Trip detection method for leakage protection switch of electric energy metering box

Abstract

The invention discloses an electric energy metering box leakage protection switch trip detection method, which comprises the following steps of: when power failure is detected, generating event context data under the constraint of limited energy based on multi-mode original data, execution energy and buffer combined arrangement; executing short window early warning analysis to obtain an early warning trigger instruction and an early warning feature set; and responding to the early warning trigger instruction, constructing a transient fingerprint and executing parallel crosstalk decoupling to obtain a transient fingerprint and decoupling conclusion. The fingerprint construction adopts logarithmic time scattering fingerprint and electric arc physical fitting, and the decoupling adopts geometric constraint sparse chromatography decoupling to identify a responsibility loop. And fusing the transient fingerprint, the decoupling conclusion and the early warning feature set, applying a multi-modal time sequence consistency constraint, and forming a tripping conclusion and a leakage protection type suggestion. According to the method, the technical problems of loss of evidence at the moment of tripping power failure and strong crosstalk of parallel loops are solved, the responsibility loop can be positioned with high confidence, and interpretable tripping mechanism diagnosis and type suggestions are provided.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of power detection methods, and particularly relates to a detection method for tripping of a leakage protection switch of an electric energy metering box. BACKGROUND

[0002] The leakage protection switch in the electric energy metering box is a key device for ensuring the safety of terminal users in power utilization, and accurate monitoring of the operating state and in-depth analysis of the tripping event are of great significance for preventing electrical fires, ensuring personal safety and improving operation and maintenance efficiency. In particular, in an intelligent power distribution network, accurate tracing of the tripping cause is the basis for realizing rapid positioning of faults and intelligent management of the power grid.

[0003] At present, for detection of tripping of the leakage protection switch, the existing technology mainly focuses on post-state reporting and simple fault recording. For example, some schemes report the on-off state of the switch after tripping through a communication module, or save low-bandwidth current waveform data for several power frequency cycles (millisecond level) before tripping locally. In addition, some technologies use conventional time-domain or frequency-domain features (such as current effective value and harmonic analysis) to analyze the steady state or slowly varying signals before tripping to make a preliminary fault type judgment.

[0004] The existing technology faces severe technical challenges when applied to this specific scenario of the electric energy metering box. This mainly boils down to two mutually coupled physical constraints: one is the loss of evidence caused by the tripping transient, and the other is the spatial crosstalk caused by the parallel and dense layout. Specifically, the essence of the tripping action is to cut off the power supply, which causes the acquisition device to lose its ability at the most critical moment (such as microseconds) when tripping occurs, and cannot capture high-frequency transient fingerprints such as arc initiation. At the same time, multiple switches in the metering box are arranged in parallel and densely, and the strong electromagnetic radiation generated by one tripping will seriously pollute the signals of adjacent circuits, and the existing technology lacks effective spatial decoupling means, making it difficult to locate the responsible circuit. Both of these factors together make it impossible to accurately identify the final tripping mechanism (such as real leakage or condensation malfunction). SUMMARY

[0005] The application aims to provide a detection method for tripping of a leakage protection switch of an electric energy metering box to solve the above problems existing in the prior art.

[0006] The technical scheme: according to one aspect of the present application, when power failure is detected, based on multi-modal raw data and device configuration, energy and buffer joint arrangement is performed to generate event context data;

[0007] Based on the event context data, short window early warning analysis is performed to obtain a warning trigger instruction and a warning feature set;

[0008] In response to the early warning trigger instruction, and based on the event context data and the device configuration, a transient fingerprint is constructed and parallel crosstalk decoupling is performed, to obtain a transient fingerprint and decoupling conclusion;

[0009] The transient fingerprint and decoupling conclusion, the early warning feature set, and the event context data are fused, and a multi-modal temporal consistency constraint is applied, to form a trip conclusion and a leakage protection type suggestion.

[0010] The event context data includes event window data, a time reference sequence, and an environmental state. The transient fingerprint and decoupling conclusion include a decoupled transient fingerprint, a responsible loop identifier, and a crosstalk indicator.

[0011] Beneficial effects: The application solves the technical problems of loss of transient evidence at the moment of trip power failure and strong crosstalk in parallel loops, can locate the responsible loop with high confidence, and provides interpretable trip mechanism diagnosis and type suggestion. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 The overall flowchart of the application.

[0013] Figure 2 The flowchart of the application for obtaining a transient fingerprint and decoupling conclusion.

[0014] Figure 3 The flowchart of the application for performing sparse chromatography decoupling on microsecond-level data segments.

[0015] Figure 4 The flowchart of the application for synthesizing a transient fingerprint.

[0016] Figure 5 The flowchart of the application for generating event context data. DETAILED DESCRIPTION

[0017] Embodiment 1 describes a general method for detecting the trip of a leakage protection switch of an electric energy metering box. The technical problems of loss of key transient evidence at the moment of trip power failure, strong electromagnetic crosstalk between parallel loops, and difficulty in distinguishing trip mechanisms (such as real leakage, condensation misoperation, and load type mismatch) in the leakage protection switch in the metering box are solved.

[0018] Step 101: Upon detection of power failure, based on multi-modal raw data and device configuration, energy and buffer joint arrangement is performed to generate event context data.

[0019] In the present embodiment, the power outage refers to the event of interruption of the mains supply of the electric energy metering box. The multi-modal raw data refers to the various sensor data continuously acquired by the device before the power outage, preferably including differential current raw data, near-end voltage raw data, near-field magnetic and electric field array raw data, acoustic and acceleration raw data, and temperature and humidity raw data. The device configuration includes the hardware parameters of the device, geometric layout information, calibration matrix, etc. Since the conventional scheme loses working energy at the time of power outage, it cannot record the key evidence at the tripping instant (especially at the microsecond level). Therefore, the present step adopts an energy and buffer combined scheduling strategy, using the weak residual energy (including the back electromotive force of the trip coil, the residual voltage of the line, and the bus capacitor) that can still be obtained by the device after the power outage, under the energy constraint of millijoules, to preferentially ensure the acquisition and solidification of the key evidence (including the microsecond-level high-bandwidth segment before triggering and the time reference). The finally generated event context data is a data set, which can specifically include: event window data (i.e. high and low bandwidth sensor data segments), time reference sequence (for subsequent multi-modal alignment), and environmental state (such as temperature and humidity).

[0020] Step 102, based on the event context data, performing short window early warning analysis to obtain early warning trigger instruction and early warning feature set.

[0021] Since the tripping event (especially the leakage caused by condensation or aging) often has electrical precursors within ten milliseconds before the real occurrence (i.e. contact separation). The present step analyzes the event context data (specifically a short time window, for example 10 ms, before tripping). By extracting the time domain and frequency domain features (including root mean square trajectory, spectral centroid, odd-even harmonic ratio, etc.) in this window, and combining the environmental state (including condensation) to form an adaptive threshold, a high-confidence early warning analysis is performed. The early warning feature set is a collection of the above-mentioned electrical features. The early warning trigger instruction is an internal control signal generated when the early warning analysis (including the pair-wise sequential detection) determines that the precursor is true. This instruction is used to trigger high-bandwidth confirmation sampling on the one hand, and can be fed back to step 101 on the other hand, to dynamically adjust the energy-buffer scheduling strategy, to ensure that valuable microsecond-level acquisition energy is used at the most critical moment.

[0022] Step 103, in response to the early warning trigger instruction, and based on the event context data and the device configuration, constructing the transient fingerprint and performing parallel crosstalk decoupling to obtain the transient fingerprint and decoupling conclusion.

[0023] This step aims to solve two technical problems: one is that the electrical signals at the moment of tripping (such as transient signal segments within 0 to 100 microseconds) are non-stationary, multi-peak ringing, and conventional fingerprints are unstable; two is that the multiple parallel circuits in the metering box are densely arranged, causing strong electromagnetic crosstalk, making it difficult to locate the responsible circuit. To solve the first problem, this step intercepts the microsecond-level data segment under the high-bandwidth confirmation trigger (step 102), and uses, for example, a logarithmic time scattering fingerprint combined with arc physical fitting to construct a time-scaled and ringing-robust transient fingerprint. To solve the second problem, this step uses the geometric layout information in the device configuration and the near-field magnetic and electric field array data in the event context data to construct geometric constraints and perform sparse tomography decoupling to separate the main source signal and radiation crosstalk. The final generated transient fingerprint and decoupling conclusion are a result set, which preferably includes: decoupled transient fingerprint, responsible circuit identification, and crosstalk index (including residual energy ratio).

[0024] Step 104, fuse the transient fingerprint and decoupling conclusion, the early warning feature set, and the event context data, and apply multi-modal temporal consistency constraints to form the tripping conclusion and the leakage protection type suggestion.

[0025] The conventional scheme relies only on a single threshold decision, which is easily disturbed and cannot distinguish mechanisms. This step performs multi-modal fusion, time-aligns the transient fingerprint and decoupling conclusion (including electrical fingerprint, responsible circuit) of step 103, the early warning feature set (including pre-tripping signs) of step 102, and other modalities (including acoustic features, magnetic field features) in the event context data. This step applies multi-modal temporal consistency constraints, for example, physically, coil excitation (magnetic), contact separation (acoustic), and arc initiation (electric) must satisfy strict time sequence. Through this constraint, the real action pattern is identified. In addition, this step also performs causal mapping based on load spectrum evidence (from event context data) to give leakage protection type suggestions (including whether the type mismatch is caused by direct current component or high frequency energy). Finally, all evidence is fused (and down-weighted in high crosstalk or low consistency), generating the final tripping conclusion (including event nature, responsible circuit, action category).

[0026] In a preferred embodiment, the method further includes the step of evidence solidification and strategy closed loop. The tripping conclusion, action pattern, leakage protection type suggestion, decoupled transient fingerprint, responsible circuit identification, crosstalk index, early warning feature set, and integrity and tamper-proof information are written into the whole event record for traceability. More preferably, the device can generate strategy updates based on statistical analysis of historical whole event records, and backfill to the energy budget of step 101 and the adaptive threshold of step 102, thereby realizing long-term adaptive evolution of the detection strategy.

[0027] At the same time of generating the event context data, the embodiment also calculates the integrity and tamper-proof information such as hash value or digital signature for the pre-window data segment, the post-window data segment and the sampling result, and indexes and associates the information with the event context data for calling by the evidence solidification step in Embodiment 6.

[0028] Embodiment 2 describes the power-off self-sustaining and evidence prioritization process as a preferred and detailed implementation of step 101 in Embodiment 1. This embodiment focuses on how the device reliably generates event context data under extremely limited energy constraints (including millijoule-level energy) through an evidence-prioritized energy and buffer joint scheduling strategy when a power-off event is detected.

[0029] Step 201, read the multi-modal raw data and device configuration, identify and evaluate candidate energy sources to establish an energy source priority table. In other words, based on the energy source priority table, read the preset system power consumption model from the device configuration, decompose the system power consumption model and allocate sampling resources according to the evidence importance, to formulate energy budget parameters and degradation strategies.

[0030] Specifically, the device continuously monitors the near-end voltage raw data in the multi-modal raw data.

[0031] Preferably, the device determines the occurrence of a power-off event by detecting a voltage drop or slope threshold crossing of the voltage signal.

[0032] After determining the power-off, the device immediately performs identification and evaluation of candidate energy sources. This process preferably includes:

[0033] First, detect and enumerate candidate energy sources. The device detects energy traces that can still be utilized after the interruption of commercial power within a preset micro-short window to form a list of candidate energy sources. These energy sources can specifically include: back-EMF of the trip coil (i.e. the instantaneous energy generated by the coil when the switch trips), line residual voltage, charge stored in the bus capacitor, and possibly super-capacitors or backup batteries.

[0034] Second, evaluate the equivalent electrical parameters. For each energy source in the list, estimate its equivalent internal resistance and pulse power capability (i.e. the maximum current that can be provided instantaneously) through small perturbation testing or historical event playback to screen out a set of compliant energy sources.

[0035] Finally, weighted scoring. For compliant energy sources, weighted scoring is performed according to available energy (total amount), response time (speed) and safety margin (reliability) to form a final energy source priority table and fallback path.

[0036] Step 202, based on the energy source priority table and the device configuration, decompose the system power consumption model and allocate sampling resources according to the evidence importance to formulate the energy budget parameters and the degradation strategy.

[0037] Firstly, based on the device configuration (including processor model, storage medium type), the device calculates the average and peak power consumption of the processor, storage write, sampling front end and clock retention to form a system power consumption model.

[0038] Secondly, based on the energy source priority table of step 201 and the power consumption model, the total "available energy" (E total ) after power failure and the minimum operating voltage are calculated to form the energy budget parameters.

[0039] It should be noted that the energy source (including back electromotive force) used by the present application may be extremely limited, for example, in the order of millijoule. The feasibility of the present application lies in allocating resources according to the importance of evidence. For example, in a preferred embodiment, the device estimates that the total available energy E total is only 5mJ. The system power consumption model determines that 1mJ is required for metadata writing, and 1.5mJ is required for millisecond-level slow sampling segments. The remaining 2.5mJ is dedicated to microsecond-level high-bandwidth segments. If the conventional scheme tries to maintain continuous high-bandwidth acquisition with 2.5mJ of energy (assuming instantaneous power consumption of 250mW), it may only support 10ms, which cannot guarantee the capture of the trip moment in terms of timing.

[0040] The 2.5mJ budget is allocated to the guaranteed retention segment according to the importance of the evidence, that is, the microsecond-level high-bandwidth segment of a fixed time length before the trigger point required by the subsequent transient fingerprint (example four). At the same time, this step formulates the degradation strategy, which provides that when the energy is insufficient, part of the data of the millisecond-level slow sampling segment is preferentially discarded (including performing sparse sampling and key point retention), to ensure the integrity of the microsecond-level evidence.

[0041] Step 203, cut into self-sustaining mode at power failure, and perform evidence prioritization according to the energy budget parameters and the degradation strategy to capture and generate event context data.

[0042] At power failure (including the voltage slope threshold of step 201), the device immediately cuts into self-sustaining mode, for example, turns off non-critical loads, only retains the sampling, processing and storage cores, and selects and activates the available energy sources in turn according to the energy source priority table and its fallback path, to automatically switch to the next candidate energy source when a certain energy source is exhausted or abnormal.

[0043] Before performing the arrangement, the device preferably also performs a pre-trigger ring buffer planning. Specifically, according to the target window length and sampling duty of step 202, a multi-channel buffer mapping table of high-bandwidth buffer + slow-sampling buffer is established for channels such as differential current, near-field magnetic field array, etc., and a write pointer strategy and an overlay strategy are defined.

[0044] On the basis of the buffer planning, the device performs the capture according to the energy budget parameter and the degradation strategy. This process preferably includes:

[0045] The microsecond-level high-bandwidth segment of a fixed length before the priority solidification trigger point is fixed as the pre-window data segment. This is an embodiment of evidence priority, and this segment (e.g. 100 μs) must be forcibly retained regardless of whether the subsequent energy is sufficient.

[0046] The post-window data segment is adaptively extended according to the energy budget parameter when the energy allows. For example, in the 5 mJ example of step 202, after the pre-window is solidified (which may only consume 0.1 mJ), the remaining 2.4 mJ of energy is used to dynamically extend the post-window, and as long as the energy allows, the capture is performed as long as possible.

[0047] When the energy is tight, the degradation strategy is performed. If the total energy ( Etotal ) is extremely low, or the post-window extension has exhausted the energy, the device will perform sparse sampling and key point retention on the millisecond-level slow-sampling segment, e.g. only retaining the peak value and peak interval, and abandoning the complete waveform.

[0048] The pre-window data segment, the post-window data segment, and the sampling result are combined and bound to a time reference sequence to generate event context data. Sparse sampling and key point retention can also be performed to obtain the sampling result. The time reference sequence ensures that data of different bandwidths and different modalities have uniform timestamps, providing a basis for multi-modal fusion in subsequent embodiments (including embodiment five).

[0049] In some preferred embodiments, the evidence priority arrangement of this step (step 203) is dynamic. The arrangement module receives a warning score from the subsequent short window warning analysis (including embodiment three). If the warning score is low, the arrangement module can actively reduce the sampling rate of the post-window or terminate the capture in advance to save energy; if the warning score is high, the integrity of the microsecond-level high-bandwidth segment is prioritized, and a rolling adjustment of energy-buffer is achieved.

[0050] Embodiment 3 describes the process of short window warning and adaptive triggering. As a preferred and detailed implementation of performing the short window warning analysis step in embodiment 1. This embodiment focuses on how the device detects electrical precursors with high confidence within ten milliseconds before the trip event actually occurs (i.e. contact separation), and solves the technical problem that the conventional solution is easily disturbed by the environment (especially condensation) and produces false positives.

[0051] Step 301: Extract the environmental state from the event context data, estimate the dew point and condensation state, and combine them with historical whole event records to generate an adaptive threshold set.

[0052] This step is a prerequisite for achieving high-confidence early warning. The event context data comes from the output of Example 2, which specifically includes the environmental conditions, namely the raw temperature and humidity data collected by the device, as well as the internal surface temperature data of the chamber installed on the inner wall of the metering chamber.

[0053] Specifically, the device first estimates the current dew point based on temperature and humidity data in the environment, and then determines the current condensation state by combining this with the surface temperature inside the chamber. The condensation state can be a graded variable, including: dry, slight condensation, and severe condensation.

[0054] Based on this, the device generates an adaptive threshold set. This process does not use fixed static thresholds, but rather performs relevant adjustments. For example, a static threshold T for warning features... static It is set to 1.0. When the condensation condition is determined to be severe, the device applies an additive correction. Ccond =0.5; simultaneously, if a large temperature difference ΔT is detected between the inside and outside of the chamber, the device applies a multiplicative correction. Mtemp =1.2.

[0055] At this point, the corrected threshold T_corrected =( Tstatic +C cond )*M temp =(1.0+0.5)*1.2=1.8.

[0056] Furthermore, to achieve long-term adaptation, the device will also utilize historical whole event records (accumulated and managed by Example 6) to modify the revised threshold (T). corrected Fine-tuning can be performed. For example, if historical data shows that a threshold of 1.8 still results in false alarms under the current operating conditions, the threshold can be further increased to finally output an adaptive threshold set.

[0057] In some alternative implementations, this step also includes extreme environment strategies. For example, when extreme environments such as high humidity and high temperature or sustained low temperature are detected, the device may enable a separate set of conservative thresholds or forcibly increase the correction margin, sacrificing some sensitivity for absolute reliability.

[0058] Step 302: Locate the early warning analysis segment before the trip from the event context data, and extract time-domain intensity features, spectral statistical features, and odd-even harmonic ratio to form an early warning feature set. Specifically:

[0059] The device locates a specific time window, such as the first 10 milliseconds, before the trip trigger point from the event window data contained in the event context data, and uses this as a warning analysis segment. Preferably, before extracting features, this segment can also be resampled and detrended to eliminate interference from power frequency fluctuations or DC offset.

[0060] The device extracts multi-dimensional features from the early warning analysis segment to form an early warning feature set. Specifically, the early warning feature set may include:

[0061] Temporal intensity characteristics: For example, the root mean square locus calculated over a short window, and the slope of the signal change (first derivative). Spectral statistical characteristics: For example, the spectral centroid and short-time energy density calculated after performing short-time spectral estimation. Odd-even harmonic ratio characteristics: For example, integrating energy at the fundamental frequency (50Hz) and each harmonic, and calculating the ratio of the sum of odd-order harmonic energy to the sum of even-order harmonic energy.

[0062] In some preferred embodiments, a normalization and drift compensation step is included before forming the final warning feature set. That is, the original features extracted above are normalized according to the statistical data (including mean and variance) of historical fault-free operation, and features that may drift (including spectral centroid) are compensated according to the current temperature to improve the stability and comparability of the features.

[0063] Step 303: Based on the early warning feature set and the adaptive threshold set, perform dual sequential detection to obtain the early warning determination.

[0064] Different leakage causes (including insulation aging and condensation conductivity) exhibit different patterns on the feature set extracted in step 302, and therefore the following processing is performed:

[0065] Specifically, this step constructs two parallel statistical streams to perform dual sequential detection:

[0066] The first approach constructs a condensation path statistic. This statistic is designed to capture patterns in early warning features that characterize low-frequency drift and intermittent conductivity. For example, this statistic might highly weight the low-frequency shift of the spectral centroid and the intermittent pulses of the time-domain intensity. The second approach constructs an aging path statistic. This statistic is designed to capture patterns in early warning features that characterize monotonic increases and rightward shifts in the spectral centroid. For example, this statistic might highly weight the monotonic growth slope of the RMS trajectory and the shift of the spectral centroid towards higher frequencies.

[0067] The device uses the adaptive threshold set generated in step 301 to perform sequential detection on the two statistics in parallel (including the CUSUM cumulative sum test or the GLRT generalized likelihood ratio test).

[0068] A key innovation of this embodiment is that, in order to suppress rare noise triggering, the sequential detection also incorporates an extreme value protection strategy. That is, when an anomaly is detected in the sequential detection, it is not triggered immediately. Instead, extreme value theory is introduced to model the statistical significance of the anomaly and make a secondary judgment, so as to effectively suppress those statistically rare noise spikes that are not caused by the target mechanism (condensation or aging), thereby significantly reducing the false alarm rate.

[0069] The two detection results (including either one confirmed to be triggered after extreme value protection, or the weighted score of the two results exceeding the total threshold) are merged to generate the final warning judgment and warning score.

[0070] Step 304: Based on the warning determination, generate a warning trigger command according to the hysteresis and debounce strategy.

[0071] When the warning determination generated in step 303 is true, this step does not immediately issue a command, but instead applies a hysteresis and debouncing strategy. For example, the true determination must last for 3 sampling points (debouncing), or the warning score must exceed the trigger threshold by more than 5% (hysteresis upper boundary) before a warning trigger command is generated. Similarly, when the warning score falls back, it must be below the trigger threshold by 10% (hysteresis lower boundary) before the command is revoked. This prevents the warning signal from oscillating frequently near the critical point.

[0072] In a preferred embodiment, the warning trigger command (or high bandwidth confirmation trigger command) generated in this step is sent to the corresponding module in Embodiment 4.

[0073] In another preferred embodiment, this step also generates high-bandwidth sampling parameters (including dynamically determining the target sampling rate or window length for subsequent microsecond-level sampling based on the warning score) and sends them to the corresponding module in Embodiment 4 to guide the high-bandwidth segment truncation in step 401 of Embodiment 4.

[0074] In another preferred embodiment, the warning score or warning determination generated in this step is also fed back to the energy-buffer joint orchestration module of Embodiment 2 to achieve energy-buffer rolling adjustment across embodiments, ensuring that energy is optimally used to capture the transients of high-scoring warning events.

[0075] The sequential detection of early warning statistics can also be achieved using Bayesian filtering or Kalman filtering frameworks to support continuous confidence assessment of slowly evolving faults.

[0076] Example 4 describes the process of transient fingerprint construction and parallel crosstalk decoupling. This is a preferred and detailed implementation of the transient fingerprint construction and parallel crosstalk decoupling steps in Example 1. This example addresses the following problems: First, the electrical signal at the moment of tripping is non-stationary and has multi-peak ringing, making conventional fingerprints (including FFT and envelope) unstable; second, multiple circuits are densely arranged in parallel within the metering box, with close conductor spacing, leading to strong near-field electromagnetic crosstalk, making it difficult to locate the responsible circuit using conventional methods.

[0077] Step 401: Respond to the warning trigger command and extract high-bandwidth microsecond-level data segments from the event context data.

[0078] After receiving the warning trigger command from Example 3 (step 304), the device immediately accesses the event window data (specifically, the high-bandwidth circular buffer) in the event context data generated in Example 2.

[0079] The device extracts a high-bandwidth, microsecond-level data segment from the buffer based on instructions (or high-bandwidth sampling parameters carried by the instructions). Preferably, the sampling rate of this segment is in the tens of megahertz range (e.g., 10 MHz), and the time window covers the vicinity of the trip trigger point, for example, from 0 to 100 microseconds. Prior to subsequent processing, this segment can preferably be subjected to DC removal and bandpass filtering (including 500 kHz to 5 MHz) and strictly aligned with the time reference sequence.

[0080] Step 402: Log-time resampling is performed on the microsecond-level data segment, and the scattering coefficient is calculated to characterize the high-frequency non-stationary texture. The rising edge physical parameters are then fitted using the arc model to synthesize the transient fingerprint.

[0081] Construct transient fingerprints robust to ringing and time scaling. This is a two-track synthesis method combining statistical texture and physical fitting.

[0082] Log-time resampling is performed on the linear time axis of the microsecond-level data segment obtained in step 401. In a preferred embodiment, the log-time mapping can be expressed as t_log=log(1+a·t), where a is a coefficient controlling the degree of time compression, and its value is preferably between 10 and 10. 5 Up to 10 6 The goal is to cover critical transients within the range of 0 to 100 microseconds. The physical purpose is that the initial physical processes of the electric arc (such as channel conduction establishment and contact bounce) mostly occur in the earliest stages of the transient (including 0 to 10 microseconds), which have the highest information density. Conventional linear sampling has sparse sampling points in this stage, while log-time mapping can nonlinearly amplify the resolution of this initial stage, obtaining logarithmic domain data segments.

[0083] Scattering coefficients are calculated on the logarithmic domain data segment to characterize high-frequency non-stationary texture. In this embodiment, the scattering coefficients are a mathematical representation based on a custom wavelet transform to describe the non-stationary texture of the signal. Specifically, for the input logarithmic domain data segment X... (t) Its first-order scattering coefficient S_1 (t,λ) It can be approximated as |X* ψ_λ(t) |, where * denotes convolution, ψ λ This represents a wavelet at a specific scale λ. In one implementation example, the number of scales corresponding to λ and γ can be 4 to 8 to balance computational complexity and the ability to distinguish non-stationary textures.

[0084] S_1 (t,λ) It can be physically interpreted as the local energy envelope of the signal at different scales λ, used to characterize the amplitude modulation texture of the signal. Its second-order scattering coefficient S_2 (t,λ,γ) It can be approximated as ||X* ψ_λ(t) |* ψ_γ(t) The coefficients can be physically interpreted as the rate of change of these envelopes themselves (the envelope of the envelope), used to characterize the frequency modulation texture of the signal. The set of scattering features composed of these two orders of coefficients is highly robust to the time scaling and transient ringing of the signal, which is not the case with conventional FFT features.

[0085] The arc-starting rising edge in a logarithmic domain data segment is detected, and a simplified arc model (such as a simplified form of the Mayr or Cassie arc model) is used to fit this rising edge. Through fitting, a set of physical parameters including the channel conductance rise rate or contact bounce interval is obtained.

[0086] The set of scattering features (characterizing statistical texture) and the set of physical parameters (characterizing physical origin) are integrated and synthesized into the final transient fingerprint.

[0087] Step 403: Based on the event context data and device configuration, construct the array spatiotemporal features and geometric calibration matrix.

[0088] Specifically, the device reads the raw data of the near-field magnetoelectric field array from the event context data and combines it with the array sensor geometric position information stored in the device configuration.

[0089] By performing geometric calibration and channel synchronization on the array data, the device calculates the arrival time difference of the transient signal (the segment in step 401) to each sensor and the amplitude and phase distribution of the signal at each sensor, which constitute the spatiotemporal characteristics of the array. The geometric information used for calibration is then organized into a geometric calibration matrix.

[0090] Step 404: Using the array spatiotemporal features and geometric calibration matrix as geometric constraints, perform sparse tomography decoupling on the microsecond-level data segment to separate the main source and radiative crosstalk, obtain the decoupled transient fingerprint, and match it to obtain the responsibility loop identifier.

[0091] Based on the geometric calibration matrix and device configuration (e.g., the physical orientation and orientation information of each loop conductor), a near-field radiation dictionary D is constructed. It should be noted that those skilled in the art will understand that near-field electromagnetic coupling is inherently highly nonlinear. However, the near-field radiation dictionary D of this invention is not based on a simplified far-field linear assumption, but rather, through the geometric calibration or offline electromagnetic simulation in step 403, it pre-constructs specific response atoms (i.e., column vectors of dictionary D) generated by each loop (source) in the near-field array (observation point), which already contain near-field effects.

[0092] A non-negative sparse reconstruction model is established. The key to solving the nonlinear problem in this invention lies in assuming that within the microsecond-level transient of the trip, only one responsible loop is the main source, and the signals observed in other loops are all radiated crosstalk from this main source. Therefore, at the signal source level, the problem can be linearly modeled as: Y = D * X + N. Where Y is the microsecond-level data segment observed in step 401 (projected onto each channel of the array), D is the aforementioned near-field radiation dictionary, N is noise, and X is the source component vector to be solved (its dimension is equal to the number of loops, for example, 12 channels). Due to the assumption of a single main source, the vector X must be sparse (i.e., only one or a very small number of elements in X are non-zero values).

[0093] This step uses dictionary D and the arrival time difference and amplitude-phase consistency in the array spatiotemporal features obtained in step 403 as hard constraints to establish the sparse reconstruction model. This hard constraint greatly compresses the solution space.

[0094] Solve the nonnegative sparse reconstruction model (e.g., through iterative optimization algorithms, such as nonnegative least squares combined with L1 norm regularization) to recover the source component X. The non-zero terms in the recovered X vector correspond to the main source and its contribution. Substituting this main source component back, we can obtain the decoupled transient fingerprint (i.e., the clean source signal separated from crosstalk).

[0095] The energy of the source component X is matched with the geometric consistency of the array spatiotemporal characteristics in step 403 (i.e., whether the dictionary atom corresponding to the component with the highest energy in X is consistent with the observed TDOA and amplitude-phase distribution) to finally confirm the responsible loop identifier. The energy of the residual N (residual energy ratio) is then calculated to define the crosstalk index.

[0096] In some preferred embodiments, this example further includes an evidence consistency verification step. This step reads the transient fingerprint (containing physical parameters) generated in step 402 and the decoupled transient fingerprint and responsibility loop identifier generated in step 404, and combines them with the warning feature set from Example 3. Consistency scores are assigned to these cross-step pieces of evidence in terms of frequency band, slope, and time alignment. For example, if the warning feature shows a condensation path, but the transient fingerprint shows a strong electric arc, the consistency score is reduced. This consistency score is output to Example 5 as a weighting criterion during its fusion decision.

[0097] In another implementation, the scattering feature set can be replaced by time-frequency texture features based on multi-scale short-time Fourier transform, or used in cascade with such features to further enhance the ability to distinguish different types of electric arcs.

[0098] Example 5 describes the implementation process of multimodal fusion and causal type recommendation. This is a preferred and detailed implementation of the steps in Example 1 for forming the trip conclusion and leakage protection type recommendation. This example focuses on how the device comprehensively utilizes all the evidence from the foregoing examples to solve the technical problems of unclear operating mechanisms (e.g., whether it is a real leakage or a mechanical impact) and the lack of a causal chain in the selection recommendation (why the trip occurred, whether the selection was incorrect).

[0099] According to one aspect of this application, the process for arriving at a tripping conclusion and a recommendation for a type of residual current protection is as follows:

[0100] Extract coil magnetic field features and contact release acoustic spectrum from event context data, and combine them with transient fingerprints and arc physical parameters from decoupling conclusions to apply a hard constraint on the time sequence of "coil excitation → contact separation → arc initiation" to identify action patterns;

[0101] The DC component, high-frequency energy ratio, and odd-even harmonic energy ratio are estimated from event context data to form a load type evidence set, and a causal mapping from the load type evidence set to type selection is constructed to generate leakage current protection type recommendations.

[0102] The system integrates operating modes, leakage protection type recommendations, transient fingerprints and decoupling conclusions, responsibility circuit identification and crosstalk indicators, and early warning feature sets. It also deweights high crosstalk or low consistency cases to infer and generate tripping conclusions.

[0103] The steps for applying a hard constraint on the time sequence of "coil excitation → contact separation → arc initiation" to identify the operating mode include:

[0104] Extract the rising edge and holding plateau of the coil magnetic field from the event context data to obtain a set of magnetic features;

[0105] The release acoustic spectrum and bounce timing of the touch point are calculated from the event context data to obtain the set of voiceprint features;

[0106] Extracting arc physical parameters from transient fingerprints and decoupling conclusions;

[0107] Using time-series hard constraints, a consistency decision is made on the magnetic feature set, acoustic feature set, and electric arc physical parameters to output action patterns and action evidence scores.

[0108] Optionally, the tripping conclusion is inferred, including:

[0109] From the event context data, estimate the DC component, the proportion of high-frequency energy, and the ratio of odd and even harmonic energy to form a load type evidence set;

[0110] Construct a cause-effect graph or rule base from the DC component and high-frequency energy ratio in the load type evidence set to the selection of leakage protection type;

[0111] Match the cause-effect graph or rule base, and combine it with the existing types in the device configuration to output leakage protection type recommendations, recommendation confidence levels, and explanatory information.

[0112] Optionally, the trip conclusion is generated by inference, including:

[0113] Read the crosstalk index in the action mode, leakage protection type recommendation, transient fingerprint and decoupling conclusion, and combine the action evidence score and recommendation confidence level;

[0114] Based on crosstalk index or evidence consistency score, set fusion weights, and reduce the weight of corresponding evidence in the case of high crosstalk or low consistency.

[0115] Inference is made based on fusion weights to generate tripping conclusions and their confidence levels;

[0116] Extract the key rules and physical parameters hit during the fusion process and encapsulate them into a traceable explanatory summary.

[0117] In a certain scenario, the data processing flow is as follows:

[0118] Step 501, Multimodal evidence preparation and alignment.

[0119] The device first aggregates evidence from different embodiments to form a fused feature library. This feature library preferably includes: decoupled transient fingerprints, responsibility loop identifiers, and crosstalk indices from Embodiment 4; arc physical parameters from the transient fingerprints of Embodiment 4; warning scores and warning feature sets from Embodiment 3; and environmental states from the event context data of Embodiment 2, original coil magnetic field feature segments clipped from the coil magnetic field measurement channel according to the time reference sequence, and contact release acoustic spectrum segments clipped from the acoustic measurement channel and obtained through short-time Fourier transform. The key to this step is to use the time reference sequence generated in Embodiment 2 to strictly align all the above evidence in time, ensuring the effectiveness of subsequent timing consistency constraints.

[0120] Step 502 applies a hard constraint on the time sequence of coil excitation → contact separation → arc initiation to identify the action mode and then distinguish the true cause of the trip through the physical mechanism. The device extracts three key time-series evidences from the fusion feature library of step 501:

[0121] First, extract the magnetic feature set: analyze the characteristic segments of the coil's magnetic field, and detect the rising edge and holding plateau of the trip coil's magnetic field. Second, extract the acoustic feature set: analyze the acoustic spectrum segments of the contact release, calculate their sub-band energy and spectral entropy, and identify the impact moment of mechanical contact separation and possible rebound timing. Third, extract the arc physical parameters: obtain parameters such as the arc initiation moment and the channel conductivity rise rate from the transient fingerprint generated in Example 4.

[0122] Based on this, the device performs timing consistency decisions. In a physically valid tripping action driven by leakage current, these three events must satisfy strict time sequence hard constraints, namely: Tmag_start < Tacou_release < Tarc_start .

[0123] For example, the timing sequence of an effective residual current circuit breaker trip might be as follows: Tmag_start At 0ms, Tacou_release At 6ms, Tarc_start At 6.5ms. Conversely, if detected Tarc_start Long before Tmag_start This may indicate an external arc flashover rather than RCD activation. If detected... Tacou_release and Tmag_start The absence of this information may indicate a malfunction caused by an external mechanical shock. By determining the degree of matching of this timing constraint, the device outputs an action mode (e.g., coil-driven leakage, manual test trip, mechanical shock malfunction, external arc) and an action evidence score (characterizing the credibility of the mode).

[0124] Step 503: Estimate the DC component, high-frequency energy ratio and odd-even harmonic energy ratio to form a load type evidence set, and construct a causal mapping from the load type evidence set to type selection to generate a leakage current protection type recommendation.

[0125] This step aims to resolve the issue of mismatched residual current device (RCD) selection and unilateral tripping caused by nonlinear loads.

[0126] First, the device estimates the DC component of the current flowing through the responsibility loop, the proportion of high-frequency energy (e.g., the proportion of energy in the frequency band above 1kHz), and the ratio of odd and even harmonic energy from the fusion feature library (especially the decoupled transient fingerprint) in step 501 within a synthesized analysis window, thereby forming a load type evidence set.

[0127] Secondly, a cause-effect graph or rule base is constructed to map the aforementioned evidence to the type selection of the RCD. Those skilled in the art will understand that conventional AC-type RCDs are insensitive to smooth DC leakage currents (e.g., greater than 6mA) or prone to magnetic saturation (failure to operate), while they may be overly sensitive to high-frequency leakage currents (false operation). The rule base of this invention is constructed based on this:

[0128] Case 1 (Rule): If the DC component is >6mA and the high-frequency energy percentage is <10%, it indicates that the load (such as a photovoltaic inverter or some variable frequency air conditioners) has generated DC leakage. An AC-type RCD may fail to operate or operate malfunctioning. In this case, the causal mapping points to a Type A or Type B RCD. Case 2 (Rule): If the high-frequency energy percentage is >30% (e.g., from a VFD inverter), the causal mapping points to a Type F or Type B RCD, because an AC-type or Type A RCD may operate malfunctioning at this frequency.

[0129] The device matches the cause-effect graph or rule base and combines it with the existing type recorded in the device configuration to output a leakage current protection type recommendation (e.g., recommend upgrading the existing type: AC to the recommended type: A), recommendation confidence level, and explanatory information (e.g., if an 8mA DC component is detected, the AC type RCD may have a risk of failure to operate or false tripping).

[0130] Step 504: Integrate the operating mode, leakage protection type recommendations, transient fingerprint and decoupling conclusions, responsibility circuit identifiers and crosstalk indicators, and early warning feature sets, and deweight high crosstalk or low consistency cases to infer and generate tripping conclusions.

[0131] The device reads all upstream evidence, including the action pattern and action evidence score in step 502, the type suggestion and suggestion confidence level in step 503, the responsibility loop identifier and crosstalk index in embodiment 4, and preferably, the consistency score generated at the end of embodiment 4 (step 404).

[0132] In this step, the device does not place equal trust in all evidence, but rather performs dynamic weighting:

[0133] High crosstalk scenario: If the crosstalk index (e.g., residual energy ratio) from Example 4 is greater than a high threshold (e.g., 0.4), it indicates severe inter-loop crosstalk. In this case, the device will downweight the evidence of the responsible loop identification and mark its confidence level as low in the final conclusion.

[0134] Low consistency scenario: If the consistency score from Example 4 is low (for example, the warning feature of Example 3 points to condensation, but the transient fingerprint of Example 4 shows a strong electric arc), it indicates a contradiction within the chain of evidence. In this case, the device will downweight the confidence level of the overall operating mode or tripping conclusion, or mark the conclusion as conflicting evidence.

[0135] Low evidence score scenario: If the action evidence score (poor temporal consistency) in step 502 or the suggestion confidence (unobvious features) in step 503 is low, the weight of the corresponding evidence in the fusion inference is also reduced.

[0136] Based on the aforementioned fusion weights, inferences are made to generate the final tripping conclusion and its confidence level. This conclusion is structured data, for example: {Event nature: real leakage (high confidence), responsible circuit: circuit 3 (low confidence, high crosstalk), operating mode: coil driven (high confidence), type recommendation: upgrade to Type A (high confidence)}.

[0137] Finally, the device extracts the key rules (e.g., DC component > 6mA in step 503) and key physical parameters (e.g., timing in step 502) encountered during the fusion process. Tacou_release This is packaged into a traceable explanatory summary for subsequent solidification and verification.

[0138] Example 6 describes the implementation process of evidence solidification and strategy closure. It explains how to securely solidify evidence and how to use these records to achieve long-term adaptive evolution of the system.

[0139] Step 601: Evidence consolidation and tracing.

[0140] In this embodiment, the device writes the tripping conclusion, conclusion confidence level, explanation summary, and key evidence relied upon in the inference process generated in Embodiment 5 into a complete event record. These key pieces of evidence preferably include: the operating mode and leakage protection type recommendation of Embodiment 5; the decoupled transient fingerprint (or its summary), responsibility loop identifier, and crosstalk index of Embodiment 4; the warning feature set of Embodiment 3; and the integrity and tamper-proof information generated in Embodiment 2 (e.g., hash values ​​or digital signatures of key data segments).

[0141] The entire event log is stored in the device's non-volatile memory. The purpose of this is to provide an irrefutable traceable record containing a complete chain of evidence for subsequent maintenance and repair (e.g., accurate replacement of the recommended upgraded RCD) or potential legal review.

[0142] Step 602, Strategy closed loop and adaptive backfeeding.

[0143] This step is crucial for the long-term optimization of this invention. The device (or a backend server analyzing data from these devices) can perform statistical analysis on massive amounts of event records to generate policy updates. These policy updates are then fed back into the relevant modules of the aforementioned embodiments.

[0144] In a preferred embodiment (reinjection to Example 3), the system, through statistical analysis, discovers that the false alarm rate of the condensation path statistics in Example 3 is high when the condensation state of a certain batch of devices is slight. Based on this, the system generates a strategy update and, in step 301 of reinjection to Example 3, fine-tunes its adaptive threshold set (e.g., increases the correction coefficient C_cond under slight condensation state) to reduce such false alarms in the future.

[0145] In another preferred embodiment (re-fed back to Embodiment 2), the system analysis reveals that for tripping of specific loads (e.g., frequency converters, corresponding to the Type F recommendation in Embodiment 5), the adaptively extended back-window data segments in step 203 of Embodiment 2 are generally too short, resulting in insufficient evidence for Embodiment 5 when analyzing high-frequency energy. Based on this, the system generates a strategy update, re-fed back to step 202 of Embodiment 2, adjusting its energy budget parameters and instructing it to allocate more energy budget to the back-window when a warning of such loads is detected.

[0146] Through this strategy feedback mechanism, the method of the present invention can achieve long-term adaptive evolution of key strategies such as threshold and energy allocation, and continuously optimize its detection accuracy and evidence quality under different environments and loads.

[0147] According to one aspect of this application, in a certain embodiment, the data processing flow is as follows:

[0148] The device reads raw data on configuration and near-end voltage, and detects observable signs of trip coil back electromotive force, line residual voltage, bus capacitance, overcapacitance, and backup battery within a preset micro-short window to obtain a list of candidate energy sources.

[0149] The list of candidate energy sources is retrieved, and small disturbance tests or historical event replays are performed on each energy source to estimate the equivalent internal resistance, maximum pulse current, minimum available voltage and temperature coefficient, and generate an energy source electrical parameter table.

[0150] Based on the electrical parameter table of the energy source and the environmental conditions, we checked the overvoltage, overcurrent, temperature resistance and insulation margin, eliminated non-compliant items, and obtained a set of compliant energy sources.

[0151] Obtain a set of compliant energy sources and an electrical parameter table of energy sources, and score them by weighting based on available energy, response time, safety margin, and lifetime impact to form an energy source priority table and fallback path.

[0152] Based on the energy source priority table and backoff path, determine the conditions for power failure entry, maintenance, and exit, and output a self-sustaining mode switching table.

[0153] Based on the device configuration, the average and peak power consumption of the statistical processor, memory write, sampling front-end, and clock hold are used to form a system power consumption model.

[0154] By using an energy source priority table and a system power consumption model, the sustainable time and minimum operating voltage are calculated, and energy budget parameters are generated.

[0155] Based on the energy budget parameters, allocate target window lengths and sampling duty cycles for microsecond-level high-bandwidth segments, millisecond-level slow sampling segments, and metadata writing, and mark "must-retain segments".

[0156] Based on the energy budget parameters, target window length, and sampling duty cycle, a sparse sampling ratio, key point retention, and writing priority are determined for energy-deficient scenarios to obtain a degradation strategy.

[0157] Read the target window length, sampling duty cycle, and degradation strategy; complete offline dry run verification; output a budget verification report; and lock the parameters.

[0158] Based on the target window length, sampling duty cycle, and device configuration, high bandwidth and slow sampling capacity are allocated to the differential current, near-end voltage, and near-field magnetoelectric field array to form a multi-channel buffer mapping table.

[0159] The multi-channel buffer mapping table and degradation strategy are invoked to set "fixed window before and variable window after" for high-bandwidth buffers and "sliding window and key point preservation" for slow sampling buffers, and write pointer strategy and overwrite strategy are generated.

[0160] Based on the time base sequence and device configuration, determine the channel start phase, sampling boundary and cross-domain alignment coefficient, and output the channel alignment parameters.

[0161] The write pointer strategy is invoked to generate rollback and retry rules for power outage fluctuations and buffer saturation, resulting in abnormal write protection rules.

[0162] Based on the write pointer strategy and channel alignment parameters, a fixed duration segment before the trigger point is solidified to obtain the preceding window data segment.

[0163] Based on the write pointer strategy and energy budget parameters, the post-term duration is dynamically extended according to the energy margin to obtain the post-term window data segment.

[0164] Read the degradation strategy and abnormal write protection rules, perform sparse sampling on the slow sampling channel and mark the peak value and the interval between peaks to obtain sparse fragments and key point annotations.

[0165] Based on the data fragments of the preceding window, the data fragments of the following window, and the sparse fragments and key point annotations, a time base sequence is bound, and event window data and write index are output; at the same time, integrity and anti-tampering information are generated.

[0166] By calling the early warning analysis segment, and after amplitude limiting and window function weighting, the root mean square trajectory and change slope are calculated to obtain the time domain intensity characteristics.

[0167] Based on the early warning analysis segment, short-time spectrum estimation is performed to calculate the spectral centroid and short-time energy density, thereby obtaining the spectral statistical characteristics.

[0168] By employing early warning analysis segments, energy is integrated at the fundamental wave and each harmonic to output the odd-even harmonic ratio characteristics.

[0169] By integrating time-domain intensity characteristics, spectral statistical characteristics, and odd-even harmonic ratio characteristics, and normalizing and compensating for temperature drift based on historical fault-free statistics, a normalized feature set is obtained.

[0170] Based on the normalized feature set, the signal-to-noise ratio and stability score are calculated, outliers are removed, and an early warning feature set is formed.

[0171] Using environmental conditions, estimate the dew point and generate the condensation state.

[0172] By combining the condensation state and the early warning feature set, the static threshold is corrected for condensation and temperature difference to obtain an adaptive threshold set.

[0173] Based on the adaptive threshold set and historical whole event records, the threshold is fine-tuned to obtain the calibrated adaptive threshold set.

[0174] Read the condensation status and environmental status, activate conservative thresholds under extreme conditions such as high humidity, high temperature or low temperature, and output special threshold strategies.

[0175] Based on the early warning feature set and historical whole event records, the fault-free distribution and covariance are estimated to obtain the fault-free model.

[0176] By reading the early warning feature set and the calibrated adaptive threshold set, the statistical sequences of condensation path and aging path are constructed to obtain the set of mechanism statistics.

[0177] By combining a set of mechanistic statistics, a fault-free model, and a special threshold strategy, sequential detection is performed and extreme value suppression is applied to upper-tail anomalies, outputting early warning trigger candidates.

[0178] The warning score is obtained by weighted fusion of the candidate warning triggers and the statistical set of mechanisms.

[0179] By combining the early warning score with historical event records, a threshold is set to balance false alarms and missed alarms, and an early warning judgment is output.

[0180] Based on microsecond-level data segments, the linear time axis is mapped to the logarithmic domain to amplify the resolution of the initial stage, outputting logarithmic domain data segments. Based on these logarithmic domain data segments, first- and second-order scattering coefficients are calculated using a custom wavelet, outputting a set of scattering features. Based on the logarithmic domain data segments, the arc initiation rising edge is detected, and saturation and spurious peaks are removed, yielding rising edge segments. Within these rising edge segments, the channel conductance rise rate and contact bounce interval are fitted using a simplified arc model, resulting in a set of physical parameters.

[0181] By integrating the set of scattering features with the set of physical parameters, a transient fingerprint is formed and a summary for solidification is generated.

[0182] The geometric calibration matrix and device configuration are read, and near-field radiation atoms for each loop are generated based on the conductor orientation and position, forming a near-field radiation dictionary. Combining the near-field radiation dictionary, array spatiotemporal characteristics, and microsecond-level data fragments, a non-negative sparse reconstruction problem is established with arrival time difference and amplitude-phase consistency as hard constraints, resulting in a constrained reconstruction model. The constrained reconstruction model is invoked, and the source components are recovered through iterative optimization, outputting the source component set and the decoupled transient fingerprint. Based on the source component set, array spatiotemporal characteristics, and device configuration, responsible loops are matched according to energy and arrival time difference consistency, and the residual energy ratio is calculated as a crosstalk index, outputting the responsible loop identifier and crosstalk index. The decoupled transient fingerprint, responsible loop identifier, and crosstalk index are summarized to generate an entry marker.

[0183] Based on the fusion feature library, the rising edge, holding plateau, and attenuation segment of the coil magnetic field are extracted, and a set of magnetic features is output. The fusion feature library is read, and the sub-band energy, spectral entropy, and bounce timing of the contact release acoustic spectrum are calculated, and a set of acoustic signature features is output.

[0184] Based on the physical parameters in the magnetic feature set, acoustic feature set, and transient fingerprint, and applying the time sequence constraint of "coil excitation → contact separation → arc initiation parameters," the action mode and action evidence score are output. The load type evidence set and device configuration are read to form a joint input of DC component, high-frequency energy ratio, and odd / even harmonic ratio, outputting the type criterion input. The type criterion input and the existing type are read, matched against the rule base, and a draft of the leakage current protection type recommendation and explanation information are generated. The draft of the leakage current protection type recommendation and historical whole event records are called, the confidence level is synthesized and calibrated, and the leakage current protection type recommendation and recommendation confidence level are output. The action mode, action evidence score, load type evidence set, leakage current protection type recommendation, recommendation confidence level, responsibility circuit identifier, crosstalk index, and consistency score are read, and high crosstalk or low consistency scenarios are weighted down, outputting a fused weight set. Based on the fused weight set and each piece of evidence, a tripping conclusion and conclusion confidence level are generated, including event nature, responsibility circuit, action category, and type recommendation. Based on the tripping conclusion and upstream evidence, key hit rules and physical parameters are extracted to generate an explanatory summary for consolidation and review.

[0185] To address the issue of evidence loss, an evidence-priority energy-buffer joint orchestration strategy is adopted. This strategy does not rely on mains power but actively identifies and models residual energy at the millijoule level, such as the back electromotive force of the trip coil. Through refined energy budgeting and degradation strategies, this extremely limited energy is prioritized for essential segments (i.e., high-bandwidth data at the microsecond level before the trigger point), and combined with trigger commands based on short-window warnings, it ensures that the most valuable transient arc fingerprint is completely solidified before energy depletion.

[0186] To address the spatial crosstalk problem, a geometrically constrained sparse electromagnetic tomography decoupling technique is employed. A near-field radiation dictionary characterizing the conductor's trajectory is constructed using the geometric calibration matrix within the device configuration. It innovatively incorporates arrival time difference and amplitude-phase consistency as hard constraints and performs sparse reconstruction of microsecond-level short-window data based on the sparse assumption of a single primary source. This approach effectively separates the primary source signal from radiated crosstalk, thereby enabling the location of the responsible loop with high confidence.

[0187] A transient fingerprint robust to non-stationary ringing was constructed by combining log-time scattering fingerprinting with arc physics fitting technology. A hard constraint for magneto-acoustic-electrical timing consistency was applied to distinguish between real leakage current and mechanical impact based on the physical mechanism (coil excitation → contact separation → arc initiation). Finally, causal type suggestions were used to indicate the correlation between tripping and load type (such as DC component).

[0188] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A method for detecting the tripping of a leakage current protection switch in an electricity metering box, characterized in that, include: Upon detecting a power failure, energy and buffer joint orchestration is performed based on multimodal raw data and device configuration to generate event context data. Based on event context data, perform short-window alert analysis to obtain alert triggering instructions and alert feature sets; In response to the warning trigger command, and based on the event context data and device configuration, a transient fingerprint is constructed and parallel crosstalk decoupling is performed to obtain the transient fingerprint and decoupling conclusion; By integrating transient fingerprints with decoupling conclusions, early warning feature sets, and event context data, and applying multimodal timing consistency constraints, tripping conclusions and leakage protection type recommendations are formed. The event context data includes event window data, time base sequence and environmental state, while the transient fingerprint and decoupling conclusions include the decoupled transient fingerprint, responsibility loop identifier and crosstalk index.

2. The method according to claim 1, characterized in that, Constructing transient fingerprints and performing parallel scrambling decoupling yields transient fingerprints and decoupling conclusions, including: In response to an alert trigger command, extract high-bandwidth microsecond-level data fragments from the event context data; Log-time resampling is performed on microsecond-level data segments, and scattering coefficients are calculated to characterize high-frequency non-stationary textures. The rising edge physical parameters are then fitted using an arc model to synthesize transient fingerprints. Based on event context data and device configuration, construct array spatiotemporal characteristics and geometric calibration matrix; By using the array's spatiotemporal features and geometric calibration matrix as geometric constraints, sparse tomography decoupling is performed on microsecond-level data segments to separate the main source and radiative crosstalk, yielding transient fingerprints and decoupling conclusions.

3. The method according to claim 2, characterized in that, Using the spatiotemporal features of the array and the geometric calibration matrix as geometric constraints, sparse tomography decoupling is performed on microsecond-level data segments, including: Based on the geometric calibration matrix and device configuration, a near-field radiation dictionary characterizing the orientation and direction of each loop conductor is constructed. A non-negative sparse reconstruction model for microsecond-level data segments is established using a near-field radiation dictionary and hard constraints such as arrival time difference and amplitude-phase consistency in the array's spatiotemporal features. Solve the nonnegative sparse reconstruction model to recover the source components and obtain the decoupled transient fingerprint; Based on the geometric consistency between the energy of the source component and the spatiotemporal characteristics of the array, the responsible loop identifier is obtained by matching, and the residual energy ratio is calculated to define the crosstalk index.

4. The method according to claim 2, characterized in that, Synthesized transient fingerprints include: The linear time axis of the microsecond-level data segment is mapped to the logarithmic domain to amplify the resolution of the initial stage, resulting in a logarithmic domain data segment. On the logarithmic domain data fragment, the first and second order scattering coefficients are calculated to form a set of scattering features characterizing the high-frequency amplitude modulation and frequency modulation textures; The arc-starting rising edge in the logarithmic domain data segment is detected, and the rising edge is fitted using a simplified arc model to obtain a set of physical parameters including the channel conductance rise rate or the contact bounce interval. By integrating the set of scattering features with the set of physical parameters, a transient fingerprint is obtained.

5. The method according to claim 1, characterized in that, Generate event context data, including: Read multimodal raw data and device configuration, identify and evaluate candidate energy sources to establish an energy source priority table; Based on the energy source priority table and device configuration, the preset system power consumption model is read and decomposed, and sampling resources are allocated according to the importance of evidence in order to formulate energy budget parameters and degradation strategies. When power fails, it switches to self-sustaining mode and performs evidence-based orchestration based on energy budget parameters and degradation strategies to capture and generate event context data.

6. The method according to claim 5, characterized in that, The steps for identifying and evaluating candidate energy sources to establish an energy source priority table include: Read the near-end voltage raw data from the multi-mode raw data, and in conjunction with the device configuration, detect and enumerate candidate energy sources including trip coil back electromotive force, line residual voltage and bus capacitance to form a candidate energy source list; Evaluate the equivalent internal resistance and pulse power supply capability of each item in the candidate energy source list to screen out the set of compliant energy sources; The set of compliant energy sources is weighted and scored according to available energy, response time, and safety margin to form an energy source priority table.

7. The method according to claim 5, characterized in that, The orchestration is prioritized based on energy budget parameters and evidence-based implementation of degradation strategies, including: Based on energy budget parameters, target window lengths and sampling duty cycles are assigned to microsecond-level high-bandwidth segments and millisecond-level slow-sampling segments according to the importance of evidence. During the capture process, a high-bandwidth microsecond segment of a fixed duration before the trigger point is prioritized and used as the data segment of the preceding window. When energy permits, the data fragments in the post-window are adaptively extended according to the energy budget parameters; When energy is scarce, a degradation strategy is implemented to sparsely sample and preserve key points for millisecond-level slow sampling segments. Combine the data fragments from the front window, the data fragments from the back window, and the sampling results, and bind them to a time base sequence to generate event context data.

8. The method according to claim 1, characterized in that, Based on event context data, short-window alert analysis is performed to obtain alert triggering instructions and alert feature sets, including: Extract environmental state from event context data, estimate dew point and condensation state, and combine with pre-stored historical whole event records to generate an adaptive threshold set; Locate the early warning analysis segment before the trip from the event context data, and extract time-domain intensity features, spectral statistical features and odd-even harmonic ratio to form an early warning feature set; Based on the early warning feature set and the adaptive threshold set, dual sequential detection is performed to obtain the early warning judgment; Based on the early warning determination, an early warning trigger command is generated according to the preset hysteresis and debounce strategy.

9. The method according to claim 8, characterized in that, Based on the early warning feature set and the adaptive threshold set, the steps for performing dual sequential detection include: For the early warning feature set, a condensation path statistic is constructed to characterize low-frequency drift and intermittent conductivity; For the early warning feature set, an aging path statistic representing monotonically increasing and rightward shift of the spectral center is constructed; Sequential detection is performed in parallel for condensation path statistics and aging path statistics. In sequential detection, extreme value modeling is used for upper tail anomalies to suppress rare noise triggering, and the detection results from both channels are fused to generate an early warning judgment.

10. The method according to claim 8, characterized in that, Combine historical event records to generate an adaptive threshold set, including: Raw temperature and humidity data are read from the environment to estimate the dew point and determine the condensation state. Based on the condensation state and the temperature difference between the inside and outside of the chamber, the preset static threshold is modified to obtain the modified threshold. Fine-tune the corrected thresholds using pre-stored historical whole event records, and output an adaptive threshold set.

Images