Electrical Fault Detection Methods and Systems

By acquiring multiple physical quantities and generating causal structure graphs on edge devices, and combining them with a Bayesian deep learning model, unified detection of DC leakage current, AC arc and DC arc is achieved. This solves the problems of unified framework and cloud dependency of detection equipment in existing technologies, and realizes efficient and interpretable fault detection.

CN123115933APending Publication Date: 2026-11-17HUADA SEMICON CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611231978.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-14
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing electrical fault detection equipment cannot meet the needs of intelligent electrical networks in the complex physical world. It lacks a unified framework, the deep model is a black box with poor interpretability, it relies too much on the cloud, and it cannot simultaneously detect DC leakage, AC arc and DC arc. Furthermore, a single threshold judgment cannot quantify the risk of false alarms and false alarms.

Method used

Through the multi-physical quantity acquisition and physical feature extraction module, the lightweight causal discovery module generates a causal structure graph between variables. Combined with the Bayesian deep learning model, fault classification is performed to achieve fault detection and edge intelligent protection. The Bayesian anomaly diagnosis module is used for uncertainty quantification decision-making, and the edge self-learning and cloudless federation module performs incremental model learning.

Benefits of technology

It achieves millisecond-level unified detection of DC leakage current, AC arc and DC arc on edge devices, reduces the false alarm rate caused by load disturbance, provides a quantitative basis for risk classification, completely eliminates dependence on cloud computing power and communication links, and is suitable for battery-powered scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115933A_ABST
    Figure CN123115933A_ABST
Patent Text Reader

Abstract

This invention provides an electrical fault detection method and system, comprising: a multi-physical quantity acquisition and physical feature extraction module extracting features from a multi-source data set and outputting a multi-dimensional physical feature vector; a lightweight causal discovery module running a lightweight causal discovery algorithm based on the multi-dimensional physical feature vector to generate a causal structure graph between variables; and a Bayesian anomaly diagnosis module training a Bayesian deep learning model incorporating causal consistency constraints based on the causal structure graph between variables, so that the Bayesian deep learning model can classify faults in the multi-source data set. Combined with Bayesian neural network uncertainty quantification and hard-rule trip protection, it achieves unified fault handling for various fault types such as AC / DC arcs and leakage currents. Furthermore, the Bayesian deep learning model based on the causal structure graph between variables overcomes the problems of lack of physical causal support, poor black-box interpretability, lack of physical constraints, and only outputting labels in existing technologies.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of electrical fault detection and edge intelligent protection technology, and in particular to an electrical fault detection method and system. Background Technology

[0002] Current faults frequently occur in electrical equipment, including various types of faults such as arcing and leakage. However, existing technologies lack a unified framework (different faults require different fault identification devices). For example, traditional Type AC leakage current protectors (IEC 61008) can only detect AC leakage and cannot cover DC leakage.

[0003] Furthermore, relying solely on traditional protection devices to detect current for alarm purposes is no longer adequate for the fault identification needs of new electrical networks. The reasons are as follows: Arc detection schemes based on current change rate thresholds have high false alarm rates and do not differentiate between loads; one-dimensional convolutional neural network waveform classification methods rely on pure data-driven approaches, lack physical causal support, and have poor black-box interpretability (no physical constraints, only output labels); multi-sensor fusion schemes lack real fault samples (data-driven model training is difficult); and most existing AI leakage current solutions are based on cloud inference, are overly dependent on the cloud, and fail when the network is disconnected. Summary of the Invention

[0004] The purpose of this invention is to provide an electrical fault detection method and system to solve the problem that existing electrical fault detection equipment cannot meet the needs of intelligent electrical networks in the complex physical world.

[0005] To solve the above-mentioned technical problems, the present invention provides an electrical fault detection method, comprising: The multi-physical quantity acquisition and physical feature extraction module extracts features from the multi-source data set and outputs a multi-dimensional physical feature vector; The lightweight causal discovery module runs a lightweight causal discovery algorithm based on the multi-dimensional physical feature vector to generate a causal structure graph between variables; and The Bayesian anomaly diagnosis module trains a Bayesian deep learning model that incorporates causal consistency constraints based on the causal structure graph between variables, and then performs fault classification on the multi-source dataset based on the Bayesian deep learning model.

[0006] Optionally, in the electrical fault detection method, the multi-physical quantity acquisition and physical feature extraction module further includes a multi-physical quantity synchronous sampling module and a parameterized physical feature extraction module. The method also includes: The multi-physical quantity synchronous sampling module synchronously collects high-frequency current, high-frequency voltage, DC leakage current, and shell temperature to form the multi-source data set; The parameterized physical feature extraction module performs deep modeling of multi-source heterogeneous time-series data from multi-source datasets to form a parameterized physical model of arc leakage current; and The modules in the local edge controller run a lightweight causal discovery algorithm, a Yeats neural network uncertainty quantification, and hard rule trip protection based on a parameterized arc leakage physical model.

[0007] Optionally, in the electrical fault detection method, the Yeesian neural network uncertainty quantification and hard rule tripping protection include: The Bayesian anomaly diagnosis module performs fault classification to obtain fault results, which include: normal state, DC leakage fault, AC arc fault, and DC arc fault; the fault results are uncertainty quantification results. The multi-fault parallel decision-making and trip protection module obtains the uncertainty quantification results from the Bayesian anomaly diagnosis module, and performs hierarchical decision-making for each fault result based on the uncertainty quantification results, and directly drives the trip actuator through hardware GPIO; The edge self-learning and cloudless federated module completes incremental model learning locally, and encrypted and synchronizes model parameters across multiple devices via local area network for local inference, training, threshold updates, and event logging. The system displays fault results, causal structure diagrams between variables, and historical records locally through edge visualization and human-computer interaction modules.

[0008] Optionally, the electrical fault detection method further includes: Multi-source heterogeneous time-series data deep modeling of synchronously acquired multi-source datasets is as follows: in This is an environmental auxiliary observation vector, whose components include load type, battery state of charge (SOC), photovoltaic irradiance, and switching status; sampling frequency. satisfy For transient arc detection, take A sliding window of length L is denoted as: The above results are input into the multi-physical quantity acquisition and physical feature extraction module, the lightweight causal discovery module, the yaws anomaly diagnosis module, the multi-fault parallel decision-making and trip protection module, and / or the edge self-learning and cloudless federation module.

[0009] Optionally, the electrical fault detection method further includes: Based on the DC leakage current equivalent model, the sampling data of the AC main current sensor and the DC leakage current sensor are fused to obtain the DC leakage current risk score, insulation resistance risk score, leakage current risk score, and temperature rise risk score: DC leakage current risk score Insulation resistance risk score Leakage risk score Temperature rise risk score It is 100kΩ. Below the threshold At that time, the DC leakage current hard rule trip protection is directly triggered; Based on the Cassie-Mayr hybrid model of AC arc, the arc conductance is obtained using half-cycles as the basic detection unit. and arc voltage The score for a single half-cycle of an AC arc is calculated based on the arc conductance and arc voltage. Among them, based on high frequency energy 0-point shoulder period , peak state abrupt change in slope Obtain the score for a single half-cycle of the alternating current arc. The number of arcs to be judged within a 1-second window of 100 half-waves is: when Output AC arc alarm; C_ac(k) represents the cumulative number of half-waves judged to be arcs in the k-th 1-second sliding window or in 100 consecutive half-waves; C_ac(k) = Σ_{j=1}^{100} b_j Where b_j ∈ {0,1} is the single half-wave arc detection result of the j-th half-wave, 1 = arc present, 0 = no arc, and the threshold C_ac(k) ≥ 14 triggers the AC arc alarm. And execute the cutoff within 40 ms; A model for the continuous combustion of a DC arc is calculated based on the absence of a natural zero-crossing point in the DC arc, while also considering arc voltage drop. Current saturation MPPT perturbation and ripple energy Obtain the series DC arc score And parallel DC arc score ; The temperature co-model was used as post-validation evidence for the thermal risk of AC and / or DC arcs. .

[0010] Optionally, the electrical fault detection method further includes: organizing the above physical quantities into a multi-dimensional feature vector: Where I_leak is the instantaneous estimate of the DC leakage current; R_iso ​​is the estimated value of insulation resistance; E_leak represents the energy of the leakage event; ρ_T is the local temperature rise rate of the shell; E_HF represents the high-frequency noise energy in the 10-100 kHz range within half a cycle; S_z represents the "zero-point shoulder period" when the current approaches zero. Duration; K_HF represents the high-frequency component peak state; D_i is the index of the sudden change in the current rise slope; C_ac is the cumulative number of arcs detected within a 1-second window. An arc is detected when C_ac ≥ 14. ΔV_dc is the DC bus voltage sag. E_ripple is the ripple energy of 1-10 kHz within a 100 ms window; A_dc_series represents the half-wave score of a DC series arc. A_dc_parallel is the score of the half-wave of the DC parallel arc; Q_arc is the number of arc-determining evidence accumulated during the confidence period in Bayesian analysis; V_drop is the cumulative value of voltage drop events; I_spike is the cumulative value of current spike events; MPPT_amp represents the oscillating current component of the photovoltaic MPPT; P_S is the system auxiliary state parameter; The NOTEARS objective function with non-cyclic constraints is computed by the lightweight causal discovery module: Where W is the weighted adjacency matrix, and h(W) represents the acyclic constraint. For matrix exponents, The L1 regularization coefficient; Obtain the causal adjacency matrix ; Safety constraints: No action. Tripping, AC arc trip. DC arc trip, Warning} False tripping .

[0011] Optionally, the electrical fault detection method further includes: inputting a multidimensional feature vector into the Bayesian anomaly diagnosis module. The causal adjacency matrix G is used to enable the Bayesian anomaly diagnosis module to calculate and output the probability distribution of the fault results; .

[0012] Optionally, the electrical fault detection method further includes using a multi-fault parallel decision-making and tripping protection module to perform hierarchical decision-making for each fault based on uncertainty quantification results, including: .

[0013] Optionally, the electrical fault detection method further includes: incremental model learning performed locally by edge self-learning and a cloudless federated module. .

[0014] The present invention also provides an electrical fault detection system, comprising: The multi-physical quantity acquisition and physical feature extraction module is configured to extract features from multi-source data sets and output multi-dimensional physical feature vectors; The lightweight causal discovery module is configured to run a lightweight causal discovery algorithm based on the multidimensional physical feature vector to generate the causal structure graph between variables; The Bayesian anomaly diagnosis module trains a Bayesian deep learning model that incorporates causal consistency constraints based on the causal structure graph between variables, and then classifies faults in multi-source datasets based on the Bayesian deep learning model.

[0015] In the electrical fault detection method and system provided by this invention, a lightweight causal discovery algorithm is used to output a causal structure graph between variables. Based on the causal structure graph, a Bayesian deep learning model that integrates causal consistency constraints is trained to classify faults in multi-source datasets. This constructs a localized multi-physical quantity fault detection architecture driven by the lightweight causal discovery module and Bayesian uncertainty diagnosis, realizing unified fault handling for various fault types such as AC / DC arcs and DC leakage. Furthermore, the Bayesian deep learning model based on the causal structure graph between variables overcomes the problems of lack of physical causal support, poor black-box interpretability, lack of physical constraints, and only outputting labels in the prior art.

[0016] Given that electrical fault evolution has a clear physical causal chain, causal discovery algorithms can eliminate spurious correlations, thereby removing non-fault-related load fluctuation interference at the model input and improving feature robustness. This invention introduces a weighted prior distribution through a Bayesian neural network and employs Monte Carlo Dropout to approximate posterior inference, enabling the characterization of knowledge gaps and noise fluctuations in a single inference iteration, thus supporting differentiated protection strategies. All inference links are embedded within the edge controller, achieving low-latency response through data flow closed-loop.

[0017] This system runs entirely locally on the edge device and does not rely on any cloud services or edge computing. Cloud-based collaborative mechanism. All model inference, training, and decision-making are completed on MCUs, DSPs, or edge GPUs, adapting to deployment scenarios such as battery power, privacy sensitivity, and low bandwidth.

[0018] The inventors of this invention discovered through research that DC leakage current detection primarily serves scenarios such as electric vehicle charging piles, energy storage cabinets, photovoltaic inverters, and battery BMS, detecting DC-side insulation damage, leakage current, and low impedance to ground faults. AC arc fault detection is applied to AC distribution boxes, the AC side of photovoltaic systems, and household appliances, addressing issues such as series / parallel arcs, poor contact, and insulation aging arcs. DC arc fault detection targets the DC side of photovoltaic systems, energy storage batteries, and high-voltage circuits in electric vehicles, detecting DC series arcs, DC parallel arcs, and MPPT anomalies. DC leakage current is one of the most dangerous safety hazards in electric vehicles, energy storage, and photovoltaic systems. DC leakage current cannot be detected in conventional AC leakage current protectors (RCD / RCBO) because RCDs only recognize AC leakage current. DC leakage current continuously heats the insulation layer, potentially leading to thermal runaway and fire. Existing IEC 61851-1 and GB / T 18487.1 standards require that Mode 4 DC charging be cut off within 30 milliseconds, while leakage current is typically between 3 and 30 milliamps, requiring the extraction of a weak DC component from the charging current in the hundreds of amperes range.

[0019] AC arc fault detection also faces many challenges. AC arc current and series AC arc current last only half a week to several weeks, making them easily confused with normal loads such as frequency converters and LED switching power supplies. The UL 1699 standard requires operation within eight and a half cycles (approximately 67 milliseconds at 60 Hz).

[0020] The difficulty in detecting DC arc faults lies in the fact that DC arcs have no zero-junction point and will continue to burn once ignited. The large current fluctuations during the operation of the photovoltaic DC-side MPPT are similar to those of an arc. Therefore, they need to be identified separately. UL 1741 and IEC 63027 standards require disconnection within 0.5 seconds.

[0021] From the current technological perspective, AC residual current devices (IEC 61008) can only detect AC leakage current and cannot cover DC leakage current. Arc detection schemes based on current change rate thresholds have high false alarm rates and do not distinguish between load types. Although one-dimensional convolutional neural network waveform classification methods have made progress in academia, they are purely data-driven and lack physical causal interpretability. Multi-sensor fusion schemes are complex and expensive, and fixed weights cannot adapt. Most existing AI leakage current solutions are based on cloud inference and fail when the network is disconnected.

[0022] In summary, existing solutions suffer from three fundamental shortcomings. First, they lack a unified framework. The three types of faults exhibit significantly different characteristics, and existing solutions essentially offer "one solution for each fault," making simultaneous detection on a single device difficult. Second, deep learning models are black-boxes with poor interpretability, only outputting labels without pinpointing the root cause, making it difficult for maintenance personnel to take targeted measures. Third, existing AI solutions rely excessively on the cloud, becoming ineffective when the network is down, making them unsuitable for scenarios such as battery-powered electric vehicle charging stations. Furthermore, relying on a single threshold cannot quantify the risk of false alarms and missed alarms.

[0023] This invention synchronously collects multi-source data such as high-frequency current, voltage, and DC leakage current. Then, it uses the NOTEARS algorithm to learn the causal relationships between variables under the condition of satisfying acyclic constraints to generate a directed acyclic graph, i.e., a causal structure graph between variables. The causal structure graph between variables is used as a prior constraint input into a Bayesian neural network. Multiple random forward propagations are performed through Monte Carlo Dropout. Finally, the fault prediction mean, cognitive uncertainty, and random uncertainty are output locally on the edge device. This achieves millisecond-level unified detection of three types of faults: DC leakage current, DC arc, and AC arc. It effectively reduces the false alarm rate caused by load disturbances and provides a quantifiable risk classification basis for each diagnostic decision. At the same time, it completely eliminates the dependence on cloud computing power and communication links. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the electrical fault detection system architecture and operation method according to an embodiment of the present invention; Figure 2 This is a schematic diagram comparing the actual AC half-wave current and high-frequency residual of the electrical fault detection method and system in one embodiment of the present invention. Figure 3 This is a schematic diagram of the standardized mean difference of the 24-dimensional half-wave characteristics of the electrical fault detection method and system in one embodiment of the present invention between samples with and without arcs. Figure 4 This is a schematic diagram of a half-wave feature matrix heatmap composed of existing data in one embodiment of the present invention; Figure 5 This is a schematic diagram of the number of arcs detected within 100 consecutive half-waves in an embodiment of the electrical fault detection method and system of the present invention. Figure 6 This is a schematic diagram of the Cassie-Mayr parameterized AC arc dynamic curve of the electrical fault detection method and system in one embodiment of the present invention. Detailed Implementation

[0025] In this invention, the various embodiments are merely intended to illustrate the solutions of the invention and should not be construed as limiting.

[0026] It should also be noted that within the scope of this invention, the terms "same", "equal", and "equal to" do not mean that the two values ​​are absolutely equal, but allow for a certain reasonable error. In other words, the terms also cover "substantially the same", "substantially equal", and "substantially equal to".

[0027] Furthermore, the numbering of the steps in the methods of the present invention does not limit the execution order of the method steps. Unless otherwise specified, the method steps may be executed in different orders.

[0028] The electrical fault detection method and system proposed in this invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The advantages and features of this invention will become clearer from the following description and claims. It should be noted that the drawings are all in a very simplified form and use non-precise proportions, and are only used to facilitate and clarify the illustration of the embodiments of this invention.

[0029] The purpose of this invention is to provide an electrical fault detection method and system to solve the problem that existing electrical fault detection equipment cannot meet the needs of intelligent electrical networks in the complex physical world.

[0030] This invention belongs to the field of electrical fault detection and edge intelligent protection technology, specifically involving a unified detection system and method for multiple scenarios, including DC leakage current detection, AC arc fault detection, and DC arc fault detection.

[0031] This invention aims to solve five core problems. First, how to simultaneously detect three types of faults—DC leakage, AC arc, and DC arc—on a single edge device. Second, how to perform fusion inference of multiple physical quantities such as current, voltage, temperature, and leakage under strict latency constraints of 30 to 100 milliseconds. Third, how to quantify the confidence level of each detection decision and achieve risk classification and early warning through Bayesian uncertainty. Fourth, how to enable edge devices to autonomously complete incremental model learning and adapt to new fault modes without cloud access. Fifth, how to enable the detection algorithm to run on resource-constrained MCUs or edge GPUs, with the model size controlled within 1MB and inference latency below 50 milliseconds.

[0032] To achieve the above objectives, embodiments of the present invention provide an electrical fault detection method and system. Figure 1 An electrical fault detection method and system according to an embodiment of the present invention is shown. The method diagram includes: a multi-physical quantity acquisition and physical feature extraction module extracting features from a multi-source data set and outputting a multi-dimensional physical feature vector; a lightweight causal discovery module running a lightweight causal discovery algorithm based on the multi-dimensional physical feature vector to generate a causal structure graph between variables; and a Bayesian anomaly diagnosis module training a Bayesian deep learning model incorporating causal consistency constraints based on the causal structure graph between variables, so that the Bayesian deep learning model can perform fault classification on the multi-source data set.

[0033] One embodiment of the present invention proposes an electrical fault detection method, including a multi-physical quantity electrical fault detection system and method based on "causal discovery guidance, uncertainty quantification decision-making, and pure edge operation".

[0034] One embodiment of the present invention proposes a multi-physical quantity electrical fault detection system based on causal discovery and Yeats uncertainty quantification, including a multi-physical quantity synchronous sampling module, a parameterized physical feature extraction module, a lightweight causal discovery module, a Yeats anomaly diagnosis module, a multi-fault parallel decision-making and trip protection module, and an edge self-learning and cloudless federation module; Figure 2 This diagram illustrates a comparison of the actual AC half-wave current and high-frequency residual of an electrical fault detection method and system according to an embodiment of the present invention; wherein the multi-physical quantity synchronous sampling module outputs X(t), and the parameterized physical feature extraction module extracts the data according to... and ρ T Formula for generating F k The lightweight causal discovery module is based on the objective function and The constraint generates a causal graph G, and the Yeesian uncertainty diagnosis module outputs according to the MCDropout formula. The multi-fault parallel trip protection module is designed according to... and The formula outputs a trip signal.

[0035] Synchronous acquisition of X k Based on the DC leakage current equivalent circuit, the AC Cassie-Mayr arc equation, and the DC continuous arc equation, F is extracted. k G is learned through NOTEARS acyclic constraints; F is then used to learn G. k And G input to the Yeats neural network and execute T mc Secondary random forward propagation; based on and hard rule R B R AC R DC Fusion All steps are completed locally on the edge device.

[0036] DC leakage current estimated as The insulation resistance is estimated to be when Greater than 30 mA for 5 ms and DC leakage protection is directly triggered when the current is less than 100 kΩ.

[0037] The alternating current arc is modeled using the Cassie-Mayr hybrid equation and based on the half-wave high-frequency energy E. HF Zero point flat shoulder period S z , peak state K HF and slope D i Calculate A ac Satisfying within 100 consecutive half-waves The alternating current arc is determined at that time.

[0038] A_ac is the "score" (dimensionless) of a single half-wave of an AC arc, calculated by normalizing and weighting four physically differentiable features: high-frequency energy E_HF, zero-point shoulder period S_z, peak state K_HF, and slope abrupt change D_i. A larger A_ac indicates that the half-wave is more arc-like. When A_ac > a threshold, b_j = 1 (the single half-wave is judged as arc-like); otherwise, b_j = 0. Similarly, A_dc_series / A_dc_parallel are the single half-wave scores for DC series / parallel arcs, respectively.

[0039] DC arc adopts and Modeling, and with Distinguish between series arcs and parallel arcs. Figure 4 A schematic diagram of a half-wave feature matrix heatmap composed of existing data is shown in one embodiment of the present invention; temperature is not used as the sole basis for tripping, but as enhanced evidence for the a posteriori risk of electric arc.

[0040] Series arc (A_dc_series high): characterized by a voltage drop ΔV_dc > 1 V and current saturation (sudden decrease in load current); Parallel arc (A_dc_parallel high): characterized by a current spike I_spike > 3 × I_normal and an accompanying voltage drop > 0.5 V. Both rely on the Cassie-Mayr DC arc sustaining combustion model, have no natural zero crossing, and are accompanied by observable phenomena such as arc voltage drop, current saturation, MPPT perturbation, and ripple energy.

[0041] The output of the Yeats uncertainty diagnosis module and And based on uncertainty, suppress low-confidence tripping.

[0042] The loss of the causal discovery module includes and It is used to prohibit physically impossible causal edges and punish predictions that violate causal graphs.

[0043] Edge self-learning and cloudless federated modules update local parameters according to differential privacy gradients and aggregate them within the local area network according to the FedAvg formula, without going through a cloud server.

[0044] Total trip delay Furthermore, the hard rule for DC leakage current has a higher priority than the output of the Yeesian deep learning model.

[0045] The model is ≤500KB, and the time for a single feature extraction and yaws diagnosis is less than 50 ms. It is suitable for charging piles, energy storage cabinets, photovoltaic inverters, battery BMS and AC distribution boxes.

[0046] An embodiment of the present invention proposes an electrical fault detection method including multi-physical quantity edge acquisition and causal feature extraction. The inputs of the multi-physical quantity acquisition and physical feature extraction module include high-frequency current (not less than 100 kHz, for arc detection), high-frequency voltage (not less than 100 kHz), DC leakage current (such as a fluxgate DC residual current sensor or a type B RCD dedicated sensor with a sensitivity of not less than 1 mA), casing temperature (for arc combustion prediction), and equipment status parameters such as load type, battery SOC, photovoltaic irradiance, etc.

[0047] The processing flow of the multi-physical quantity acquisition and physical feature extraction module proposed in one embodiment of the present invention is divided into six stages. The first stage is multi-source signal synchronous sampling, with hardware synchronization of all channels and timestamps accurate to 1 microsecond. The second stage is DC leakage current component extraction, in which DC leakage current is extracted from the AC main current through low-pass filtering and DC integration, and fused with the original reading of the DC leakage current sensor to output leakage current estimate and insulation resistance estimate. The third stage is AC arc feature extraction, in which high-frequency components above 10 kHz are extracted from the 100 kHz current waveform, the randomness and abrupt rise slope of the "flat shoulder period" when the current is close to zero are calculated, and the arc noise energy within half a cycle is extracted. The fourth stage is DC arc feature extraction, in which the current ripple energy from 1 to 10 kHz within a 100-millisecond window is calculated, and voltage drop events (drops exceeding 1 volt and lasting less than 50 milliseconds) are extracted. The fifth stage is temperature-coordinated feature extraction, in which the local shell temperature rise rate is calculated. During normal operation, it is approximately 0.1 degrees Celsius per minute, and can reach 5 degrees Celsius per minute near the arc. The sixth step is causal discovery. The lightweight causal discovery module runs the lightweight NOTERAS algorithm on the edge device, pre-computes the causal graph offline, and only makes fine adjustments online. It learns causal relationships such as "DC leakage leads to temperature rise" and "electric arc leads to abnormal temperature rise", and outputs a dynamic causal graph G and a 16-dimensional feature vector F. Figure 3 This diagram illustrates the standardized mean difference of the 24-dimensional half-wave characteristics of the electrical fault detection method and system in an embodiment of the present invention between samples with and without arcs. An embodiment of the present invention proposes a multi-physical quantity acquisition and physical feature extraction module that ultimately outputs the original waveform segment (last 100 milliseconds), a 16-dimensional feature vector (including leakage current, insulation resistance, arc energy, voltage drop, current spike, temperature rise rate, etc.), and a causal graph G.

[0048] The 16-dimensional fault feature vector F is composed of 18-dimensional multi-source features. Figure 1 Remove P_S Auxiliary state parameters The physical observables retained after redundancy features constitute Specifically including The graph F has 16 components: I_leak, R_iso, E_leak, ρ_T, E_HF, S_z, K_HF, D_i, C_ac, ΔV_dc, E_ripple, A_dc_series, A_dc_parallel, Q_arc, V_drop, and I_spike. The dimensions of F are constrained and filtered by the causal graph G output by NOTEARS. See step S2 .

[0049] One embodiment of the present invention proposes an electrical fault detection method including Bayesian deep learning anomaly diagnosis incorporating causal priors. Specifically, the Bayesian anomaly diagnosis module receives the feature vector F from the multi-physical quantity acquisition and physical feature extraction module and the causal graph G from the lightweight causal discovery module. The Bayesian neural network structure consists of a 16-dimensional input layer, passing through a 64-dimensional fully connected layer (with Gaussian weights), a 128-dimensional LSTM layer, and another 64-dimensional fully connected layer, ultimately outputting a 4-dimensional vector corresponding to the probabilities of four fault types: normal, DC leakage, AC arc, and DC arc. Inference employs the MC Dropout method, performing 30 forward propagations per inference iteration to obtain 30 prediction distributions, outputting the prediction mean, cognitive uncertainty, and accidental uncertainty. During training, a causal consistency loss is introduced: if the causal graph G indicates that A is the cause of B, the model should not predict B as abnormal when A is normal. The loss function is the cross-entropy loss plus a causal penalty term.

[0050] One embodiment of the present invention proposes a Bayesian anomaly diagnosis module that employs a tiered mechanism for decision-making regarding operational uncertainty. When the mean DC leakage probability exceeds 0.7 and the cognitive uncertainty is below 0.2, a high-confidence DC leakage alarm is triggered. When the mean DC leakage probability exceeds 0.3 but the cognitive uncertainty is below 0.4, it is marked as pending review, requiring the accumulation of more sampling evidence. All other cases are considered normal.

[0051] One embodiment of the present invention proposes an electrical fault detection method including multi-fault parallel decision-making and trip protection. Specifically, the multi-fault parallel decision-making and trip protection module receives the diagnostic results from the Bayesian anomaly diagnosis module and performs cross-validation based on the rapid rules of three types of faults. At the hard rule level, the triggering condition for DC leakage current is that the DC leakage current is greater than 30 mA for more than 5 milliseconds, while the estimated insulation resistance is less than 100 kΩ and the temperature rise rate is more than 2 degrees Celsius per minute as auxiliary verification. The hard rule for AC arc is that the arc noise energy exceeds the threshold and occurs more than 5 times within 5 half cycles, and the auxiliary condition is that the current change rate exceeds 50 amperes per microsecond and the interval is less than 100 milliseconds. The electrical characteristics of DC series arc are mainly manifested as follows: in the steady-state combustion stage, the DC bus voltage does not have a significant steady-state drop, but is superimposed with broadband high-frequency noise and chaotic ripple in the 10 kHz to 1 MHz frequency band; the current waveform exhibits non-periodic distortion and high-frequency pulsation, and due to the introduction of the arc dynamic nonlinear resistance, the steady-state current amplitude may be lower than the expected value under arc-free conditions. The aforementioned high-frequency noise and waveform distortion characteristics are strongly correlated with load type, line parameters, and arc gap status, constituting the core criteria distinguishing it from normal operating conditions and other fault types. Auxiliary conditions include MPPT oscillation and temperature rise. The hard rule for DC parallel arcing is that the current spike exceeds three times the normal value and the voltage drop exceeds 0.5 volts. Decision fusion employs a Bayesian model and a hard rule voting mechanism; any confidence level exceeding 0.9 triggers a trip, with a trip signal delay of less than 10 milliseconds, directly driving a solid-state relay or contactor via hardware GPIO. The protection action timing is as follows: 0 to 30 milliseconds are initially determined by a hardware comparator for overcurrent or differential faults; 30 to 100 milliseconds are finely determined by an AI model; and 100 to 200 milliseconds are used to execute the disconnection action. The module ultimately outputs the trip signal, fault type and location, root cause information, and event records on the local SD card.

[0052] (1) DC leakage When I_dc > 30 mA for > 5 ms and R_iso ​​< 100 kΩ, the leakage current hard rule is triggered directly. Cut off within 30 ms (2) Alternating current arc E_HF exceeds the threshold and there are ≥5 instances of single half-wave arc detection within 5 and a half cycles. auxiliary conditions Current change rate > 50 A / μs and interval < 100 ms Cut off within 40 ms (3) The electrical characteristics of the electric arc are mainly manifested as follows: During the steady-state combustion stage, the DC bus voltage does not drop significantly, but is superimposed with broadband noise and chaotic ripple in the 10kHz to 1MHz frequency band. The current waveform exhibits non-periodic distortion and high-frequency pulsation, and due to the introduction of the arc's dynamic nonlinear resistance, the steady-state current amplitude may be lower than the expected value under arc-free conditions. The above-mentioned high-frequency noise and waveform distortion characteristics are strongly correlated with the load type, line parameters, and arc gap state, constituting the core criteria that distinguish it from normal operating conditions and other fault types. Support MPPT oscillation + temperature rise (4) DC parallel arc I_spike>3×I_normal and ΔV_dc>0.5 V;(5) DC arc confirmation A DC arc is confirmed when any parallel / series condition is met and ρ_T > 2 °C / min. Cut off within 2 seconds.

[0053] One embodiment of the present invention proposes an electrical fault detection method that includes edge self-learning and cloudless federation. Specifically, the edge self-learning and cloudless federation modules are responsible for processing continuously running data and maintenance tags. For local incremental learning, the Bayesian neural network is fine-tuned daily (no more than once) with new data. The fine-tuning data comes from local historical waveforms and manually labeled fault cases on-site. When multiple devices (such as multiple charging piles) are deployed on-site, model parameters are encrypted and synchronized via a local area network (Wi-Fi or Ethernet), completely bypassing any cloud server, and implemented using a pure local area network with the FedAvg algorithm. For local differential privacy, Laplacian noise (ε=1.0) is added to the model gradient to ensure that user electricity consumption data for a single charging pile is not leaked. During cold start, model pre-training is completed before leaving the factory, and the model is automatically loaded upon device startup, requiring no network connection for initial use.

[0054] One embodiment of the present invention proposes an electrical fault detection method that includes edge visualization and human-machine interaction. Specifically, the edge visualization and human-machine interaction module aggregates the outputs of all upstream modules. It can display fault type, confidence level, waveform, root cause, and cause-effect graph in real time on a 4.3-inch touchscreen, supporting historical data scrolling and local parameter configuration (leakage threshold, trip delay, etc.). Local alarms are implemented via LED indicators (green / yellow / red) and a buzzer; an optional SIM card module can be used to send SMS messages. Outputs include local user operation logs and maintenance work orders that can be manually exported to a USB flash drive.

[0055] One embodiment of the present invention also proposes an electrical fault detection system, which is a multi-physical quantity electrical fault detection system based on causal discovery and Bayesian uncertainty quantification. The system includes: a multi-physical quantity edge acquisition and causal feature extraction module, used to simultaneously acquire multi-source data such as high-frequency current, high-frequency voltage, DC leakage current, and casing temperature, and run a lightweight NOTEARS causal discovery algorithm locally on the edge device to learn the causal structure graph between variables, outputting a multi-dimensional feature vector based on the causal graph; a Bayesian anomaly diagnosis module, connected to the feature extraction module, using a Bayesian deep learning model incorporating causal consistency constraints to classify the feature vectors for faults (normal / DC leakage / AC arc / DC arc) and perform uncertainty quantification; a multi-fault parallel decision-making and trip protection module, connected to the Bayesian anomaly diagnosis module, performing hierarchical decisions for each fault based on the uncertainty quantification results, and directly driving the trip actuator through hardware GPIO; an edge self-learning and cloudless federated module, used to complete incremental model learning locally and encrypt and synchronize model parameters among multiple devices via a local area network, without relying on any cloud server; and an edge visualization and human-computer interaction module, used to locally display diagnostic results, causal graphs, and historical records.

[0056] An embodiment of the present invention proposes an electrical fault detection system that implements a multi-physical quantity electrical fault detection method based on causal discovery and Bayesian uncertainty quantification, including the following steps: S1, synchronously acquiring high-frequency current (not less than 100) on the edge device. (Multi-source data including) high-frequency voltage, DC leakage current, and casing temperature.

[0057] Let the sequence of multiple physical quantities acquired synchronously be: in This is an environmental auxiliary observation vector, whose components include load type, battery state of charge (SOC), photovoltaic irradiance, and switching status; sampling frequency. satisfy For transient arc detection, take A sliding window of length L is denoted as: The above results are input into the multi-physical quantity acquisition and physical feature extraction module, the lightweight causal discovery module, the yaws anomaly diagnosis module, the multi-fault parallel decision-making and trip protection module, and / or the edge self-learning and cloudless federation module.

[0058] S2 runs the lightweight NOTEARS causal discovery algorithm locally at the edge, constructs a directed acyclic causal graph G, and extracts a 16-dimensional fault feature vector F based on G.

[0059] Physical quantities are organized into multidimensional feature vectors: The meanings of the symbols in the above 18-dimensional multi-source physical quantity characteristic vector are as follows: I_leak — instantaneous estimated value of DC leakage current (A); R_iso ​​— estimated value of insulation resistance (Ω); E_leak — leakage event energy (J); ρ_T — local temperature rise rate of the casing (°C / min); E_HF — high-frequency noise energy of 10-100 kHz within half a cycle (10 ms); S_z — duration of the "zero-point flat shoulder period" when the current approaches zero; K_HF — high-frequency component kurtosis; Di — index of abrupt change in current rise slope; C_ac — cumulative number of arcs detected within a 1-second window (100 half-wave), the threshold of this patent is C_ac≥14 when an arc is detected (C = Count); ΔV_dc — DC bus voltage drop value (V); E_ripple — 1-10 kHz ripple energy within a 100 ms window; A_dc_series — DC series arc half-wave score (A = Arc score); A_dc_parallel — DC parallel arc half-wave score; Q_arc — Number of arc-identifying evidence accumulated at confidence level in Bayesian method (Q = Queue); V_drop — Cumulative value of voltage drop events; I_spike — Cumulative value of current spike events; I_MPPT — PV MPPT oscillating current component; P_S — System auxiliary state parameters (load type / SOC / irradiance / switching state).

[0060] Causal discovery uses the NOTEARS objective function with non-cyclic constraints: .

[0061] S3 inputs F into a pre-trained Bayesian deep learning diagnostic model, uses the MC Dropout method for inference, and outputs the fault prediction mean and cognitive uncertainty.

[0062] Yeats model input Given the causal adjacency matrix G, output the probability distribution of four states {normal, DC leakage, AC arc, DC arc}. .

[0063] S4. Make tiered decisions based on Bayesian model output and hard rule cross-validation: Immediately trip and locate the root cause when Bayesian confidence is high (mean greater than 0.7 and uncertainty less than 0.2); accumulate evidence and perform secondary confirmation when Bayesian confidence is medium; mark suspicious cases and continue monitoring when Bayesian confidence is low. S5. Directly drive solid-state relays or contactors via GPIO to execute tripping actions, with a total latency of less than 100 milliseconds.

[0064] μ is the mean of the predicted probabilities after 30 forward propagations of MC Dropout. σ represents the cognitive uncertainty of the prediction probability. Standard deviation This patent decision-making rule (1) High confidence If μ > 0.7 and σ < 0.2, trip immediately and locate the root cause. (2) Zhongzhixin 0.3 < μ < 0.7 or 0.2 < σ < 0.4 → Cumulative evidence Q_arc, secondary confirmation (3) Low confidence The rest → Mark as suspicious and continue monitoring. (Multi-fault parallel decision fusion) Tripping is triggered if the confidence level of any hard rule or Bayesian source is greater than 0.9.

[0065] The multi-fault parallel decision-making and tripping protection algorithm is as follows: S6 performs incremental model learning locally on the edge device and encrypts and synchronizes model parameters across multiple devices via a local area network. S7 saves fault diagnosis results, cause-effect graphs, and trip logs locally, and supports exporting to a USB drive. All steps are completed locally on the edge device, without relying on cloud services.

[0066] further The electrical fault detection system provided in the embodiments of the present invention further includes: DC leakage current detection, which includes extracting DC leakage current from the AC main current through low-pass filtering and DC integration, and using a dedicated DC leakage current sensor. The readings of the fluxgate current sensor are fused, and the judgment rule is that when the DC leakage current is greater than 30 mA for more than 5 milliseconds and the estimated insulation resistance is less than 100 kΩ, DC leakage is confirmed and the leakage is cut off within 30 milliseconds.

[0067] B+ DC leakage current equivalent model: Complementary fusion of AC main current and B+ dedicated leakage current sensor yields DC leakage current, insulation resistance, and leakage risk. It is 100kΩ. When the current is below the threshold, the DC leakage current hard rule is triggered directly; AC arc fault detection involves extracting the high-frequency components from 10 to 100 kHz in the 100 kHz current waveform, calculating the high-frequency noise energy within half a cycle (10 milliseconds), and confirming an AC arc when the noise energy exceeds the threshold and occurs more than 5 times within 5 half cycles, and cutting it off within 40 milliseconds.

[0068] The Cassie-Mayr hybrid model for alternating current arcs uses a half-cycle as the basic detection unit, and the arc conductance... Arc voltage ; Figure 5 This diagram illustrates the number of arc detections within 100 consecutive half-waves in an electrical fault detection method and system according to an embodiment of the present invention; as shown... Figure 5 As shown, the score for a single half-cycle of the alternating current arc. The number of arcs determined by high-frequency energy, zero-point flat shoulder period, peak state, and slope abrupt change is: [The number of arcs determined within a 1-second window is missing from the original text.] when Output AC arc alarm; DC arc fault detection includes calculating the 1 to 10 kHz current ripple energy within a 100-millisecond window, distinguishing between series and parallel arcs, and determining that a DC arc is confirmed when any condition is met and the temperature rise exceeds 2 degrees Celsius per minute, and is cut off within 2 seconds.

[0069] DC arc continuous combustion model: DC arc has no natural zero crossing point, and arc voltage drop, current saturation, MPPT disturbance and ripple energy are modeled simultaneously; Causal discovery employs the lightweight NOTEARS algorithm, with a local convergence time of less than 50 milliseconds on edge devices. The generated causal graph G is used to constrain the loss function of the Bayesian model during training: L = L_CE + α × L_causal, where L_causal penalizes predictions that contradict G.

[0070] L represents the total training loss of the Bayesian model. L_CE is the cross-entropy loss between the predicted probability and the true label. L_causal is the causality consistency penalty term. Punishment contradicts the prediction of the causal graph G G indicates that A is the reason for B. Therefore, if A is normal, B should not be predicted to be abnormal. α is the weighting coefficient of L_causal. This patent pre-sets hyperparameters Can be optimized through training Typical value: 0.1-1.0 .

[0071] The Bayesian deep learning diagnostic model is a 4-layer BNN with 16-dimensional input, 64-128-64-dimensional hidden layers, and 4-dimensional output (normal / DC leakage / AC arc / DC arc). The model size is less than 500KB, and the single inference time is less than 50 milliseconds, making it suitable for deployment on edge NPUs or DSPs.

[0072] Bayesian neural networks have a 4-layer structure. 16-dimensional input layer Corresponding 16-dimensional fault feature vector F First hidden layer: 64-dimensional fully connected layer Second hidden layer: 128-dimensional LSTM layer The third hidden layer is a 64-dimensional fully connected layer. 4D output layer Model size < KB Single MC Dropout Inference < ms Suitable for MCU deployment.

[0073] Edge self-learning and cloudless federated modules encrypt and synchronize model parameters across multiple devices via a local area network (Wi-Fi / Ethernet), using the FedAvg algorithm with differential privacy (ε=1.0 Laplace noise), without passing through any cloud server.

[0074] ε=1.0 is the differential privacy budget parameter. The smaller the ε value, the stronger the privacy protection. Typical applications use values ​​of 0.1-10. During gradient updates, Laplace noise with a mean of 0 and a scale of Δf / ε is applied. Δf is the gradient sensitivity. This makes training data from a single device unrecoverable. This patent (ε=1.0) balances privacy protection with model availability.

[0075] This fault detection method, applied in scenarios such as electric vehicle charging piles, energy storage battery cabinets, photovoltaic inverters, battery BMS, and household distribution boxes, completes all steps locally on edge devices. It boasts capabilities including a trip response time of less than 100 milliseconds, a model size of less than 1 MB, inference time of less than 50 milliseconds, low power consumption (suitable for battery-powered applications), and privacy protection (no data upload). The system includes a primary feature triggering circuit that only wakes up the main AI inference module when abnormal signs are detected (current surge rate exceeding a threshold or DC leakage current greater than 6 mA for more than 10 milliseconds), achieving an average power consumption of less than 50 milliwatts, suitable for long-term battery-powered scenarios.

[0076] The primary feature trigger circuit is used to wake up the main AI inference module. Reduce average power consumption <50 mW Awakening conditions (1) The current mutation rate di / dt exceeds the preset threshold. Typical 50 A / μs (2) Or, the DC leakage current I_dc > 6mA for > 10 ms. The system will wake up if either of these conditions is met. After being triggered, the main AI inference module runs at full speed. <50 ms inference .

[0077] The embodiments of this invention construct a localized multi-physical quantity fault detection architecture driven collaboratively by a lightweight causal discovery module and a Bayesian uncertainty diagnosis module. Specifically, the system synchronously collects multi-source data such as high-frequency current, voltage, and DC leakage current. Then, it uses the NOTEARS algorithm to learn the causal relationships between variables under the condition of satisfying acyclic constraints to generate a directed acyclic graph. This causal graph structure is used as a priori constraint input into a Bayesian neural network. Multiple random forward propagations are performed through Monte Carlo Dropout, and finally, the fault prediction mean, cognitive uncertainty, and accidental uncertainty are output locally at the edge device. Figure 6 A schematic diagram of the Cassie-Mayr parameterized AC arc dynamic curve of the electrical fault detection method and system according to an embodiment of the present invention is shown. In the figure, i_ac(t) is the instantaneous value of the AC arc current (A), u_arc(t) is the instantaneous value of the arc voltage (V), and R_arc(t) = 1 / g(t) is the transient resistance of the arc. Ω) Where g(t) is the Cassie-Mayr arc conductance. S This invention parameterizes the above equations into the feature extraction stage. This invention is used to generate differentiable physical features such as half-wave high-frequency energy E_HF and zero-point shoulder period S_z. In embodiments of this invention, the equivalent circuit of DC leakage current, the Cassie-Mayr hybrid arc equation, and the DC continuous arc equation are directly parameterized into the feature extraction stage, generating a multi-dimensional physical feature vector including half-wave high-frequency energy, zero-point shoulder period, and insulation resistance estimates. This endows the fault classifier with a close connection to physical interpretability, significantly improving the extraction accuracy of weak early fault signals and providing a highly discriminative structured input for subsequent causal learning and uncertainty quantification. Based on the characteristics of arc discharge plasma and insulation degradation heat... The electrical coupling mechanism transforms empirical formulas into differentiable computational nodes, enabling the feature space to be directly mapped to the physical essence of the fault. Based on existing technology, this type of physical prior injection can effectively overcome the adaptation bottleneck of pure data-driven models in scenarios with scarce samples, thereby ensuring the determinism of feature generation.

[0078] This invention designs a multi-fault parallel tripping protection module that logically fuses the probability and uncertainty quantities output by Bayesian analysis with rule quantities based on physical thresholds to construct a hierarchical decision-making mechanism. When the probability exceeds the set threshold and the uncertainty falls below the safety boundary, immediate tripping is triggered; otherwise, evidence accumulation or continuous monitoring begins. This balances the speed and selectivity of fault clearing, preventing erroneous or non-operational actions due to instantaneous data fluctuations and insufficient model confidence, achieving dynamic matching between protection actions and risk levels. It integrates the flexibility of probabilistic reasoning with the security of hard rules, utilizing uncertainty indicators as confidence gates. Based on existing technology, this soft-rule and hard-rule collaborative architecture can adaptively adjust protection sensitivity, ensuring the deterministic boundary of protection logic under complex operating conditions.

[0079] The embodiments of the present invention deploy a model self-learning unit locally on the edge device, which supports continuous optimization based on real-time monitoring data and incremental updates, and configures a local area network federation synchronization interface to realize encrypted transmission of model parameter gradients between multiple nodes.

[0080] Technical benefits: The system can continuously adapt to new failure modes and equipment aging trends in environments with no network or restricted privacy, extending the effective lifespan of the model, and achieving swarm intelligence enhancement without requiring the original data to leave the domain.

[0081] Technical Principle: This design leverages online learning algorithms and locally new samples for weight fine-tuning, combined with a parameter aggregation mechanism within a federated learning framework. Based on existing technology, this design, while meeting data compliance requirements and ensuring complementary features from multiple data sources, alleviates the problem of single-point device sample distribution bias, thereby maintaining the model's diagnostic efficiency over long periods of operation. The improvement of this invention does not lie in the abstract use of AI models, but in incorporating the electromagnetic / thermal physical equations of the three types of faults into feature generation, causal structure learning, Yeats uncertainty calculation and tripping rules, so that each functional module corresponds to a computable mathematical constraint, forming a quantitative process that can be implemented on MCU or edge NPU.

[0082] The technical problems solved by this invention include: uniformly detecting DC leakage current, AC arc, and DC arc in a single edge device; completing multi-physical quantity fusion inference within a protection time limit of 30 ms to 100 ms; outputting the mean probability, cognitive uncertainty, and random uncertainty for each judgment; using a causal graph constraint model to avoid misjudging load disturbances as faults; and still being able to perform local incremental learning under network outage and low power consumption conditions.

[0083] The beneficial effects of this invention are reflected in six aspects. First, it unifies the detection of three types of faults: DC leakage, AC arcing, and DC arcing, simultaneously covering a single edge device, reducing hardware costs by approximately 60%. Second, the tripping response time is less than 100 milliseconds, meeting the requirement of GB / T 18487.1 Mode 4 for disconnecting charging within 30 milliseconds. Third, it possesses interpretable causal reasoning capabilities, not only providing alarms but also identifying which sensor or circuit is malfunctioning. Fourth, it operates completely offline, working without cloud or network access, making it suitable for sensitive scenarios such as electric vehicle charging stations and energy storage. Fifth, it has high resource efficiency, with a model size of less than 500KB and an inference time of less than 50 milliseconds, making it suitable for MCUs such as Cortex-M7. Sixth, it has strong small-sample learning capabilities; causal priors combined with a Bayesian framework allow new fault types to be adapted with only 50 to 100 samples.

[0084] In this invention, the modules of the system according to the invention can be implemented using software, hardware, firmware, or a combination thereof. When a module is implemented using software, its function can be implemented through computer program flow. For example, the module can be implemented using code segments (such as code segments in languages ​​like C and C++) stored in a storage device (such as a hard disk, memory, etc.), wherein the corresponding function of the module can be implemented when the code segment is executed by a processor. When a module is implemented using hardware, its function can be implemented by setting a corresponding hardware structure. For example, the module's function can be implemented by hardware programming a programmable device such as a field-programmable gate array (FPGA), or by designing an application-specific integrated circuit (ASIC) that includes multiple transistors, resistors, capacitors, and other electronic devices. When a module is implemented using firmware, the module's function can be written into a read-only memory such as an EPROM or EEPROM in the form of program code, and the corresponding function of the module can be implemented when the program code is executed by a processor. In addition, some functions of the module may need to be implemented by separate hardware or by working in cooperation with the hardware. For example, the detection function is implemented by corresponding sensors (such as current sensors, voltage sensors, leakage current sensors, temperature sensors, arc sensors, etc.), the signal transmission function is implemented by corresponding communication devices (such as Bluetooth devices, infrared communication devices, baseband communication devices, Wi-Fi communication devices, etc.), the output function is implemented by corresponding output devices (such as displays, speakers, etc.), and so on.

[0085] While some embodiments of the present invention have been described in this application, those skilled in the art will understand that these embodiments are merely illustrative. Numerous variations, alternatives, and improvements will arise in those skilled in the art under the teachings of this invention without departing from its scope. The appended claims are intended to define the scope of the invention, and the methods and structures within the scope of its equivalents.

Claims

1. An electrical fault detection method, characterized in that, include: The multi-physical quantity acquisition and physical feature extraction module extracts features from the multi-source data set and outputs a multi-dimensional physical feature vector; The lightweight causal discovery module runs a lightweight causal discovery algorithm based on the multidimensional physical feature vector to generate a causal structure graph between variables; as well as The Bayesian anomaly diagnosis module trains a Bayesian deep learning model that incorporates causal consistency constraints based on the causal structure graph between variables, and then performs fault classification on the multi-source dataset based on the Bayesian deep learning model.

2. The electrical fault detection method as described in claim 1, characterized in that, The multi-physical quantity acquisition and physical feature extraction module further includes a multi-physical quantity synchronous sampling module and a parameterized physical feature extraction module. The method also includes: The multi-physical quantity synchronous sampling module synchronously collects high-frequency current, high-frequency voltage, DC leakage current, and shell temperature to form the multi-source data set; The parameterized physical feature extraction module performs deep modeling of multi-source heterogeneous time-series data from multi-source datasets to form a parameterized physical model of arc leakage current; and The modules in the local edge controller run a lightweight causal discovery algorithm, a Yeats neural network uncertainty quantification, and hard rule trip protection based on a parameterized arc leakage physical model.

3. The electrical fault detection method as described in claim 2, characterized in that, Yesse neural network uncertainty quantification and hard rule tripping protection include: The Bayesian anomaly diagnosis module performs fault classification to obtain fault results, which include: normal state, DC leakage fault, AC arc fault, and DC arc fault; the fault results are uncertainty quantification results. The multi-fault parallel decision-making and trip protection module obtains the uncertainty quantification results from the Bayesian anomaly diagnosis module, and performs hierarchical decision-making for each fault result based on the uncertainty quantification results, and directly drives the trip actuator through hardware GPIO; The edge self-learning and cloudless federated module completes incremental model learning locally, and encrypted and synchronizes model parameters across multiple devices via local area network for local inference, training, threshold updates, and event logging. The system displays fault results, causal structure diagrams between variables, and historical records locally through edge visualization and human-computer interaction modules.

4. The electrical fault detection method as described in claim 1, characterized in that, Also includes: Multi-source heterogeneous time-series data deep modeling of synchronously acquired multi-source datasets is as follows: in This is an environmental auxiliary observation vector, whose components include load type, battery state of charge (SOC), photovoltaic irradiance, and switching status; sampling frequency. satisfy ; For transient arc detection, take A sliding window of length L is denoted as: The above results are input into the multi-physical quantity acquisition and physical feature extraction module, the lightweight causal discovery module, the yaws anomaly diagnosis module, the multi-fault parallel decision-making and trip protection module, and / or the edge self-learning and cloudless federation module.

5. The electrical fault detection method as described in claim 4, characterized in that, Also includes: Based on the DC leakage current equivalent model, the sampling data of the AC main current sensor and the DC leakage current sensor are fused to obtain the DC leakage current risk score, insulation resistance risk score, leakage current risk score, and temperature rise risk score: DC leakage current risk score Insulation resistance risk score Leakage risk score Temperature rise risk score It is 100kΩ. Below the threshold At that time, the DC leakage current hard rule trip protection is directly triggered; Based on the Cassie-Mayr hybrid model of AC arc, the arc conductance is obtained using half-cycles as the basic detection unit. and arc voltage The score for a single half-cycle of an AC arc is calculated based on the arc conductance and arc voltage. Among them, based on high frequency energy 0-point shoulder period , peak state abrupt change in slope Obtain the score for a single half-cycle of the alternating current arc. The number of arcs to be judged within a 1-second window of 100 half-waves is: when Output AC arc alarm; C_ac(k) represents the cumulative number of half-waves judged to be arcs in the k-th 1-second sliding window or in 100 consecutive half-waves; C_ac(k) = Σ_{j=1}^{100} b_j Where b_j ∈ {0,1} is the single half-wave arc detection result of the j-th half-wave, 1 = arc present, 0 = no arc, and the threshold C_ac(k) ≥ 14 triggers the AC arc alarm. And execute the cutoff within 40 ms; A model for the continuous combustion of a DC arc is calculated based on the absence of a natural zero-crossing point in the DC arc, while also considering arc voltage drop. Current saturation MPPT perturbation and ripple energy Obtain the series DC arc score And parallel DC arc score ; The temperature co-model was used as post-validation evidence for the thermal risk of AC and / or DC arcs. 。 6. The electrical fault detection method as described in claim 5, characterized in that, Also includes: The above physical quantities are organized into a multidimensional feature vector: Where I_leak is the instantaneous estimate of the DC leakage current; R_iso ​​is the estimated value of insulation resistance; E_leak represents the energy of the leakage event; ρ_T is the local temperature rise rate of the shell; E_HF represents the high-frequency noise energy in the 10-100 kHz range within half a cycle; S_z represents the "zero-point shoulder period" when the current approaches zero. Duration; K_HF represents the high-frequency component peak state; D_i is the index of the sudden change in the current rise slope; C_ac is the cumulative number of arcs detected within a 1-second window. An arc is detected when C_ac ≥ 14. ΔV_dc is the DC bus voltage sag. E_ripple is the ripple energy of 1-10 kHz within a 100 ms window; A_dc_series represents the half-wave score of a DC series arc. A_dc_parallel is the score of the half-wave of the DC parallel arc; Q_arc is the number of arc-determining evidence accumulated during the confidence period in Bayesian analysis; V_drop is the cumulative value of voltage drop events; I_spike is the cumulative value of current spike events; MPPT_amp represents the oscillating current component of the photovoltaic MPPT; P_S is the system auxiliary state parameter; The NOTEARS objective function with non-cyclic constraints is computed by the lightweight causal discovery module: Where W is the weighted adjacency matrix, and h(W) represents the acyclic constraint. For matrix exponents, The L1 regularization coefficient; Obtain the causal adjacency matrix ; Safety constraints: No action. Tripping, AC arc trip, DC arc trip, Warning} False tripping 。 7. The electrical fault detection method as described in claim 6, characterized in that, Also includes: Input multidimensional feature vectors into the Bayesian anomaly diagnosis module The causal adjacency matrix G is used to enable the Bayesian anomaly diagnosis module to calculate and output the probability distribution of the fault results; 。 8. The electrical fault detection method as described in claim 7, characterized in that, It also includes using a multi-fault parallel decision-making and tripping protection module to perform hierarchical decisions for each fault based on uncertainty quantification results, including: 。 9. The electrical fault detection method as described in claim 8, characterized in that, This also includes: incremental model learning performed locally by edge self-learning and cloudless federated modules. 。 10. An electrical fault detection system, characterized in that, include: The multi-physical quantity acquisition and physical feature extraction module is configured to extract features from multi-source data sets and output multi-dimensional physical feature vectors; The lightweight causal discovery module is configured to run a lightweight causal discovery algorithm based on the multidimensional physical feature vector to generate the causal structure graph between variables; The Bayesian anomaly diagnosis module trains a Bayesian deep learning model that incorporates causal consistency constraints based on the causal structure graph between variables, and then classifies faults in multi-source datasets based on the Bayesian deep learning model.