortion, and abrupt changes in electrical parameters.

[0036] Finally, the multi-layer deep fault feature maps are integrated and encapsulated into tensors to unify the data layout format and dimensional order, generating a fault deep feature tensor with fixed specifications. This tensor is three-dimensional structured data, corresponding to the number of feature channels, feature map height, and feature map width, respectively. In the example, the final generated fault deep feature tensor has a dimension of 8×8×16. Each value inside the tensor corresponds to a refined fault feature representation, fully carrying the core implicit features of the line fault and providing accurate data support for subsequent dual-branch fault identification operations.

[0037] The deep feature tensor of the fault is input into the fault type classification branch and the fault location branch respectively. The classification branch outputs the probability distribution of four types of fault types—single-phase grounding, two-phase short circuit, two-phase grounding, and three-phase short circuit—through the fully connected layer, generating a preliminary fault type determination result. The core of this step is to construct a dual-branch parallel inference architecture. Relying on the refined feature data of the deep fault feature tensor, it performs intelligent classification of fault types and preliminary calculation of fault locations. Through feature mapping operations of fully connected layers, it outputs standardized fault type probability results, achieving preliminary accurate determination of fault types. The specific implementation method is as follows: This model employs a dual-branch parallel computation structure. The two branches share the deep fault feature tensor extracted by the front end, simultaneously performing fault classification and fault location calculations without interference and with data from the same source. This effectively ensures the consistency and synchronization of the two types of fault judgment results. Furthermore, the parallel computation mode significantly improves fault identification efficiency, meeting the millisecond-level response requirements for rapid self-healing of power distribution line faults. After feature extraction of the deep fault feature tensor, it is completely and losslessly input into the fault type classification branch and the fault location branch through the model's internal feature splitting module. The splitting process does not change the tensor's dimension or numerical characteristics; it only completes the splitting and transmission of the data path.

[0038] The fault type classification branch is a dedicated inference unit for fault pattern recognition. It consists of three fully connected layers connected in series, with the number of neurons in each fully connected layer decreasing sequentially to achieve gradual compression of feature dimensions and precise mapping of feature weights. The first fully connected layer has 128 neurons, which are used to integrate and summarize the massive and refined features of the deep fault feature tensor. The second fully connected layer has 64 neurons, which are used to filter core fault classification features and remove invalid interference features. The third fully connected layer has 4 neurons, which correspond one-to-one with the four preset fault types.

[0039] The fully connected layer operates with a softmax activation function. The core function of this function is to normalize the original feature values ​​output by the fully connected layer into probability values ​​between 0 and 1, ensuring that the sum of the probability values ​​of the four neurons is always 1. This accurately represents the matching probability of various faults corresponding to the current line state. The model pre-defines four types of faults: single-phase grounding fault, two-phase short-circuit fault, two-phase grounding fault, and three-phase short-circuit fault. Single-phase grounding is the most common minor fault in distribution lines; two-phase short-circuit is a phase-to-phase metallic short-circuit fault; two-phase grounding is a composite fault involving phase-to-phase faults and grounding; and three-phase short-circuit is the most dangerous symmetrical short-circuit fault. These four types of faults cover more than 99% of fault scenarios in distribution lines, enabling full-scenario fault identification.

[0040] After feature mapping and probability transformation by the activation function in the fully connected layer, the classification branch outputs four independent probability values, forming a complete fault type probability distribution. In the example, the probability distribution for a certain fault identification is: single-phase ground fault probability 0.9215, two-phase short circuit fault probability 0.0326, two-phase ground fault probability 0.0402, and three-phase short circuit fault probability 0.0057. The probability values ​​intuitively reflect the matching degree of each type of fault. The model follows the maximum probability matching principle, selecting the fault type with the highest probability value as the output result. In the above example, the single-phase ground fault has the highest probability value, so the preliminary fault type determination result is a single-phase ground fault, completing the preliminary intelligent identification of the fault type.

[0041] In the ranging branch, the fault depth feature tensor is convolved with the line parameter matrix, and the distance from the fault point to the measurement point is calculated by combining the traveling wave method principle. The fault distance estimate is output, and the fault section location result is generated. The core of this step is to rely on the coupling relationship between the inherent electrical parameters of the line and the fault characteristics, combined with the classic traveling wave ranging principle, to achieve accurate fault distance calculation through matrix convolution operation. This overcomes the shortcomings of traditional ranging methods, such as low accuracy and susceptibility to interference, and accurately locates the physical location of the fault, generating the corresponding fault segment location result. The specific implementation method is as follows: The fault location branch is a model-specific location calculation unit, distinct from the feature classification logic of the classification branch. This branch relies on the fusion calculation of inherent line parameters and fault dynamic features to achieve distance measurement. The core inputs are the deep fault feature tensor and a pre-stored line parameter matrix. The line parameter matrix is ​​a standardized two-dimensional matrix pre-calibrated offline and stored locally. All matrix parameters are inherent and fixed parameters of the target distribution line, unchanging with the line's operating state. The matrix contains core electrical and spatial parameters such as line wave velocity v_1, positive-sequence impedance per unit length Z_1, zero-sequence impedance per unit length Z_0, total line length L_1, and line capacitance per unit length C_1. All parameters are calibrated through actual line construction measurements, ensuring accurate and unbiased values. The typical value for line wave velocity v_1 is 2.98 × 10^8 m / s, representing the propagation speed of the fault traveling wave in the distribution line. The total line length L_1 is set according to the actual line section; in this example, it is 2000m, representing the total length of the currently monitored line.

[0042] This method employs adaptive convolution operations between the deep fault feature tensor and the line parameter matrix. Unlike conventional feature convolution, this convolution operation is a parameter fusion convolution. Its core function is to deeply couple the time-domain and frequency-domain variation patterns of the fault transient features with the inherent propagation characteristics and impedance characteristics of the line. This allows for the calculation of the propagation time difference characteristic of the fault traveling wave, accurately capturing the time difference Δt_1 from the fault point to the measurement point of the pole-mounted circuit breaker after the fault occurs. The calculation accuracy of the time difference Δt_1 can reach 0.1 microseconds, effectively capturing subtle time differences in the traveling wave.

[0043] After the calculation is completed, the fault distance is calculated based on the core principle of the traveling wave method. The core logic of the traveling wave method is that a high-frequency transient traveling wave is generated at the moment of a fault. The traveling wave propagates along the line at a fixed wave speed. By measuring the propagation time difference of the traveling wave, the fault distance can be accurately estimated. The corresponding distance calculation formula is fault distance S_1 = v_1 × Δt_1 / 2. The division by 2 in the formula is because the traveling wave will reflect between the fault point and the measurement point, and the distance error of the round-trip propagation needs to be eliminated to ensure the accuracy of the distance measurement. In the example, the measured traveling wave propagation time difference Δt_1 is 10.06 microseconds. Substituting the wave speed parameter, the fault distance S_1 can be calculated to be 1500m. This value is the estimated straight-line distance from the fault point to the measurement point of the pole-mounted circuit breaker.

[0044] After obtaining the estimated fault distance, the fault segment location is completed by combining it with the preset line segment division rules. The line segment is divided equally into units of 500m each. In the example, the 2000m line is divided into four segments: 0 to 500m, 500 to 1000m, 1000 to 1500m, and 1500 to 2000m. The fault distance of 1500m calculated in this case corresponds to the end of the third segment of the line. Therefore, the final fault segment location result is that the fault is in the 1000 to 1500m section of the line, accurately locating the line segment where the fault occurred and providing a location basis for subsequent fault isolation.

[0045] Based on the preliminary fault type determination results, fault section location results, and three-phase current change phase information, the final fault phase is determined using a majority voting fusion rule, generating a fault determination result that includes fault type, fault distance, and fault phase.

[0046] The core of this step is the multi-dimensional fault information fusion and verification. It integrates three core data categories: fault type, fault location, and electrical phase characteristics. Through majority voting fusion rules, it eliminates the errors and interference of single-dimensional judgment, accurately determines the fault phase, and finally integrates all valid information to generate a complete and accurate integrated fault judgment result. The specific implementation method is as follows: The three types of core data required for multi-dimensional fusion each possess independent fault characterization capabilities, complementing and verifying each other to comprehensively reconstruct the true state of line faults. Specifically, the preliminary fault type determination result clarifies the fault mode, the fault section location result clarifies the spatial location of the fault, and the three-phase current mutation phase information clarifies the electrical phase in which the fault occurred. The three-phase current mutation phase information is feature information extracted based on the difference in the mutation amplitude of the three-phase current at the time of the fault. The preset current mutation threshold is 10% of the rated current. When the current mutation amplitude of one or more phases exceeds this threshold, it can be determined that there is a fault anomaly in the corresponding phase. In the example, the current mutation amplitude of phase A at the time of the fault far exceeds the threshold, while the currents of phases B and C show no obvious mutations. Therefore, the preliminary determination of the fault anomaly phase is phase A.

[0047] The majority voting fusion rule is the core algorithm of multi-source information fusion. This rule adopts the logic of independent judgment of three-source data and selection of the best by vote count. It outputs an independent set of fault phase judgment conclusions based on fault type characteristics, fault section electrical parameter characteristics, and three-phase current change characteristics. The three sets of conclusions are independent of each other and do not interfere with each other. Finally, the vote count of the three sets of conclusions is counted, and the phase with the most votes is selected as the final fault phase. This effectively avoids judgment errors caused by line harmonics, load fluctuations, and environmental interference due to single features, and greatly improves the accuracy of fault phase judgment.

[0048] In the specific calculation process, firstly, based on the initially determined single-phase grounding fault type, the conventional phase characteristics of single-phase grounding faults are matched, and the first set of phase identification conclusions is output as phase A grounding fault. Then, based on the historical fault data of the 1000-1500m fault section, the electrical parameter characteristics of the section, and the feature matching degree of this fault, the second set of phase identification conclusions is output as phase A fault. Finally, based on the measured data of the three-phase current mutation, the third set of phase identification conclusions is output as phase A fault. If the three sets of voting conclusions are consistent, the final fault phase is determined as phase A through the majority voting fusion rule. If there are differences among the three sets of conclusions, the phase with the highest percentage of votes is taken as the final result. If the votes are tied, the conclusion based on the measured data of the three-phase current mutation phase is adopted first to ensure that the judgment result closely matches the actual electrical operating state.

[0049] After the final fault phase identification is completed, all valid fault information is integrated and summarized. Standardized core fault parameters are sorted out, and the preliminary fault type identification results, fault distance estimation, fault section location results, and final fault phase identification information are integrated. Redundant intermediate calculation data is removed, and the data output format is unified to generate a complete fault identification result. In the example, the final fault identification result generated is as follows: the fault type is a single-phase ground fault, the fault distance is 1500m from the measurement point, the fault section is the 1000 to 1500m interval of the line, and the fault phase is phase A. This result covers all core information such as fault mode, fault location, and fault phase. The data is accurate and complete, and can be directly used for subsequent adversarial verification and self-healing strategy matching calculations.

[0050] S203, The fault determination result is input into the locally deployed adversarial verification network. The credibility of the fault determination result is evaluated through the game between the generator and the discriminator, and the valid fault result confirmed by the verification is output. Specifically, the fault determination result can be encoded into a condition vector, concatenated with the original fault feature vector, and then input into the generator of the adversarial verification network. The generator attempts to generate a pseudo feature vector that is consistent with the distribution of real fault features, thereby generating fitted pattern data. The core of this step is to complete the digital encoding and data fusion of the fault determination results. Using real fault determination information as constraints, the generator is driven to simulate the characteristic distribution patterns of real line faults, generating simulated feature data that closely matches the actual fault scenario. This provides positive and negative sample data support for subsequent credibility verification. The specific implementation method is as follows: The fault determination result is a comprehensive output from the lightweight convolutional neural network model, containing three core types of information: specific fault type, fault distance value, and fault phase. This is structured semantic information and cannot be directly input into an adversarial verification network for computation. Therefore, it requires digital conversion using a vector encoding algorithm to generate standardized condition vectors. The vector encoding employs a fusion encoding method combining one-hot encoding and numerical normalization. One-hot encoding is used for discrete fault types and fault phases, while maximum-minimum value normalization encoding is used for continuous fault distance values, unifying data units and dimensions. In practical applications, the condition vector is set to a fixed 16-dimensional dimension. The first 4 dimensions represent the encoding information of four fault types, corresponding to single-phase grounding, two-phase short circuit, two-phase grounding, and three-phase short circuit faults, respectively. The value of the dimension corresponding to a single fault type is set to 1, and the other dimensions are set to 0. In the example, if the fault type is a single-phase grounding fault, the values ​​of the first 4 dimensions of the condition vector are 1, 0, 0, and 0. The middle 8 dimensions represent the normalized value of the fault distance. The original fault distance ranges from 0m to 10000m, and after normalization, it is mapped to the interval between 0 and 1. If the detected fault distance is 2500m, the corresponding dimension value after normalization is 0.25. The last 4 dimensions represent the fault phase information, corresponding to phase A, phase B, phase C, and multi-phase faults. The value of the dimension corresponding to a single fault phase is set to 1, and the other dimensions are set to 0. If the fault phase is phase A grounding, the values ​​of the last 4 dimensions are 1, 0, 0, and 0. Finally, the 16-dimensional standardized condition vector is generated.

[0051] The original fault feature vector is a set of time-domain and frequency-domain fault features extracted in the early stage through a sliding time window. It is a multi-dimensional numerical vector with fixed dimensions. In this method, the original fault feature vector is uniformly set to 32 dimensions, which includes all core fault features such as the amplitude of three-phase current surge, harmonic distortion rate, voltage drop depth, and effective value of zero-sequence current. The values ​​of each dimension are standardized values ​​after preprocessing, and the value range is stable in the range of 0 to 1, thus completely preserving the original data features and transient change patterns of line faults.

[0052] Vector concatenation is the core operation for fusing conditional information with original feature information. Using a dimensional concatenation algorithm, the 16-dimensional conditional vector and the 32-dimensional original fault feature vector are concatenated end to end to generate a 48-dimensional fusion input vector. The concatenation process does not change the original value of any dimension, but only achieves the superposition and fusion of data dimensions. This preserves both the constraints of the fault judgment result and the complete original fault features of the line, ensuring the integrity of the input data information.

[0053] The generator of the adversarial verification network is a lightweight generative network structure, adapted to the local computing power deployment requirements of edge computing nodes. The core function of the network is to learn the feature distribution patterns of real line faults based on the input fusion vector, and simulate and generate fault feature data that closely matches actual operating conditions. The real fault feature distribution refers to the feature value distribution patterns formed by the statistics of massive historical real line fault samples. It covers the numerical range, fluctuation range and correlation patterns of various feature quantities such as current and voltage under different fault types, different fault distances and different fault phases. It is a standardized feature distribution model of line faults on power grid poles.

[0054] After receiving a 48-dimensional fused input vector, the generator uses multi-level linear transformations and activation function operations to fit the numerical patterns of the real fault feature distribution. It outputs a 32-dimensional pseudo-feature vector with the same dimensions as the original fault feature vector. This pseudo-feature vector is virtual fault feature data simulated by the algorithm, not actual data collected from the line, but its numerical distribution and feature correlations highly closely resemble real fault scenarios. After generating multiple sets of pseudo-feature vectors in batches, they are integrated to form a complete fitted pattern data set. This fitted pattern data is a standardized fault feature dataset with a unified data format and dimensional standards, which can be directly input into the subsequent discriminator for reliability verification, providing sufficient simulated sample data support for fault result verification.

[0055] The fitted pattern data output by the generator is input into the discriminator along with samples from the historical real fault sample library. The discriminator calculates the probability score of each sample belonging to a real fault and outputs the discrimination confidence score. The core of this step is to use a discriminator to differentiate between real and fake fault samples, quantify the similarity between simulated fitted pattern data and real fault samples, and quantify the reliability of the fault determination result using a standardized confidence score, providing a quantitative basis for subsequent threshold comparison. The specific implementation method is as follows: The historical real-world fault sample library is a sample database built offline and updated in real time on local edge computing nodes. All samples in the library are derived from real fault data collected during actual line operation. After preprocessing, feature extraction, and manual verification, the data is standardized and stored. The sample library covers all fault types, including single-phase grounding, two-phase short circuit, two-phase grounding, and three-phase short circuit. It also encompasses various fault samples under different fault distances, load conditions, and environmental interference. The total sample capacity can be dynamically expanded according to the line's operating time, with a basic reserve of no less than 5000 sets of samples. Each real sample is a 32-dimensional fault feature vector, completely consistent with the pseudo-feature vector in terms of dimension, data format, and numerical range, ensuring the consistency of the discriminator's input sample dimensions and avoiding discrimination errors caused by dimensional differences.

[0056] The fitted pattern data is generated in batches by the generator. Each verification process generates 20 sets of pseudo-feature vectors to form a single batch of fitted pattern data. All pseudo-feature vectors are generated based on the conditional constraints of the current fault judgment result, corresponding to the fault type and fault segment characteristics detected in this test. They are specifically matched to the current fault result to be verified, thereby improving the accuracy of the judgment and verification.

[0057] The discriminator is the core discriminant unit of the adversarial verification network. It adopts a lightweight binary classification network structure, adaptable to the real-time computing needs of the edge computing side. The core algorithm is a probability normalization discriminant algorithm, which can classify each set of input feature samples as true or false faults. The input layer of the discriminator simultaneously receives pseudo-feature samples from the fitted pattern data and real samples from the historical real fault sample library. The two types of samples are mixed and input without distinguishing labels. The network autonomously learns the feature differences and completes the discriminant operation.

[0058] The discriminator performs deep feature extraction on each set of input samples, compares the fit between the sample feature distribution and the baseline real fault feature distribution, and calculates a probability score in the range of 0 to 1. This probability score is used to characterize the matching probability of a single set of samples as real line fault data. The closer the value is to 1, the higher the fit between the sample features and the real fault features. The closer the value is to 0, the higher the probability that the sample is simulated pseudo fault data.

[0059] After calculating the probability scores of all samples in a single batch, the discriminator uses a weighted average algorithm to integrate the scores. It performs a weighted summation and average of the probability scores of 20 pseudo-feature samples, with the weights uniformly set to equal weights to eliminate the interference of random errors from a single sample. Finally, it outputs a unique discrimination confidence score, with the score range fixed between 0 and 1. This score can accurately quantify the degree of matching between the feature data corresponding to the current fault judgment result and the real fault scenario. The score directly corresponds to the credibility of the fault judgment result, providing a quantitative indicator for subsequent credibility judgment.

[0060] The confidence score is compared with the confidence threshold. If the score is higher than the threshold, the fault determination result is determined to be credible. Otherwise, it is determined to be unreliable and needs to be retested to generate a credibility assessment conclusion. The core of this step is to use a fixed threshold to quantify the discrimination criteria, thereby completing the reliability screening of fault judgment results, distinguishing between valid fault judgment results and invalid judgment results caused by misjudgment or interference, and realizing the initial verification and screening of fault results. The specific implementation method is as follows: The confidence threshold is a fixed verification standard threshold determined based on training and simulation verification using massive amounts of power grid line fault samples. It is adapted to the operating conditions of pole-mounted circuit breaker lines and is used to define the valid and invalid boundaries of fault judgment results. In this method, the pre-set confidence threshold is 0.85. This value has been verified through multi-scenario fault simulation and can effectively avoid fault misjudgment caused by line noise interference, transient fluctuations, and sampling errors, while ensuring that no real minor fault signals are missed. The threshold value of 0.85 represents the judgment benchmark. When the confidence score reaches this value or above, it proves that the current fault characteristics match the real fault scenario, and the judgment result is valid. When the score is lower than this value, it proves that the current characteristics are mostly caused by interference data and data fluctuations, and are not real line faults.

[0061] The score comparison employs precise numerical comparison logic, with no ambiguity intervals or compromise judgment rules. There are only two judgment results, and the judgment logic is simple and stable, adapting to the needs of high-speed real-time verification of edge nodes. During the comparison process, the system automatically retrieves the judgment confidence score generated in this calculation and compares it bit by bit with the locally fixed 0.85 confidence threshold. The entire process is independently calculated by the local edge computing node without the need for the main station's participation, ensuring the real-time nature of the verification.

[0062] When the confidence score is higher than the 0.85 threshold, it indicates that the fault characteristics simulated by the generator are highly consistent with the historical real fault samples. The fault type, fault distance, and fault phase identification results identified by the lightweight convolutional neural network model in the early stage are consistent with the actual fault conditions of the line. There are no problems such as sampling interference, model misidentification, or data anomalies. Therefore, the fault judgment result is reliable.

[0063] When the confidence score is below the 0.85 threshold, it indicates that the extracted fault features and the fault results determined by the model deviate significantly from the actual fault feature distribution. This is most likely due to misidentification caused by non-fault interference factors such as line instantaneous noise, transformer sampling fluctuations, and load mutations. The fault determination result is not of reference value and is determined to be unreliable. It is necessary to trigger the line secondary sampling and detection process to re-collect line data, extract fault features, and complete the fault determination.

[0064] After completing the numerical comparison and result judgment, the system automatically generates a standardized credibility assessment conclusion. The conclusion includes three core types of information: the confidence score of this judgment, the threshold standard, and the credibility judgment result. It is stored locally in a standardized data format, providing a clear basis for subsequent output of valid fault results or triggering of secondary detection processes, ensuring that every fault verification process is traceable and verifiable.

[0065] For fault determination results that are determined to be reliable, the results are directly output as valid fault results. For results that are determined to be unreliable, secondary sampling is triggered to re-execute the fault determination process, and finally, a valid fault result that has been verified and confirmed is output.

[0066] The core of this step is to execute differentiated processing logic based on the credibility assessment conclusion, complete the final verification and screening of fault results, eliminate invalid and misjudged results, compensate for the error defects of single detection through a secondary detection mechanism, and finally output accurate and reliable valid fault results to support the subsequent execution of self-healing strategies. The specific implementation method is as follows: For fault determination results deemed reliable by the credibility assessment, the system directly completes data encapsulation and output, marking the complete fault determination result, including fault type, fault distance, and fault phase, as valid data and defining it as a valid fault result. This result undergoes adversarial network game verification, eliminating the possibility of data interference and model misjudgment, and can accurately reflect the current fault state of the line. It eliminates the need for repeated testing and can be directly used for matching subsequent fault self-healing strategies and generating control commands, ensuring the real-time performance and efficiency of fault handling.

[0067] For fault determinations deemed unreliable by the credibility assessment, the system immediately triggers a secondary sampling mechanism at the edge computing nodes, initiating a new round of line data acquisition and fault determination. The secondary sampling employs the same sampling rate of 256 points per cycle as the initial detection, maintaining consistent sampling accuracy. Simultaneously, the sliding time window acquisition duration is extended, adding two cycles of steady-state data acquisition to the original window spanning the first half of the fault cycle to the last five cycles, further mitigating the impact of instantaneous data fluctuations and improving the stability of the sampled data.

[0068] After the second sampling is completed, the newly acquired original high-frequency time-series data stream is reprocessed with denoising and normalization. Median filtering is used to remove sudden spike interference again, and the current and voltage amplitude data are re-normalized. Then, a new data block sequence is extracted through a sliding time window, and a new round of time-domain and frequency-domain fault feature vectors are extracted. The new fault feature vectors are input into a lightweight convolutional neural network model to re-complete fault type identification, fault segment location, and fault phase determination, generating a new fault determination result.

[0069] The newly generated fault determination result is input again into the adversarial verification network to complete a new round of credibility assessment. This involves repeating the entire verification process of encoding, concatenation, generating simulated data, scoring, and threshold comparison until a credible fault determination result is obtained. The method is configured to perform a maximum of two detection and verification checks per fault scenario to avoid infinite loop detection affecting system efficiency. If a credible result is obtained within two checks, it is directly output. If both checks determine it to be untrustworthy, it is assumed that there is no real fault in the current line, and a valid fault-free result is output, terminating the current fault determination process.

[0070] After passing a single reliable judgment or a second re-inspection and verification, the final output of valid fault results are all accurate results that have been verified by adversarial game, eliminated interference errors, and closely match the actual working conditions of the line. The results fully retain the core information of fault type, fault distance, and fault phase. The data accuracy and reliability meet the control requirements of local fault self-healing of pole-mounted circuit breakers, providing accurate and reliable data support for subsequent self-healing strategy matching, fault isolation and power restoration operations.

[0071] S204. Based on the valid fault result, call the locally stored fault self-healing strategy library, match the action logic corresponding to the current fault type, and generate a fault isolation and recovery control instruction containing a tripping instruction, a reclosing instruction, or a blocking instruction. Specifically, it can parse the fault type, fault distance and fault phase information in the valid fault results, extract the key fields for strategy matching, and generate a fault feature tag set; The core of this step is to perform refined data analysis on the valid fault results completed by the edge computing node's local verification, filter and refine the core fault feature information that is suitable for matching the fault self-healing strategy, and transform the composite fault judgment data into standardized, lightweight tagged data. This provides a unique index basis for subsequent accurate retrieval and matching in the strategy library, avoiding redundant data from interfering with strategy judgment. The specific implementation method is as follows: The valid fault result is the final fault judgment data output after identification and localization by a lightweight convolutional neural network and credibility verification by an adversarial verification network. Unlike the initial model judgment result, this result has undergone authenticity verification and error correction, possessing the validity and reliability to be directly used for field equipment control. The main data includes three core technical parameters: fault type, fault distance, and fault phase, along with a small amount of auxiliary log data for tracing. This step only analyzes the three core control parameters, automatically filtering out auxiliary data without strategy matching value. The fault type is the core parameter defining the fault mechanism and impact form of line faults, including four high-frequency fault types in distribution lines: single-phase grounding, two-phase short circuit, two-phase grounding, and three-phase short circuit. The four types of faults differ significantly in their fault severity, fault duration characteristics, and handling priorities, serving as the primary core basis for self-healing strategy classification and matching. In field operation, single-phase grounding faults are mostly caused by line insulation damage or foreign object contact, while two-phase and three-phase short circuit faults are mostly caused by line phase-to-phase breakdown or equipment failure. Different fault types correspond to completely independent self-healing action logic. Fault distance is a key quantitative parameter characterizing the spatial location of a fault. It refers to the distance of the line extension from the fault location to the measurement point of the pole-mounted circuit breaker, with the unit being kilometers. The data analysis accuracy can reach 0.01 kilometers. This parameter is used to distinguish between near-end and far-end faults. Near-end faults have concentrated fault energy and a faster spread speed, resulting in a greater impact on line equipment and a higher priority for self-healing response. Far-end faults have significantly attenuated fault energy and a lower risk of equipment damage. In this example, the fault distance value obtained in this analysis can be set to 1.65 kilometers to define the fault location as belonging to the middle fault range of the line. Fault phase is a core parameter for accurately locating the power phase sequence of the fault. Distribution lines are divided into three independent power supply phase sequences: A, B, and C. Fault phase can be classified into three categories: single-phase fault, two-phase fault, and three-phase fault. It directly determines the phase control parameters of phase-controlled opening and closing operations and is a key basis for achieving refined, impact-free self-healing operations. In this example, the fault phase obtained in this analysis can be set to a single-phase fault of phase B.

[0072] After accurately analyzing the three core parameters, the system extracts key fields for specific strategy matching from the complete fault result data according to preset strategy matching filtering rules. Field extraction follows the principles of precise adaptation, no redundancy, and full coverage, retaining only three core fields: fault type identifier, fault distance quantification value, and fault phase identifier, while removing all auxiliary data unrelated to the self-healing logic. After extraction, the system standardizes and organizes the discrete key fields, unifying field naming conventions, numerical precision, and identifier formats to eliminate matching errors caused by data format differences, ultimately generating a structured and standardized fault feature tag set. The fault feature tag set is a standardized data unit integrating all core fault features. Each tag set uniquely corresponds to a line fault scenario and can be directly used as a retrieval index for the local strategy library. In this example, the generated fault feature tag set specifically includes the fault type tag: single-phase ground fault, the fault distance tag: 1.65 km, and the fault phase tag: phase B fault, fully covering all core features required for strategy matching.

[0073] Using the fault feature tag set as an index, query the locally stored fault self-healing strategy library. This strategy library is organized in a decision tree structure. Traverse and match the action logic entries corresponding to the current fault type to generate matching strategy records. The core of this step is to rely on a locally deployed decision tree architecture fault self-healing strategy library. Using a standardized fault feature label set as the retrieval index, and through a hierarchical traversal and feature comparison matching mechanism, it accurately locates the standardized self-healing action logic adapted to the current fault scenario, and generates a complete strategy record specific to this fault. This provides compliant and standard rule support for subsequent action instruction parsing. The specific implementation method is as follows: The fault self-healing strategy library is a dedicated set of handling rules pre-embedded in the local storage unit of the edge computing node of the pole-mounted circuit breaker. All strategy data is stored and called locally, without relying on remote data interaction from the power master station. It can achieve millisecond-level strategy response, effectively avoiding the problem of delayed self-healing handling caused by remote communication latency. The strategy library adopts a decision tree structure for data organization and logical arrangement. The decision tree is a hierarchical and progressive logical discrimination architecture that is adapted to the accurate matching scenario of multi-dimensional fault characteristics. Compared with the traditional linear retrieval method, it has the advantages of high retrieval efficiency, accurate scenario adaptation, and clear logical hierarchy, and is fully adapted to the judgment requirements of multi-feature coupling of distribution line faults. The decision tree structure is divided into three progressive discrimination levels: the first level is the fault type discrimination layer, the second level is the fault distance interval discrimination layer, and the third level is the fault phase discrimination layer. The three levels converge layer by layer, realizing accurate positioning of the entire scenario from macro fault type to micro fault location.

[0074] The first level of the decision tree divides branches according to four standard fault types: single-phase grounding, two-phase short circuit, two-phase grounding, and three-phase short circuit. Each branch corresponds to a basic self-healing framework for a specific fault type and serves as the core primary criterion for strategy matching. All subsequent branch logic is based on the results of this level. The second level, based on the fault types in the first level, divides branches into intervals according to preset fault distance thresholds. The system presets 0-2 km as the near-end fault interval, 2-5 km as the mid-end fault interval, and above 5 km as the far-end fault interval. Different distance intervals correspond to different fault handling priorities and action sequences. In the example, the fault distance of 1.65 km belongs to the near-end fault interval and corresponds to the near-end fault secondary branch under the single-phase grounding fault branch. The third level, based on the secondary branches, completes the final branch location based on the fault phase label, distinguishing refined handling parameters for different fault phases to achieve a unique and accurate match for the fault scenario.

[0075] After completing the decision tree hierarchical positioning, the system initiates a global traversal matching mechanism, traversing all stored action logic entries layer by layer along the decision tree branches corresponding to the fault feature labels. Each logic entry is a pre-defined, standardized self-healing handling rule bound to a specific combination of fault features. Each entry fully stores all core content related to the corresponding scenario, including action sequence, instruction type, safety threshold, and phase parameters. During the traversal, the system performs a comprehensive comparison in real time between the fault feature label set and the fault feature parameters bound to the entry, accurately matching target entries with perfectly matching features and automatically filtering redundant entries with mismatched features to prevent misuse of cross-scenario strategies. After accurate matching, the system fully integrates and encapsulates all handling rules and parameter data within the target entry, removing invalid placeholder data and generating a standardized matching strategy record. This record completely retains all self-healing action logic adapted to this fault scenario, with no missing rules or parameters, providing a complete rule basis for subsequent action instruction parsing.

[0076] Parse the action logic sequence from the matched strategy record, including the reclosing sequence corresponding to transient faults, the tripping and blocking instructions corresponding to permanent faults, and the phase angle parameters of phase-controlled tripping and closing, to generate the original action instruction set; The core of this step is to deeply disassemble the closed-loop self-healing action logic in the matching strategy record, distinguish the differentiated handling logic between transient and permanent faults, accurately extract the three core parameters of timing control, instruction type, and phase modulation, and integrate and encapsulate them into a raw action instruction set that has not undergone safety verification. This provides raw data support for subsequent instruction safety verification and hardware signal conversion. The specific implementation method is as follows: The action logic sequence stored in the matching strategy record is a complete closed-loop handling process designed for specific fault scenarios. Based on the duration characteristics of power distribution line faults, faults are divided into two main categories: transient faults and permanent faults. These two types of faults correspond to completely different self-healing handling logics, adaptable to the fault handling scenarios of most outdoor pole-mounted lines. Transient faults are short-lived faults without substantial damage to line equipment, often caused by lightning flashovers, air arcing, or momentary contact with foreign objects. These faults have a short duration, and the fault location can self-recover insulation, requiring no long-term power outage isolation. The core handling logic is delayed reclosing after fault isolation to quickly restore normal power supply and ensure power reliability. Permanent faults are persistent faults with substantial physical damage, often caused by line breaks, insulator breakdown, equipment burnout, or permanent short circuits. These faults cannot be eliminated on their own. The core handling logic is rapid fault isolation followed by blocking reclosing to prevent secondary damage to equipment and expansion of the line fault caused by reverse reclosing.

[0077] Reclosing timing is a core timing control parameter for the self-healing of transient faults. It consists of three core parameters: tripping response delay, reclosing waiting delay, and closing status monitoring delay. All timing parameters are measured in seconds, with an accuracy of up to 0.01 seconds. This precise control of the circuit breaker's operating rhythm prevents self-healing failure due to timing deviations. The tripping response delay is the interval between fault confirmation and the execution of the tripping operation. It is used to stabilize transient fault data and avoid malfunctions caused by transient fluctuations. The standard value is 0.2 seconds. The reclosing waiting delay is the waiting time from when the circuit breaker is tripped to when the reclosing operation is initiated. Its core function is to wait for the line fault arc to completely extinguish and the line insulation to recover. It is a key parameter for successful self-healing of transient faults. The standard value is 0.5 to 1.5 seconds depending on the line operating conditions. In this example, the reclosing waiting delay for the near-end single-phase ground fault is set to 0.7 seconds. The closing status monitoring delay is the duration for which the line status is continuously monitored after the reclosing action is completed. It is used to determine whether the fault has been completely eliminated, and the normal value is 2.0 seconds.

[0078] The tripping lockout command is the core handling command for permanent faults. This command includes three core logics: a fast tripping control flag, a reclosing function lockout flag, and a fault continuous monitoring flag. The fast tripping control flag is used to drive the circuit breaker to perform a tripping action instantaneously, quickly cutting off the power supply to the faulty line and preventing the fault from spreading. The reclosing function lockout flag is used to lock the circuit breaker's reclosing control circuit, prohibiting the equipment from automatically reclosing, thus avoiding the safety risks of closing the circuit with a fault from a hardware logic perspective. The fault continuous monitoring flag is used to trigger the line's normalized monitoring logic, continuously collecting line electrical data, and waiting for remote intervention from the master station.

[0079] The phase angle parameter for phase-controlled opening and closing is a core electrical parameter for achieving refined and low-impact operation of circuit breakers. It is measured in degrees with an accuracy of up to 0.1 degrees. This parameter is designed based on the phase characteristics of power frequency signals. By controlling the circuit breaker to complete the opening and closing actions at the optimal phase nodes when the voltage and current cross zero, it can significantly suppress operational overvoltages and inrush currents, protecting the safety of lines and distribution equipment. Different fault phases correspond to specific optimal phase angles. In this example, the optimal phase angle for phase-controlled closing corresponding to a single-phase fault in phase B is 0 degrees, meaning the circuit breaker completes the closing operation at the moment the phase B voltage crosses zero, achieving impact-free closing.

[0080] The system analyzes all action logic in the strategy record line by line and parameter by parameter, strictly distinguishes the exclusive action rules corresponding to the fault type, accurately extracts the complete set of timing parameters for reclosing, the logic of permanent fault tripping and blocking instructions, and the phase angle parameters of phase control tripping and closing. It integrates the parameters according to the execution order of fault handling, removes duplicate and redundant parameters, and supplements missing logic identifiers. Finally, it generates a complete set of original action instructions with complete parameters and closed logic. This set of instructions contains the original parameters of all executable self-healing actions under the current fault scenario. Only the logic integration is completed, and no safety compliance verification is carried out. It provides the original instruction data source for subsequent steps.

[0081] The original action command set is subjected to safety verification to check whether the command conforms to the current state of the circuit breaker and the energized conditions of the line. After the verification is passed, it is converted into specific executable signals of the circuit breaker operating mechanism, and fault isolation and recovery control commands containing tripping command, reclosing command or blocking command are generated.

[0082] The core of this step is to build a dual safety verification system for equipment status and line operating conditions. This system comprehensively verifies the compliance and on-site safety of the execution of the original action command set, intercepts risky commands that do not conform to the on-site operating conditions, and then converts standardized logic parameters into hardware-recognizable electrical signal commands. Finally, it generates standardized control commands that can directly drive the equipment to execute, ensuring the safe, accurate, and effective implementation of fault self-healing operations. The specific implementation method is as follows: After the original action instruction set is integrated, the system immediately starts the local fully automatic safety verification process. The verification work is divided into two core dimensions: circuit breaker real-time status verification and line live operation condition verification. Only when both verifications are qualified can the instruction conversion stage be entered. If any verification fails, the current original instruction set will be discarded and the strategy matching process will be retried to prevent safety hazards such as equipment malfunction and illegal operation of live lines. The circuit breaker current status verification mainly verifies the real-time operating condition of the equipment, including three core indicators: opening and closing position status, operating mechanism energy storage status, and equipment fault lockout status. The opening and closing position status is divided into three states: closed in place, open in place, and action transition. When the equipment is in the action transition state, the mechanical structure of the mechanism is in motion and cannot respond to new operation commands, so the system directly determines that the command is invalid. The operating mechanism energy storage status is divided into two states: energy storage complete and energy storage incomplete. The circuit breaker's opening and closing actions depend on the energy stored in the mechanism. When energy storage is incomplete, the operating power is insufficient and the standard opening and closing operation cannot be completed, so the command is determined to be unexecutable. The equipment fault lockout status is the lockout state after the equipment's own protection mechanism is triggered. In the lockout state, all external operation commands are prohibited. In the example, the circuit breaker status verified in this case is closed in place, energy storage is fully completed, and there is no fault lockout signal, which meets the basic conditions for command execution.

[0083] The line energization condition verification mainly verifies the real-time electrical parameters of the line after fault handling, including three core parameters: three-phase voltage amplitude, three-phase load current, and residual fault current. All parameters are judged against the local preset safety thresholds. The normal range of three-phase voltage amplitude is 0.9 to 1.1 times the line rated voltage. A voltage below 0.9 times is considered an undervoltage anomaly, and a voltage above 1.1 times is considered an overvoltage anomaly. Closing operations are prohibited under abnormal voltage conditions. The three-phase load current must be stable within the rated load range of the equipment, without overload or sudden changes. The residual fault current must be close to zero to ensure that there is no potential for continuous faults in the line. In this example, the line verification data shows that the three-phase voltage is stable at 1.0 times the rated voltage, the load current is normal without overload, and there is no residual fault current, which fully meets the line operation safety conditions.

[0084] After completing dual safety verifications and ensuring all parameters are compliant, the system initiates the command signal conversion process. This process transforms standardized logical command parameters into standardized electrical execution signals that the device controller can accurately recognize, according to the circuit breaker's operating mechanism's hardware communication protocol. This ensures consistent signal voltage amplitude, duration, and encoding format, guaranteeing zero-error command transmission and parsing. Based on the fault type and verification results, the system generates differentiated final control commands. For transient fault scenarios that pass verification, a reclosing command with precise timing parameters is generated. For permanent fault scenarios that pass verification, a fast tripping command and a long-term reclosing interlocking command are generated. For scenarios with abnormal verification, no valid operation command is generated. After command conversion, the system standardizes and encapsulates the final command, labeling the command type, execution parameters, and verification status. This generates compliant and valid fault isolation and recovery control commands, which can be directly output through the edge computing node interface to drive the circuit breaker to complete a full set of self-healing operations for fault isolation and power restoration.

[0085] S205, execute the fault isolation and recovery control command, drive the operating mechanism of the pole-mounted circuit breaker to complete the opening and closing operation, and collect the line status for self-check after the operation, and generate a self-healing success or report to the main station.

[0086] Specifically, fault isolation and recovery control commands can be output to the circuit breaker's operating mechanism controller through the I / O interface of the edge computing node. After the controller parses the commands, it drives the permanent magnet or spring mechanism to perform opening and closing actions, generating a circuit breaker action execution record. The core of this step is to complete the hardware transmission, parsing, execution, and action process recording of fault control commands. This involves converting the digital control commands generated by the edge computing module into the mechanical actions of the circuit breaker entity, while simultaneously recording the complete action process data to provide the original execution basis for subsequent self-healing result verification. The specific implementation method is as follows: The I / O interface on the edge computing node serves as a dedicated transmission channel connecting the algorithm processing unit and the circuit breaker hardware execution unit. It possesses industrial-grade real-time transmission and anti-interference capabilities, adapting to the complex electromagnetic environment of outdoor power distribution systems and effectively avoiding command transmission distortion caused by line harmonics and electromagnetic radiation. This interface employs a point-to-point real-time transmission mechanism, with command transmission latency consistently controlled within 20ms, ensuring timely fault handling and preventing the fault range from expanding due to excessive delays. The interface supports digital command signal output, accurately identifying three core control signals: tripping commands, reclosing commands, and interlocking commands. It also features a command verification mechanism that automatically identifies incomplete, garbled, and abnormal commands, preventing erroneous triggering.

[0087] The fault isolation and recovery control commands are standardized digital signal commands, with each command corresponding to a unique signal code identifier. Specifically, the tripping command corresponds to the circuit breaker's fault isolation action, the reclosing command corresponds to the line power restoration action for transient faults, and the interlocking command corresponds to the mechanism locking and anti-maloperation action for permanent faults. All commands include timing control parameters and action priority parameters to ensure the orderly execution of actions. In the example, the reclosing command generated for a single-phase ground fault on the line has a built-in action delay parameter of 0.5s. This parameter represents a 0.5s delay after the circuit breaker completes tripping isolation before executing the reclosing action, adapting to the arc-extinguishing recovery characteristics of transient faults in distribution lines.

[0088] The operating mechanism controller is the core control unit for the mechanical actions of the circuit breaker. It is specifically designed to receive control commands transmitted through the I / O interface and perform command decoding, parameter parsing, and action drive scheduling. The controller has a built-in dedicated command parsing algorithm that can accurately extract core information such as action type, execution sequence, and action interlocking authority from the commands. At the same time, it reads the current equipment status parameters of the circuit breaker in real time, including the energy storage status of the mechanism, the initial opening and closing position, and the temperature conditions of the mechanism, to ensure that the commands match the equipment operating status and prevent actions from being executed with faults or abnormalities.

[0089] Circuit breakers are equipped with two main types of drive mechanisms: permanent magnet mechanisms and spring mechanisms. These two mechanisms are suited to different fault operation scenarios, and their response speed and accuracy are subject to standardized parameter specifications. The permanent magnet mechanism offers faster response, with a single opening / closing action completion time consistently between 30ms and 50ms, making it suitable for rapid fault isolation scenarios. The spring mechanism boasts stronger energy storage stability and excellent resistance to outdoor environmental interference, with a single opening / closing action completion time consistently between 60ms and 80ms, making it suitable for permanent fault blocking and conventional reclosing scenarios. In this example, the fault handling utilizes a spring-operated mechanism to perform the opening / isolation action. After receiving the drive signal, the mechanism releases its pre-stored mechanical potential energy, driving the circuit breaker contacts to complete the separation operation. The entire process is smooth and without jamming, meeting the mechanical action standards for power distribution line fault isolation.

[0090] Throughout the entire process of the circuit breaker performing opening and closing actions, the operating mechanism controller collects real-time action status data, including the command reception time, action initiation time, contact movement duration, action completion time, energy consumption of the mechanism, and voltage and current fluctuations during the action. All data is recorded continuously in chronological order, without any breaks or omissions. After the action is completed, the controller integrates and summarizes all chronological data to generate a standardized circuit breaker action execution record. The record includes core information such as a unique action number, command type, execution start and end times, mechanism operating parameters, and action completion status. The data accuracy is uniformly retained to two decimal places, providing complete execution traceability data for subsequent line status self-inspection and self-healing result determination.

[0091] After the circuit breaker operates, the three-phase current and voltage data of the line are immediately collected again through the edge computing node to check whether the line has been restored to normal power supply and whether there is no fault current residue, and generate the line status self-test data after the operation. The core of this step is to complete a second, accurate acquisition of the electrical quantities of the line immediately after the circuit breaker has completed its operation. By comparing these quantities with normal power supply standards, residual faults are identified, and standardized self-test data is generated. This provides accurate and effective data support for determining the self-healing effect. The specific implementation method is as follows: Once the circuit breaker's mechanical action is fully completed and the contact state is stable, the high-frequency acquisition mechanism of the edge computing node is immediately triggered without any additional delay. This ensures that the acquired data accurately reflects the real-time operating status of the line after fault handling, avoiding data distortion caused by secondary changes in the line's state. The edge computing node reuses the line's original high-precision acquisition transformer, maintaining a high-frequency sampling rate of 256 points per cycle. This sampling parameter can completely capture transient and steady-state changes in the line's electrical quantities, accurately identify weak residual fault signals, and eliminate the problem of missed fault features due to low sampling rates.

[0092] The core data collected this time are the full data of three-phase current and three-phase voltage of the line. The three phases are defined as phase A, phase B and phase C, and the corresponding collected parameters are phase A current I_A, phase B current I_B, phase C current I_C, phase A voltage U_A, phase B voltage U_B and phase C voltage U_C. All electrical quantity data are measured in per-unit value, and the value range is uniformly normalized to the range of 0 to 1, which is consistent with the data standard of previous fault feature extraction to ensure the consistency of data comparison.

[0093] The baseline parameters for normal power supply are the rated operating parameters of the distribution line. For a conventional 10kV distribution line, the rated phase voltage per unit is 1.0, and the rated phase current per unit is 1.0. Under normal operating conditions, the deviation of the three-phase voltage and three-phase current values ​​should not exceed ±0.05 per unit. This deviation threshold is the standard for judging the normal steady-state operation of the line. Residual fault current specifically refers to abnormal current signals such as zero-sequence current, harmonic current, and sudden inrush current present in the line. The threshold for residual zero-sequence current is set at 0.02 per unit. When the effective value of the zero-sequence current exceeds this value, it is determined that the line has residual fault and has not returned to normal operating conditions.

[0094] During data acquisition, the edge computing node continuously collects electrical quantity data for at least three complete power frequency cycles. The power frequency cycle is fixed at 0.02s, and the cumulative acquisition time for the three cycles is 0.06s. Multi-cycle data acquisition avoids instantaneous fluctuation interference and improves self-test accuracy. After acquisition, following the preprocessing logic, a median filtering algorithm is used to remove spike pulse interference from the acquisition process. This algorithm effectively filters out abnormal extreme values ​​generated by outdoor electromagnetic interference, retaining the true electrical quantities of the line operation. Simultaneously, the data is normalized and calibrated to unify the data format and accuracy.

[0095] After data preprocessing, the system performs item-by-item checks on the three-phase voltage and three-phase current data. On one hand, it verifies whether the three-phase voltage has recovered to the rated steady-state range, with no voltage drops or deviations. On the other hand, it checks whether the three-phase current is in a balanced steady-state state. Simultaneously, it calculates the effective value of the zero-sequence current and the current harmonic distortion rate to investigate for residual fault current and latent fault characteristics. In the example, after fault handling, electrical quantity data for three power frequency cycles are collected. The per-unit values ​​of the three-phase voltage are 0.98, 0.99, and 0.98, respectively; the per-unit values ​​of the three-phase current are 0.97, 0.98, and 0.97, respectively; and the effective value of the zero-sequence current is 0.01 per-unit, below the residual threshold, indicating a preliminary determination that there is no residual fault in the line.

[0096] All collected, preprocessed, and tested electrical quantity data, combined with intermediate test results, are integrated according to a standardized data structure to generate post-operation line status self-test data. The data fully records all information such as acquisition time, sampling frequency, electrical quantity values ​​of each phase, filtering preprocessing parameters, and fault residue detection results. The data storage precision is uniformly retained to four decimal places to ensure the integrity and accuracy of the self-test data.

[0097] The success of self-healing is determined based on the self-test data of the line status after the operation. If the line voltage recovers and the current is balanced, the self-healing is considered successful. If the fault characteristics still exist, the self-healing is considered to have failed. A judgment flag for success or failure of self-healing is generated. The core of this step is to rely on standardized line self-inspection data, combined with the normal operation judgment criteria of power distribution lines, to accurately determine the self-healing effect of faults and generate a unique and identifiable status judgment mark, providing a core judgment basis for the generation of subsequent handling reports. The specific implementation method is as follows: The core criteria for judging the self-healing effect are the line voltage recovery state and the three-phase current balance state. Both conditions must be met simultaneously for self-healing to be considered successful; failure to meet either condition results in self-healing failure. The judgment logic strictly adheres to industry operating standards for distribution network fault self-healing. The judgment process relies on steady-state electrical quantity parameters from the self-test data, discarding instantaneous fluctuation data to ensure the stability and accuracy of the judgment results.

[0098] The line voltage recovery determination has clear quantitative standards. Using the rated phase voltage per unit value of 1.0 as the benchmark, the acceptable voltage recovery range is set between 0.95 and 1.05 per unit. When all three phase voltage values ​​are stable within this range, and there are no abnormal phenomena such as continuous voltage drops, voltage spikes, or three-phase voltage imbalances, the line voltage is considered to have recovered normally. If any phase voltage is lower than 0.95 per unit or higher than 1.05 per unit, and the duration exceeds one power frequency cycle (0.02s), the voltage is considered not to have recovered, and the line still has a fault. In the example, the self-test data shows that all three phase voltage values ​​are within the acceptable range of 0.95 to 1.05 per unit, and there are no abnormal fluctuations for three consecutive power frequency cycles, thus meeting the voltage recovery determination conditions.

[0099] Three-phase current balance determination uses a three-phase current imbalance algorithm for quantitative judgment. The current imbalance calculation formula is the ratio of the maximum difference between the three-phase currents to the rated current. The industry-standard acceptable threshold is set at 3%. When the three-phase current imbalance is ≤3%, the three-phase currents are considered to be in a balanced state; when the three-phase current imbalance is >3%, the current is considered unbalanced, and the line is in an abnormal operating state. Simultaneously, combined with fault residual detection results, if the self-inspection data shows fault characteristics such as excessive zero-sequence current, excessively high harmonic distortion rate, or sudden current fluctuations, regardless of whether the voltage and current balance meet the standards, the line is determined to have an unresolved fault problem. In the example, the maximum difference between the three-phase currents is 0.01 per unit, and the rated current is 1.0 per unit. The calculated current imbalance is 1%, which is less than the 3% threshold, meeting the current balance judgment condition, and there are no residual fault characteristics.

[0100] When the line voltage fully recovers to the acceptable range, the three-phase current remains balanced, and no fault characteristics remain, the system determines that the self-healing process has been successfully executed and generates a self-healing success flag. This flag is a standardized digital identifier used by the system backend to quickly identify the handling result. When the line voltage fails to recover, the three-phase current is unbalanced, or any fault characteristics are detected, the system determines that the self-healing operation cannot eliminate the line fault, the self-healing process fails, and a self-healing failure flag is generated.

[0101] After the judgment flag is generated, the system will bind and associate the flag with the corresponding self-inspection data and action execution records to ensure that each judgment flag can be traced back to the complete line operation data and equipment operation data, avoid the flag from being misaligned with the actual operating status, ensure the authenticity and validity of the judgment results, and provide accurate status identifiers for the classification and generation of subsequent handling reports.

[0102] A handling report is generated based on the judgment flags. When self-healing is successful, the action time, number of actions, and recovery time are recorded and stored locally. When self-healing fails, the fault type, operation records, and self-test data are packaged and uploaded to the main station to request remote intervention. Finally, a handling report is generated to indicate whether self-healing was successful or reported to the main station.

[0103] The core of this step is to classify and execute data archiving or remote reporting processes based on the status flags of the self-healing determination, integrate the fault handling data of the entire process to generate a standardized handling report, and complete the local fault self-healing closed loop or trigger the main station's remote handling mechanism. The specific implementation method is as follows: The system identifies self-healing judgment indicators based on fixed logical rules, distinguishes between two handling scenarios: successful self-healing and failed self-healing, and executes differentiated report generation and data processing processes. The data organization, storage, and reporting rules for both scenarios are standardized and regulated to ensure the integrity of closed-loop management of fault handling.

[0104] For scenarios where self-healing is successful, the system automatically extracts core timing and operational data from the entire process, records key parameters, and stores them locally. The core recorded parameters include three categories of key indicators: action time, number of actions, and recovery time. Each parameter has a clear definition and measurement standard. Action time is the moment the circuit breaker first executes the fault isolation command, recorded using a millisecond-level timestamp to accurately mark the fault handling initiation time. The number of actions is the total number of opening and reclosing operations performed by the circuit breaker during this fault self-healing process; each effective mechanical action is counted as one, accurately reflecting the operational complexity of fault handling. Recovery time is the total time from the initiation of the fault isolation action to the complete restoration of the line's electrical quantities to normal steady-state operation, measured in milliseconds; this time value directly reflects the efficiency of the self-healing process.

[0105] In the example, in this successful self-healing scenario, the timestamp of the action time is 20260520102536452, representing that the fault action was initiated at 10:25:36:452 on May 20, 2026; the number of actions was 2, namely the fault tripping isolation action and the fault reclosing recovery action; the recovery time was 1200ms, that is, it took 1.2 seconds from the start of the action to the line fully restoring normal power supply. After all parameters are extracted, the edge computing node integrates the data with the action execution record, self-test data, and fault judgment results to generate a self-healing success handling report. The report fully records the entire fault handling process information and is then stored in the local solid-state storage unit of the edge computing node. The local storage data retention period is 90 days, which is convenient for subsequent operation and maintenance traceability and data analysis.

[0106] In response to scenarios where self-healing fails, the system automatically triggers a data packaging and main station reporting mechanism. First, it extracts the core characteristics and handling data of the fault, including three types of core data: the fault type determined in the early stage, the complete circuit breaker action execution record, and the line status self-inspection data after the operation. The fault type is clearly marked as one of the four specific types: single-phase grounding, two-phase short circuit, two-phase grounding, and three-phase short circuit. The action execution record includes the timing of all mechanical actions and equipment operating parameters. The self-inspection data includes the full-dimensional electrical quantity detection results and the fault residue judgment results.

[0107] The data packaging process employs a standardized encapsulation format, integrating and compressing the three types of data in the order of fault basic information, equipment operation information, and line status information. Invalid and redundant data is eliminated, while all valid core information is retained. The compressed data ensures efficient and complete remote transmission. After packaging, the edge computing node uploads the data packet to the power distribution master station system via a dedicated power distribution communication channel. The upload process includes breakpoint resumption and data verification mechanisms to prevent data loss and garbled characters. Simultaneously, a remote intervention request signal is sent to the upper-level master station, prompting master station maintenance personnel to intervene and handle any remaining line faults.

[0108] After data upload is complete, the system integrates and packages the data, upload logs, and remote intervention request status to generate a handling report for self-healing failure reported to the main station. The report clearly indicates the reason for self-healing failure, the entire handling process data, the reporting time, and the intervention request status, fully recording the entire fault handling process. Finally, through differentiated data processing logic, a self-healing success handling report and a main station reporting handling report are generated separately, fully realizing the closed-loop handling process for local fault self-healing of pole-mounted circuit breakers.

[0109] Another embodiment of the present invention provides a local fault self-healing system for pole-mounted circuit breakers based on edge computing, see [link to relevant documentation]. Figure 3 The system may include: The acquisition module 301 is used to acquire the three-phase current, voltage and circuit breaker status data of the line in real time through the edge computing node deployed on the side of the pole-mounted circuit breaker, and to extract fault feature quantities using a sliding time window. The identification module 302 is used to identify the fault type and locate the fault segment based on the fault feature quantity using a lightweight convolutional neural network model, and generate a fault determination result including fault type, fault distance and fault phase. Evaluation module 303 is used to input the fault determination result into a locally deployed adversarial verification network, evaluate the credibility of the fault determination result through the game between the generator and the discriminator, and output a valid fault result that has been verified. The matching module 304 is used to call the locally stored fault self-healing strategy library according to the valid fault result, match the action logic corresponding to the current fault type, and generate fault isolation and recovery control instructions including tripping instructions, reclosing instructions or blocking instructions. The self-healing module 305 is used to execute the fault isolation and recovery control command, drive the operating mechanism of the pole-mounted circuit breaker to complete the opening and closing operation, and collect the line status for self-checking after the operation, and generate a self-healing success or reporting to the main station.

[0110] This invention also provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above method embodiments when running.

[0111] This invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0112] Specifically, the aforementioned electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the aforementioned processor, and the input / output device is connected to the aforementioned processor.

[0113] The above description, based on the embodiments shown in the figures, details the structure, features, and effects of the present invention. The above description is only a preferred embodiment of the present invention, but the present invention is not limited to the scope of implementation shown in the figures. Any changes made in accordance with the concept of the present invention, or equivalent embodiments modified to have equivalent changes, that do not exceed the spirit covered by the specification and figures, should be within the protection scope of the present invention.

Claims

1. A method for local fault self-healing of pole-mounted circuit breakers based on edge computing, characterized in that, The method includes: The three-phase current, voltage and circuit breaker status data of the line are collected in real time by deploying computing nodes on the side edge of the pole-mounted circuit breaker, and fault characteristic quantities are extracted by using a sliding time window. Based on the aforementioned fault features, a lightweight convolutional neural network model is used for fault type identification and fault segment location, generating a fault determination result that includes fault type, fault distance, and fault phase. Specifically, the fault feature vector is input into the input layer of the lightweight convolutional neural network model, which employs a depthwise separable convolutional structure to reduce the number of parameters. Deep fault feature maps are extracted through convolution and pooling operations, generating a deep fault feature tensor. This deep fault feature tensor is then input into both the fault type classification branch and the fault location branch. The classification branch outputs probability distributions for four types of faults—single-phase grounding, two-phase short circuit, two-phase grounding, and three-phase short circuit—through a fully connected layer, generating a preliminary fault type determination result. In the location branch, the deep fault feature tensor is convolved with the line parameter matrix, and the distance from the fault point to the measurement point is calculated using the traveling wave method, outputting a fault distance estimate and generating a fault segment location result. Combining the preliminary fault type determination result, the fault segment location result, and the three-phase current change phase information, a majority voting fusion rule is used to determine the final fault phase, generating a fault determination result that includes fault type, fault distance, and fault phase. The fault determination result is input into the locally deployed adversarial verification network. The credibility of the fault determination result is evaluated through the game between the generator and the discriminator, and the valid fault result after verification is output. Based on the valid fault results, the locally stored fault self-healing strategy library is invoked to match the action logic corresponding to the current fault type and generate fault isolation and recovery control instructions containing tripping instructions, reclosing instructions, or blocking instructions. Execute the fault isolation and recovery control command, drive the operating mechanism of the pole-mounted circuit breaker to complete the opening and closing operation, and collect the line status for self-check after the operation, and generate a self-healing success or report to the main station.

2. The method according to claim 1, characterized in that, The method involves real-time acquisition of three-phase current, voltage, and circuit breaker status data via computational nodes deployed on the edge of the pole-mounted circuit breaker, and extraction of fault characteristic quantities using a sliding time window, including: Edge computing nodes are deployed on the pole-mounted circuit breaker side. The three-phase current, three-phase voltage, circuit breaker opening and closing position and energy storage status data of the line are collected in real time through the current transformer at a sampling rate of 256 points per cycle, generating the original high-frequency time-series data stream. The original high-frequency time-series data stream is denoised and normalized. Median filtering is used to remove spike pulse interference, and the current and voltage amplitudes are normalized to the range of 0 to 1 to generate preprocessed standard data frames. Based on the preprocessed standard data frame, a fixed-length sliding time window is set, which covers the complete transient process from the first half of the fault occurrence to the last five cycles after the fault occurrence. Data segments are truncated by sliding half a cycle in steps to generate a sliding window data block sequence. For each data block in the sliding window data block sequence, time-domain and frequency-domain features are extracted, and the amplitude of the sudden change in three-phase current, harmonic distortion rate, voltage drop depth, and effective value of zero-sequence current are calculated to generate a fault feature vector.

3. The method according to claim 2, characterized in that, The step of inputting the fault determination result into a locally deployed adversarial verification network, evaluating the credibility of the fault determination result through a game between the generator and the discriminator, and outputting a verified and confirmed valid fault result includes: The fault determination result is encoded into a condition vector, which is then concatenated with the original fault feature vector and input into the generator of the adversarial verification network. The generator attempts to generate a pseudo feature vector that is consistent with the distribution of real fault features, thereby generating fitted pattern data. The fitted pattern data output by the generator is input into the discriminator along with samples from the historical real fault sample library. The discriminator calculates the probability score of each sample belonging to a real fault and outputs the discrimination confidence score. The confidence score is compared with the confidence threshold. If the score is higher than the threshold, the fault determination result is determined to be credible. Otherwise, it is determined to be unreliable and needs to be retested to generate a credibility assessment conclusion. For fault determination results that are determined to be reliable, the results are directly output as valid fault results. For results that are determined to be unreliable, secondary sampling is triggered to re-execute the fault determination process, and finally, a valid fault result that has been verified and confirmed is output.

4. The method according to claim 3, characterized in that, Based on the valid fault result, the system calls the locally stored fault self-healing strategy library, matches the action logic corresponding to the current fault type, and generates fault isolation and recovery control instructions containing tripping instructions, reclosing instructions, or blocking instructions, including: The fault type, fault distance, and fault phase information in the valid fault results are analyzed, and key fields for strategy matching are extracted to generate a fault feature tag set. Using the fault feature tag set as an index, query the locally stored fault self-healing strategy library. This strategy library is organized in a decision tree structure. Traverse and match the action logic entries corresponding to the current fault type to generate matching strategy records. Parse the action logic sequence from the matched strategy record, including the reclosing sequence corresponding to transient faults, the tripping and blocking instructions corresponding to permanent faults, and the phase angle parameters of phase-controlled tripping and closing, to generate the original action instruction set; The original action command set is subjected to safety verification to check whether the command conforms to the current state of the circuit breaker and the energized conditions of the line. After the verification is passed, it is converted into specific executable signals of the circuit breaker operating mechanism, and fault isolation and recovery control commands containing tripping command, reclosing command or blocking command are generated.

5. The method according to claim 4, characterized in that, The execution of the fault isolation and recovery control command drives the operating mechanism of the pole-mounted circuit breaker to complete the opening and closing operations. After the operation, the line status is collected for self-checking, and a self-healing success or a handling report to be sent to the main station is generated, including: The fault isolation and recovery control commands are output to the circuit breaker's operating mechanism controller through the I / O interface of the edge computing node. After parsing the commands, the controller drives the permanent magnet or spring mechanism to perform opening and closing actions, and generates a circuit breaker action execution record. After the circuit breaker operates, the three-phase current and voltage data of the line are immediately collected again through the edge computing node to check whether the line has been restored to normal power supply and whether there is no fault current residue, and generate the line status self-test data after the operation. The success of self-healing is determined based on the self-test data of the line status after the operation. If the line voltage recovers and the current is balanced, the self-healing is considered successful. If the fault characteristics still exist, the self-healing is considered to have failed. A judgment flag for success or failure of self-healing is generated. A handling report is generated based on the judgment flags. When self-healing is successful, the action time, number of actions, and recovery time are recorded and stored locally. When self-healing fails, the fault type, operation records, and self-test data are packaged and uploaded to the main station to request remote intervention. Finally, a handling report is generated to indicate whether self-healing was successful or reported to the main station.

6. A local fault self-healing system for pole-mounted circuit breakers based on edge computing, characterized in that, The system includes: The data acquisition module is used to collect three-phase current, voltage and circuit breaker status data in real time through the edge computing nodes deployed on the side of the pole-mounted circuit breaker, and to extract fault feature quantities using a sliding time window. The identification module is used to identify fault types and locate fault segments based on the fault feature quantities using a lightweight convolutional neural network model, generating a fault determination result including fault type, fault distance, and fault phase. Specifically, the fault feature vector is input into the input layer of the lightweight convolutional neural network model, which employs a depthwise separable convolutional structure to reduce the number of parameters. Deep fault feature maps are extracted through convolution and pooling operations to generate a deep fault feature tensor. This deep fault feature tensor is then input into the fault type classification branch and the fault ranging branch, respectively. The classification branch outputs data through a fully connected layer. The probability distributions of four fault types—single-phase grounding, two-phase short circuit, two-phase grounding, and three-phase short circuit—are used to generate a preliminary fault type determination result. In the ranging branch, the deep fault feature tensor is convolved with the line parameter matrix, and the distance from the fault point to the measurement point is calculated using the traveling wave method. The estimated fault distance is output, generating the fault segment location result. Combining the preliminary fault type determination result, the fault segment location result, and the phase information of the three-phase current change, the final fault phase is determined using a majority voting fusion rule, generating a fault determination result that includes the fault type, fault distance, and fault phase. The evaluation module is used to input the fault determination result into the locally deployed adversarial verification network, evaluate the credibility of the fault determination result through the game between the generator and the discriminator, and output the valid fault result that has been verified. The matching module is used to call the locally stored fault self-healing strategy library according to the valid fault result, match the action logic corresponding to the current fault type, and generate fault isolation and recovery control instructions including tripping instructions, reclosing instructions, or blocking instructions. The self-healing module is used to execute the fault isolation and recovery control commands, drive the operating mechanism of the pole-mounted circuit breaker to complete the opening and closing operations, and collect the line status for self-checking after the operation, and generate a self-healing success or reporting to the main station.

7. The system according to claim 6, characterized in that, The acquisition module is specifically used for: Edge computing nodes are deployed on the pole-mounted circuit breaker side. The three-phase current, three-phase voltage, circuit breaker opening and closing position and energy storage status data of the line are collected in real time through the current transformer at a sampling rate of 256 points per cycle, generating the original high-frequency time-series data stream. The original high-frequency time-series data stream is denoised and normalized. Median filtering is used to remove spike pulse interference, and the current and voltage amplitudes are normalized to the range of 0 to 1 to generate preprocessed standard data frames. Based on the preprocessed standard data frame, a fixed-length sliding time window is set, which covers the complete transient process from the first half of the fault occurrence to the last five cycles after the fault occurrence. Data segments are truncated by sliding half a cycle in steps to generate a sliding window data block sequence. For each data block in the sliding window data block sequence, time-domain and frequency-domain features are extracted, and the amplitude of the sudden change in three-phase current, harmonic distortion rate, voltage drop depth, and effective value of zero-sequence current are calculated to generate a fault feature vector.

8. A storage medium, characterized in that, The storage medium stores a computer program, wherein the computer program is configured to execute the method of any one of claims 1-5 when it is run.

9. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method of any one of claims 1-5.

Images