Electrical fault identification system and method

Abstract

The application provides an electrical fault identification system and method, comprising: a physical information contrast learning encoder module based on physical information guided contrast learning, which extracts transient fault features and enhances transient features for multi-modal input signals, and outputs a low-dimensional fault feature vector; a safety constraint reinforcement learning decision module receiving the low-dimensional fault feature vector, based on a safety constraint Jarovoy decision process for adaptive decision and action optimization, generating a preliminary action suggestion; a super-low-delay edge inference and hardware acceleration module integrating a lightweight version of the encoder module and the decision module, performing control barrier function safety checking on the preliminary action suggestion, and outputting a final control instruction, to perform full-process edge local pipeline processing from signal acquisition to control instruction generation, deeply integrating physical information contrast learning, safety constraint reinforcement learning, multi-modal signal fusion including current, voltage, leakage, temperature and arc light, and real-time edge intelligence.

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 identification system and method. 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 identification system and method 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 identification system, comprising: The physical information contrastive learning encoder module is configured to perform contrastive learning based on physical information, extract instantaneous fault features and enhance transient features from multimodal input signals, and output a low-dimensional fault feature vector. A safety-constrained reinforcement learning decision module is configured to receive the low-dimensional fault feature vector, perform adaptive decision-making and action optimization based on a safety-constrained Gorkov decision process, and generate preliminary action suggestions; and The ultra-low latency edge inference and hardware acceleration module is configured as a lightweight version of the integrated encoder module and decision module. After performing a control barrier function security check on the preliminary action suggestion, it outputs the final control command to perform full-process edge local pipeline processing from signal acquisition to control command generation.

[0006] Optionally, in the aforementioned electrical fault identification system, The physical information contrastive learning encoder module includes a physical information fault generation module and a TCN-Transformer contrastive learning encoder module, wherein: The physical information fault generation module is configured to generate local samples according to the Cassie-Mayr AC arc equation, the photovoltaic DC arc and MPPT disturbance equation, and the DC leakage current equivalent circuit equation. The local samples include high-speed sampling current waveform data, high-speed sampling voltage waveform data, and DC leakage current sensor data. The TCN-Transformer contrastive learning encoder module is configured to train a framework for contrastive learning of local physical model generated data and physical consistency constraint loss, so as to convert high-speed sampled current waveform data and high-speed sampled voltage waveform data into low-dimensional feature vectors. The security constraint reinforcement learning decision module includes a policy network and a control barrier function security verification module, wherein: The policy network is connected to the TCN-Transformer contrastive learning encoder module to receive the low-dimensional feature vector and system state. The policy network is trained using a local safety-constrained Markov decision process based on the low-dimensional feature vector and system state to output preliminary action suggestions. The control barrier function security verification module performs formal security verification on the preliminary action suggestion based on the control barrier function and outputs the final control command. The ultra-low latency edge inference and hardware acceleration module includes an edge low-power pipeline module and a local fine-tuning module. The edge low-power pipeline module integrates a lightweight version of the TCN-Transformer contrastive learning encoder module and the safety constraint reinforcement learning decision module. It performs full pipeline processing on the signal acquisition of high-speed sampled current waveform data and high-speed sampled voltage waveform data and the generation of control instructions for high-speed sampled current waveform data and high-speed sampled voltage waveform data, and is equipped with a primary feature triggering circuit to reduce average power consumption. The ultra-low latency edge inference and hardware acceleration module runs locally on the edge device.

[0007] Optionally, in the aforementioned electrical fault identification system, the physical information comparison learning encoder module locally generates three types of physical samples, including AC arc samples x. ac DC arc sample x dc and DC leakage current sample x B ; Pair the real samples with the physical samples to form positive and negative samples; Train the TCN-Transformer contrastive learning encoder to obtain a low-dimensional feature vector z; input the low-dimensional feature vector z and the system state s into the policy network. θ ; Screening based on the security verification formula of the control barrier function and output All steps are performed locally on the edge device. Let the input multimodal segment be: Window W is 1 second, sampling frequency or .

[0008] Optionally, in the aforementioned electrical fault identification system, the AC arc sample is generated by the Cassie-Mayr mixed differential equation and determined by the arc length l, separation speed, and load impedance Z. L The sample family is formed by the medium parameter χ and the random perturbation ζ. Cassie equations in the high current / high power region Mayr equations for low current / low power regions Hybrid switching conditions: when Use the Cassie equation; when When using Mayr equations Arc pressure equation: Arc length update discretization Medium parameter correction therefore DC arc sample by , Generate and explicitly add To distinguish between MPPT disturbances and real DC arcs; Photovoltaic IV characteristics (single diode model) DC arc voltage drop MPPT perturbation model: Series arc criterion: And current saturation Parallel arc criterion: and DC leakage current sample from And the temperature rise equation is generated; .

[0009] Optionally, in the aforementioned electrical fault identification system, local data enhancement and noise superposition include: Contrast learning loss is InfoNCE Contrast Loss: in τ is the temperature coefficient Physical distance constraint loss: Upper limit of positive sample distance: in Distance using physical parameters The formula for constraining the distance between positive samples in the embedding space and the physical information contrastive learning formula includes: Positive samples are those that meet the same physical fault type, similar physical parameters, and similar safety consequences; negative samples include normal MPPT disturbances, startup inrush current, LED switching power supply noise, and / or non-faulty load switching.

[0010] Optionally, in the aforementioned electrical fault identification system, the action space of the safety constraint reinforcement learning decision module includes no action, DC leakage trip, AC arc trip, DC arc trip, and early warning. The reward function includes a reward for correct action, a penalty for false trip, a penalty for refusal to operate, and a delay penalty, and satisfies... Safety constraints; Fault action decision modeling is a safety-constrained Markov decision process, and the safety-constrained reinforcement learning formula includes: Optimization goal: Safety constraints: Annual false trip rate (st) DC leakage detection delay AC arc detection delay DC arc detection delay Lagrange duality Safety constraints: No action. Tripping, AC arc trip, DC arc trip, Warning} False tripping .

[0011] Optionally, in the aforementioned electrical fault identification system, the control barrier function safety verification module defines hB, and hT, and according to Filtering the set of safety actions ; When the policy network output contradicts hard security constraints, the security barrier overrides the policy output; the security verification rules for the control barrier function include: Define leakage current safety barrier Temperature safety barrier Inrush protection window Safety Action Set Filtering Rules: like Right now If the failure persists for more than 5ms, a full trip will be forcibly executed. Without AI decision-making; like Right now If the current is exactly 30mA and lasts for more than 5ms, an alert will be triggered. And start a confirmation timer to prevent inaction. ; like Right now If the leakage current barrier is not activated, the policy network can output normally. Do nothing; like Right now Then the main circuit breaker will be forced to trip. ; like Right now During the inrush current period at the start of charging, tripping actions are prohibited. , , Only allowed or ; If safety action set If a conservative backoff action is executed, the entire circuit breaker will trip. ; like If empty, then Take conservative tripping or early warning actions.

[0012] Optionally, in the aforementioned electrical fault identification system, the edge low-power pipeline satisfies Average power consumption The model size is less than 500KB; the low-latency edge pipeline formula includes: The three-level triggering is as follows: L1 hardware comparator is normally open, detecting di / dt, Ileak, and dT / dt; L2 encoder is woken up on demand to confirm fault symptoms; L3 policy network and CBF are activated only in case of an anomaly and output GPIO trip.

[0013] Optionally, in the aforementioned electrical fault identification system, the local fine-tuning module, based on... Update the encoder and policy network, and perform encryption parameter synchronization within the local area network; the local fine-tuning and local area network synchronization formulas include: The present invention also provides an electrical fault identification method, comprising: Based on physical information-guided contrastive learning, instantaneous fault features are extracted and transient features are enhanced from multimodal input signals, and a low-dimensional fault feature vector is output. Based on the low-dimensional fault feature vector, adaptive decision-making and action optimization are performed using a safety-constrained Morkov decision process to generate preliminary action suggestions; and A lightweight version integrating the encoder module and decision module outputs the final control command after performing a control barrier function safety check on the preliminary action suggestion, so as to perform full-process edge local pipeline processing from signal acquisition to control command generation.

[0014] In the electrical fault identification system and method provided by this invention, a low-dimensional fault feature vector is obtained by the physical information-informed contrastive learning encoder module through contrastive learning of instantaneous fault feature extraction and transient feature enhancement. The safety-constrained reinforcement learning decision module performs adaptive decision-making and action optimization based on the low-dimensional fault feature vector and the safety-constrained reinforcement learning model. The ultra-low latency edge inference and hardware acceleration module performs ultra-low latency edge inference and hardware acceleration based on adaptive decision-making and control commands. It deeply integrates physical information-informed contrastive learning, safety-constrained reinforcement learning, multimodal signal fusion (current, voltage, leakage current, temperature, arc light) and real-time edge intelligence. It aims to solve the problem of accurate detection and safety decision-making under difficult conditions such as weak instantaneous fault signals, non-stationary, strong noise interference environment, cloud network incompatibility, and DC leakage incompatibility. The response time is less than 100 milliseconds.

[0015] Specifically, instantaneous feature extraction and instantaneous feature enhancement to obtain a safety constraint reinforcement learning model can solve the problems of weak, non-stationary, and strongly noisy environments with instantaneous fault signals; ultra-low latency edge inference and hardware acceleration solve the problem of existing solutions being overly dependent on the cloud and failing when the network is disconnected; the TCN-Transformer contrastive learning encoder module trains a contrastive learning framework between locally generated physical model data and physical consistency constraint loss to convert high-speed sampled current waveform data and high-speed sampled voltage waveform data into low-dimensional feature vectors, solving the problem of DC leakage incompatibility; ultra-low latency edge inference and hardware acceleration solve the problem of existing solutions being overly dependent on the cloud and failing when the network is disconnected.

[0016] The inventors of this invention discovered through research that the core challenge in DC leakage current detection lies in the fact that DC leakage current cannot be detected in conventional AC residual current devices (RCDs). DC leakage current continuously heats the insulation layer, easily leading to thermal runaway. The IEC 61851-1 Mode 4 DC charging standard requires disconnection within 30 milliseconds and needs to extract 3 to 30 milliamps of leakage current from a mains current of hundreds of amperes.

[0017] The challenge in detecting AC arc faults lies in the short duration of the arc current (half a cycle to several weeks), which is highly similar to the characteristics of normal loads such as frequency converters and LED switching power supplies. The UL 1699 standard requires the AFCI to operate within 2.5 half cycles after detecting a series arc (approximately 21 milliseconds at 60Hz), but the false alarm rate is relatively high.

[0018] The unique characteristic of DC arc fault detection is that DC arcs have no zero-junction point; once ignited, they continue to burn, and their characteristics are similar to the fluctuations in the operating current of an MPPT (Multi-Level Photodetector). UL1741 and IEC63027 standards require that the arc be cut off within 2 seconds.

[0019] Existing technologies include one-dimensional convolutional neural network arc detection, random forest DC arc recognition, mathematical morphology frequency band energy recognition, multimodal collaborative diagnosis, and Yeats localization, but none of these fully solve the problem of instantaneous fault identification in intelligent electrical networks in the complex physical world. This invention does not focus on the general concept of "using AI to identify faults," but rather distinguishes itself by its edge-locally computable physical generation model, physical consistency comparative learning, CMDP safety constraint strategy optimization, and control barrier functions. It operates entirely locally on edge devices, without relying on cloud services or requiring an edge-cloud collaborative mechanism. All model inference, training, decision-making, and protection actions are completed on MCUs, DSPs, or edge NPUs. By integrating technologies such as instantaneous electrical fault identification, edge intelligent protection, physical information machine learning, comparative learning, and safety reinforcement learning, it achieves a highly sensitive, reliable, and rapid system and method for multimodal instantaneous fault identification, including DC leakage faults, AC arc faults, and DC arc faults. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the operation method of an electrical fault identification system 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 an electrical fault identification 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 feature of the electrical fault identification 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 electrical fault identification system according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the parameterized AC arc dynamic curve of the Cassie-Mayr electrical fault identification system in one embodiment of the present invention. Detailed Implementation

[0021] In this invention, the various embodiments are merely intended to illustrate the solutions of the invention and should not be construed as limiting.

[0022] 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".

[0023] 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.

[0024] The electrical fault identification system and method 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.

[0025] This invention addresses five key issues. First, how to simultaneously detect three types of faults—DC leakage, AC arcing, and DC arcing—on a single edge device. Second, how to perform multimodal fusion inference under latency constraints of 30 to 100 milliseconds. Third, how to ensure the edge device can still function even when permanently offline (no cloud, no network, battery powered). Fourth, how to perform adaptive model learning locally at the edge without relying on cloud GPUs. Fifth, how to train a model with strong generalization ability using a very small number of real samples (50 to 100).

[0026] To address the aforementioned issues, this invention proposes a multimodal transient fault identification system that integrates "physical information enhancement representation, security reinforcement learning decision-making, and real-time edge response." The core system comprises three main modules, all running locally on edge devices, to solve the problem that existing electrical fault detection equipment cannot meet the needs of intelligent electrical networks in the complex physical world.

[0027] To achieve the above objectives, embodiments of the present invention provide an electrical fault identification system. Figure 1 A schematic diagram of the operation method of an electrical fault identification system according to an embodiment of the present invention is shown, as follows: Figure 1 As shown, it includes: a physical information contrastive learning encoder module 1, configured to extract instantaneous fault features and enhance transient features based on physical information guidance, to obtain a safety constraint reinforcement learning model; a safety constraint reinforcement learning decision module 2, configured to perform adaptive decision-making and action optimization based on the safety constraint reinforcement learning model, and generate control commands; and an ultra-low latency edge inference and hardware acceleration module 3, configured to perform ultra-low latency edge inference and hardware acceleration according to the safety constraint reinforcement learning model and the control commands.

[0028] Figure 2This invention illustrates a schematic diagram comparing the actual AC half-wave current and high-frequency residual in an electrical fault identification system according to an embodiment of the present invention. In this embodiment, an electrical fault identification system is provided, wherein the physical information comparison learning encoder module includes a physical information fault generation module and a TCN-Transformer comparison learning encoder module, wherein: The physical information fault generation module is configured to generate local samples according to the Cassie-Mayr AC arc equation, the photovoltaic DC arc and MPPT disturbance equation, and the DC leakage current equivalent circuit equation. The local samples include high-speed sampling current waveform data, high-speed sampling voltage waveform data, and DC leakage current sensor data. The TCN-Transformer contrastive learning encoder module trains a contrastive learning framework between local physical model generated data and physical consistency constraint loss to convert high-speed sampled current waveform data and high-speed sampled voltage waveform data into low-dimensional feature vectors. The security constraint reinforcement learning decision module includes a policy network and a control barrier function security verification module, wherein: Figure 3 The diagram illustrates the standardized mean difference of 24-dimensional half-wave features of an electrical fault identification system in an embodiment of the present invention between samples with and without arcs. The policy network is connected to the TCN-Transformer contrastive learning encoder module to receive the low-dimensional feature vector and system state. The policy network is trained using a local safety-constrained Markov decision process based on the low-dimensional feature vector and system state to output preliminary action suggestions. The control barrier function security verification module performs formal security verification on the preliminary action suggestion based on the control barrier function and outputs the final control command. 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; the ultra-low latency edge inference and hardware acceleration module includes an edge low-power pipeline module and a local fine-tuning module. The edge low-power pipeline module integrates a lightweight version of the TCN-Transformer contrastive learning encoder module and the safety constraint reinforcement learning decision module. The entire process pipeline processes the signal acquisition of high-speed sampled current waveform data and high-speed sampled voltage waveform data to the generation of control commands, and is equipped with a primary feature triggering circuit to reduce average power consumption. The ultra-low latency edge inference and hardware acceleration module runs locally on the edge device.

[0029] The embodiments of this invention provide a unified definition for parameters in the formulas that have specific meanings under this patent, engineering constraints, or require understanding in conjunction with this application. Well-known mathematical symbols in the art, such as integrals, summations, exponential functions, norms, gradients, softmax, sigmoid, Fourier transforms, convolutions, means, and variances, will not be explained again. Unless otherwise stated, current is measured in A, voltage in V, time in s, temperature in degrees Celsius, and power in W.

[0030] Figure 5 This diagram illustrates the number of arc detections within 100 consecutive half-waves in an electrical fault identification system according to an embodiment of the present invention. In this electrical fault identification system, the physical information comparison learning encoder module locally generates AC arc samples x. ac DC arc sample x dc DC leakage current sample x B Three types of physical samples; Pair the real samples with the physical samples to form positive and negative samples; Train the TCN-Transformer encoder to obtain a low-dimensional feature vector z; input the low-dimensional feature vector z and the system state s into the policy network. θ ; Screening based on the security verification formula of the control barrier function and output All steps are performed locally on the edge device. Let the input multimodal segment be: Window W is 1 second, sampling frequency or .

[0031] X_k=[i(t),v(t),i_leak(t),T(t),L_arc(t)]: The k-th input multimode segment; i(t) is the AC or DC main circuit current, v(t) is the main circuit voltage, i_leak(t) is the DC leakage current, T(t) is the temperature of the housing or power device, and L_arc(t) is the arc / optical sensor signal.

[0032] t_k: End time of the kth sampling window; W: Window length, which is 1 second in this embodiment; f_s: Sampling frequency, which is not less than 100kHz in the AC waveform embodiment and not less than 1MHz in the embodiment involving high-frequency physical modeling; Δt: Time step of discrete simulation.

[0033] B_r: the r-th circular buffer; D_MA: moving average or digital moving average operator; X_k=B_r[n:n+W] means extracting an input segment of length W from the circular buffer.

[0034] Figure 6 This diagram illustrates the dynamic curve of a Cassie-Mayr parameterized AC arc in an embodiment of the electrical fault identification system of the present invention. The horizontal axis represents time (ms), covering at least two complete AC cycles (40ms @ 50Hz); the vertical axis (left) represents the arc current iarc (A); the vertical axis (right) represents the arc voltage Uarc (V); the labeled areas are: Cassie region (│iarc│ > Ith): high current range; Mayr region (│iarc│ ≤ Ith): near zero crossing; arc extinguishing / reigniting occurs at the current zero crossing point; parameter labels: arc length l, time constant τ. c / τ m Power parameter P c / P m In the aforementioned electrical fault identification system, AC arc samples are generated by the Cassie-Mayr mixed differential equation and determined by factors such as arc length l, separation velocity, and load impedance Z. L The sample family is formed by the medium parameter χ and the random perturbation ζ. g: Arc equivalent conductance; g_dot: Rate of change of arc conductance with time; u_arc: Arc voltage; i_arc: Arc current; P_c, P_m: Arc power scaling parameters for the Cassie high-power region and the Mayr low-power region, respectively; τ_c, τ_m: Arc thermal inertia / dynamic time constants for the two operating regions, respectively; g_c, g_m: Steady-state conductance scaling for the corresponding operating regions.

[0035] U_c: Basic arc voltage parameter in the Cassie region; R_c: Equivalent dynamic resistance in the Cassie region; κ_l: Gain coefficient of arc length on arc voltage; l: Arc length; Z_L: Equivalent impedance of the load; v_s: Equivalent voltage on the power supply side before the arc; ξ_u, ξ_l: Random disturbances acting on arc voltage and arc length, respectively.

[0036] v_sep: Electrode or contact separation speed; l_k, l_{k+1}: Arc length at adjacent discrete moments; P_c0, τ_c0: Cassie power parameters and time constants under reference conditions; χ_p, χ_τ: Correction coefficients for power parameters and time constants based on medium parameters; χ: Medium / environmental parameter, representing the influence of gas medium, pressure, or equivalent heat dissipation conditions on arc dynamics.

[0037] θ_ac=[l,v_sep,Z_L,P_c,P_m,τ_c,τ_m]: AC arc simulation parameter vector; n_real(t): background noise estimated from real collected data; x_ac(t;θ_ac): AC arc sample generated from the above parameters and superimposed with real noise.

[0038] DC arc sample by , Generate and explicitly add To distinguish between MPPT disturbances and real DC arcs; I_pv, V_pv: photovoltaic array operating current and terminal voltage; V_bus: DC bus voltage; V_arc: voltage drop caused by DC arc; V_mppt: voltage disturbance introduced by maximum power point tracking control.

[0039] V_a: DC arc base voltage drop; r_a: arc equivalent dynamic resistance; κ_l: arc length voltage gain coefficient; l_arc: DC arc length; I_ph: photocurrent; I_0: diode reverse saturation current; R_s, R_sh: photovoltaic equivalent series resistance and parallel resistance; a: diode ideality factor; V_T: thermal voltage.

[0040] A_m: MPPT perturbation amplitude; f_m: MPPT perturbation frequency; φ_m: perturbation initial phase; b_m: sawtooth wave perturbation amplitude; saw(t): normalized sawtooth wave; ξ_m(t): MPPT control noise. θ_dc=[I_pv,V_bus,E_ripple,ΔV]: DC arc sample parameters / observation vector; E_ripple: DC ripple energy; ΔV: voltage drop caused by the arc.

[0041] DC leakage current sample from And the temperature rise equation is generated; .

[0042] V_B+: DC bus positive terminal voltage; V_body: voltage of the human body, casing, or ground reference point; R_iso: insulation equivalent resistance; R_sense: leakage current sensor or sampling resistor; ξ_L: leakage current measurement noise; ε: a small positive quantity that avoids division by zero and represents measurement uncertainty.

[0043] C_th: Equivalent heat capacity; R_path: Equivalent resistance of the leakage current heating path; h: Heat dissipation coefficient; T_amb: Ambient temperature; T_k, T_{k+1}: Temperature at adjacent discrete moments; θ_B=[I_leak,R_iso,dT / dt,V_B+]: DC leakage current physical sample parameters / observation vector.

[0044] The safety exceptions in this application are: protection is directly triggered when |I_leak| exceeds 30mA and lasts for 5ms; the circuit breaker is directly tripped when the temperature exceeds 90 degrees Celsius; and leakage current is confirmed when the insulation resistance is below 100kΩ. These values ​​are safety strategy parameters of this application and are not general mathematical constants.

[0045] Optionally, in the aforementioned electrical fault identification system, local data enhancement and noise superposition include: p(θ): Joint sampling distribution of physical parameters; p(l), p(v_sep), p(Z_L), p(χ), p(A_m), p(f_m): Edge sampling distribution of arc length, separation speed, load impedance, medium, MPPT amplitude and MPPT frequency.

[0046] T_amp, T_time, and T_freq represent the amplitude-domain, time-domain, and frequency-domain enhancement operators, respectively; α and β are the superposition weights of real noise and colored noise; n_real is the real background noise; and n_colored is the colored noise constructed according to the target spectrum.

[0047] ε(f): Frequency domain random noise; F^{-1} and F: Inverse Fourier transform and Fourier transform; f_c: Colored noise cutoff frequency; H(f)=1 / (1+f / f_c): First-order spectrum shaping function used in this application. y_phys, y_aug: Physical sample label and enhanced label; c_aug=[θ,fault_type,load_type,noise_level]: Metadata of enhanced sample parameters, fault type, load type, and noise level.

[0048] Contrast learning loss is in Distance using physical parameters The formula for constraining the distance between positive samples in the embedding space and the physical information contrastive learning formula includes: z: Low-dimensional normalized feature vector output by the encoder; z_i, z_j: Embeddings of the i-th and j-th samples, respectively; z_i^+: Embeddings of positive samples with similar physical parameters to the anchor z_i; B: Contrast learning batch; τ: InfoNCE temperature coefficient, which controls the sharpness of the similarity distribution.

[0049] R: Number of TCN dilated convolutional layers; W_r^(l): Dilated convolutional kernel of layer l and r; d_l=2^l: Dilation coefficient of layer l; b^(l): Bias of layer l; Q, K, V: Transformer query, key and value matrices; W_Q, W_K, W_V: Corresponding learnable projection matrices; d_k: Key vector dimension.

[0050] r_+(θ_i,θ_j): Upper limit of the allowed embedding distance for positive samples; r_0: Upper limit of the base distance; c_l, c_Z, c_m: Weights of arc length, load impedance, and amplitude difference on the upper limit of the distance; A_m,i, A_m,j: MPPT perturbation amplitudes of the two samples; m_-: Minimum interval of negative samples.

[0051] L_InfoNCE: Instance-level contrast loss; L_phys: Physical distance constraint loss; L_cls: Fault classification loss; L_energy: Energy / power consumption regularization term; α, β, γ: Weights of the four types of losses, where this application specifically refers to the weights of the physical constraint, classification, and energy terms.

[0052] Positive samples are those that meet the same physical fault type, similar physical parameters, and similar safety consequences; negative samples include normal MPPT disturbances, startup inrush current, LED switching power supply noise, and / or non-faulty load switching.

[0053] Optionally, in the aforementioned electrical fault identification system, the action space of the safety constraint reinforcement learning decision module includes no action, DC leakage trip, AC arc trip, DC arc trip, and early warning. The reward function includes a reward for correct action, a penalty for false trip, a penalty for refusal to operate, and a delay penalty, and satisfies... Safety constraints; Fault action decision modeling is a safety-constrained Markov decision process, and the safety-constrained reinforcement learning formula includes: No action. Tripping, AC arc trip, DC arc trip, Warning} False tripping .

[0054] CMDP=(S,A,P,R,C,γ): Security-Constrained Markov Decision Process; S is the state space, A is the action space, P is the state transition probability, R is the reward function, C is the constraint cost, and γ is the long-run reward discount factor.

[0055] s_t=[z_t,I_leak,t,R_iso,t,E_HF,t,ΔV_t,dT / dt_t,load_t,SOC_t,history_t]: This is a specific state vector of this application; E_HF is the high-frequency noise energy; SOC is the battery state of charge; and history is the historical action / detection result.

[0056] A={a_0,a_B,a_AC,a_DC,a_W}: Action set; a_0 means no action, a_B means general trip, a_AC means AC arc trip, a_DC means DC arc trip, and a_W means warning but no power interruption.

[0057] R(s_t, a_t): Immediate reward; in this embodiment, the reward for a correct action is 100, the penalty for a false trip is 1000, the penalty for refusing to act is 10000, and c_d in the delay penalty c_d·delay_t is the delay penalty coefficient. J_R(π): Discounted cumulative reward of strategy π; C_m: The m-th type of security cost; d_m: The upper limit allowed for this security cost.

[0058] C_1: Annual false trip probability constraint, d_1=0.001 / year; C_2, C_3, C_4: Total detection delay, AC arc detection delay and DC arc detection delay constraints, respectively, with upper limits of 30ms, 100ms and 2s in the example.

[0059] π_θ(a|s): Policy network with parameter θ; θ*: Policy parameter that maximizes cumulative reward while satisfying security constraints; λ_m: Non-negative constraint multiplier in CPO / Lagrange relaxation.

[0060] In the aforementioned electrical fault identification system, the control barrier function safety verification module defines h. B , and h T and in accordance with Filtering the set of safety actions ; When the policy network output contradicts hard security constraints, the security barrier overrides the policy output; the security verification formula for the control barrier function includes: The control barrier function security verification module defines hB. and hT, and according to Filtering the set of safety actions ; When the policy network output contradicts hard security constraints, the security barrier overrides the policy output; the security verification rules for the control barrier function include: Define leakage current safety barrier Temperature safety barrier Inrush protection window Safety Action Set Filtering Rules: like Right now If the failure persists for more than 5ms, a full trip will be forcibly executed. Without AI decision-making; like Right now If the current is exactly 30mA and lasts for more than 5ms, an alert will be triggered. And start a confirmation timer to prevent inaction. ; like Right now If the leakage current barrier is not activated, the policy network can output normally. Do nothing; like Right now Then the main circuit breaker will be forced to trip. ; like Right now During the inrush current period at the start of charging, tripping actions are prohibited. , , Only allowed or ; If safety action set If a conservative backoff action is executed, the entire circuit breaker will trip. ; like If empty, then Take conservative tripping or early warning actions.

[0061] h_j(s): The j-th security barrier function; h_j^f(s,a): The security barrier value of the next state after executing action a; α_h: The K-type decay coefficient of the control barrier function; A_safe(s): The set of safe actions that satisfy all barrier constraints in the current state; a_final: The final action actually executed after security verification.

[0062] h_B(s)=30mA-|I_leak|: The leakage current hard safety barrier of this application; h_T(s)=90°CT: The temperature hard safety barrier of this application; h_inrush(s)=t-500ms: Charging start-up inrush current protection window, which prohibits tripping due to inrush current within 500ms after start-up.

[0063] When A_safe(s) is empty, the conservative strategy of this application is adopted: execute the total trip or warning action; the backoff rule is used to prevent the policy network from outputting an unverified action when all candidate actions violate the hard constraints.

[0064] Optionally, in the aforementioned electrical fault identification system, the edge low-power pipeline satisfies Average power consumption The model size is less than 500KB; the low-latency edge pipeline formula includes: t_ADC, t_DMA, t_L1, t_encoder, t_policy, t_CBF, and t_GPIO represent the delays for ADC sampling, DMA transfer, L1 hardware triggering, encoder inference, policy inference, CBF verification, and GPIO output, respectively. In this embodiment, t_L1 ≤ 1ms, t_encoder ≤ 30ms, t_policy ≤ 10ms, and t_CBF ≤ 5ms, with a total delay t_total ≤ 50ms.

[0065] P_L1, P_L2, P_L3: Power consumption of the three-level trigger circuit, encoder and policy / security barrier stages; ρ_2, ρ_3: Percentage of time L2 and L3 are woken up; N_alarm / N_window: Percentage of abnormal wake-up within the window; P_avg: Average power consumption calculated based on the wake-up percentage, which is required to be no higher than 50mW in this embodiment of the application.

[0066] S_TCN, S_Transformer, S_policy: storage size of TCN, Transformer, and policy network; Model_size is the total model size of the three, which is required to be no more than 500KB in this embodiment.

[0067] q_int8: quantized 8-bit integer; s_x: input quantization scaling factor; z_x: input quantization zero point; clip(·,-128,127) indicates that the quantization result is restricted to the signed INT8 range.

[0068] Level 3 triggering is as follows: L1 hardware comparator is normally open, detecting di / dt and I. leak And dT / dt; L2 encoder wakes up on demand to confirm fault signs; L3 policy network and CBF are only activated and output GPIO trip when there is an abnormality.

[0069] Optionally, in the aforementioned electrical fault identification system, the local fine-tuning module, based on... Update the encoder and policy network, and perform encryption parameter synchronization within the local area network; the local fine-tuning and local area network synchronization formulas include: D_t: Local dataset for round t; (X_k, z_k, a_final, y_human): Input fragment, embedding, final action, and human-labeled record; θ_E, θ_π: Encoder and policy network parameters; η: Learning rate; λ_q: Weights of quantization-preserving regularization term; Q(θ_E): Quantized encoder parameters.

[0070] θ_global: Global parameters for local area network synchronization; N: Number of edge devices participating in synchronization; θ_i: Local parameters of the i-th device; Lap(0,Δ / ε_dp): Zero-mean Laplace noise, Δ is the parameter update sensitivity, and ε_dp is the differential privacy budget. In this embodiment, ε_dp = 1.0.

[0071] Acc_new, FNR_new, T̄_new: new model accuracy, false negative rate, and average detection latency; Acc_old, FNR_old, T̄_old: corresponding metrics of the old model; δ_a, δ_f, δ_t: allowable performance variation tolerance; the old model is only replaced when the new model meets the three update conditions.

[0072] The above parameter definitions apply only to the physical model, fault-safe threshold, edge deployment constraints, and local synchronization mechanism appearing in this application; unlisted basic mathematical operators and general machine learning symbols are to be understood according to their common meaning in the art.

[0073] Embodiments of the present invention also provide an electrical fault identification method, comprising: guided by physical information, comparing and learning the extraction of instantaneous fault features and the enhancement of transient features to obtain a safety constraint reinforcement learning model; performing adaptive decision-making and action optimization based on the safety constraint reinforcement learning model to generate control commands; and performing ultra-low latency edge inference and hardware acceleration according to the safety constraint reinforcement learning model and the control commands.

[0074] Furthermore, in the aforementioned electrical fault identification system, the method of using the electrical fault identification system includes: Parameterized physical models of three types of faults—DC leakage, AC arc, and DC arc—are locally established on edge devices. Generate diverse simulated fault waveform data locally based on the parametric physical model; A contrastive learning framework is constructed based on diverse simulated fault waveform data. Using simulated fault waveform data and real data as a basis, a temporal encoder is trained by applying physical consistency constraints, enabling it to extract essential features that are robust to noise and distinguishable by fault type, thus forming fault identification decisions. The fault identification decision is modeled as a security-constrained Markov decision process. The security-constrained Markov decision process is used to train a policy network on the edge local device, so that the policy network maximizes the long-term reward under the constraints of preset false action rate and refusal rate. A lightweight encoder and policy network are deployed locally on edge devices to extract features and make decision suggestions on real-time acquired waveforms. The decision recommendations output by the policy network on the edge local device are verified in real time through independent control barrier function security verification. If the transient security constraints are violated, a conservative action is executed; otherwise, the recommendation is executed. Incremental training of the model is completed on the edge local device, and model parameters are encrypted and synchronized among multiple devices via local area network.

[0075] In one embodiment of the present invention, the physical information comparison learning encoder module inputs the raw current and voltage waveforms of a 12-bit ADC with a sampling frequency of not less than 1 MHz, the output of a dedicated DC leakage current sensor, the housing temperature, and a small amount of real fault labeling data. The physical mechanisms of AC arc faults, DC arc faults, and DC leakage faults are modeled locally at the edge. AC arc faults are parametrically modeled using the Cassie-Mayr equations to describe the dynamic changes in arc resistance and arc voltage. The first stage of the DC arc adopts a continuous combustion model without zero intersection points, and introduces MPPT oscillation characteristics as an interference term. The DC leakage current adopts an equivalent circuit model of DC leakage current plus insulation resistance.

[0076] The second step involves physical mechanism data augmentation, which generates simulated fault waveforms offline on edge devices, randomly alters arc parameters (arc length, medium, load) in batches, and superimposes real noise. All data is generated entirely locally, eliminating any risk of privacy breaches.

[0077] The third step involves constructing the contrastive learning framework. The encoder employs a Temporal Convolutional Network (TCN) plus a Transformer architecture, taking a 1-second waveform segment as input. During training, anchor samples are used as the basis, with samples generated under the same physical model parameters treated as positive samples and other fault types or normal samples treated as negative samples. The loss function is the InfoNCE loss plus a physical consistency constraint term. The physical consistency constraint ensures that the distance between samples generated under the same physical parameters in the feature space is less than a given threshold.

[0078] The fourth aspect involves the specific application of physical consistency constraints. In the DC leakage current scenario, the DC leakage current and insulation resistance must satisfy Ohm's law. In the AC arc scenario, the arc voltage-current characteristics must satisfy the Cassie equation. In the DC arc scenario, voltage sag and current saturation must conform to the characteristics of the photovoltaic IV curve.

[0079] The fifth step is small-sample fine-tuning, which uses 50 to 100 real fault samples accumulated locally at the edge for fine-tuning. The encoder size is kept within 200KB, and the inference time is less than 30 milliseconds.

[0080] In one embodiment of the present invention, adaptive decision-making and action optimization based on safety constraint reinforcement learning are implemented. Safety constraint reinforcement learning decision module 2 receives the feature vector and system state (load type, historical decisions, noise level) from module 1. Fault identification decision is modeled as a safety constraint Markov decision process (CMDP). State st is the current feature vector plus the system state. Action space at includes five options: no action, DC leakage trip, AC arc trip, DC arc trip, and warning without power interruption. The reward function is set to reward 100 points for correct identification and action, penalize 1000 points for false trip, and penalize 10000 points (most severe) for refusal to act. Safety constraints C include an annual false trip probability not exceeding 0.1%, DC leakage detection delay not exceeding 30 milliseconds, AC arc detection delay not exceeding 100 milliseconds, and DC arc detection delay not exceeding 2 seconds.

[0081] Constrained Policy Optimization (CPO) training is performed locally at the edge, using the Lagrange relaxation method, with training time kept under 4 hours, and utilizing a local CPU or NPU. The output policy network π(a|s) is less than 100KB in size.

[0082] The three-level decision architecture is the essence of this module. The L1 primary trigger circuit is implemented in hardware, immediately waking up L2 when the current change rate exceeds 50 amps per microsecond or the DC leakage current exceeds 6 mA, with a latency of less than 1 millisecond. L2 uses the encoder from Module 1 plus a 4D classifier for Bayesian inference, taking less than 30 milliseconds. L3 consists of a policy network and a security barrier, taking less than 10 milliseconds. The total latency of the entire link is controlled within 50 milliseconds.

[0083] The hard rules include prohibiting tripping during the inrush current period at the start of charging (0 to 500 milliseconds), tripping directly without AI when the DC leakage current exceeds 30 mA, and tripping immediately when the temperature exceeds 90 degrees Celsius. When the AI ​​decision conflicts with the hard rules, the hard rules are enforced unconditionally, prioritizing safety.

[0084] The edge-local online strategy fine-tuning is automatically executed every 24 hours with new data, running entirely locally without cloud or network access, and retaining 30 days of local operation logs.

[0085] In one embodiment of the invention, the ultra-low latency edge inference and hardware acceleration module receives the raw signal stream acquired by a high-speed ADC. Neural Architecture Search (NAS) is performed offline during the development phase (using a workstation GPU), searching for the optimal network structure for target hardware such as Cortex-M7, K210, or edge NPU, with constraints including a model size of less than 500KB and inference time of less than 50 milliseconds. The optimal architecture obtained from the search is deployed on the edge device in a fixed network structure form, with constraints including a model size of less than 500KB, inference time of less than 50 milliseconds, and power consumption of less than 100 milliwatts. Model lightweighting is achieved through INT8 quantization, 50% channel pruning, and knowledge distillation. Knowledge distillation uses a local edge-based teacher network to train the student network; the entire process does not rely on external computing power.

[0086] The hardware pipelined parallel architecture connects ADC sampling, DMA transfer, ring buffer, trigger circuit, lightweight encoder forward inference, policy network with safety barriers, and GPIO trip output into a single pipeline. A three-level triggering mechanism is key to achieving ultra-low power consumption. L1 is normally open, consuming microwatts, and uses only a hardware comparator to check current change rate, DC leakage current, and temperature rise. L2 is wake-up-on-demand, consuming milliwatts, and its symptoms are confirmed by the TCN encoder. L3 is activated upon confirmation of an anomaly, with the policy network making the decision and triggering a trip. The overall average power consumption is controlled below 50 milliwatts, suitable for battery power for more than 5 years.

[0087] Embodiments of the present invention also provide a method for using an electrical fault identification system, comprising the following steps: S1, establishing parameterized physical models of three types of faults—DC leakage, AC arc, and DC arc—locally on an edge device, and using these models to generate diverse simulated fault waveform data locally. S2, constructing a comparative learning framework, based on simulated fault waveform data and real data, training a time-series encoder by imposing physical consistency constraints, enabling it to extract essential features that are robust to noise and distinguishable by fault type. S3, modeling fault identification decisions as a safety-constrained Markov decision process, and using a constrained policy optimization method to train a policy network locally on the edge, maximizing long-term rewards while satisfying preset false tripping and refusal-to-tripping constraints. S4, deploying a lightweight encoder and policy network locally on the edge device, performing feature extraction and decision suggestions on the real-time acquired waveforms. S5, using an independent control barrier function safety barrier, performing real-time verification of the decision suggestions output by the edge policy network; if instantaneous safety constraints are violated (such as tripping during charging inrush current), a conservative action is executed; otherwise, the suggestion is executed. S6 performs incremental model training locally on edge devices (without uploading to the cloud) and encrypts and synchronizes model parameters across multiple devices via a local area network. All steps are completed locally on the edge devices, without passing through any cloud servers.

[0088] DC leakage current detection involves extracting the DC leakage current from the AC main current through low-pass filtering and DC integration, and fusing it with the readings from a dedicated DC leakage current sensor (such as a fluxgate-type DC residual current sensor or a type B RCD sensor with a sensitivity of not less than 1mA). The judgment rule is that a DC leakage current greater than 30 mA lasting for more than 5 milliseconds and an estimated insulation resistance value less than 100 kΩ is confirmed as a DC leakage current, and the leakage is cut off within 30 milliseconds. This judgment is not processed by an AI model, but is directly triggered by hard rules to ensure the highest priority.

[0089] AC arc fault detection involves extracting the 10-100 kHz high-frequency components from a 100 kHz current waveform and calculating the high-frequency noise energy within a half-cycle (10 milliseconds). The criterion is that an AC arc is confirmed when the noise energy exceeds a threshold and occurs more than five times within five half-cycles, and is interrupted within 40 milliseconds. Training data is generated locally at the edge using the Cassie-Mayr physics model.

[0090] DC arc fault detection involves calculating the 1-10 kHz current ripple energy within a 100-millisecond window and distinguishing between series and parallel arcs. The judgment rule is that a DC arc is confirmed when any condition is met and the temperature rise exceeds 2 degrees Celsius per minute, and it is cut off within 2 seconds. Training data is generated locally using a photovoltaic MPPT equivalent circuit plus a DC arc physical model.

[0091] Pure local incremental training is implemented using a federated learning algorithm over a local area network. Multiple edge devices synchronize model parameters within a local Wi-Fi or Ethernet LAN via TLS 1.3 encryption and differential privacy (ε=1.0 Laplace noise), without going through the cloud at all.

[0092] The primary feature triggering circuit only wakes up the main AI inference module when it detects a current surge rate exceeding 50 amps per microsecond or a DC leakage current greater than 6 mA for more than 10 milliseconds, achieving an average power consumption of less than 50 milliwatts, which is suitable for long-term battery power supply scenarios (more than 5 years).

[0093] This method is applied to fault detection of single edge devices such as electric vehicle charging piles (mode 4 DC charging), energy storage battery cabinets, photovoltaic inverters, household distribution boxes, and battery BMS. All steps are completed locally on the edge device, with trip response times of less than 100 milliseconds, model size of less than 1 MB, inference time of less than 50 milliseconds, average power consumption of less than 50 milliwatts, offline operation (no cloud required), and privacy protection (no data upload).

[0094] The edge inference hardware acceleration module employs a three-stage pipeline (from primary triggering to lightweight encoder to policy network plus security barrier) and supports fully offline local model adaptive training. The Bayesian model and policy network are fine-tuned daily on edge devices using new data, employing local operation logs as the training set, without relying on any external servers.

[0095] This invention generates positive and negative samples through physical equations, defines the contrast learning distance using physical parameters, defines action benefits and risks using CMDP, and performs formal security verification using CBF, which differs from existing solutions that only collect current waveforms and input them into CNN, random forest, or ordinary pattern recognition models. Its advantages include: adaptation to fewer samples, robustness to strong noise, offline operation at the edge, verifiable hard security constraints, constrainable probability of false tripping, and computable power consumption and latency.

[0096] Specifically, the beneficial effects of this invention are reflected in six aspects. First, it provides unified detection of three types of faults: DC leakage, AC arc, and DC arc, all simultaneously covered on a single edge device. Second, the tripping response time is less than 100 milliseconds. Third, it operates completely offline, functioning without cloud or network access, making it suitable for sensitive scenarios such as electric vehicle charging stations and energy storage. Fourth, it boasts extremely high sample efficiency; the physical model combined with comparative learning allows new fault types to be adapted with only 50 to 100 samples. Fifth, its safety can be formally verified; the control barrier function and hard rules ensure an annual false tripping rate of less than 0.1%. Sixth, it features ultra-low power consumption; the three-level triggering system with an average power consumption of less than 50 milliwatts makes it suitable for long-term battery power supply.

[0097] 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 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, 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.

[0098] 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, including methods and structures within its equivalent scope.

Claims

1. An electrical fault identification system, characterized in that, include: The physical information contrastive learning encoder module is configured to perform contrastive learning based on physical information, extract instantaneous fault features and enhance transient features from multimodal input signals, and output a low-dimensional fault feature vector. The safety constraint reinforcement learning decision module is configured to receive the low-dimensional fault feature vector, perform adaptive decision-making and action optimization based on the safety constraint Gorkov decision process, and generate preliminary action suggestions; as well as The ultra-low latency edge inference and hardware acceleration module is configured as a lightweight version of the integrated encoder module and decision module. After performing a control barrier function security check on the preliminary action suggestion, it outputs the final control command to perform full-process edge local pipeline processing from signal acquisition to control command generation.

2. The electrical fault identification system as described in claim 1, characterized in that, The physical information contrastive learning encoder module includes a physical information fault generation module and a TCN-Transformer contrastive learning encoder module, wherein: The physical information fault generation module is configured to generate local samples according to the Cassie-Mayr AC arc equation, the photovoltaic DC arc and MPPT disturbance equation, and the DC leakage current equivalent circuit equation. The local samples include high-speed sampling current waveform data, high-speed sampling voltage waveform data, and DC leakage current sensor data. The TCN-Transformer contrastive learning encoder module is configured to train a framework for contrastive learning of local physical model generated data and physical consistency constraint loss, so as to convert high-speed sampled current waveform data and high-speed sampled voltage waveform data into low-dimensional feature vectors. The security constraint reinforcement learning decision module includes a policy network and a control barrier function security verification module, wherein: The policy network is connected to the TCN-Transformer contrastive learning encoder module to receive the low-dimensional feature vector and system state. The policy network is trained using a local safety-constrained Markov decision process based on the low-dimensional feature vector and system state to output preliminary action suggestions. The control barrier function security verification module performs formal security verification on the preliminary action suggestion based on the control barrier function and outputs the final control command. The ultra-low latency edge inference and hardware acceleration module includes an edge low-power pipeline module and a local fine-tuning module. The edge low-power pipeline module integrates a lightweight version of the TCN-Transformer contrastive learning encoder module and the safety constraint reinforcement learning decision module. It performs full pipeline processing on the signal acquisition of high-speed sampled current waveform data and high-speed sampled voltage waveform data and the generation of control instructions for high-speed sampled current waveform data and high-speed sampled voltage waveform data, and is equipped with a primary feature triggering circuit to reduce average power consumption. The ultra-low latency edge inference and hardware acceleration module runs locally on the edge device.

3. The electrical fault identification system as described in claim 1, characterized in that, The physical information contrast learning encoder module locally generates three types of physical samples, including AC arc samples x. ac DC arc sample x dc and DC leakage current sample x B ; Pair the real samples with the physical samples to form positive and negative samples; Train the TCN-Transformer contrastive learning encoder to obtain a low-dimensional feature vector z; input the low-dimensional feature vector z and the system state s into the policy network. θ ; Screening based on the security verification formula of the control barrier function and output All steps are performed locally on the edge device. Let the input multimodal segment be: Window W is 1 second, sampling frequency or .

4. The electrical fault identification system as described in claim 3, characterized in that, The AC arc sample is generated by the Cassie-Mayr mixed differential equation and by considering the arc length l, separation velocity, and load impedance Z. L The sample family is formed by the medium parameter χ and the random perturbation ζ. Cassie equations in the high current / high power region Mayr equations for low current / low power regions Hybrid switching conditions: when Use the Cassie equation; when When using Mayr equations Arc pressure equation: Arc length update discretization Medium parameter correction therefore DC arc sample by , Generate and explicitly add To distinguish between MPPT disturbances and real DC arcs; Photovoltaic IV characteristics (single diode model) DC arc voltage drop MPPT perturbation model: Series arc criterion: And current saturation Parallel arc criterion: and DC leakage current sample from And the temperature rise equation is generated; 。 5. The electrical fault identification system as described in claim 4, characterized in that, Local data augmentation and noise overlay include: Contrast learning loss is InfoNCE Contrast Loss: in τ is the temperature coefficient Physical distance constraint loss: Upper limit of positive sample distance: in Distance using physical parameters The formula for constraining the distance between positive samples in the embedding space and the physical information contrastive learning formula includes: Positive samples are those that meet the same physical fault type, similar physical parameters, and similar safety consequences; negative samples include normal MPPT disturbances, startup inrush current, LED switching power supply noise, and / or non-faulty load switching.

6. The electrical fault identification system as described in claim 5, characterized in that, The action space of the safety constraint reinforcement learning decision module includes no action, DC leakage trip, AC arc trip, DC arc trip, and early warning. The reward function includes a reward for correct action, a penalty for false trip, a penalty for refusal to act, and a delay penalty, and satisfies the following conditions: Safety constraints; Fault action decision modeling is a safety-constrained Markov decision process, and the safety-constrained reinforcement learning formula is... include: Optimization goal: Safety constraints: Annual false trip rate (st) DC leakage detection delay AC arc detection delay DC arc detection delay Lagrange duality Safety constraints: No action. Tripping, AC arc trip, DC arc trip, Warning} False tripping 。 7. The electrical fault identification system as described in claim 6, characterized in that, The control barrier function security verification module defines hB. and hT, and according to Filtering the set of safety actions ; When the policy network output contradicts hard security constraints, the security barrier overrides the policy output; the security verification rules for the control barrier function include: Define leakage current safety barrier Temperature safety barrier Inrush protection window Safety Action Set Filtering Rules: like Right now If the failure persists for more than 5ms, a full trip will be forcibly executed. Without AI decision-making; like Right now If the current is exactly 30mA and lasts for more than 5ms, an alert will be triggered. And start a confirmation timer to prevent inaction. ; like Right now If the leakage current barrier is not activated, the policy network can output normally. Do nothing; like Right now Then the main circuit breaker will be forced to trip. ; like Right now During the inrush current period at the start of charging, tripping actions are prohibited. , , Only allowed or ; If safety action set If a conservative backoff action is executed, the entire circuit breaker will trip. ; like If empty, then Take conservative tripping or early warning actions.

8. The electrical fault identification system as described in claim 7, characterized in that, Edge low-power pipeline meets Average power consumption The model size is less than 500KB; The low-latency edge pipeline formula includes: The three-level triggering is as follows: L1 hardware comparator is normally open, detecting di / dt, Ileak, and dT / dt; L2 encoder is woken up on demand to confirm fault symptoms; L3 policy network and CBF are activated only in case of an anomaly and output GPIO trip.

9. The electrical fault identification system as described in claim 8, characterized in that, Local fine-tuning module based on Update the encoder and policy network, and perform encryption parameter synchronization within the local area network; The formulas for local fine-tuning and LAN synchronization include:

10. An electrical fault identification method, characterized in that, include: Based on physical information-guided contrastive learning, instantaneous fault features are extracted and transient features are enhanced from multimodal input signals, and a low-dimensional fault feature vector is output. Based on the low-dimensional fault feature vector, adaptive decision-making and action optimization are performed using a safety-constrained Morkov decision process to generate preliminary action suggestions; as well as A lightweight version integrating the encoder module and decision module outputs the final control command after performing a control barrier function safety check on the preliminary action suggestion, so as to perform full-process edge local pipeline processing from signal acquisition to control command generation.

Images