Fault arc detection system and method based on multi-source fusion and federated self-learning

Abstract

The invention discloses a fault arc detection system and method based on multi-source fusion and federated self-learning. The system comprises a sensor, a data processing and feature extraction module, an edge AI module, a decision logic module, a cloud server and a communication module. The method comprises the following steps: firstly, carrying out system starting and self-inspection, then carrying out multi-source data synchronous acquisition, then carrying out data preprocessing, multi-modal feature generation and lightweight edge AI model reasoning, finally carrying out fault judgment and rapid breaking control, and carrying out logic judgment according to an AI model output result. And when the fault arc is judged to be true, the breaking action of the circuit breaker is quickly executed, and an event is recorded and reported. Through innovation in three aspects of multi-source information fusion, real-time AI identification and federal self-learning updating, full-working-condition, high-precision, fast-response and evolvable fault arc detection is realized, the problem of pain points in the prior art is solved, and the method can be widely applied to the field of low-voltage power distribution safety protection and has remarkable practical value and popularization prospect.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of power distribution safety monitoring and artificial intelligence, and relates to a method and system for intelligent detection of fault arcs, in particular, a fault arc real-time detection and rapid power-off protection system and method combining multi-sensor data fusion and federal self-learning algorithm. BACKGROUND

[0002] A fault arc is an abnormal discharge phenomenon that occurs under conditions such as loose electrical connections, poor contact, aging degradation, etc., and can cause local high temperature and electric sparks, which is an important factor inducing electrical fires. Inside a low-voltage distribution box, partial fault arcs are often easily generated due to poor maintenance, loose screws, aging insulation, improper installation, etc. If not detected and cut off in time, arc discharge may develop into an electrical fire, causing damage to power distribution equipment, and even causing serious personnel injury accidents.

[0003] Current fault arc detection schemes and products on the market mostly rely on a single electrical signal threshold or template matching to identify arc faults. For example, traditional arc fault circuit interrupters (AFCI / AFDD) mainly monitor the current waveform of the circuit, and when the current abnormal fluctuation exceeds a certain fixed threshold or matches a high-frequency pattern, it is determined as an arc fault. However, in actual complex electrical environments, different load conditions may produce current fluctuation signals similar to arcs, and if only a single current feature is relied on for detection, false positives and false negatives are likely to occur. On the one hand, the current and voltage distortion amplitude caused by some weak arc discharge is small, and the single threshold method is difficult to detect in time, leading to missed detection; on the other hand, a large number of nonlinear loads (such as devices containing rectifiers and power electronics) in modern low-voltage distribution systems will introduce rich harmonics and random fluctuations, and the current may also have large waveform distortion during normal operation, and if the threshold is set too low, it will also frequently trigger false positives. In addition, the difference between arc signals and normal electrical noise is sometimes not obvious, and it is difficult to accurately distinguish fault arcs from other interference based on current signals alone.

[0004] In terms of international standards, standards such as American UL1699 and International Electrotechnical Commission IEC 62606 have made provisions for the action requirements of fault arc protection devices, such as requiring the circuit to be cut off within 0.5 seconds after an arc fault occurs. However, these standards do not provide specific details of the detection algorithm, only specifying the action time and test conditions. This makes manufacturers mainly rely on experience-based designed algorithms, which are not adaptable to new loads. Traditional AFCI devices have the following technical defects when dealing with complex scenarios: (1) Single feature dimension limitation: Only single electrical signal features such as current or voltage are used, which cannot fully capture the diverse manifestations of arc faults. For example, when an arc occurs, the line current may exhibit a clear "zero point pause" phenomenon (the waveform appears flat near the zero crossing point) and an increase in multi-pulse and high-frequency noise components; it may also be accompanied by voltage waveform distortion, sudden power factor changes, etc. If the detection means is limited to a single channel signal, many details with recognition cannot be utilized.

[0005] (2) High model complexity and poor real-time performance: Some new research attempts to use machine learning models (such as CNN convolutional neural networks, SVM support vector machines, etc.) to identify arcs. However, ordinary deep learning models have a large number of parameters and are difficult to deploy on edge devices such as MCUs with limited resources in distribution boxes. Models with high computational overhead can also result in excessive detection delays, which cannot meet the requirements of millisecond-level fast tripping.

[0006] (3) Lack of online learning and self-adaptation: Existing devices usually use algorithms with fixed parameters, and the performance of the algorithm will decrease significantly when faced with different on-site environments and load types, i.e., the "domain shift" problem. For example, a certain type of AFCI has a very low false alarm rate in laboratory tests, but when installed in actual field, it may frequently misreport due to new loads (such as variable frequency air conditioners, photovoltaic inverters, etc.) that produce special waveforms. Traditional devices lack the ability to evolve online and cannot continuously optimize the model based on on-site data, nor can they share new knowledge learned by one device with other devices.

[0007] Therefore, there is an urgent need for a new method and system for detecting fault arcs that integrates multi-dimensional electrical features, has embedded real-time intelligent recognition and collaborative evolution capabilities between devices, to improve detection accuracy, reduce false positives and false negatives, and adapt to new types of loads. SUMMARY

[0008] In view of the deficiencies in the prior art, the present application provides a fault arc detection system and method based on multi-source fusion and federal self-learning, to significantly improve the accuracy and adaptability of fault arc detection, and realize intelligent real-time identification and rapid breaking of series arcs and parallel arcs.

[0009] The technical scheme of the system of the present application is as follows: a fault arc detection system based on multi-source fusion and federal self-learning, comprising a sensor, a data processing and feature extraction module, an edge AI inference module, a decision logic module, a cloud server and a communication module; The sensor is arranged in the distribution box for monitoring the state of the distribution line, and simultaneously collecting the current, voltage waveform and high-frequency components generated during the operation of the circuit, as well as the light radiation, sound wave and temperature rise signals generated by the fault arc; The data processing and feature extraction module is arranged in an edge computing unit of the distribution box, synchronously samples and pre-processes each raw signal collected by the sensor, and extracts multi-dimensional feature parameters representing arc fault features from the multi-source signals, including time domain waveform features, frequency domain harmonic features, phase difference features, zero point residence time features, power factor changes, and wavelet time-frequency features. The edge AI inference module is arranged in the edge computing unit, receives the multi-dimensional features output by the data processing and feature extraction module, and locally and in real time infers whether the current distribution line working condition belongs to a normal or arc fault state. The decision logic module is configured to make a circuit breaking control decision according to the output of the edge AI inference module, and trigger the switching element in the distribution protection device to cut off the circuit power supply in a predetermined time sequence when a fault arc is determined to occur. The communication module is configured to communicate between devices in the system and between the devices and the cloud server.

[0010] Preferably, the sensor includes a Rogowski coil current sensor, a high-resistance voltage divider, an EMI antenna, a photodiode, an acoustic sensor, and a temperature sensor. Preferably, the edge AI inference model is an initial version of model parameters obtained through offline training, and an online distillation self-learning mechanism with artificial feedback is introduced during the training process: when the decision result ArcScore falls into the gray area and the field operation personnel confirm through the mobile application that the event is not a fault, the multi-source raw data of the corresponding time window of the event and the artificial confirmation label are written into the ring sample buffer area built in the edge computing unit; when the buffer area accumulates to a predetermined capacity, an online knowledge distillation update process is triggered: in this process, the edge AI inference model freezes the existing weights of the convolution layer and the first several layers of the LSTM network, and only fine-tunes the self-attention fusion layer and the output classification layer in the latter part of the model; during fine-tuning, a small batch gradient descent method is used, and a knowledge distillation loss function is used to consider the original model output and the new artificial label, both of which work together under a set temperature coefficient to gradually correct the model's ability to distinguish non-fault modes under new loads.

[0011] Preferably, the system further includes a federal self-learning module, an event aggregator, a model coordination module, and an OTA update server. The federal self-learning module is arranged in the edge computing unit, and is configured to update and optimize the edge AI model online using artificially confirmed samples during system operation, and extract and compress the update information of the local model and send it to the cloud server for federal learning aggregation through the communication module. The event aggregation and model coordination module is arranged on the cloud server, and is used for receiving model update information sent by multiple edge devices, adopting a federated average algorithm to aggregate the updates of different devices, and combining a differential privacy mechanism to ensure the safe anonymity of the data of each device; after the aggregation is completed, the updated global model is packaged as a differential firmware and is sent to the OTA update server of each device through the communication module; The OTA update server is used for storing and managing the firmware version of the edge AI model, and performing firmware signature verification and online update replacement when the cloud server sends a new model differential patch.

[0012] As preferred, the federated self-learning module realizes the collaborative evolution of the multi-device model through a mechanism combining local training and cloud aggregation; specifically including the following processes: after each online distillation update is completed, the edge device compresses the gradient of the local model parameter update; a power iterative random gradient compression algorithm is used to perform low-rank decomposition on the gradient matrix, and the factor vectors after decomposition are quantized by 4 bits, so as to generate a low-rank gradient package with only dozens of bytes; in each day, during the set low-traffic load period, the gradient package is sent to the cloud federated service through LoRaWAN communication uplink; the communication adopts a preset spreading factor to ensure that the data transmission of 50 bytes or less in a single uplink is safe and reliable and has negligible impact on normal business; after receiving the gradient updates from multiple terminal devices, the event aggregation and model coordination module adopts a federated average algorithm to weight and aggregate all the gradients, and then applies differential privacy noise to the aggregation result, so as to ensure that the sensitive data of a single device cannot be inferred from the global update; after the aggregation is completed, the cloud server compares the new model parameters under differential privacy protection with the current model to generate firmware differential data, divides the firmware differential data into several blocks, signs the blocks, and sends the blocks to the OTA update server of each device; after verifying the signature, each device synthesizes the upgraded model and replaces the old model parameters.

[0013] The technical scheme adopted by the method of the application is: a fault arc detection method based on multi-source fusion and federated self-learning, comprising the following steps: Step 1: system startup and self-checking; Step 2: simultaneously collecting current, voltage, high-frequency EMI, optical, acoustic and temperature signals in a t-second sliding window, and buffering the data length of one sliding window for subsequent analysis; wherein t is a preset value; Step 3: filtering, denoising and aligning the collected original signals, and extracting multi-dimensional feature representations, including converting part of the signals into image-form features; Step 4: performing lightweight edge AI reasoning by using an edge AI reasoning model to obtain an evaluation value of the probability of occurrence of a fault arc in the current window; Step 5: According to the output result of the edge AI inference model, logical decision is made, and when the fault arc is determined to exist, the breaker breaking action is executed, and the event is recorded and reported.

[0014] As preferred, in step 3, first, the length of the half-cycle of the original time-domain current sequence collected is applied to the signal segment, and the voltage sequence of the same window length is mapped to the voltage-current phase plane coordinate to obtain a two-dimensional trajectory graph reflecting the current-voltage relationship; then, the two-dimensional trajectory graph is converted into a grayscale image and scaled to a predetermined pixel size, and the histogram equalization processing is performed on the grayscale image to enhance the contrast of the details at the sharp corners of the trajectory, thereby highlighting the abnormal waveform features caused by the fault arc; then, the short-time Fourier transform is performed on the high-frequency current signal by the data processing and feature extraction module, and the current high-frequency component is converted into a time-frequency energy spectrum to obtain a time-frequency spectrum frame with a size of 128x64; at the same time, for the optical and acoustic signals, the time difference between the triggering time of the optical sensor and the acoustic sensor in the fault event is compared, and the time delay At of the optical-acoustic signal is calculated, and the amplitude ratio r of the optical signal and the acoustic signal is obtained, which is used to represent the coupling strength of the arc light energy and the acoustic energy; the optical-acoustic pulse time difference At and the amplitude ratio r are encoded together with the temperature rise gradient AT / At measured by the temperature sensor before and after the event to form a five-dimensional statistical feature vector, which is used to describe the related features of the changes of various physical quantities when the fault arc occurs; thus, the data of each time window after processing form a feature set containing the following modes: time-domain current and voltage numerical sequence, two-dimensional phase trajectory grayscale image, two-dimensional time-frequency spectrum image, and a set of multi-dimensional statistical quantity features. As preferred, in step 4, the edge AI inference model is used for lightweight edge AI inference, and all the feature groups in the current time window are packaged as a tensor input X; the one-dimensional time series signal features in X are subjected to multi-layer one-dimensional convolution operation by the convolutional neural network to obtain an intermediate tensor C containing local peak patterns; C is sent to the bidirectional long short-term memory network to extract the bidirectional context features of the time series, and the time series context tensor H is output; then, the tensors C and H are connected and fused in the channel dimension to form a comprehensive time series feature tensor F; at the same time, for the two-dimensional image features contained in X, the spatial feature representation thereof is extracted by the lightweight convolutional neural network to obtain an image feature tensor M; then, the time series feature tensor F and the image feature tensor M are sent to the multi-head self-attention fusion module, which contains several heads of self-attention mechanism, and each head calculates a weight matrix a i, for measuring the contribution degree of each modal feature under the current window to the fault classification result; the features of the corresponding modal are adaptively weighted through the attention weights to obtain a fused weighted tensor Z; finally, the tensor Z is input into a classification output layer to generate a four-dimensional fault state probability vector through Softmax normalization wherein P represents the probability of normal working condition, P represents the probability of series type fault arc, P represents the probability of parallel type fault arc, P represents the probability of non-fault interference; define a fault arc score , that is, taking the greater value representing the possibility of arc fault as the comprehensive score.

[0015] As preferred, in step 5, the decision logic includes a multi-level decision threshold and a fast breaking control mechanism; when the ArcScore of the continuous three frames of inference results are all not lower than the preset threshold A, it is determined that the fault arc is established, and the edge computing unit immediately drives the safety breaking circuit through the decision logic module to first trigger the high-speed semiconductor switch SiC-MOSFET to turn off the current, and then cut off the main contact of the insulated gate bipolar transistor within a preset time, thereby realizing the fast power-off of the fault arc current; when the current frame ArcScore is lower than the preset threshold B, it is determined as normal state without action; for the gray area frame with ArcScore between B and A, the system starts a secondary confirmation mechanism, preferentially analyzes the high-frequency EMI signal in this window to confirm whether there is hidden arc feature, and then decides whether to trip.

[0016] As preferred, in step 5, after completing the fault decision, the edge computing unit writes the multi-modal original waveform data of this event together with the decision result label into the local non-volatile memory through the event aggregator and model coordination module, and sends a JSON format data packet to the local or remote monitoring station through the communication module using the MQTT protocol, wherein the data packet contains the event type, occurrence time, ArcScore value, breaking delay, peak arc current and arc energy fields, so as to enable the back-end system to perform centralized statistical analysis and evidence tracing.

[0017] Compared with the prior art, the present application has the following advantages: (1) The detection accuracy is greatly improved: the present application uses a multi-source fusion feature extraction method, and the system can extract more rich arc features than single current signal, thereby realizing comprehensive perception of fault arc. According to tests, the comprehensive detection accuracy of the system of the present application under various load conditions can reach more than 98%, which is much higher than that of the traditional single modal scheme; at the same time, the false alarm rate is less than 1%, effectively solving the problem of easy false tripping of ordinary AFCI.

[0018] (2) Fast response speed, meeting high real-time performance: Due to the use of light model inference on the embedded AI chip and the special fast circuit breaking hardware, the total delay of the system from detection to arc breaking is within 20 milliseconds, which is much better than the standard requirement of 500 milliseconds threshold. This means that the arc fault is killed in the initial millisecond development stage, greatly reducing the risk of electrical fire.

[0019] (3) Strong model self-adaptive ability: Through online distillation self-learning, the device of the present application can continuously optimize its model according to the new working conditions appearing on site. For example, after adding a certain electrical load, the system can learn its normal working mode within a few minutes to avoid misidentifying it as a fault arc. According to experiments, a single online update can reduce the false alarm rate of a specific new load scenario by several percentage points, significantly improving the adaptability of the model to changing environments.

[0020] (4) Federal intelligent collaboration: With the help of federal learning architecture, the learning achievements of multiple devices are fed back to all devices after being aggregated on the cloud, gradually improving the fault detection capability of the entire network. Each round of federal aggregation is equivalent to integrating the data experience of different sites, so the newly released model is more universal for various loads. In actual tests, the weekly model federal update has steadily improved the average detection F1 score of the entire network, proving the effectiveness of collaborative evolution.

[0021] (5) Safe and reliable data mechanism: The system introduces multiple security designs in data communication and model updating, such as using low-rank gradient compression to reduce bandwidth occupation, using differential privacy to protect user data from leakage, and using signed OTA update to ensure firmware security. These measures ensure that the system improves intelligence while keeping the operation and maintenance costs and information security risks under control.

[0022] In summary, the present application provides a fault arc detection system that can work in all conditions, with high precision, fast response, and evolution, which solves the pain points in the prior art and can be widely used in low-voltage power distribution safety protection field, with significant practical value and promotion prospect. BRIEF DESCRIPTION OF DRAWINGS

[0023] The technical solutions of the present application are further illustrated by the following examples and specific embodiments. In addition, some drawings are also used in the process of explaining the technical solutions. For those skilled in the art, other drawings and the intention of the present application can be obtained without creative labor based on these drawings.

[0024] Figure 1 The system structure block diagram of the present embodiment is shown in Figure 1. Figure 2 The method flowchart of the present embodiment is shown in Figure 2. Figure 3 Edge AI inference model structure schematic diagram of the present embodiment. DETAILED DESCRIPTION

[0025] In order to facilitate those skilled in the art to understand and implement the present application, the present application will be further described in detail below in conjunction with the drawings and examples. It should be understood that the implementation examples described herein are only used to illustrate and explain the present application, and are not used to limit the present application.

[0026] See Figure 1 The fault arc detection system based on multi-source fusion and federal self-learning provided by the present embodiment includes a sensor, a data processing and feature extraction module, an edge AI inference module, a decision logic module, a cloud server, a communication module, a federal self-learning module, an event aggregator and a model coordination module, an OTA update server and a monitoring dashboard. The sensor module is arranged in the distribution box and includes various sensors for monitoring the state of the distribution circuit, including a current sensor (such as a Rogowski coil), a voltage sensor (such as a high-resistance voltage divider), a high-frequency electromagnetic interference (EMI) antenna, a photoelectric sensor, an acoustic sensor (such as a MEMS microphone), and a temperature sensor, for synchronously collecting the current, voltage waveform and high-frequency components generated during circuit operation, as well as the light radiation, sound waves and temperature rise signals that may be generated by the fault arc; the current sensor (such as a Rogowski coil) is sleeved on the live wire or branch of the monitored circuit to sense the circuit current; the voltage sensor uses a high-resistance voltage divider network and is connected to the two ends of the protected circuit to obtain the circuit voltage; the EMI antenna is placed in the distribution box to sense the electromagnetic radiation and radio frequency interference signals generated by arc discharge; the photoelectric sensor (such as a photodiode or a phototransistor) faces the inside of the distribution box to detect the light flash of the arc spark; the MEMS microphone is installed in a proper position inside the distribution box housing to pick up the high-frequency crackling sound when the arc is drawn; the temperature sensor is attached to the key parts (such as near the guide rail or cable) inside the distribution box to monitor the ambient temperature and the local temperature rise caused by the arc. During normal operation, the signals of all sensors are connected to the edge processing unit (which can be a high-performance microcontroller MCU or an embedded processor) after being conditioned by the front-end circuit. The edge processing unit integrates software functions such as the data processing and feature extraction module, the edge AI inference module, the communication module, and the federal self-learning module. After the system is powered on, the self-checking initialization of step 1 is first performed, and after the modules are confirmed to be normal, the monitoring cycle is entered.

[0027] The data processing and feature extraction module is connected with the sensor module, arranged on an edge computing unit (MCU or DSP) of the distribution box, performs synchronous sampling and preprocessing on each raw signal collected by the sensor, and extracts multi-dimensional feature parameters representing arc fault characteristics from the multi-source signals, including time domain waveform features, frequency domain harmonic features, phase difference features, zero point residence time features, power factor changes, wavelet time-frequency features, etc. The edge AI inference module is deployed in the edge computing unit and includes a pre-trained lightweight arc fault recognition model, which is used to receive the multi-dimensional features output by the feature extraction module and locally infer in real time whether the current distribution line working condition belongs to a normal or arc fault state. The decision logic module is connected with the edge AI inference module and is used to make circuit breaking control decisions according to the fault probability or score output by the AI model. When a fault arc is determined to occur, the switching elements (including high-speed SiC-MOSFET and main circuit breaking IGBT) in the distribution protection device are triggered to cut off the circuit power supply in a predetermined time sequence. The communication module is used for communication between devices and between devices and the cloud, including uploading local detection events to the monitoring end or cloud server, and receiving model updates from the cloud during a safe period. This module can use low-power wide-area network communication protocols (LoRaWAN), MQTT, etc. to realize data interaction between devices and remote servers / monitoring stations. The federated self-learning module is embedded in the edge computing unit and is used to update and optimize the edge AI model online using artificially confirmed samples during device operation, and extract and compress the update information of the local model and send it to the cloud for federated learning aggregation through the communication module. The event aggregation and model coordination module is set on the cloud server and is used to receive model update information sent by multiple edge devices, aggregate updates from different devices using a federated averaging algorithm, and ensure the security and anonymity of data from each device using a differential privacy mechanism. After aggregation, the updated global model is packaged as a differential firmware and sent to the OTA update server of each device through the communication module. The OTA update server communicates with the model coordination module and is responsible for storing and managing the firmware version of the edge AI model. When a new model differential patch is issued by the cloud, it performs firmware signature verification and online update replacement. The monitoring dashboard is used to centrally display the running status of multiple distribution devices and arc fault detection event information, receives event logs uploaded from the communication module of each device, and provides a human-machine interaction interface to allow operation and maintenance personnel to view alarms, confirm false positives, and configure system parameters.

[0028] In an embodiment, the edge AI inference model is an initial version of model parameters trained offline; during the model training process after the device is put into operation, an online distillation self-learning mechanism with artificial feedback is introduced: when the decision result ArcScore falls into the gray area and the on-site operation and maintenance personnel confirm through the mobile application that the event is not a fault, the multi-source original data of the event corresponding time window and the artificial confirmation label are written into the ring sample buffer area built in the edge computing unit; when the buffer area accumulates to a predetermined capacity (for example, 128 frames of samples), an online knowledge distillation update process is triggered: in this process, the device freezes the existing weights of the convolutional layers and the first several layers of the LSTM network in the edge AI model, and only fine-tunes the self-attention fusion layer and the output classification layer in the latter part of the model. When fine-tuning, the small batch gradient descent method is used, and the knowledge distillation loss function is used to consider the original model output and the new artificial label, both of which work together under a set temperature coefficient to gradually correct the model's ability to distinguish non-fault modes under new loads. A single distillation optimization lasts about a few minutes, which can reduce the false alarm rate of the model under new load scenarios from about 6% to below 1%, greatly improving the adaptive accuracy of the device to the changing conditions on site.

[0029] In an embodiment, the federal self-learning module realizes the coordinated evolution of multi-device models through a mechanism combining local training with cloud aggregation, including the following processes: after each online distillation update, the edge device compresses the gradient of the local model parameter update: the PowerSGD algorithm is used to perform low-rank decomposition on the gradient matrix, and the factor vectors after decomposition are quantized to 4 bits, thereby generating a low-rank gradient package of only a few dozen bytes; during the set low-traffic period every day, the gradient package is sent to the cloud federal service through LoRaWAN communication uplink, and the communication uses a preset spreading factor (such as SF7) to ensure that the transmission of 50 bytes or less in a single uplink is safe and reliable and has negligible impact on normal business. After receiving the gradient updates from multiple terminal devices, the cloud event aggregation and model coordination module uses the FedAvg algorithm to weight and aggregate all gradients, and then applies differential privacy noise to the aggregation result to ensure that sensitive data from a single device cannot be inferred from the global update (the risk of single-point reconstruction is less than one in ten thousand). After aggregation is complete, the cloud compares the new model parameters under differential privacy protection with the current model to generate firmware differential data, divides it into several blocks (each block is about 1 KB in size), signs it, and encrypts it to be sent to the OTA update server of each device. After verifying the signature by each device, the upgraded model is synthesized and replaces the old model parameters. According to experimental statistics, performing the above federal aggregation update once a week can improve the overall F1 score of the arc fault detection model of the entire network by an average of about 0.6 percentage points, achieving good results for device collaborative learning.

[0030] SeeFigure 2 The embodiment provides a fault arc detection method based on multi-source fusion and federal self-learning, and comprises the following steps. Step 1: system startup and self-checking; after power-on, the system enters an initialization process, and self-checking is performed on each sensor module, an edge AI inference module and a communication module, and after confirming that the hardware state is normal, the system enters a monitoring mode; Step 2: multi-source data synchronous acquisition; within a preset sliding time window t, current, voltage, high-frequency EMI, optical, acoustic and temperature signals are simultaneously acquired, and a sliding window of data length is cached for subsequent analysis, wherein t is a preset window length; The edge processing unit reads each sensor signal in parallel at a fixed high-frequency sampling rate, for example, the sampling rate of the current and voltage channels can be set to tens to hundreds of kHz to capture the high-frequency components generated by the arc; the light, sound and temperature channels select appropriate sampling frequencies according to their dynamic characteristics. All channels of data are cached and processed in a sliding window manner, for example, a 1 millisecond window length is used (which can be set according to the actual power frequency period, for example, 1 ms is about 20 degrees of electrical angle at 50 Hz), and data processing and inference are triggered once every window time. The sliding window length t and the step length can be adjusted to balance the response speed and feature integrity.

[0031] Step 3: data preprocessing and multi-modal feature generation; filtering, denoising and synchronization alignment are performed on the acquired original signals, and multi-dimensional feature representations are extracted on this basis, including converting some signals into image-form features; In one embodiment, in each sliding window, for the current and voltage signals, first, digital-to-analog conversion is performed to obtain a discrete sampling point sequence, then a band-pass filter is used to filter out the fundamental wave (to highlight the high-frequency spark signal) and the direct current component, and necessary denoising processing is performed. Next, a half-cycle length waveform segment is extracted for subsequent analysis: specifically, the voltage zero-crossing point is taken as the starting reference, and two adjacent zero-crossing points are found on the voltage signal to determine the time range of the current half-cycle, and the corresponding current signal segment is extracted within the same time range. A Blackman window is applied to the current signal segment to reduce the influence of edge discontinuity on spectral analysis while ensuring smooth transition of the truncated time sequence. Then, the current waveform and the voltage waveform of the corresponding length are plotted as a V-I phase trajectory curve in the rectangular coordinate system, that is, a two-dimensional trajectory with the voltage as the horizontal axis and the current as the vertical axis. This trajectory can reflect the abnormality of the voltage-current relationship during the arc, such as the sharp angle and lag caused by the instability of the arc. In order to facilitate subsequent convolutional neural network processing, the trajectory curve is converted into a two-dimensional grayscale image (such as Figure 3: The trajectory coordinate range is normalized and mapped onto a grid of, for example, 48x48 pixels, with the pixels through which the curve passes assigned a certain brightness, and a histogram equalization algorithm is applied to the resulting initial grayscale image to enhance the brightness contrast at the trajectory turning points (sharp corners). The resulting phase-plane grayscale image highlights the distortion characteristics of the V-I relationship when an arc occurs, distinguishing it from the smooth trajectory under normal load. At the same time, a fast Fourier transform (FFT) is performed on the current signal in the half cycle to obtain its frequency spectrum, but considering the non-stationary randomness of the arc signal, the present application further calculates the short-time Fourier transform (STFT) of the current signal: with a short window length (for example, 128 sampling points, corresponding to several microseconds), the frequency spectrum amplitude of each window is calculated, and the time-frequency spectrum is formed by superimposing the frequency spectrum over time. In specific implementation, the half-cycle current signal can be zero-padded or truncated to a fixed length (such as 2.5 ms, about 128 points of FFT window), and the frequency spectrum matrix is calculated and the amplitude is taken to form a two-dimensional matrix (for example, size 128x64, corresponding to 128 time steps and 64 frequency bins) on the time axis x frequency axis, which can be regarded as a grayscale image. The horizontal axis represents time subdivision, the vertical axis represents frequency (from several kHz to several tens of kHz or more), and the grayscale intensity represents the energy of the frequency component. Fault arcs usually exhibit intermittent high-frequency pulses, and thus appear as speckle textures or local highlights on the time-frequency spectrum that are different from normal states. The time-frequency spectrum matrix is scaled to a uniform size (which can be the same as the phase trajectory image, for example, 64x64 or 48x48 pixels) and used as a time-frequency feature map for the second image input branch of the model.

[0032] For the optical and acoustic signal channels, since both arc light and sound are pulsed transients, in the embodiments of the present application, a method of comparing the outputs of the two sensors is used to extract features: once the light sensor detects a significant light pulse (rapid rise above a set threshold), the start time stamp of the light pulse is recorded; similarly, when the microphone detects a sudden sound pulse, the time stamp of the sound pulse is recorded. The time difference between the two pulses is calculated which can be used as an optical-acoustic delay feature, reflecting the delay from the generation of light to the generation of sound, and helping to distinguish arcs (which usually produce light and sound almost simultaneously) from general noise interference. The ratio of the peak amplitude of the light pulse to the peak amplitude of the sound pulse is calculated which can reflect the relative strength of the light energy and sound energy generated by the arc, and different types of arcs may have different ratios. Further, the temperature sensor provides temperature gradient information, such as calculating the temperature rise within a short period of time (for example, 1 second) after the arc event occurs Divided by time The temperature rise rate is obtained. These features from non-electrical sensors are numerical quantities themselves, and the present invention combines them with other statistical information into a fixed-length feature vector. For example, in one embodiment, a five-dimensional statistical feature is constructed: including the photoacoustic delay , the photoacoustic amplitude ratio r, the temperature rise rate , and the light pulse count and the sound pulse count (the number of light / sound pulses detected within a given window). Of course, this is just an example, and the types and dimensions of statistical features can be adjusted as needed in actual implementation. This multi-dimensional statistical vector, together with the above-mentioned phase trajectory grayscale map and time-frequency spectrum, and the one-dimensional current-voltage numerical sequence, collectively constitute the comprehensive feature data of the current sliding window.

[0033] Step 4: Lightweight edge AI inference model inference; using the artificial intelligence model deployed at the edge to perform real-time inference on the features generated in step 3, obtaining the evaluation value of the probability of arc fault occurrence under the current window; See Figure 3 In one embodiment, the lightweight edge AI inference model receives multi-modal feature data, including: time-domain sequence vectors of voltage and current, representing numerical features of waveforms changing over time; phase trajectory grayscale maps generated from voltage-current relationships and time-frequency spectrum grayscale maps generated from current STFT, which together serve as image modal features; and numerical vector features composed of sensor statistics such as light, sound, and temperature. The model processes these different forms of data through different network branches, and then makes a high-level fusion decision.

[0034] Specifically, in the embodiment, for time-domain sequence data (current and voltage sequences, which can be concatenated or connected in parallel as a multi-channel sequence), the model uses a stack of one-dimensional convolutional neural networks (1D-CNN) to extract local pattern features: for example, using 3 layers of 1D convolution, each with a convolution kernel length that can cover several sampling points, so as to identify local morphologies such as current sequence spikes and current zero-crossing shoulders. The convolution layer is followed by a pooling layer and a normalization layer to compress the feature length and enhance stability, and finally outputs a feature tensor C with several channels. Then, C is input into a bidirectional LSTM network for time series modeling. Bidirectional LSTM can understand sequences from both forward and backward directions, obtaining context information spanning multiple power frequency cycles. This is very important for capturing features of arcs that appear continuously or intermittently over multiple half cycles. After Bi-LSTM processing, the forward and backward hidden states at the last time step are output, concatenated to form a time series feature tensor H, which can be regarded as a compressed representation containing the entire window time series dynamics. Then, the convolution output C (which can be taken from the last convolution layer or the convolution + pooling output) and the LSTM output H are merged in the channel dimension to obtain a fused time series feature tensor F, which contains both local details and global time series patterns.

[0035] For the features of image modalities, i.e. the phase trajectory gray scale image and the time-frequency spectrum gray scale image, the embodiment adopts a pre-trained lightweight convolutional neural network for processing. In a preferred implementation, the MobileNet-V3 Small network is selected as the feature extraction backbone. MobileNet-V3 is an optimized lightweight CNN containing modules such as depthwise separable convolution, which can extract high-quality image features with less computation. In the present invention, the two gray scale images can be regarded as a two-channel input to feed into the MobileNet network (or the two images can also be separately extracted by a small network and then spliced), obtaining the feature tensor M of the image modalities. This tensor contains the pattern information in the voltage-current trajectory and the current spectrum, such as the distorted shape of the trajectory caused by the arc, the distribution of high-frequency spots in the time-frequency image, etc.

[0036] Next is the key innovative part of the model - the multi-modal self-attention fusion mechanism. As shown in Figure 3 , the time series feature tensor F and the image feature tensor M are first adjusted to the same feature dimension through linear transformation, and then sent into the self-attention module together. The module includes 8 parallel attention heads (of course, other numbers are also possible), and each attention head calculates the attention weight matrix through self-attention calculation on the fused features, i.e. by setting the query (Query), key (Key) and value (Value) matrices. In the context of the present invention, it can be understood that the model automatically learns how to allocate the weights of each modal feature in the final decision. For example, in some cases, the current time series features best represent the arc, so the model will give F a higher weight; while in other cases, such as when the environmental noise interference is large, the optical acoustic signal may be more discernible, and the model will enhance the role of M. Each head independently produces a set of weight matrices, and multiple heads enable the model to make a comprehensive judgment from different angles. After the outputs of all attention heads are weighted and fused, the final fused feature tensor Z is obtained through the gating unit. It can be considered that Z is already a highly abstract feature that contains all the modal information of the current window and has been optimally weighted.

[0037] Finally, the fused feature Z is sent to the output layer for classification decision. The output layer can be a fully connected neural network, which maps to 4 neuron outputs corresponding to the 4 kinds of discrimination results set in the embodiment: normal (Normal), series arc (Series Arc), parallel arc (Parallel Arc) and noise interference (Noise). Through the Softmax function, the 4 outputs are normalized to obtain the probability distribution , which respectively represent the probability values of the model determining that the current time window belongs to each state. Define As a comprehensive arc fault score, it is used for subsequent decision logic. In actual implementation, only two categories of arc and non-arc can also be output, but the present application helps further analysis and targeted action by subdividing series arc and parallel arc states. It is worth mentioning that due to the lightweight design and optimized inference process of the model, its running efficiency is very high. On the experimental hardware (embedded ARM Cortex-M7 MCU with hundreds of MHz frequency), the model consumes about 1-2 milliseconds for inference per frame, which fully meets the real-time requirements.

[0038] Step 5: Fault decision and rapid breaking control; logical decision is made according to the output results of the edge AI model, and when the fault arc is determined to be true, the breaker breaking action is quickly executed, and the event is recorded and reported; In an embodiment, a hierarchical threshold decision logic is introduced to balance false positives and false negatives, and to ensure rapid breaking: A preset upper threshold A (such as 0.8) and a lower threshold B (such as 0.4) are introduced. If the ArcScore of three consecutive frames (i.e. three consecutive sliding window periods) is higher than A, it means that the model strongly believes that an arc fault has occurred multiple times, and it can be almost certain that there is an arc. At this time, the system immediately enters the fault handling process and executes the power-off action. The requirement of three consecutive frames is to avoid false action caused by single-frame accidental anomaly, but the number of frames can also be adjusted according to requirements.

[0039] If the ArcScore of the current frame is lower than B, it means that the model believes that there is basically no arc, and the system remains in normal monitoring state without any action and clears the previous suspected fault count (equivalent to resetting the continuous count).

[0040] If the ArcScore falls between B and A, i.e. the model is somewhat suspicious but not sure enough, it is considered as a gray area. For the gray area, the present application does not immediately trip, but enters a secondary confirmation process: the system will mobilize specific features for further analysis to reduce uncertainty. In the present embodiment, high-frequency EMI signals are selected for secondary confirmation - because high-frequency pulses captured by EMI antennas are particularly sensitive to arcs and are not easily disturbed by general power noise. When the gray area appears, the processor immediately performs higher-resolution spectral analysis or pattern matching on the EMI raw waveform cached in the current frame (compared to the rough AI features, this step is more detailed but only when necessary). If the analysis result supports the existence of arc (for example, the characteristic arc radio frequency pattern is detected), it is determined that the fault is true, and the system immediately executes the breaking; otherwise, it is considered as a false trigger, and the system ignores this alarm and returns to normal monitoring. Through this secondary confirmation, the possibility of false tripping due to model uncertainty is greatly reduced.

[0041] Once the final decision of arc fault is made, the system will quickly cut off the circuit power supply. The specific action is controlled and driven by the decision logic module: first, turn off the semiconductor power electronic switch (such as SiC-based MOSFET), which is in series with the circuit, and after turning off, it can block the current and limit the arc current in microseconds; then within a few milliseconds, drive the IGBT main contact in the circuit breaker to open, completely isolating the circuit from the power supply. This multi-level breaking scheme combines speed and reliability, with MOSFET responsible for ultra-fast response and IGBT responsible for carrying isolation. According to experimental device testing, the average time from first detection to current interruption is about 12ms, which is much faster than the maximum time of 0.5 seconds allowed by UL and IEC standards. Fast cutting can prevent the arc from growing and reduce fault energy. In addition, due to the high accuracy of the algorithm, the inconvenience caused by frequent false alarms in traditional devices is avoided.

[0042] After completing the tripping action, the system performs event recording and reporting. The edge computing unit stores the relevant data of this event in the local storage medium, including all original sensor waveforms for several seconds before and after the event, as well as the model output decision results (series arc / parallel arc, ArcScore curve, etc.) and the executed actions (such as whether to cut off, arc duration, etc.). At the same time, through the communication module, the event information is sent to the monitoring center or the cloud. One implementation is to send a JSON format data packet, which includes: event unique ID, timestamp, decision type (series arc or parallel arc), ArcScore peak value, circuit breaking delay time, estimated energy released by this arc, etc. These data can be presented on the monitoring dashboard interface, making it easy for maintenance personnel to analyze the details of each fault arc afterwards, and can be used to optimize the maintenance strategy of the power distribution system.

[0043] In one embodiment, after completing the fault decision, the edge computing unit writes the multi-modal original waveform data of this event (5 seconds before and after the event) along with the decision result label to the local non-volatile memory (such as eMMC storage chip) through the event aggregation module, and sends a JSON format data packet to the local or remote monitoring station through the communication module using the MQTT protocol, which contains fields such as event type, occurrence time, ArcScore value, breaking delay, peak arc current, arc energy, etc. for centralized statistical analysis and evidence tracing by the backend system.

[0044] Step 6: Federal self-learning model update; during the long-term operation of the system, adaptive learning and model optimization are performed on gray area samples, and multi-device collaborative evolution is achieved through cloud aggregation, continuously improving the global detection performance of the model.

[0045] The great improvement of the present embodiment over the prior art lies in its federated self-learning model updating mechanism, which enables the device to become "smarter" over time. The mechanism includes two parts: local online learning and cloud federated aggregation.

[0046] (1) Online self-learning (local distillation update): Although the model initially deployed in the present embodiment has been trained offline and has a high accuracy, it may still encounter new waveform patterns that the model has not seen before, resulting in misjudgments. To further reduce the false alarm rate, the system allows people to "participate" in the learning process in the loop: when the device judges the gray area or suspicious situation, it will alarm on the monitoring dashboard. If the manual confirmation of the alarm is a false alarm (i.e., there is no actual fault, only the model misjudges), it can be fed back to the device through the monitoring interface. This is equivalent to labeling the data in a certain time window as "normal", correcting the model's understanding. The device will store the labeled data segment in the internal sample buffer. The buffer uses a ring structure and has a fixed capacity (e.g., stores the last 128 frames of labeled data) to ensure that it does not consume too much memory. When the buffer is full, an online update process of the model is triggered.

[0047] Online update is different from traditional direct continuous training. To prevent a small number of new data samples from causing the model to forget existing knowledge or overfit, the present embodiment adopts a knowledge distillation strategy for gentle update. Specifically, during the update process, most layers of the model are frozen, and only the last few layers (especially the self-attention fusion layer and the output layer) are unfrozen, allowing their weights to be slightly adjusted according to the new data. The original weights of the underlying feature extraction part, such as the convolution and LSTM layers, are kept unchanged because the underlying features are usually still common when new data appears. The loss function during training is designed as a distillation loss: on the one hand, the cross-entropy loss between the model's output for new data and the manually confirmed label is calculated, and on the other hand, the difference between the model's output for new data and its predicted output under the old model (this is achieved by adding a temperature coefficient to the old model's output to create a soft label) is also calculated, and then the two are weighted and summed. This approach ensures that the model fits the new samples while not deviating too far from the original decision. The online distillation process can be performed during the device's idle period, such as after a trip or during low load periods, and can be completed in a few minutes. Since only a few parameters are fine-tuned, this update has low computational resource requirements. Multiple field tests have shown that this online learning can significantly optimize the model's adaptability to special loads. For example, after introducing a power tool load on a device, the initial model's false alarm rate once reached 6%, and after maintenance personnel labeled about 100 frames of normal samples and experienced a distillation update, the false alarm rate dropped to less than 1%. Online self-learning ensures that the device becomes more accurate over time and continuously adapts to the actual situation.

[0048] (2) Federated Learning Aggregation (Cross-device Collaboration): The new knowledge learned by a single device can also be shared to other devices through the cloud. This embodiment uses the Federated Learning framework to achieve this. Unlike directly uploading raw data to the cloud for centralized training, Federated Learning only allows devices to upload the update of model parameters, which reduces the communication volume and protects user data privacy. In practice, this embodiment chooses to upload the gradient (i.e., the amount of change in model weights) instead of the complete weights to further reduce the upload size. The specific process is as follows: after a device completes an online distillation update, it extracts the model weight difference Delta W generated by this update. Since Delta W is a high-dimensional matrix (although only the output layer and attention layer are involved, there can still be thousands of parameters), direct transmission is not practical. This invention uses the PowerSGD technique to compress the gradient: this is an efficient gradient compression algorithm in recent years, the basic idea is to decompose the matrix into several low-rank matrix approximations, so that the original gradient matrix is represented by a small number of parameters. In this embodiment, rank r = 2 is selected for matrix approximation, and the resulting matrix components are quantized to 4 bits to further reduce the number of bytes. After processing, the amount of data required to upload each update is only a few dozen bytes, which is very small. The system packages these compressed gradients, adds the device's anonymous ID (or encrypted signature), and sends them out through the LoRa wireless module. LoRaWAN is a low-power wide-area network protocol, which has a low data rate but is sufficient to transmit a few dozen bytes of load. To not affect normal monitoring, this embodiment sets the device to automatically report at low-usage periods such as early morning every day, such as a maximum of once an hour. Since the data uploaded each time is very small, the time for a 50-byte data packet to be reliably delivered using LoRa's spreading factor 7 is only a few tens of milliseconds, and the interference to the entire network can be ignored.

[0049] The cloud is deployed with a Federated Learning Aggregation Server (corresponding Figure 1The server continuously listens to the uplink gradients from each device (through the event aggregation and model coordination module in the server). Once enough devices have updated or a predetermined time has been reached, the server begins to perform model aggregation. The server first verifies the data signatures uploaded by each device, filters out abnormal or duplicate values, and then decompresses and aggregates all the gradients. The federated averaging algorithm is used, that is, the weighted average of all received gradients is calculated (the weight can be set according to the amount of data represented by each gradient or the reliability of the device). After obtaining the average gradient, differential privacy technology can be selected to protect the influence of a single device from being inferred: a certain amplitude of Gaussian noise is added to the average gradient, so that the change of any single device gradient has little effect on the final result, so that an attacker cannot infer the training data of an individual device from the global model. This measure ensures that federated learning improves the model while not leaking personal privacy data (such as the power usage habits of a device user). Finally, the server adds the average gradient to the current global model parameters to obtain the updated global model weights. Since sending the complete model directly can be large in size (even if the model of the present application is already small, sending the full weights can still be hundreds of KB), the server calculates a model differential patch here: that is, compares the new and old models and only records their differences, and then packages it into a firmware upgrade file. For example, the patch output by the present application is about several KB in size after compression and signature, and can be safely distributed over a narrowband network. The OTA module of each device receives the patch distributed by the cloud, verifies the signature to confirm the source is reliable, and then applies the patch to the existing model to complete the upgrade.

[0050] Through the above federated learning process, the present embodiment achieves the goal of co-evolution of all devices in the network: the experience uploaded by each device is reflected in the new global model, and then benefits back to all devices. In other words, the more devices, the smarter the system. According to the test data collected by the present embodiment, after running in a small federated network consisting of 20 devices for 3 months, the overall performance of the model is significantly improved compared to the initial: the average detection accuracy is improved from 98% to more than 99%, and the F1-score is improved by an average of 0.6 percentage points; In particular, for some rare load scenarios that a single device has not experienced, the new model can also correctly identify. This fully verifies the effectiveness and originality of the federated self-learning mechanism in the application of fault arc detection.

[0051] The present application has undergone simulation experiments and actual scene tests, and the results show that it has excellent performance and significant technical effects in fault arc detection. The following will be explained in combination with specific experimental data: (1) Laboratory simulation of fault arc scenarios; In a laboratory environment, a set of arc fault test circuit is built, including a controllable series arc generator and several typical household / industrial loads (such as incandescent lamps, motors, power adapters, etc.), and the detection device of the invention is installed to monitor arc faults. According to the test requirements in UL1699 and IEC62606 standards, the arc generator is allowed to produce series arc and parallel arc at different current levels and voltages in the experiment, and the detection and action performance of the device is recorded. The results show that: the system of the invention can detect abnormalities and determine faults within about 3-4 milliseconds after arc discharge, and send a trip command within 5 milliseconds, and the entire circuit is cut off within 20 milliseconds after arc generation, thereby effectively extinguishing the arc. This response speed is much faster than the 0.5 second action time required by the standard, reaching an internationally leading protection level. In all simulated tests of dozens of arc events, the device successfully detected and cut off 100% of the time without a single false alarm or refusal to act; in contrast, some commercially available traditional AFCIs failed to trip in time or misjudged normal load changes as faults under the same conditions. It can be seen that multi-source feature fusion and intelligent decision strategy effectively improve the reliability of the system.

[0052] (2) Long-term testing of actual power distribution systems; Several devices of the invention are installed in an industrial power distribution cabinet and a residential distribution box, and have been running continuously for several months to verify their practical performance. During this period, the device has experienced various load switching and several real faults (including a series arc caused by a loose terminal in the power distribution cabinet). Statistical data shows that the system of the invention maintains an average detection accuracy of >98% and a false alarm rate of <1%, basically eliminating the common false alarm annoyance of traditional AFDD. For some special power use scenarios (such as spark interference signals generated during high-power motor startup), the initial model occasionally judges them as gray areas, but after the operation personnel confirm normal on the instrument panel and trigger online learning, the model quickly adjusts and no misjudgment occurs subsequently. This proves that the online adaptive ability of the invention effectively makes up for the shortcomings of the prior training data of the model. The federated learning of multiple devices also shows benefits: after a period of time, the versions of each device are upgraded, and some complex load signals that have appeared can be directly and correctly classified without going through the learning process for each device. This reflects the overall intelligence improvement brought by federated collaboration, saving the cost of manual intervention.

[0053] In summary of the experiment, the fault arc detection system and method based on multi-source fusion and federal self-learning of the application successfully solves the pain points and problems of the prior art, and realizes higher detection accuracy, faster response speed and continuously evolving model. Through multi-source data fusion, the miss detection rate is greatly reduced; through edge AI intelligent judgment and high-speed switch, millisecond-level safe power-off protection is achieved; through the federal self-learning mechanism, the system can continuously maintain excellent performance in various new environments and new loads. This technical scheme has great significance in ensuring power distribution safety and preventing electrical fires, and has broad application prospects.

[0054] The main technical solutions and innovative points of the application include the following aspects: First, at the data acquisition level, multi-source sensing signals are fused to enrich the fault feature dimension. The device of the application introduces multiple sensors such as voltage, high-frequency EMI, electromagnetic radiation, optics, acoustics and temperature on the basis of traditional current sensing, and constructs a multi-modal data acquisition system. Through the comprehensive utilization of multi-dimensional electrical characteristics such as current, voltage, harmonic, phase difference, zero point residence time, power factor and wavelet time-frequency energy, the problem of missing detection caused by only focusing on a single signal in the prior art is solved. The information fusion of multiple sensing sources can more comprehensively capture the triggering signs of arc fault. For example, the shoulder feature near the current zero point when the arc is lengthened or extinguished, the extension of "zero rest time", the EMI peak of a certain frequency band generated by arc spark, and the arc light flicker and sound impact are all monitored synchronously and used as decision basis. Through multi-source fusion at the feature level, the robustness and sensitivity of fault arc identification are greatly improved.

[0055] Secondly, at the embedded intelligent detection level, a lightweight edge AI model is used to realize high-speed and high-precision local inference. The application designs a deep learning model suitable for deployment on MCU-level hardware, which combines one-dimensional convolutional neural network, bidirectional LSTM network and self-attention mechanism, and is specially optimized for the feature mode of fault arc signal. On the model structure, it includes a convolutional layer for extracting transient peak features, a Bi-LSTM layer for capturing cross-cycle timing patterns, a multi-head self-attention mechanism for fusing different modal features, and a classification output layer. In order to balance accuracy and real-time performance, the model selects a small parameter architecture (such as MobileNet series network for image feature extraction), and is optimized through quantization compression and model pruning, so that the model parameter size is only about 280KB in 16-bit floating point format, and the single inference delay on the edge device is not more than 2 milliseconds after 8-bit quantization. This small model high performance design ensures that the device can complete fault detection and action decision in much less time than the standard requires, realizing millisecond-level response.

[0056] Again, at the decision logic level, the application sets up adaptive multi-level decision and fast breaking control strategy. For detection results of different confidence levels, the system divides into three states of "normal", "gray area" and "fault", and introduces a secondary confirmation algorithm for the gray uncertain area to further reduce the misjudgment rate. When the model output indicates a high probability of arc fault, the device immediately triggers fast breaking: it preferentially cuts off the current through solid-state devices, and then opens the main relay contacts within a few milliseconds. The entire arc breaking process takes an average of about 0.012 seconds in the experiment, which is much better than the 0.5 second action requirement specified in IEC 62606. Compared with the situation that the traditional AFCI may not move or misjump frequently due to the dependence on a single threshold, the multi-level decision logic of the application significantly improves the reliability and speed of decision-making.

[0057] Finally, at the model evolution level, a federal self-learning mechanism is introduced to give the device the ability of online evolution and group intelligence cooperation. The system of the application can continuously learn during operation: when new conditions that are difficult for the model to distinguish appear (for example, a certain window is judged as a gray area), the true situation can be confirmed by the operation and maintenance personnel through the monitoring dashboard, and the device will cache these samples verified by manual for online retraining of the model. Through a lightweight online learning method based on knowledge distillation, the device quickly absorbs new sample experience without forgetting the original knowledge, and after updating the model, it can greatly reduce the false alarm and missed alarm in such new conditions. At the same time, the model updates extracted by each device will be uploaded to the cloud for safe fusion through the federal learning framework: the federal average algorithm is combined with differential privacy to aggregate incremental knowledge from different sites. The cloud then distributes the updated global model to the devices, so that the learning of a single device can serve the entire network. This federal self-learning mechanism fully protects the privacy of data from different sites, and realizes the sharing of knowledge across devices, solving the problem of static model and poor domain adaptability in the prior art.

[0058] It should be understood that the above-described embodiments are part of the embodiments of the application, rather than all the embodiments. In addition, the technical features in each embodiment or single embodiment provided by the application can be combined with each other in any way to form a feasible technical solution, and such combination is not restricted by the sequence of steps and / or structure composition mode, but must be based on the realization by a person of ordinary skill in the art, and when the combination of technical solutions appears contradictory or unachievable, it should be considered that such combination of technical solutions does not exist and is not within the protection scope required by the application.

[0059] It should be understood that the above description is merely a detailed explanation of the preferred embodiments and is not intended to limit the patent protection scope of the present application. Any modification or alternation made by those skilled in the art without departing from the scope of the present application shall fall within the patent protection scope of the present application. The patent protection scope of the present application shall be subject to the appended claims.

Claims

1. A fault arc detection system based on multi-source fusion and federated self-learning, characterized in that: It includes sensors, data processing and feature extraction modules, edge AI inference modules, decision logic modules, cloud servers, and communication modules; The sensor is installed in the distribution box to monitor the status of the power distribution line, and synchronously collect the current and voltage waveforms and their high-frequency components generated when the circuit is working, as well as the light radiation, sound waves and temperature rise signals generated by the fault arc. The data processing and feature extraction module is deployed in the edge computing unit of the distribution box. It synchronously samples and preprocesses the raw signals collected by the sensors and extracts multi-dimensional feature parameters that characterize the fault arc from the multi-source signals, including time-domain waveform features, frequency-domain harmonic features, phase difference features, zero-point dwell time features, power factor changes and wavelet time-frequency features. The edge AI inference module is deployed in the edge computing unit and is used to receive multi-dimensional features output by the data processing and feature extraction module, and to infer in real time whether the current power distribution line is in a normal or fault arc state. The decision logic module is used to make a circuit breaker control decision based on the output of the edge AI inference module. When a fault arc is detected, the switching element in the power distribution protection device is triggered to cut off the circuit power supply in a predetermined sequence. The communication module is used for communication between devices in the system and between the devices and the cloud server.

2. The fault arc detection system based on multi-source fusion and federated self-learning according to claim 1, characterized in that: The sensors include a Rogowski coil current sensor, a high-resistance voltage divider, an EMI antenna, a photodiode, an acoustic sensor, and a temperature sensor.

3. The fault arc detection system based on multi-source fusion and federated self-learning according to claim 1, characterized in that: The edge AI inference model is an initial version of the model parameters obtained through offline training. During training, an online distillation self-learning mechanism with human feedback is introduced: when the judgment result ArcScore falls into the gray area and the on-site maintenance personnel confirm that the event is non-faulty through a mobile application, the multi-source original data of the time window corresponding to the event and the human confirmation label are written into the circular sample buffer built into the edge computing unit. When the buffer reaches a predetermined capacity, an online knowledge distillation update process is triggered: during this process, the device freezes the existing weights of the convolutional layers and the first few layers of the LSTM network in the edge AI inference model, and only fine-tunes the self-attention fusion layer and the output classification layer in the latter half of the model. During fine-tuning, the mini-batch gradient descent method is used, and the knowledge distillation loss function is used to consider both the original model output and the new human label. The two work together under a set temperature coefficient to gradually correct the model's ability to distinguish non-faulty modes under new loads.

4. The fault arc detection system based on multi-source fusion and federated self-learning according to any one of claims 1-3, characterized in that: The system also includes a federated self-learning module, an event aggregator, a model coordination module, and an OTA update server; The federated self-learning module is deployed in the edge computing unit and is used to update and optimize the edge AI model online using manually verified samples during system operation. The update information of the local model is extracted, compressed, and sent to the cloud server through the communication module for federated learning aggregation. The event aggregation and model coordination module is set on the cloud server. It is used to receive model update information sent by multiple edge devices, aggregate the updates of different devices using a federated averaging algorithm, and combine differential privacy mechanism to ensure the security and anonymity of data from each device. After aggregation, the updated global model is packaged into differential firmware and sent to the OTA update server of each device through the communication module; The OTA update server is used to store and manage the firmware versions of edge AI models, and performs firmware signature verification and online update replacement when the cloud server issues a new model differential patch.

5. The fault arc detection system based on multi-source fusion and federated self-learning according to claim 4, characterized in that: The federated self-learning module achieves multi-device model co-evolution through a mechanism combining local training and cloud aggregation. Specifically, it includes the following process: After each online distillation update, the edge device compresses the gradient of the local model parameter update: a power iterative stochastic gradient compression algorithm is used to perform low-rank decomposition on the gradient matrix, and the decomposed factor vectors are quantized with 4 bits, thereby generating a low-rank gradient packet of only tens of bytes in size; during the daily low-load period, this gradient packet is sent to the cloud federated service via LoRaWAN communication uplink. This communication uses a preset spreading factor to ensure 50 bytes per uplink. Data transmission within the specified range is secure and reliable with negligible impact on normal business operations. The event aggregation and model coordination module receives gradient updates from multiple terminal devices, uses a federated averaging algorithm to weight and aggregate all gradients, and then applies differential privacy noise to the aggregation result to ensure that sensitive data of a single device cannot be inferred from the global update. After aggregation, the cloud server compares the new model parameters under differential privacy protection with the current model to generate firmware differential data, divides it into several blocks, signs them, encrypts and sends them to the OTA update server of each device, and then each device verifies the signature to synthesize the upgrade model and replace the old model parameters.

6. A fault arc detection method based on multi-source fusion and federated self-learning, applied to the system described in any one of claims 1-4; characterized in that, Includes the following steps: Step 1: System startup and self-test; Step 2: Simultaneously acquire six signals—current, voltage, high-frequency EMI, optical, acoustic, and temperature—within a sliding window of t seconds, and buffer the data for one sliding window length for subsequent analysis; where t is a preset value. Step 3: Filter, denoise, and align the acquired raw signals from each channel to extract multi-dimensional feature representations, including features that convert some signals into image form; Step 4: Use an edge AI inference model to perform lightweight edge AI inference and obtain an evaluation value of the probability of fault arc occurrence under the current window; Step 5: Make a logical decision based on the output of the edge AI inference model. When it is determined that a fault arc has occurred, execute the circuit breaker disconnection action and record and report the event.

7. The fault arc detection method based on multi-source fusion and federated self-learning according to claim 6, characterized in that: In step 3, firstly, a Blackman window of half a power frequency cycle is applied to the acquired original time-domain current sequence to extract a signal segment. The synchronously extracted and phase-aligned long voltage sequences within the same window are then mapped onto the voltage-current phase plane coordinates to obtain a two-dimensional trajectory diagram reflecting the current-voltage relationship. Next, the two-dimensional trajectory diagram is converted into a grayscale image and scaled to a predetermined pixel size. Histogram equalization is then performed on the grayscale image to enhance the contrast of details at the sharp corners of the trajectory, thereby highlighting the abnormal waveform features caused by the fault arc. Then, the data processing and feature extraction module performs a short-time Fourier transform on the high-frequency current signal to convert the high-frequency components of the current into a time-frequency energy spectrum to obtain a time-spectrum frame with a size of 128×64. Meanwhile, for the optical and acoustic signals, the trigger time difference between the optical sensor and the acoustic sensor during the fault event is compared to calculate the time delay Δt of the optical-acoustic signal, and the ratio r of the amplitude of the optical signal to the acoustic signal is obtained to characterize the coupling strength between the arc light energy and the acoustic energy. The optical-acoustic pulse time difference Δt and the amplitude ratio r are encoded together with the temperature rise gradient ΔT / Δt measured by the temperature sensor before and after the event to form a five-dimensional statistical feature vector, which is used to characterize the relevant features of the changes of various physical quantities when the fault arc is generated. Thus, the data of each time window is processed to form a feature set containing the following modes: time-domain current and voltage numerical sequence, two-dimensional phase trajectory grayscale image, two-dimensional time spectrum image, and a set of multi-dimensional statistical features.

8. The fault arc detection method based on multi-source fusion and federated self-learning according to claim 6, characterized in that: In step 4, a lightweight edge AI inference model is used to perform edge AI inference. All features within the current time window are packaged into a tensor input X. A convolutional neural network performs multi-layer one-dimensional convolution operations on the one-dimensional time-series signal features in X to obtain an intermediate tensor C containing local spike patterns. C is then fed into a bidirectional long short-term memory network to extract bidirectional contextual features of the time series, outputting a time-series context tensor H. Tensors C and H are then fused along the channel dimension to form a comprehensive time-series feature tensor F. Simultaneously, for the two-dimensional image features contained in X, their spatial feature representation is extracted using a lightweight convolutional neural network to obtain an image feature tensor M. Next, the time-series feature tensor F and the image feature tensor M are fed into a multi-head self-attention fusion module. This module contains several heads with self-attention mechanisms, each head calculating and outputting a weight matrix α based on the query-key-value pair. i This is used to measure the contribution of each modal feature in the current window to the fault classification result; These attention weights are used to adaptively weight the features of the corresponding modalities to obtain the fused weighted tensor Z. Finally, the tensor Z is input into the classification output layer and normalized by Softmax to generate a four-dimensional fault state probability vector. ,in This represents the probability of normal operating conditions. This represents the probability of a series-type fault arc. This represents the probability of a parallel fault arc. Represents the probability of non-fault interference; defines the fault arc score. The larger value representing the probability of an arc fault occurring is taken as the comprehensive score.

9. The fault arc detection method based on multi-source fusion and federated self-learning according to claim 6, characterized in that: In step 5, the decision logic includes multi-level decision thresholds and a fast disconnection control mechanism. When the ArcScore of the inference results for three consecutive frames is not lower than the preset threshold A, the fault arc is determined to be established. The edge computing unit immediately drives the safety disconnection circuit through the decision logic module. First, it triggers the high-speed semiconductor switch SiC-MOSFET to turn off the current, and then cuts off the main contacts of the insulated gate bipolar transistor within a preset time to achieve rapid power disconnection of the fault arc current. When the ArcScore of the current frame is lower than the preset threshold B, it is judged to be in a normal state and no action is taken. For gray-zone frames where the ArcScore is between B and A, the system starts a secondary confirmation mechanism, prioritizes further analysis of the high-frequency EMI signal in the window to confirm whether there are hidden arc characteristics, and then decides whether to trip.

10. The fault arc detection method based on multi-source fusion and federated self-learning according to any one of claims 6-9, characterized in that: In step 5, after the fault determination is completed, the edge computing unit writes the multimodal raw waveform data of this event along with the determination result label into the local non-volatile memory through the event aggregator and model coordination module, and sends a JSON format data packet to the local or remote monitoring station through the communication module using the MQTT protocol. The data packet contains the fields of event type, occurrence time, ArcScore value, breaking delay, peak arc current and arc energy, so that the backend system can perform centralized statistical analysis and evidence traceability.

Images