Deep learning based precise detection technology for fault arc

Abstract

The application discloses a kind of based on deep learning's fault arc accurate detection method and system, belong to low-voltage distribution safety and electrical fire early warning technical field.The present application is aimed at the problem that fault arc is prone to occur, difficult to detect in low-voltage distribution system, and the high false detection and missed detection rate of traditional method, a high-precision, fast detection scheme of fusion current time-frequency domain feature and deep learning model is proposed.The present application quickly locates current disturbance by improving CUSUM algorithm, greatly reduces the amount of calculation;Automatic extraction current time-frequency domain joint feature, input Stacking double-layer integrated deep learning model to realize high-precision fault classification;Combined with double exponential arc model to expand samples, improve generalization ability.The detection accuracy of the present application reaches 99.06%, and the detection time is only 10% of the traditional method, can effectively identify series, parallel, ground fault arc, adapt to residential, commercial, industrial and other complex multi-load scenarios, and support multi-branch line fault positioning, with high precision, low delay, strong robustness, easy deployment and other advantages, can be widely used in arc fault circuit interrupter, electrical fire monitoring system, to prevent electrical fire from the source.

Description

Technical Field

[0001] This invention belongs to the fields of low-voltage power distribution safety, electrical fire early warning, intelligent monitoring and protection technology. Specifically, it relates to a method and system for fault arc detection, identification, location, and linkage protection based on deep learning, time-frequency domain feature fusion, and improved cumulative sum algorithms. This invention is applicable to 220V / 50Hz low-voltage AC power distribution systems and can be widely used in residential buildings, commercial complexes, industrial and mining enterprises, rail transit, underground utility tunnels, and renovation of old residential areas. It can perform early, rapid, and accurate detection of series arc faults, parallel arc faults, and grounding arc faults. It can be directly integrated into arc fault circuit interrupters (AFDD / AFCI), electrical fire monitoring systems, intelligent power safety management systems, low-voltage power distribution terminals, and other devices to prevent electrical fires at the source. Background Technology

[0002] With the acceleration of urbanization, the continuous increase in residential electricity load, and the continuous expansion of the low-voltage distribution network in my country, a large number of distribution lines have been operating at full or overload for a long time. Problems such as aging insulation, loose wiring, poor contact, mechanical crushing damage, overheating, moisture corrosion, insect and rodent damage, and non-standard construction are becoming increasingly prominent. The resulting intermittent, low-current, high-impedance, and highly random fault arcs have become the most important, dangerous, and difficult-to-detect core cause of electrical fires.

[0003] Once a fault arc occurs, its local temperature can instantly reach over 3000℃, enough to melt copper and aluminum wires, ignite insulation layers, and surrounding flammable materials. Furthermore, the arc can ignite irreversible combustion in just a few hundred milliseconds. According to fire statistics from the National Fire and Rescue Bureau over the past decade, electrical fires consistently account for more than 30% of all fires. In major fires resulting in mass casualties, over 60% of accidents are directly caused by electrical arcing, causing enormous loss of life and incalculable property damage. The State Council's Safety Production Committee has repeatedly emphasized the need to strengthen the management of electrical fire sources and enhance the research and application of early detection and warning technologies for fault arcs. However, in existing technologies and engineering applications, fault arc detection still faces a series of insurmountable technical bottlenecks and safety shortcomings:

[0004] 1. Traditional protective devices are completely unable to detect arc faults: Traditional low-voltage power distribution protective devices, such as miniature circuit breakers (MCBs), residual current devices (RCDs / RCCBs), molded case circuit breakers, and fuses, are all based on overcurrent, short-circuit, and residual current triggering principles, and are only effective against high-current faults. However, arc faults, especially series arcs, have characteristics such as small fault current, high circuit impedance, and weak characteristics. Their current is often close to or even smaller than the normal load current. Traditional protection devices are in a state of non-operation, non-response, and missed detection failure for a long time, and cannot provide any protection.

[0005] 2. Conventional electrical quantity detection methods have poor adaptability and serious false alarms and missed alarms. Traditional detection methods mostly rely on fixed threshold criteria such as current amplitude, harmonic content, abrupt change slope, and zero rest interval. They can barely work under purely resistive loads, but in real power environments where a large number of nonlinear loads, inductive loads, and mixed loads coexist, such as motors, frequency converters, LED drivers, switching power supplies, dimming equipment, and frequency conversion home appliances, the load current itself has serious fluctuations and harmonic interference, which leads to frequent false alarms and a large number of missed alarms by fixed threshold methods, and cannot meet the requirements of engineering use.

[0006] 3. Traditional machine learning relies on manually designed features, resulting in severely insufficient generalization ability. Recent machine learning-based arc detection methods rely on manually designed time-domain and frequency-domain features, and on expert experience to extract feature components such as wavelet entropy, harmonic ratio, and correlation coefficient. These methods have limited feature representation capabilities, poor model convergence, and weak cross-scenario transfer capabilities. They are prone to feature drift and recognition failure under different line lengths, load types, and fault locations, making stable operation in complex environments difficult.

[0007] 4. Classical arc models are not suitable for low-voltage, low-current scenarios. The lack of samples is a problem. Traditional arc models such as the Cassie model and the Mayr model are derived from the high-current breaking arc of high-voltage circuit breakers. Their physical assumptions and parameter ranges are fundamentally different from the low-current, free-burning, and air-borne arcs in low-voltage distribution lines. The simulated waveforms deviate greatly from the real arcs and have low similarity. It is impossible to generate high-quality, high-coverage training samples, resulting in insufficient training and poor generalization of the intelligent model.

[0008] 5. The high detection delay caused by full-time and full-data calculations cannot meet the requirements of real-time protection. Most existing algorithms calculate the continuous current signal cycle by cycle, resulting in large data volume, high computational redundancy, and long processing time. However, the Arc Fault Circuit Breaker (AFDD) standard requires rapid action within hundreds of milliseconds. Traditional methods cannot meet the real-time requirements and have obvious response lag.

[0009] 6. Lack of Detection and Location Capabilities in Complex Multi-Branch and Multi-Load Scenarios: Real-world low-voltage power distribution systems often employ trunk-type or radial multi-branch structures, with mixed load types and multiple devices operating simultaneously. Faults can occur on the main bus or in any branch. Existing technologies mostly only support single-load, single-line test scenarios, lacking the ability to identify and locate faults in multiple branches. After a fault occurs, the faulty branch cannot be quickly located, impacting power supply reliability and fault diagnosis efficiency.

[0010] 7. Insufficient sensitivity in detecting weak and early-stage electric arcs: The energy of an electric arc is weak, its characteristics are subtle, its duration is short, and it is highly intermittent in the early stages. It is easily submerged in load noise and background fluctuations. Existing methods are not sensitive enough and have poor early identification capabilities. They often only respond when the arc energy increases and there is already a risk of fire, thus losing their early warning significance.

[0011] In summary, existing fault arc detection technologies cannot simultaneously meet the engineering requirements of high precision, low latency, strong anti-interference, full-scenario adaptability, weak arc identification, and multi-branch positioning. The market urgently needs a new generation of deep learning detection technology based on artificial intelligence that can automatically learn fault characteristics, quickly locate disturbances, accurately identify arcs, and adapt to complex loads. Summary of the Invention

[0012] The deep learning-based arc fault detection system described in this invention adopts an integrated architecture of real-time edge sensing, rapid front-end preprocessing, deep learning intelligent recognition, and linked protection execution. The overall hardware and software are co-designed, featuring high-precision data acquisition, low-latency computation, strong anti-interference capabilities, high-reliability output, and multi-scenario compatibility. It can operate independently or be embedded in existing arc fault circuit interrupters (AFDDs), electrical fire monitoring equipment, and smart electricity monitoring terminals to form an integrated product. The system structure is as follows: Figure 1 As shown, the various units work together through electrical isolation, synchronous sampling, high-speed data exchange, and multi-level fault-tolerant logic, as detailed below:

[0013] Current acquisition and signal conditioning unit:

[0014] This unit serves as the sensing front-end of the entire system, responsible for non-contact, high-precision, and low-distortion acquisition of line current signals. It includes a high-precision Hall current sensor employing a 0.1-class wide-range closed-loop Hall sensor with a measurement range of 0-100A. This sensor features electrical isolation, low temperature drift, wide frequency response, and short response time, completely preserving the high-frequency arc characteristics of the current without waveform distortion or phase delay. The signal conditioning circuit includes differential amplification, anti-aliasing low-pass filtering, 50Hz power frequency notch filtering, DC bias adjustment, and overvoltage and overcurrent protection, suppressing grid harmonics, electromagnetic interference, and spike pulse noise to ensure clean, stable, and distortion-free signals sent to the acquisition card. The synchronous sampling circuit uses a multi-channel synchronous sampling topology, supporting synchronous acquisition of bus current and multiple branch currents with a time synchronization error of less than 1μs, providing a foundation for subsequent multi-branch fault location.

[0015] High-speed data acquisition and storage unit:

[0016] This unit is responsible for converting analog current signals into digital signals and providing temporary buffering. It includes a 16-bit high-precision synchronous AD acquisition chip with a single-channel maximum sampling rate of 40kHz, supporting multi-channel parallel sampling to meet the requirements for acquiring high-frequency characteristics of electric arcs. A hardware clock and cycle marking module uses the 50Hz power frequency zero-crossing point as a reference to accurately mark the current cycle, ensuring accurate data segmentation for each cycle. High-speed cache and local storage: Built-in RAM caches the most recent 3-5 seconds of current waveform in real time. In the event of a fault, it automatically latches the complete waveform of 2 seconds before the fault plus 3 seconds after the fault for fault tracing and remote uploading.

[0017] Edge preprocessing and disturbance localization unit:

[0018] This unit is the core of the system's speed-up and noise reduction, running the improved CUSUM algorithm to achieve rapid hardware-level disturbance localization. It includes: an embedded MCU / DSP hardware acceleration unit responsible for real-time current cycle segmentation, RMS value calculation, normalization, and sliding window (sliding window displacement example diagram shown in the figure). Figure 2 (As shown) Statistics, offset cumulative sum calculation, standard deviation and kurtosis (current-standard deviation-kurtosis sequence plot as shown) Figure 3 (As shown) The calculation is performed. The dual sliding window disturbance detection module uses W1 as the stable reference window and W2 as the real-time detection window to dynamically track the effective value of the current, identifying disturbances and locating the starting point in milliseconds. The data filtering and pruning module automatically filters out stable and abnormal current cycles, retaining only the disturbance starting point and several subsequent cycles for deep learning inference, reducing the amount of subsequent calculation data to 10% of the original. The electromagnetic compatibility (EMC) protection module adopts multi-level anti-interference design including opto-isolation, magnetic isolation, and power supply filtering to ensure stable operation in complex electromagnetic environments such as industrial, commercial, and residential environments.

[0019] Deep learning intelligent detection and fault identification unit:

[0020] This unit is the core decision-making layer of the system, responsible for time-frequency domain feature extraction and Stacking ensemble model inference. It includes: an automatic feature extraction accelerator, and hardware-accelerated computation of time-domain features (zero rest time, peak-to-peak value, RMS value, standard deviation, kurtosis) and frequency-domain features (harmonic amplitude, spectral centroid, spectral standard deviation). The Stacking ensemble deep learning inference engine incorporates a pre-trained two-layer heterogeneous ensemble model. An algorithm diagram is shown below. Figure 4 As shown, the base learners are Decision Tree (DT), K-Nearest Neighbors (KNN), and Support Vector Machine (SVM); the meta-learner is a Logistic Regression (LR) model deployed on an embedded device with fixed-point quantization and a lightweight structure, independent of the cloud, with an inference time of <10 ms. The fault confidence assessment module outputs the fault probability and confidence level, supports multi-level threshold adjustment, and allows for flexible sensitivity settings according to the scenario. The main algorithms of the decision tree include ID3, C4.5, and CART, and the corresponding information entropy and information gain calculation formulas are as follows: , , , Represents the current set Information entropy express The percentage of category elements in collection X express Category elements in a collection The proportion in Represents conditional entropy. This represents the conditional entropy. The core of the KNN algorithm lies in distance calculation. Commonly used distance metrics include Manhattan distance, Euclidean distance, and Minkowski distance, as shown in the formula: ,when The formula for representing Manhattan distance is as follows: The formula for Euclidean distance is given by time. The formula represents the Minkowski distance for any other value.

[0021] Fault decision-making, protection output and human-machine interface unit:

[0022] This unit is responsible for final fault determination, tripping execution, local notification, and remote uploading. It includes: a multi-cycle fault-tolerant decision module employing a continuous N-cycle (N ≥ 2) consistent decision mechanism to avoid false triggering caused by transient interference, lightning surges, or switch actions. The relay / optocoupler isolation output unit provides passive normally open / normally closed contacts to directly drive AFDD circuit breaker tripping, external alarms, and fire linkage modules. The local audible and visual indicator module displays the following: Normal: solid green light; Fault: flashing red light + buzzer alarm; Communication: flashing yellow light. The remote communication unit supports multiple interfaces such as UART, RS485, CAN, WiFi, and LoRa, and can upload: fault type (series / parallel / grounding arc); fault time; fault branch / location; current waveforms before and after the fault; current RMS value, harmonics, and disturbance intensity.

[0023] Power supply and system monitoring unit:

[0024] To further enhance system reliability and engineering practicality, this system also includes: a wide-range auxiliary power supply circuit supporting AC / DC 85V~265V wide voltage input to adapt to different field power supply conditions; over-temperature, over-voltage, and under-voltage monitoring to monitor the core chip temperature and power supply voltage in real time, automatically entering protection mode in case of abnormalities; and a hardware watchdog and operation self-test system with power-on self-test, periodic self-test, and fault self-reset to ensure long-term stable operation without disconnection.

[0025] The working principle of this invention lies in fully utilizing the inherent nonlinearity, intermittency, time-varying nature, and zero-rest characteristics of low-voltage fault arcs, combined with signal processing technology and deep learning methods, to achieve full-process capture and accurate identification of fault arcs from "weak early disturbances" to "clear fault characteristics." The system continuously acquires line current waveforms through a current acquisition module. Utilizing the physical law that a fault arc inevitably causes current disturbances, the system improves the CUSUM algorithm to quickly locate abnormal sections, filtering out invalid data under stable operating conditions, significantly reducing system computation and detection latency. Based on this, it automatically extracts joint time-domain and frequency-domain features from suspected disturbance sections, quantifying and enhancing essential characteristics such as the zero-rest phenomenon of the arc, amplitude fluctuations, waveform distortion, high-frequency harmonic growth, and spectral centroid shift. Subsequently, the multi-dimensional features are input into a lightweight Stacking integrated deep learning model, utilizing multi-learner collaborative decision-making to distinguish between normal load fluctuations and fault arc characteristics, achieving highly robust and accurate intelligent identification. Finally, a multi-cycle fault-tolerant decision mechanism confirms the fault, driving the protection equipment to quickly trip, completing the early warning and protection process. Overall, this invention is based on the physical and electrical characteristics of fault arcs, improves detection speed through rapid perturbation screening, enhances detection sensitivity through time-frequency domain features, and improves detection accuracy and generalization ability through deep learning, thereby achieving "fast, accurate, stable, and reliable" fault arc detection. It fundamentally solves the technical problems of traditional methods, such as easy false detection, difficulty in missing detection, poor real-time performance, and weak adaptability.

[0026] The beneficial effects of this invention are: significantly improving the accuracy and reliability of arc fault detection, effectively reducing false detection and missed detection rates; significantly accelerating detection speed, reducing computation time, and meeting the requirements of high-speed real-time protection; strong anti-interference capability, and stable adaptation to various complex loads and multi-scenario operating conditions; good generalization ability, allowing direct deployment without on-site retraining; accurate identification of early weak arcs, preventing electrical fires at the source; support for multiple types of arcs (series, parallel, and grounding) and multi-branch fault location, with a wider range of applications; lightweight, low-power, and easily embedded in various protection devices, with strong engineering practicality, comprehensively improving the safety protection level, power supply reliability, and intelligent operation and maintenance capabilities of low-voltage power distribution systems, while reducing equipment costs and modification difficulty, possessing outstanding social safety benefits and promotional value, effectively making up for the shortcomings of existing arc detection technologies in terms of accuracy, speed, adaptability, and practicality, and providing a more reliable, efficient, and stable intelligent solution for the source prevention and control of electrical fires. Attached Figure Description

[0027] Figure 1 Diagram of a deep learning-based accurate arc fault detection system

[0028] Figure 2 Example of sliding window displacement

[0029] Figure 3Current-Standard Deviation-Kujit Series Plot

[0030] Figure 4 Stacking integrated deep learning algorithm diagram

[0031] Figure 5 Arc fault detection step diagram Detailed Implementation

[0032] Those skilled in the art can fully and accurately implement this invention by referring to the following description. The specific embodiments described herein are only for explaining the invention and are not intended to limit the scope of protection of the invention. The deep learning-based method for accurate arc fault detection of this invention includes the following arc fault detection steps: Figure 5 As shown, the complete implementation should follow these seven steps:

[0033] Multi-channel synchronous current signal acquisition and electrical isolation:

[0034] A high-precision, wide-range, electrically isolated Hall current sensor is used to collect the current of the bus and branch lines of low-voltage power distribution lines in a non-contact, real-time manner. The sensor output signal is processed by a signal conditioning circuit to perform differential amplification, anti-aliasing filtering, power frequency notch filtering, overvoltage and overcurrent protection, and level boosting to suppress electromagnetic interference and spike noise in the power grid. Subsequently, a 16-bit high-precision, high-speed data acquisition card completes multi-channel synchronous AD conversion at a sampling rate of 40kHz to obtain a distortion-free, low-noise, and time-synchronized digital current sequence, ensuring that weak arc characteristics are not lost.

[0035] Current signal preprocessing and precise period segmentation:

[0036] The digital current signal is preprocessed, including removing DC drift, eliminating outliers, and smoothing and filtering to reduce noise. Based on the voltage zero-crossing point, the power frequency cycle is accurately aligned and segmented, and the continuous current is divided into independent and regular single-cycle data segments to ensure that the data length of each cycle is consistent and the phase is aligned, providing a standardized data format for subsequent disturbance identification, feature calculation and model inference.

[0037] Improved CUSUM current disturbance fast location and data clipping:

[0038] An improved CUSUM cumulative sum algorithm with dual sliding windows is adopted for real-time online screening of current: the front window W1 is used as a stable reference interval to calculate the mean effective value of the current and establish a normal fluctuation interval; the effective value of the current in the back detection window W2 is normalized to eliminate the influence of load size on disturbance judgment; the positive offset cumulative sum and the negative offset cumulative sum are calculated in real time, and when the cumulative sum exceeds the dynamic threshold, the current is immediately judged to have an abnormal disturbance; further, the joint criteria of sudden increase in current standard deviation and sudden drop in kurtosis are combined to accurately locate the disturbance starting point; the system automatically filters out all stable current segments before the disturbance point, and only retains the disturbance starting point and several subsequent cycles to enter the intelligent recognition stage, which greatly reduces the amount of subsequent calculation data to 10% of the traditional method.

[0039] Automatic extraction of joint time-frequency features of the perturbation segment:

[0040] For identified suspected disturbance sections, multi-dimensional time-frequency domain joint features are automatically extracted, transforming the subtle physical characteristics of arc faults into quantifiable and distinguishable feature vectors.

[0041] Time-domain characteristics include: zero rest time, peak-to-peak current, effective period value, standard deviation of effective value, and kurtosis of effective value;

[0042] Frequency domain features include: the amplitude of the 2nd to 21st harmonics, the centroid of the spectrum, and the standard deviation of the spectrum obtained by using the Hanning window short-time Fourier transform; through feature extraction, the intermittent, nonlinear, and distortion characteristics of the electric arc are amplified and enhanced, significantly improving the separability between normal current and electric arc fault.

[0043] Intelligent Fault Classification Based on Stacking Ensemble Deep Learning:

[0044] The constructed time-frequency domain feature vectors are input into a lightweight Stacking two-layer heterogeneous deep learning model: the first layer consists of a base learner composed of decision tree (DT), k-nearest neighbor (KNN), and support vector machine (SVM), which perform preliminary discrimination of fault features from different dimensions and output the initial classification probability; the second layer uses logistic regression (LR) as a meta learner, which performs weighted fusion and decision-making on the results of each base learner and outputs the final fault confidence. The model is trained on a mixture of real arc data and simulation data of a double exponential arc model, and can adaptively distinguish between normal load fluctuations and real fault arcs, achieving high-precision intelligent classification.

[0045] Multi-cycle fault-tolerant decision and fault reliability verification:

[0046] To avoid false triggering caused by transient interference, load switching, lightning surges, etc., the system adopts a multi-cycle consistent confirmation mechanism: only when the model determines that an arc fault has occurred for N consecutive power frequency cycles (N ≥ 2) and the fault confidence is higher than the set threshold, is the actual fault finally confirmed. At the same time, the reliability is verified by combining the stability of fault characteristics, the persistence of disturbance, and the harmonic variation law, which further improves the reliability of the judgment.

[0047] Protection action output, fault recording and remote upload:

[0048] Upon fault confirmation, the system immediately outputs an isolation drive signal to control the arc fault circuit breaker AFDD to trip quickly and cut off the power supply to the fault circuit. At the same time, a local audible and visual alarm is activated. The system automatically latches the complete current waveform before and after the fault, the fault time, fault type, fault branch, disturbance intensity, characteristic parameters, and other information. The fault information is then uploaded to the monitoring platform via RS485, WiFi, or LoRa communication interfaces, enabling the fault to be monitored, warned, traced, and managed.

[0049] Based on the aforementioned deep learning-based accurate fault arc detection technology, the following embodiments are provided:

[0050] Example 1: Arc detection of series faults under mixed loads in residential buildings

[0051] This embodiment applies to a 220V / 50Hz residential indoor power distribution circuit. The circuit simultaneously connects resistive, inductive, and nonlinear loads, forming a typical mixed household power consumption scenario. Specifically, it includes: an 1800W electric kettle (resistive load), a 50W electric fan (inductive load), and a 136W desktop computer (nonlinear load). The system hardware installation is as follows:

[0052] A 0.1-level 100A Hall current sensor is installed at the circuit inlet. The signal is conditioned by a circuit and then fed into a high-speed acquisition unit with a sampling rate of 40kHz. The edge computing unit runs the detection algorithm described in this invention, and the output is connected to an arc fault circuit interrupter (AFDD).

[0053] After the system is powered on, all loads operate stably, the line current is stable, the waveform is regular, and there are no obvious disturbances or distortions. At this time, the improved CUSUM unit in the algorithm continuously monitors the effective value of the current. Because the current is stable without sudden changes, the offset accumulation remains within the threshold and does not trigger disturbance location. When a terminal of the line experiences poor contact due to long-term heating, loosening, or aging, a series fault arc begins to occur. At this time, the current waveform shows obvious intermittent zero rest, amplitude fluctuations, steep waveform distortion, and is accompanied by a sudden increase in high-frequency components.

[0054] The improved CUSUM algorithm detected a significant shift in the effective current value within 10ms, with positive shifts accumulating and rapidly exceeding limits. Simultaneously, the current standard deviation increased sharply, and the kurtosis decreased significantly. The system accurately located the disturbance initiation point and automatically filtered out all previously stable data. Subsequently, the system automatically extracted time-frequency domain features from the current in the disturbance section.

[0055] In the time domain, the zero-down time has expanded from less than 1ms in the normal state to 2~12ms, the peak-to-peak current fluctuation amplitude exceeds 30%, and the effective value standard deviation and kurtosis both show obvious anomalies; in the frequency domain, the content of the 2nd to 21st harmonics has increased significantly, the spectral centroid has shifted significantly to the high-frequency band, the spectral standard deviation has increased significantly, and the eigenvectors show typical arc fault characteristics.

[0056] After the aforementioned feature vectors are input into the Stacking deep learning model, the three base learners—decision tree, K-nearest neighbors, and support vector machine—all output high fault probabilities. After fusion via logistic regression by the meta-learner, the final fault confidence score reaches over 0.99. To avoid false alarms, the system performs three consecutive consistent decision cycles, all of which determine it as an arc fault, thus immediately confirming the fault occurrence. Within 10ms, the system trips the AFDD circuit breaker, cutting off the circuit power supply, simultaneously activating the audible and visual alarms, and uploading the fault type (series arc), fault time, disturbance intensity, and current waveform to the backend monitoring platform.

[0057] This embodiment achieves early, rapid, and accurate detection of weak series arcs in a mixed load and strong interference environment, with no missed detections and no false alarms, fundamentally preventing fires caused by arcs.

[0058] Example 2: Parallel Arc Fault Location and Detection in Industrial Multi-Branch Trunk Circuits

[0059] This case study pertains to a low-voltage power distribution system in an industrial production workshop. Specifically, it employs a "trunk-style" power distribution architecture, encompassing one main power supply circuit and four branch lines. It primarily serves industrial scenarios such as factory production workshops and warehouses, with load types including machine tools (inductive loads), high-frequency welding machines (non-linear loads), assembly line motors (inductive loads), and office electrical equipment (non-linear loads). Daily power loads fluctuate significantly, placing extremely high demands on fault location accuracy and power supply continuity. The core requirement is to achieve "accurate fault location without affecting normal production," while simultaneously addressing the issues of misjudgment, missed detection, and ambiguous fault location associated with traditional detection methods.

[0060] To achieve fault location, this embodiment installs current acquisition modules on the bus and each branch line to achieve multi-channel synchronous sampling, with a time synchronization error of less than 1μs for each channel. Data from each channel is synchronously sent to the edge computing unit, where the detection algorithm described in this invention runs independently. During normal system operation, the load on each branch line is stably connected, the bus current is the sum of the currents in each branch line, the current is stable without sudden changes, and there are no disturbances triggering any channel. When a parallel arc fault occurs in the internal wiring of branch line 2 due to insulation damage or exposed wires, intermittent discharge occurs at the fault point, causing a sharp current distortion, intensified oscillations, and a surge in harmonics, while simultaneously creating a significant disturbance to the bus current. The results of the multi-channel synchronous detection are as follows:

[0061] The currents in branch lines 1, 3, and 4 remain stable without disturbance or distortion; the current in the main line shows obvious disturbance; the current in branch line 2 shows severe distortion, prominent zero-rest phenomenon, and frequent pulse transients, which are the most significant characteristics.

[0062] The improved CUSUM algorithm simultaneously detects strong disturbances on both the bus and branch 2, pinpointing the fault initiation time. The system extracts time-frequency domain features from the disturbance segment of branch 2, revealing a significantly prolonged zero-rest time, drastic peak-to-peak value jumps, a substantial rightward shift of the spectral centroid, and a sharp increase in high-frequency harmonics—characteristics consistent with a parallel arc fault. The Stacking model infers the feature vector of branch 2, outputting a fault confidence score as high as 0.998, thus identifying it as a parallel arc fault.

[0063] The system further identifies the faulty branch based on multi-channel correlation, disturbance synchronicity, and feature similarity. The judgment logic is: only branch 2 exhibits fault characteristics synchronously with the bus, while the other branches show no abnormalities. Therefore, the fault is accurately located in branch 2. After fault confirmation, the system drives the branch 2 circuit breaker to trip independently, without affecting the normal power supply to other branches, achieving "cutting off the faulty branch." Simultaneously, the system uploads complete fault information: fault location (branch 2), fault type (parallel arc), fault time, current waveforms of each channel, and fault characteristic parameters.

[0064] In this case, the system successfully achieved accurate identification of parallel arc faults, precise fault location, and regional power outage protection. This ensured the continuity of industrial production and eliminated fire hazards caused by arcs at the source. At the same time, through multi-channel synchronous acquisition and analysis, it solved the problems of traditional detection methods being unable to locate faults and having a high false alarm rate. This significantly improved the level of power safety and production efficiency in industrial scenarios. It is especially suitable for multi-branch, high-interference industrial power distribution environments and fully meets the core requirements of industrial production for power supply reliability and safety protection.

Claims

1. A method for accurate detection of fault arcs based on deep learning, characterized in that, include: Collect current signals from low-voltage AC distribution lines and quickly locate current disturbance sections using an improved CUSUM algorithm; Automatically extract the joint time-frequency domain features of the current in the disturbed section; The features are input into a pre-trained Stacking deep learning model, which outputs a normal or arc fault classification result. After the multi-cycle judgment rules are met, the fault is confirmed and the linkage protection is executed.

2. The method according to claim 1, characterized in that, The improved CUSUM algorithm includes: A dual sliding window is used to normalize the effective value of the current; Calculate the cumulative sum of positive and negative offsets to identify current abrupt changes; By combining the standard deviation and kurtosis criteria, the starting point of the disturbance can be accurately located, and the stable current segment can be filtered out.

3. The method according to claim 1, characterized in that, The joint time-frequency domain features include: Time-domain characteristics: zero rest time, peak-to-peak value, RMS current, standard deviation, kurtosis; Frequency domain characteristics: amplitude of 2nd to 21st harmonics, centroid of spectrum, standard deviation of spectrum.

4. The method according to claim 1, characterized in that, The Stacking deep learning model is a two-layer heterogeneous integrated structure: Base learners include decision tree (DT), k-nearest neighbor (KNN), and support vector machine (SVM); The meta-learner is logistic regression (LR); The model is trained using a mixture of real electric arc data and simulation data from a double exponential model.

5. The method according to claim 1, characterized in that, Also includes: Simulation samples are generated based on a double exponential arc model, covering the three stages of arc initiation, arc burning, and arc extinction, and together with real data, they form a training set.

6. The method according to claim 1, characterized in that, The fault determination rules are as follows: If the model determines that an arc fault has occurred for N consecutive power frequency cycles, then the arc fault is confirmed to have occurred, where N≥2.

7. The method according to claim 1, characterized in that, It can identify fault types including series arc, parallel arc, and ground arc.

8. The method according to claim 1, characterized in that, It is suitable for resistive, inductive, nonlinear, and mixed load scenarios.

9. A deep learning-based system for precise detection of fault arcs, characterized in that, include: Current acquisition module, edge preprocessing unit, deep learning detection module, protection execution module; The system executes the method described in any one of claims 1 to 8 to achieve real-time detection, early warning and tripping protection of fault arcs.

Images