A fault diagnosis method for an electric energy metering box

Abstract

This invention discloses a fault diagnosis method and system for electricity metering boxes. The method includes: synchronously collecting electrical and non-electrical quantity data within the metering box; running an improved LMS adaptive filtering algorithm in an edge computing node to extract electrical quantity features, while simultaneously extracting non-electrical quantity frequency band energy entropy features through wavelet packet transform; constructing a spatiotemporal attention fusion network to generate a fused fault feature vector; constructing a digital twin in the cloud, and using multi-physics coupling simulation to deduce the fault evolution path and remaining lifespan; setting a dynamic alarm threshold based on the fault evolution path and executing self-healing control. This invention integrates multi-source heterogeneous information and combines federated distillation learning to ensure data privacy, achieving early warning and predictive maintenance of metering box faults, exhibiting strong robustness and physical interpretability.

Description

Technical Field

[0001] This invention relates to the field of power system automation and intelligent manufacturing technology, specifically to a method for diagnosing faults in an energy metering box. Background Technology

[0002] As a core terminal node of the smart grid, the electricity metering box undertakes important functions such as electricity metering, load control, and safety isolation. The stability of its operation directly affects the accuracy of trade settlement and the reliability of power supply. However, existing fault diagnosis technologies for electricity metering boxes still have many shortcomings, making it difficult to meet the high reliability and high intelligence requirements of modern power systems.

[0003] First, the sensing dimension is limited, leading to a high false negative rate. Existing diagnostic methods, such as those based on LMS filtering, primarily rely on electrical quantity data such as voltage and current. However, extensive field operation data shows that early faults such as poor contact, loose screws, and insulation aging exhibit extremely weak changes in electrical quantities (such as the effective values ​​of current and voltage) in their initial stages, sometimes even completely masked by normal fluctuations. These faults often manifest in the early stages as accompanying mechanical vibrations, localized temperature rises, or ultrasonic emissions. Monitoring methods based on a single electrical quantity cannot capture these early signs, resulting in diagnostic delays. Often, the fault is only discovered when it has already caused significant consequences (such as fire or power outage).

[0004] Secondly, there are data silos and privacy and security risks. To achieve big data analytics, traditional solutions typically require uploading massive amounts of waveform data from various distribution areas and metering boxes to cloud servers in real time. This not only consumes significant communication bandwidth and increases operating costs, but more importantly, user-side electricity consumption data constitutes highly sensitive personal privacy information. According to the "Electricity Data Security Specification" (GB / T 36572-2018) and related laws and regulations, centralized storage and cross-regional transmission of raw electricity consumption data pose serious compliance risks.

[0005] Secondly, static thresholds have poor adaptability. Existing systems mostly use fixed thresholds for alarm judgment (for example, setting an alarm when the current exceeds 100A). However, the operating environment of metering boxes is complex and variable; changes in ambient temperature, humidity, and load characteristics can significantly affect the physical state of the equipment. For example, in the high-temperature environment of summer, the temperature rise caused by the same current load will be much higher than in winter. Fixed thresholds will either lead to frequent false alarms at high temperatures or missed alarms at low temperatures.

[0006] Finally, there is a lack of predictive maintenance capabilities. Most existing technologies remain at the level of "post-incident alerts," meaning they only issue warnings when a fault has occurred or is about to occur. The system cannot predict the evolution of the equipment's health status over a future period, nor can it assess the Remaining Useful Life (RUL). This results in maintenance work always being in a reactive "repair" mode, unable to achieve proactive "preventive maintenance," leading to low maintenance efficiency and high costs.

[0007] In summary, there is an urgent need for a new diagnostic method and system that can integrate multi-source information, protect data privacy, have environmental adaptability, and achieve fault prediction, in order to solve the aforementioned pain points in existing technologies. Summary of the Invention

[0008] The present invention aims to solve the above-mentioned technical problems and provide an intelligent solution that shifts from "passive diagnosis" to "active prediction".

[0009] To achieve the above objectives, the present invention is implemented through the following technical solution: A method for diagnosing faults in an electricity metering box, characterized by comprising the following steps: Step S1: Synchronously collect electrical and non-electrical quantity data in the metering box. The electrical quantity data includes at least three-phase voltage, three-phase current and zero-sequence current. The non-electrical quantity data includes at least box vibration acceleration, ultrasonic partial discharge signal, infrared thermal imaging temperature and ambient temperature and humidity. Step S2: In the edge computing nodes deployed in the distribution transformer area, the improved LMS adaptive filtering algorithm is run to perform power frequency tracking and noise filtering on the electrical quantity data, and extract the fundamental amplitude, phase and harmonic distortion rate features; at the same time, wavelet packet transform is performed on the non-electrical quantity data to extract the frequency band energy entropy features. Step S3: Construct a spatiotemporal attention fusion network, using the electrical quantity features as query vectors and the non-electrical quantity features as key vectors and value vectors, calculate the feature association weight matrix, and generate a fused fault feature vector; Step S4: Construct a digital twin on the cloud server that is consistent with the physical structure of the metering box, map the fused fault feature vector to the digital twin in real time, drive the twin model to update its state, and deduce the fault evolution path and remaining life through multi-physics coupling simulation. Step S5: Based on the fault evolution path, set a dynamic alarm threshold that changes with environmental parameters and output the diagnostic results; when the diagnostic results indicate a serious fault, perform load switching through the self-healing control module.

[0010] The invention further includes the following configuration: at midnight each day, the edge computing node controls an analog switch to cut off the input of the external current transformer and switch to the internal reference voltage source path; using the internal high-precision reference voltage source Vref, a standard excitation signal Sstandard is generated through a precision voltage divider resistor network; the ADC is controlled to acquire the standard excitation signal to obtain the measured value Smeasured; the gain calibration coefficient G is calculated using the formula: G=Smeasured / Sstandard, where Sstandard is the nominal value of the standard excitation signal and Smeasured is the actual value acquired by the ADC; The calibration coefficient G is stored in non-volatile memory, and the original sampled value is multiplied by G to complete the compensation in all subsequent data acquisition processes.

[0011] The present invention further includes, in step S2, the improved LMS adaptive filtering algorithm comprising: introducing a normalized variable step size factor μ(n), the update formula of which is defined as follows: , where β is the smoothing coefficient, with a value range of [0.9, 0.99]; e(n) is the error signal of the nth iteration; The square of the Euclidean norm of the input signal vector; To prevent extremely small constants with a denominator of zero; The algorithm adaptively adjusts the convergence speed based on the input signal power, increasing the step size for fast tracking during signal abrupt changes and decreasing the step size to reduce offset noise during steady-state conditions.

[0012] The present invention further specifies that the cross-modal spatiotemporal attention fusion in step S3 specifically includes: Time alignment: Cubic spline interpolation is used to resample vibration signals, ultrasonic signals and electrical signals with different sampling frequencies to a unified time reference; Spatial coding: Generate a spatial location coding vector Ppos based on the physical coordinates (X, Y, Z) of the sensor within the metering box; Attention calculation: The fused features are input into the multi-head attention mechanism, and the calculation formula is as follows:

[0013] The spatial location encoding enhances the correlation between sensor features that are physically close to each other.

[0014] The present invention further specifies that the digital twin construction method in step S4 includes: Import the CAD drawings of the metering box based on BIM technology to generate a three-dimensional mesh model that includes circuit breakers, busbars, terminal blocks and screws; Define the electromagnetic-thermal-mechanical coupling equations in the finite element analysis software: ;in R represents the Joule heat source; R' represents the contact resistance, which varies with temperature and pressure; H represents the magnetic field strength vector, in A / m; J represents the conduction current density vector, calculated from real-time acquired phase current data, in A / m. 2 D is the electric displacement vector, with units of C / m. 2 ρ represents the density of the busbar material, in kg / m³. 3 ; σ is the specific heat capacity of the busbar material, in J / (kg·K); T is the temperature field distribution function, in K; k is the thermal conductivity of the busbar material, in W / (m·K); σ is the stress tensor, in Pa; E is the elastic modulus matrix, in Pa; ε is the strain tensor, dimensionless. Real-time data collected at the edge is injected into the 3D model via the OPC UA protocol, driving dynamic changes in the model's color, deformation, and temperature field.

[0015] The present invention further specifies that the dynamic alarm threshold setting method in step S5 is as follows: Define a health index HI(t), with an initial value of 1, which decays over time and with the degree of failure. The alarm threshold Th(t) varies with ambient temperature. Dynamic adjustment, the formula is: ; in The baseline threshold is denoted by k, which is the temperature compensation coefficient (ranging from 0.001 to 0.003). For reference temperature; When HI(t) < 0.8, an early warning is triggered; when HI(t) < 0.5, an alarm is triggered and self-healing control is activated.

[0016] The invention further includes a federated distillation learning mechanism: Multiple edge computing nodes train a lightweight STAFN model locally, and only upload the soft labels output by the model to the cloud server; The cloud server uses a federated averaging algorithm to aggregate the soft labels of all edge nodes and generate global model parameters; The cloud distributes the updated global model parameters to each edge node to achieve collaborative model evolution, without transmitting the original user electricity consumption data, thus ensuring data privacy.

[0017] The present invention also provides a system for implementing the above method, comprising: Multimodal sensing layer: includes a wide dynamic range zero flux transformer, a MEMS triaxial vibration sensor, a 40kHz ultrasonic sensor, an infrared thermal imaging array, and a temperature and humidity sensor. The sensors are connected to the edge gateway via RS485 or LoRa bus. Edge computing layer: It adopts an industrial edge gateway equipped with an NPU, with a built-in ARM Cortex-A72 processor, a computing power of no less than 4 TOPS, and is equipped with SDRAM cache and security encryption chip; Digital twin layer: A microservice cluster deployed in the cloud, including a model management engine, a multiphysics simulation solver, and a database; Application interaction layer: including web visualization dashboard, mobile APP and operation and maintenance work order system.

[0018] The present invention further provides that the multimodal sensing layer also has a sensor self-calibration function: The gateway has a built-in high-precision reference voltage source and reference acceleration source; At midnight every day, the system automatically switches to self-test mode, applies a known standard signal to the measurement link, and calculates the calibration coefficient. For the accelerometer, it generates an equivalent electrostatic excitation by writing to its built-in self-test register, reads the output offset and compares it with the theoretical value to verify the integrity of the channel.

[0019] The invention further includes a self-healing control module: When the diagnosis result is "poor contact" or "transient overcurrent", the system automatically sends a command to the intelligent circuit breaker to perform the tripping operation; after a 3-second delay, it checks whether the fault has disappeared. If it has disappeared, it automatically recloses; if the fault persists, it switches to the backup power circuit and locks the faulty circuit until it is manually reset.

[0020] The beneficial effects of this invention are as follows: By integrating non-electrical characteristics such as vibration and ultrasound, the detection rate of early-stage latent faults (such as loose screws) is significantly improved. Employing a federated distillation learning mechanism, raw data remains on-site; only model parameters are uploaded, perfectly complying with national power data security regulations. Through multiphysics simulation using digital twins, fault development trends can be predicted hours to days in advance, shifting from "reactive repair" to "prevention." Sensor self-calibration and dynamic threshold mechanisms effectively eliminate the influence of temperature drift and environmental interference, significantly improving system robustness. Attached Figure Description

[0021] The accompanying drawings, which are provided to further illustrate this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application.

[0022] In the attached diagram: Figure 1 This is an overall flowchart of an embodiment of the present invention.

[0023] Figure 2 This invention presents the variable step size response characteristics of the improved LMS adaptive filtering algorithm.

[0024] Figure 3 This is the topology of the edge gateway sensor self-calibration circuit of the present invention. Detailed Implementation

[0025] The following will describe in detail the implementation of this application with reference to the accompanying drawings and embodiments, so that the implementation process of how this application uses technical means to solve technical problems and achieve technical effects can be fully understood and implemented accordingly. Example

[0026] The hardware architecture and selection for this system are as follows: The multimodal sensing layer mainly consists of an electrical quantity acquisition unit: This unit uses a zero-flux current transformer with a rated primary current of 50A and a rated secondary output current of 5mA. This signal is converted into a voltage signal via a sampling resistor (100Ω, 0.1% accuracy) and then input to an analog-to-digital converter (ADC). The ADC uses an AD7606 chip with 16-bit resolution, synchronously sampling seven channels: three-phase voltage, three-phase current, and zero-sequence current, at a sampling frequency of 6.4kHz. The non-electrical quantity acquisition unit... Vibration monitoring: An ADI ADXL1002 MEMS accelerometer was selected and fixed to the housing near the fastening bolts of the A, B, and C three-phase input terminals with epoxy resin. It was used to capture mechanical vibrations caused by loose connections. Its X, Y, and Z axis signals were conditioned and then input into the ADC.

[0027] Partial discharge monitoring: An ultrasonic sensor (UAT-40) with a center frequency of 40kHz is selected and mounted on the top plate inside the metering box, facing the busbar area.

[0028] Temperature monitoring: The MLX90640 infrared thermal imaging array sensor is selected and installed on the inside of the door. Driven by a micro stepper motor, it can periodically scan the main electrical connection points inside the box.

[0029] All sensors are connected to the edge computing gateway’s RS485 bus or analog input interface via shielded cables.

[0030] Edge computing layer: An industrial edge gateway (model: Huawei AR502H) with an integrated Neural Processing Unit (NPU) is adopted. Its core configuration includes: an ARM Cortex-A72 quad-core processor, 4GB DDR4 memory, 32GB eMMC storage, and an NPU with 4 TOPS INT8 computing power. This gateway is responsible for running the data acquisition, filtering, feature extraction, and fusion algorithms described in claims 1-3 and 5-6.

[0031] Cloud-based digital twin layer and application layer: The cloud server adopts a microservice architecture and deploys services required for use cases, such as digital twin model management, multiphysics simulation solving (based on COMSOL Multiphysics Runtime), and federated learning aggregation. The application layer provides web and mobile visualization interfaces.

[0032] To achieve the aforementioned sensor self-calibration function, a dedicated calibration circuit was designed inside the edge gateway, with the following topology: Figure 3 As shown.

[0033] Every day at 00:00, the system automatically executes the following process: Switching to the calibration path: The microcontroller (MCU) issues a command to control the analog switch (such as ADG1419) to disconnect from the external current transformer and simultaneously connect the internal reference voltage source path. The reference voltage source is REF5025, with an output Vref=2.500V (accuracy ±0.05%).

[0034] Generate standard excitation: Vref passes through a voltage divider network consisting of high-precision low-temperature drift resistors R1 (95kΩ) and R2 (5kΩ) (both with an accuracy of 0.01%) to generate the theoretical standard voltage value Vstandard: Vstandard = Vref × R2 / (R1 + R2) = 2.5 × 5 / (95 + 5) = 0.125V. This Vstandard is Sstandard.

[0035] Acquisition and Calculation: The ADC continuously samples Vstandard (e.g., 256 points), discards invalid data from switching transients, calculates the average code value, and converts it to the actual measured voltage Vmeasured. Then, the gain calibration coefficient G is calculated using the following formula: G=Vstandard / Vmeasured=Sstandard / Smeasured; Storage and Compensation: The calculated G value is stored in non-volatile memory (FRAM). During the calibration period, all raw sampled values ​​Sraw of this acquisition channel are compensated in real time according to Scalibrated = G × Sraw to eliminate gain drift of the sensor and signal link.

[0036] Since static electronic equipment cannot generate known acceleration excitation, this system uses an electrical simulation self-test method to verify the integrity of the acceleration channel: at the daily zero-point calibration time, the self-test register (SELF_TEST bit) built into the ADXL1002 accelerometer is read, and 0x01 is written to the configuration register 0x3D to enable the internal electrostatic excitation, causing the sensor to produce a known, fixed output change (the datasheet specifies an equivalent offset of 100mV / g). The ADC reads this offset; if the measured value deviates from the theoretical value by ±10%, the channel is deemed faulty and an alarm is generated.

[0037] The specific implementation of the improved LMS adaptive filtering algorithm: Within the edge gateway, the acquired A-phase current signal ia(n) is subjected to improved LMS filtering to extract fault characteristics. The specific steps are as follows: Algorithm initialization: Set the filter order L=32, the initial weight vector ω(0) to zero, the initial step size μ(0)=0.01, the smoothing coefficient β=0.98, and the constant ϵ=10. −6 Reference signal generation: A precise power frequency phase is extracted from the grid voltage signal using a digital phase-locked loop (PLL) to generate a pure sinusoidal reference signal d(n) with the same frequency and phase.

[0038] Iterative filtering and weight update: For each new sampling point x(n) (i.e., ia(n)), perform the following calculations: Construct the input vector: x(n) = [x(n), x(n−1), ..., x(n−L+1)] T .

[0039] Calculate the filter output: y(n) = ω T (n)x(n).

[0040] Calculation error: e(n) = d(n) − y(n). This error signal e(n) is the residual containing fault characteristics such as harmonics and interharmonics after filtering out the power frequency component.

[0041] Update normalization step size ; Let x(n) be the square of the Euclidean norm of the input vector x(n).

[0042] Update filter weights: ω(n+1)=ω(n)+2μ(n)e(n)x(n).

[0043] Under a signal-to-noise ratio (SNR) of 20 dB, the standard LMS (fixed step size μ = 0.01) and this improved LMS are compared: Standard LMS: steady-state offset error -25dB, convergence time 0.2 seconds, tracking time after signal abrupt change 0.15 seconds.

[0044] Improved LMS: steady-state offset error -38dB, convergence time 0.08 seconds, tracking time after signal abrupt change 0.04 seconds.

[0045] like Figure 2 As shown, when a sudden load change causes distortion of the current waveform, the variable step size mechanism automatically increases μ(n), and the algorithm quickly tracks the signal change; in steady state, μ(n) automatically decreases, effectively suppressing output noise, thereby extracting weak features characterizing early faults more accurately.

[0046] The construction and operation of the cross-modal spatiotemporal attention fusion network is as follows: a lightweight spatiotemporal attention fusion network (STAFN) is deployed on the NPU of the edge gateway.

[0047] Feature extraction and time alignment: Perform an FFT on the residual e(n) after LMS filtering to extract the amplitudes of its fundamental, 3rd, 5th and 7th harmonics, forming a 4-dimensional electrical feature vector Fe.

[0048] The vibration acceleration signal was subjected to three-level wavelet packet decomposition (db4 wavelet) to calculate the energy entropy of eight sub-bands, forming an eight-dimensional vibration feature vector Fv. Since the update rate of Fe is 1Hz and the update rate of Fv is 100Hz, cubic spline interpolation was used to resample Fv to a 1Hz timestamp to achieve time alignment with Fe.

[0049] Spatial location encoding: Define three-dimensional coordinates (x, y, z) (unit: meters) for each sensor in the metering box. For example, the coordinates of the A-phase current transformer are (0.05, 0.20, 0.15). Generate a location encoding vector Ppos for each coordinate using sine and cosine functions, and add it to the corresponding feature vector to inject spatial information into the feature.

[0050] Attention fusion calculation: The scrambled electrical feature vector Fe is used as the query matrix Q.

[0051] After concatenating the scrambled non-electrical feature vectors such as vibration and temperature, they are used as the key matrix K and value matrix V. Scaled dot product attention is then calculated.

[0052] ; is the dimension of the feature vector. The weight matrix output by the softmax function reflects the spatiotemporal correlation strength between electrical features and various non-electrical features.

[0053] Example: When the A-phase terminal is loose, the A-phase harmonic quantity in the electrical feature Fe is abnormal, and at the same time, the frequency band energy entropy corresponding to the A-phase coordinate in the vibration feature Fv increases. The attention mechanism will assign high weights to these two features, making the fused feature vector significantly point to "A-phase mechanical fault".

[0054] The construction of digital twins and multiphysics coupling simulation involves building a high-fidelity digital twin for each physical metering box in the cloud. Geometric and physical modeling is based on the CAD drawings of the metering boxes, generating precise geometric models containing circuit breakers, busbars, and terminals in 3D modeling software.

[0055] In COMSOL Multiphysics, define material properties (such as electrical conductivity, density, specific heat capacity, and thermal conductivity of copper) and boundary conditions (such as convective heat dissipation coefficient) for the model.

[0056] Establish the multiphysics coupling control equations: in R represents the Joule heat source; R' represents the contact resistance, which varies with temperature and pressure; H represents the magnetic field strength vector, in A / m; J represents the conduction current density vector, calculated from real-time acquired phase current data, in A / m. 2 D is the electric displacement vector, with units of C / m. 2 ρ represents the density of the busbar material, in kg / m³. 3 ; σ is the specific heat capacity of the busbar material, in J / (kg·K); T is the temperature field distribution function, in K; k is the thermal conductivity of the busbar material, in W / (m·K); σ is the stress tensor, in Pa; E is the elastic modulus matrix, in Pa; ε is the strain tensor, dimensionless. Fault Simulation: Assume the edge system diagnoses "abnormally increased contact resistance at phase A terminal". In the digital twin, the contact resistance parameter at the corresponding location is modified from the normal value of 50μΩ to the simulated fault value of 500μΩ. A transient simulation is initiated, inputting the current load current, to solve for the temperature and stress field distributions over a future period. The simulation results show that the temperature at this fault point will exceed the safety threshold of 180℃ after 4 hours. Based on this, the system outputs a predictive maintenance instruction: "Phase A terminal has an overheating risk; estimated remaining usable time (RUL) is approximately 4 hours."

[0057] The operation flow of the federated distillation learning mechanism, the model training process adopts the federated distillation architecture, and the flow is as follows: Local training: The edge gateways of each distribution area independently train a lightweight STAFN model using locally stored historical data (such as normal and fault data for the past month). After training, the model's prediction results for a batch of samples are obtained—soft labels (Logits), which are the probability distribution vectors of each category (such as normal, phase A fault, phase B fault, etc.).

[0058] Parameter Upload: Each gateway uploads the encrypted soft tags (Logits) and the corresponding number of samples to the cloud-based federated learning server. During this process, the user's original current and voltage waveform data remain locally and are not transmitted, thus meeting data privacy protection requirements.

[0059] Cloud-based aggregation and distillation: The cloud server collects encrypted information from all participating nodes and aggregates these soft tags using the FedAvg algorithm to form a more comprehensive "global model." Subsequently, knowledge distillation technology is used to transfer the knowledge from the global model to a new, efficient, and lightweight model.

[0060] Model Deployment: The cloud encrypts and distributes the optimized, lightweight model parameters to all edge gateways. The gateways update their local models, thus achieving global knowledge sharing and collaborative performance improvement without sharing the original data.

[0061] This embodiment also includes the execution logic for dynamic threshold diagnosis and self-healing control. The system's diagnosis and control logic is as follows: ; Dynamic threshold setting: The alarm threshold Th(t) is not a fixed value. For example, for a temperature alarm, a baseline threshold is set. =80℃, temperature compensation coefficient k=0.002 / ℃, reference temperature =25℃. Therefore, when the ambient temperature... At 40℃, the dynamic threshold is: Th(t)=80×[1+0.002×(40−25)]=82.4℃ This mechanism avoids false alarms caused by changes in ambient temperature.

[0062] Health assessment and diagnosis: The system integrates real-time data and digital twin prediction results to calculate the equipment health index HI(t)∈[0,1]. A warning is triggered when HI(t)<0.8, and an alarm is triggered when HI(t)<0.5.

[0063] Self-healing control execution: When a recoverable fault such as "transient overcurrent" is diagnosed, the self-healing control module automatically executes the following sequence: t0: The edge gateway sends a trip command to the smart circuit breaker of this loop.

[0064] t0+3s: The gateway sends a closing command to the circuit breaker.

[0065] After the circuit breaker is closed, the current is immediately checked. If the current returns to normal, the process ends and is recorded as "self-healing successful"; if the fault still exists, it is determined to be a permanent fault, the circuit breaker is tripped and locked again, and a maintenance work order is generated and pushed to the maintenance personnel, thus realizing the rapid isolation and recovery of momentary faults.

[0066] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention. As used in the specification and claims, certain terms refer to specific components. Those skilled in the art should understand that hardware manufacturers may use different names to refer to the same component. This specification and claims do not distinguish components based on differences in name, but rather on differences in function. The term "comprising" as used throughout the specification and claims is an open-ended term and should be interpreted as "comprising but not limited to". "Approximately" means that within an acceptable error range, those skilled in the art can solve the technical problem and substantially achieve the technical effect within a certain error range.

[0067] The foregoing description illustrates and describes several preferred embodiments of the present invention. However, as previously stated, it should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the inventive concept described herein through the foregoing teachings or techniques or knowledge in related fields. Any modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for diagnosing faults in an electricity metering box, characterized in that, Includes the following steps: Step S1: Synchronously collect electrical and non-electrical quantity data in the metering box. The electrical quantity data includes at least three-phase voltage, three-phase current and zero-sequence current. The non-electrical quantity data includes at least box vibration acceleration, ultrasonic partial discharge signal, infrared thermal imaging temperature and ambient temperature and humidity. Step S2: In the edge computing nodes deployed in the distribution transformer area, the improved LMS adaptive filtering algorithm is run to perform power frequency tracking and noise filtering on the electrical quantity data, and extract the fundamental amplitude, phase and harmonic distortion rate features; at the same time, wavelet packet transform is performed on the non-electrical quantity data to extract the frequency band energy entropy features. Step S3: Construct a spatiotemporal attention fusion network, using the electrical quantity features as query vectors and the non-electrical quantity features as key vectors and value vectors, calculate the feature association weight matrix, and generate a fused fault feature vector; Step S4: Construct a digital twin on the cloud server that is consistent with the physical structure of the metering box, map the fused fault feature vector to the digital twin in real time, drive the twin model to update its state, and deduce the fault evolution path and remaining life through multi-physics coupling simulation. Step S5: Based on the fault evolution path, set a dynamic alarm threshold that changes with environmental parameters and output the diagnostic results; when the diagnostic results indicate a serious fault, perform load switching through the self-healing control module.

2. The method according to claim 1, characterized in that, At midnight each day, the edge computing node controls the analog switch to cut off the input of the external current transformer and switch to the internal reference voltage source path; using the internal high-precision reference voltage source Vref, a standard excitation signal Sstandard is generated through a precision voltage divider resistor network; the ADC is controlled to acquire the standard excitation signal to obtain the measured value Smeasured; the gain calibration coefficient G is calculated using the formula: G=Smeasured / Sstandard, where Sstandard is the nominal value of the standard excitation signal and Smeasured is the actual value acquired by the ADC; The calibration coefficient G is stored in non-volatile memory, and the original sampled value is multiplied by G to complete the compensation in all subsequent data acquisition processes.

3. The method according to claim 1, characterized in that, The improved LMS adaptive filtering algorithm in step S2 includes: introducing a normalized variable step size factor μ(n), whose update formula is defined as follows: , where β is the smoothing coefficient, with a value range of [0.9, 0.99]; e(n) is the error signal of the nth iteration; The square of the Euclidean norm of the input signal vector; To prevent extremely small constants with a denominator of zero; The algorithm adaptively adjusts the convergence speed according to the input signal power, increasing the step size to track quickly when the signal changes abruptly, and decreasing the step size to reduce offset noise when in steady state.

4. The method according to claim 1, characterized in that, The cross-modal spatiotemporal attention fusion in step S3 specifically includes: Time alignment: Cubic spline interpolation is used to resample vibration signals, ultrasonic signals and electrical signals with different sampling frequencies to a unified time reference; Spatial coding: Generate a spatial location coding vector Ppos based on the physical coordinates (X, Y, Z) of the sensor within the metering box; Attention calculation: The fused features are input into the multi-head attention mechanism, and the calculation formula is as follows: The spatial location encoding enhances the correlation between sensor features that are physically close to each other.

5. The method according to claim 1, characterized in that, The digital twin construction method in step S4 includes: Import the CAD drawings of the metering box based on BIM technology to generate a three-dimensional mesh model that includes circuit breakers, busbars, terminal blocks and screws; Define the electromagnetic-thermal-mechanical coupling equations in finite element analysis software: ;in Here, R is the Joule heat source, H is the contact resistance varying with temperature and pressure, J is the magnetic field strength vector (A / m), and J is the conduction current density vector (calculated from real-time acquired phase current data, A / m). 2 D is the electric displacement vector, with units of C / m. 2 ρ represents the density of the busbar material, in kg / m³. 3 ; σ is the specific heat capacity of the busbar material, in J / (kg·K); T is the temperature field distribution function, in K; k is the thermal conductivity of the busbar material, in W / (m·K); σ is the stress tensor, in Pa. E is the elastic modulus matrix, with units of Pa; ε is the strain tensor, which is dimensionless. Real-time data collected at the edge is injected into the 3D model via the OPC UA protocol, driving dynamic changes in the model's color, deformation, and temperature field.

6. The method according to claim 1, characterized in that, The method for setting the dynamic alarm threshold in step S5 is as follows: Define a health index HI(t), with an initial value of 1, which decays over time and with the degree of failure. The alarm threshold Th(t) varies with ambient temperature. Dynamic adjustment, the formula is: ; in The baseline threshold is denoted by k, which is the temperature compensation coefficient (ranging from 0.001 to 0.003). For reference temperature; When HI(t) < 0.8, an early warning is triggered; when HI(t) < 0.5, an alarm is triggered and self-healing control is activated.

7. The method according to claim 1, characterized in that, It also includes a federated distillation learning mechanism: Multiple edge computing nodes train a lightweight STAFN model locally, and only upload the soft labels output by the model to the cloud server; The cloud server uses a federated averaging algorithm to aggregate the soft labels of all edge nodes and generate global model parameters; The cloud distributes the updated global model parameters to each edge node to achieve collaborative model evolution, without transmitting the original user electricity consumption data, thus ensuring data privacy.

8. A system for implementing the method of any one of claims 1 to 6, characterized in that, include: Multimodal sensing layer: includes a wide dynamic range zero flux transformer, a MEMS triaxial vibration sensor, a 40kHz ultrasonic sensor, an infrared thermal imaging array, and a temperature and humidity sensor. The sensors are connected to the edge gateway via RS485 or LoRa bus. Edge computing layer: It adopts an industrial edge gateway equipped with an NPU, with a built-in ARM Cortex-A72 processor, a computing power of no less than 4 TOPS, and is equipped with SDRAM cache and security encryption chip; Digital twin layer: A microservice cluster deployed in the cloud, including a model management engine, a multiphysics simulation solver, and a database; Application interaction layer: including web visualization dashboard, mobile APP and operation and maintenance work order system.

9. The system according to claim 8, characterized in that, The multimodal sensing layer also has a sensor self-calibration function: The gateway has a built-in high-precision reference voltage source and reference acceleration source; At midnight every day, the system automatically switches to self-test mode, applies a known standard signal to the measurement link, and calculates the calibration coefficient. For the accelerometer, it generates an equivalent electrostatic excitation by writing to its built-in self-test register, reads the output offset and compares it with the theoretical value to verify the integrity of the channel.

10. The system according to claim 8, characterized in that, It also includes a self-healing control module: When the diagnosis result is "poor contact" or "transient overcurrent", the system automatically sends a command to the intelligent circuit breaker to perform the tripping operation; after a 3-second delay, it checks whether the fault has disappeared. If it has disappeared, it automatically recloses; if the fault persists, it switches to the backup power circuit and locks the faulty circuit until it is manually reset.

Images