A power distribution box fault early warning method and system

Abstract

This invention discloses a method and system for early warning of distribution box faults based on multi-physics energy fingerprinting and federated learning. The method first constructs a multi-physics energy fingerprint by deploying a piezoelectric energy harvester, electromagnetic induction coil, and micro-thermal radiation sensor within the distribution box, collecting resistive loss, hysteresis / eddy current loss, and thermal radiation energy in real time. Next, time-frequency analysis is performed on the energy signal to extract 15-dimensional energy features, which are then reduced to 8-dimensional principal features using principal component analysis. Subsequently, a lightweight CNN-LSTM model is trained at the edge based on compressed features to output the local fault probability, and a federated learning framework is used to achieve cross-regional model collaboration and dynamic threshold adjustment. Finally, a Bayesian network is used to locate the root cause of the fault and generate maintenance work orders. This invention achieves early, accurate, and low-power early warning of distribution box faults, effectively solving problems such as single data dimension, strong dependence on sensor power supply, and data privacy conflicts in traditional methods.

Description

Technical Field

[0001] This application relates to power equipment condition monitoring and fault early warning technology, specifically to a method and system for early warning of faults in distribution boxes. Background Technology

[0002] Traditional distribution box fault early warning technologies mainly rely on threshold alarm mechanisms based on single parameters such as temperature and current. This monitoring method has inherent drawbacks, including limited data dimensions, inability to capture early-stage hidden faults, difficulties in powering sensors, and data privacy conflicts across regions. In actual operation, internal faults in distribution boxes, such as poor contact or insulation aging, often manifest as abnormal energy dissipation in multiple physical fields, such as resistance loss, hysteresis loss, and eddy current loss, in their early stages. However, current technologies only focus on macroscopic electrical parameters and fail to incorporate these key energy characteristics into the analysis system, leading to the neglect of early fault signals. Simultaneously, sensor power supply issues severely restrict the practicality of monitoring systems. Existing solutions require external wiring or periodic battery replacements, which is particularly problematic in remote distribution box scenarios, making long-term maintenance-free operation difficult. Frequent maintenance not only increases costs but may also lead to monitoring failure due to power outages. Furthermore, the data from cross-regional distribution boxes is dispersed and sensitive. Centralized data processing is prone to privacy leaks, and the aging levels, environmental conditions, and material properties of equipment vary significantly across different regions. Centralized training models cannot adapt to individual needs, resulting in decreased early warning accuracy. Despite recent advancements in multi-sensor data-driven methods, existing technologies still fail to effectively integrate energy fingerprints from multiple physical fields, including resistive loss vibration energy, hysteresis eddy current alternating magnetic fields, and thermal radiation energy. They also exhibit weak self-powering capabilities and lack distributed training mechanisms that balance data privacy with model generalization. Therefore, improvements to existing technologies are urgently needed to address these issues. Summary of the Invention

[0003] The purpose of this application is to provide a method, system and storage medium for early warning of power distribution box faults, which can realize comprehensive monitoring of energy characteristics of multiple physical fields to effectively capture early hidden faults; solve the problem of sensor power supply difficulties through a self-powered mechanism; and achieve cross-regional collaborative early warning while protecting data privacy by utilizing federated learning.

[0004] To achieve the above objectives, the present invention proposes the following technical solution: A method for early warning of power distribution box faults includes the following steps: S1. Multi-physics field energy fingerprint acquisition: A piezoelectric energy harvester, an electromagnetic induction coil, and a micro-thermal radiation sensor are deployed in the distribution box to collect the resistance loss vibration energy when there is poor contact, the hysteresis / eddy current loss alternating magnetic field when the insulation is aging, and the thermal radiation energy in the abnormal temperature rise zone, respectively, to construct a multi-physics field energy fingerprint that includes resistance loss, hysteresis loss, eddy current loss, sound wave energy, and thermal radiation energy. S2. 15-Dimensional Energy Feature Extraction and Compression: Time-frequency domain analysis of the energy fingerprint is performed to extract and compress the following 15-dimensional energy features. Resistive characteristics of resistance loss: proportion of high-frequency component energy, variance of resistance loss energy fluctuation, attenuation rate of fundamental frequency component energy, and correlation between resistance and vibration energy. Hysteresis loss related characteristics: hysteresis loss spectrum kurtosis, high-frequency magnetic energy ratio, hysteresis loop area change rate, magnetic energy-temperature rise coupling coefficient; Eddy current loss characteristics: eddy current loss energy abrupt change rate, vibration-eddy current energy delay, eddy current frequency bandwidth; Sound wave energy related characteristics: Mel frequency cepstral coefficients MFCC-2, sound wave energy pulse interval; Characteristics related to thermal radiation energy: temperature rise gradient, spatial distribution entropy of thermal radiation energy; Principal component analysis was used to retain the first 8 principal features. S3, Federated Learning Collaboration Warning: At the edge, a lightweight CNN-LSTM model is trained based on compressed features to output the local fault probability. ; The regional cloud aggregates the model gradients of edge nodes using the FedAvg algorithm to update the regional model parameters. Global cloud platform generates differentiated early warning thresholds based on distribution box type. ; Calculate the dynamic adjustment threshold at the edge ,in Where α = 0.05 is the aging coefficient. The fatigue limit of the contact material. The dielectric constant of the insulating layer, The accumulated energy loss is calculated using the retained features; S4. Full lifecycle fault tracing: Locate the root cause of the fault by inferring the energy fingerprint source through Bayesian network and generate a work order with maintenance priority.

[0005] The invention further includes a piezoelectric energy harvester using a PZT-5A piezoelectric ceramic sheet, which is attached to the busbar joint and circuit breaker contacts. The mechanical energy of vibration is converted into electrical energy through a rectifier bridge, with an output voltage of 3.3V. The electromagnetic induction coil is made of 200 turns of 0.1mm diameter enameled wire and is installed in the gap between the insulator and the housing to sense the alternating magnetic field generated by hysteresis / eddy current loss.

[0006] The present invention further specifies that the micro thermal radiation sensor is an uncooled infrared focal plane detector, deployed on the surface of the enclosure, with a sampling frequency of 1Hz, communicating with the MCU through an SPI interface, detecting thermal radiation energy in the 8-14μm band, locating abnormal temperature rise areas, and the abnormal temperature rise area is determined by: area ≥ 2cm² and temperature gradient > 2℃ / s.

[0007] The present invention further provides that the self-powered edge computing node integrates a supercapacitor charging and discharging management circuit, which charges the capacitor when there is sufficient external energy input, and maintains MCU operation by discharging the capacitor when there is no input.

[0008] The present invention further specifies that in the CNN-LSTM model, the CNN layer uses 3×3 convolutional kernels, with a number of 32, a stride of 1, and is padded with the same kernel to extract spatial features; the LSTM layer has an input dimension of 32, hidden layer nodes of 64, and an output dimension of 1, and is trained by the Adam optimizer with a loss function of binary cross-entropy and a local training cycle of 100 rounds.

[0009] The present invention further provides that, in the federated learning framework, differential privacy technology is used when edge nodes upload gradients, and Gaussian noise is added to prevent the leakage of original data; after the gradients are aggregated in the regional cloud, the knowledge of the regional model is transferred to the global model through model distillation, thereby improving the generalization ability in small sample scenarios.

[0010] The present invention further provides that, in the dynamic threshold adaptive mechanism, the cumulative energy loss The sum of the absolute values ​​of the key energy features retained after PCA dimensionality reduction; Fitted through accelerated life testing is the dielectric constant of XLPE insulating material.

[0011] The present invention further provides that the graded early warning includes: like This triggers an emergency alert: the edge device calls the camera to detect personnel, and if no one is present, it immediately disconnects the shunt trip unit and sends an "Immediate Repair" work order. like This triggers a medium-level warning: the cooling fan or dehumidifier (10W power) is activated, and a "check within 48 hours" work order is pushed out. like This triggers a low-risk warning, calculates the remaining lifetime L=L0·exp(−β·cumulative energy loss) based on the Weibull distribution model, and generates a remaining lifetime prediction report, where L0=36 months and β=0.01.

[0012] The present invention also provides a distribution box fault early warning system for implementing the method, comprising: Multiphysics sensing module: piezoelectric energy harvester, electromagnetic induction coil, micro-thermal radiation sensor and matching energy harvesting circuit; Self-powered edge computing module: low-power MCU, MPPT energy management chip, supercapacitor and local inference model; Federated learning collaboration module: communication interfaces and model aggregation algorithms for edge computing, regional cloud, and global cloud; Fault tracing module: Bayesian network inference engine and maintenance work order generation unit; Active protection module: camera, cooling fan, dehumidifier and circuit breaker shunt trip unit.

[0013] The present invention also provides a storage medium storing a computer program that, when executed by a processor, implements the method described herein.

[0014] The beneficial effects of this invention are as follows: This solution breaks through the bottlenecks of traditional power distribution box early warning systems, which suffer from "single data dimension, power supply difficulties, and privacy conflicts," by adopting a novel technical approach of "energy fingerprint perception - self-powered edge processing - federated learning collaboration." It achieves a leap from "parameter monitoring" to "energy mechanism perception," from "centralized training" to "distributed privacy protection," and from "fault early warning" to "root cause localization," providing a new solution for intelligent operation and maintenance of power distribution boxes that is low-power, highly generalizable, and highly privacy-oriented. Attached Figure Description

[0015] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating an embodiment of the present invention. Detailed Implementation

[0016] 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.

[0017] like Figure 1 As shown, the present invention provides a method for early warning of power distribution box faults, comprising the following steps: S1. Multi-physics field energy fingerprint acquisition: A piezoelectric energy harvester, electromagnetic induction coil, and micro-thermal radiation sensor are deployed within the distribution box to collect the resistance loss vibration energy during poor contact, the alternating magnetic field of hysteresis / eddy current loss during insulation aging, and the thermal radiation energy in the abnormal temperature rise zone, respectively. This constructs a multi-physics field energy fingerprint encompassing resistance loss, hysteresis loss, eddy current loss, acoustic energy, and thermal radiation energy. The specific deployment is as follows: Three types of sensors are deployed to accurately capture early-stage energy dissipation during faults: Piezoelectric energy harvester: attached to the contact / busbar to collect arc vibration energy (1-10kHz), which is converted into electrical energy for the node through a rectifier bridge; Electromagnetic induction coil: installed in the gap of insulators, it induces hysteresis / eddy current loss magnetic field (100Hz-1MHz), and energy harvesting is optimized by MPPT circuit; Micro thermal radiation sensor: monitors thermal radiation of 8-14μm on the surface of the chamber and locates abnormal temperature rise areas (area ≥2cm² and gradient >2℃ / s).

[0018] S2. 15-Dimensional Energy Feature Extraction and Compression: Time-frequency domain analysis of the energy fingerprint is performed to extract and compress the following 15-dimensional energy features. Resistive characteristics of resistance loss: proportion of high-frequency component energy (the proportion of the energy of the 3-5kHz high-frequency component of the piezoelectric signal to the total energy), variance of resistance loss energy fluctuation (the time domain variance of resistance loss energy within 10 minutes), attenuation rate of fundamental frequency component energy (the attenuation ratio of the energy of the 50Hz fundamental frequency component of the current to the initial energy), and correlation between resistance and vibration energy (the Pearson correlation coefficient between resistance loss energy and piezoelectric vibration energy). Hysteresis loss related characteristics: hysteresis loss spectrum kurtosis (kurtosis value of the FFT spectrum of electromagnetic signal), high frequency magnetic energy ratio (the proportion of high frequency component energy of electromagnetic signal in 100kHz-1MHz to the total energy), hysteresis loop area change rate (the ratio of the reconstructed hysteresis loop area of ​​electromagnetic signal to the initial area), magnetic energy-temperature rise coupling coefficient (the linear regression coefficient of hysteresis loss energy and micro-thermal radiation temperature rise). Eddy current loss related characteristics: eddy current loss energy mutation rate (first-order difference mean of eddy current component energy of electromagnetic signal), vibration-eddy current energy delay (time difference between vibration signal peak and eddy current energy peak), eddy current frequency bandwidth (main lobe -3dB bandwidth of eddy current component spectrum of electromagnetic signal). Sound wave energy related characteristics: Mel frequency cepstral coefficients MFCC-2 (the second-order coefficients extracted from the sound wave signal by MFCC transformation), sound wave energy pulse interval (the average pulse interval when the sound wave energy exceeds the threshold). Characteristics related to thermal radiation energy: temperature rise gradient (the rate of change of temperature in the abnormal temperature rise zone over time), and spatial distribution entropy of thermal radiation energy (the information entropy of the energy distribution of the thermal radiation sensor array). Among them, the proportion of high-frequency component energy of resistive loss is normally ≤5%, but rises to 20%-40% during faults; the spectral kurtosis of hysteresis loss is normally ≤3, but ≥6 during faults; the energy mutation rate of eddy current loss is normally ≤0.2J / s, but ≥1.5J / s during faults; the acoustic energy MFCC-2 is normally ≤0.1, but ≥0.5 during faults; and the thermal radiation temperature rise gradient is normally ≤0.5℃ / s, but ≥2℃ / s during faults.

[0019] Principal component analysis was used to retain the first 8 principal features. S3, Federated Learning Collaboration Warning: At the edge, a lightweight CNN-LSTM model is trained based on compressed features to output the local fault probability. ; The regional cloud aggregates the model gradients of edge nodes using the FedAvg algorithm to update the regional model parameters. Global cloud platform generates differentiated early warning thresholds based on distribution box type. ; Calculate the dynamic adjustment threshold at the edge ,in Where α = 0.05 is the aging coefficient. The fatigue limit of the contact material. The dielectric constant of the insulating layer, The accumulated energy loss is calculated using the retained features; S4. Full lifecycle fault tracing: Locate the root cause of the fault by inferring the energy fingerprint source through Bayesian network and generate a work order with maintenance priority.

[0020] In practical applications, multiphysics energy fingerprinting can be understood as a set of energy features extracted from multiple physical fields, with the primary purpose of capturing energy dissipation anomalies in early latent faults. For example, similar functionality can be achieved by deploying other types of energy harvesting devices, such as photoelectric converters or thermoelectric generators. Furthermore, piezoelectric energy harvesters can employ other forms of vibration energy conversion devices, such as devices that convert mechanical vibrations into electrical energy using the principle of electromagnetic induction. For electromagnetic induction coils, different magnetic field strengths and frequency ranges can be adapted by changing the winding method or number of turns, for example, by using flat coils or multi-layer winding structures. Micro-thermal radiation sensors can be implemented by selecting detectors in different wavelength bands, such as short-wave infrared detectors or long-wave infrared detectors, to adapt to different thermal radiation characteristics.

[0021] The process of 15-dimensional energy feature extraction and compression is primarily aimed at focusing on key fault information and eliminating redundant data. Specifically, feature extraction can be achieved through other signal processing methods, such as wavelet transform or Fourier transform, to analyze the time-frequency characteristics of the signal. Principal component analysis can be replaced with other dimensionality reduction methods, such as linear discriminant analysis or independent component analysis, to achieve feature compression.

[0022] The core of the federated learning collaborative early warning mechanism lies in preventing data leakage through gradient aggregation rather than raw data transmission. As a preferred implementation, the lightweight model at the edge can be implemented using other machine learning algorithms, such as support vector machines or random forests. The model aggregation algorithm for the regional cloud can also employ other distributed optimization methods, such as asynchronous stochastic gradient descent or distributed alternating direction multiplier method. In the formula for dynamically adjusting the threshold, the aging coefficient α and the fatigue limit of the contact material are used. Dielectric constant of insulating layer The specific calculation method for cumulative energy loss Q can be adjusted according to the actual application scenario.

[0023] The full lifecycle fault tracing uses Bayesian networks to infer the source of energy fingerprints, aiming to locate the root cause component of the fault and generate a repair work order. Specifically, Bayesian networks can be replaced with other inference models, such as decision trees or neural networks, to achieve similar fault location functions.

[0024] The innovation of this application lies in its systematic solution to the problems of single data dimension, early latent fault detection, sensor power supply difficulties, and cross-regional data privacy conflicts through the collaborative design of multi-physics energy feature acquisition, feature compression, federated learning collaborative early warning, and fault tracing. Specifically, the construction of multi-physics energy fingerprints breaks through the limitations of single parameters, enabling accurate capture of energy dissipation anomalies in early latent faults such as poor contact and insulation aging. The feature extraction and compression process focuses on key fault information, improving the sensitivity and computational efficiency of early fault identification. The federated learning collaborative early warning mechanism avoids centralized leakage of cross-regional data through gradient aggregation, while dynamic thresholds combined with equipment aging parameters achieve adaptive adjustment. The fault tracing process achieves accurate fault location and maintenance guidance by inferring the source of energy fingerprints, forming a closed loop for full lifecycle management.

[0025] The working principle of this application embodiment is as follows: A piezoelectric energy harvester, an electromagnetic induction coil, and a micro-thermal radiation sensor are deployed in the distribution box to collect the vibration energy of resistance loss when there is poor contact, the alternating magnetic field of hysteresis / eddy current loss when the insulation is aging, and the thermal radiation energy in the abnormal temperature rise zone, thereby constructing a multi-physics energy fingerprint that includes resistance loss, hysteresis loss, eddy current loss, acoustic energy, and thermal radiation energy. Further, by performing time-frequency domain analysis on the energy fingerprint, 15-dimensional energy features are extracted and compressed. The resistance loss-related features include the proportion of high-frequency component energy, the variance of resistance loss energy fluctuation, the attenuation rate of fundamental frequency component energy, and the correlation between resistance and vibration energy; the hysteresis loss-related features include the kurtosis of the hysteresis loss spectrum, the proportion of high-frequency magnetic energy, the rate of change of the hysteresis loop area, and the magnetic energy-temperature rise coupling coefficient; the eddy current loss-related features include the eddy current loss energy mutation rate, the vibration-eddy current energy time delay, and the eddy current frequency bandwidth; the acoustic energy-related features include the Mel frequency cepstral coefficient MFCC-2 and the acoustic energy pulse interval; and the thermal radiation energy-related features include the temperature rise gradient and the spatial distribution entropy of thermal radiation energy. Principal component analysis is used to retain the first 8 principal features in order to focus on key fault information and improve the sensitivity and computational efficiency of early fault identification.

[0026] Among these methods, gradient aggregation is used instead of raw data transmission to avoid centralized leakage of cross-regional data. At the same time, dynamic thresholds combined with equipment aging parameters enable adaptive adjustment to adapt to the individual differences of different distribution boxes.

[0027] Furthermore, by inferring the source of energy fingerprints through Bayesian networks, the root cause component of the fault is located, and work orders with maintenance priorities are generated, thereby achieving full lifecycle fault tracing. Specifically, the acquisition of multi-physics energy fingerprints overcomes the limitations of single parameters, accurately capturing energy dissipation anomalies in early latent faults such as poor contact and insulation aging. Simultaneously, piezoelectric energy harvesters and electromagnetic induction coils utilize environmental vibrations and magnetic fields to convert energy into electricity, reducing reliance on external power supply. The feature extraction and compression process focuses on key fault information, eliminating redundant data and improving the sensitivity and computational efficiency of early fault identification. The federated learning collaborative early warning mechanism avoids data leakage through gradient aggregation, and dynamic thresholds combined with equipment aging parameters achieve adaptive adjustment. The fault tracing process uses Bayesian network inference to achieve accurate fault location and maintenance guidance, forming a closed loop for full lifecycle management. All steps work closely together to construct a self-powered, high-precision, and low-privacy-risk fault early warning system.

[0028] This invention further proposes that the piezoelectric energy harvester uses a PZT-5A piezoelectric ceramic sheet, which is attached to the busbar joint and circuit breaker contacts. The mechanical energy of vibration is converted into electrical energy through a rectifier bridge, with an output voltage of 3.3V. The electromagnetic induction coil is made of 200 turns of 0.1mm diameter enameled wire and installed in the gap between the insulator and the enclosure, inducing an alternating magnetic field generated by hysteresis / eddy current losses. In practical applications, a piezoelectric energy harvester refers to a device that converts mechanical energy into electrical energy using the piezoelectric effect. It can be implemented using different types of piezoelectric materials, such as PZT series piezoelectric ceramics and PVDF piezoelectric films. Its purpose is to provide a solution that can efficiently capture weak vibration energy. The PZT-5A piezoelectric ceramic sheet is a material with a high electromechanical coupling coefficient. It can be understood as a piezoelectric element that can generate a large charge output when subjected to mechanical stress, specifically used to capture weak vibration signals caused by poor contact inside the distribution box. An electromagnetic induction coil is an induction device designed based on Faraday's law of electromagnetic induction. Its magnetic field induction performance can be optimized by changing parameters such as the number of coil turns and the diameter of the wire, aiming to sensitively capture weak alternating magnetic field changes generated during insulation aging. Specifically, this solution effectively solves key defects in the energy harvesting process by precisely designing the physical characteristics and deployment strategy of the piezoelectric energy harvester and the electromagnetic induction coil. Using PZT-5A piezoelectric ceramic sheets as the core material, its high electromechanical coupling coefficient efficiently converts the weak vibration mechanical energy generated by poor contact into electrical energy, avoiding the insufficient energy conversion efficiency of general-purpose materials. This allows for the capture of significant vibration energy signals in the early stages of a fault. The piezoelectric energy harvester is attached to the busbar joints and circuit breaker contacts; these locations are the main areas where resistance loss vibration occurs. Direct deployment at these locations maximizes the detection of fault source vibration, reduces energy loss in the signal transmission path, and improves the targeting and sensitivity of vibration energy harvesting. The vibration mechanical energy is converted into electrical energy and stably outputs a 3.3V voltage through a rectifier bridge circuit, solving the problem of large AC power fluctuations in piezoelectric elements. This provides a compatible DC power supply for subsequent low-power electronic devices, supporting the long-term operation of self-powered edge computing nodes without external power dependence. The electromagnetic induction coil uses 200 turns of 0.1mm diameter enameled wire, achieving a high number of turns within a limited space, enhancing magnetic field induction capability. It can sensitively capture the weak alternating magnetic field changes generated by hysteresis / eddy current losses during insulation aging, avoiding the weak induction signal defect of low-turn-count coils. The electromagnetic induction coil is installed in the gap between the insulator and the enclosure, a location in the area of ​​most significant magnetic field changes with minimal interference, ensuring accurate sensing of fault-related magnetic field characteristics and improving the signal-to-noise ratio and reliability of magnetic energy acquisition. The coil is specifically designed to sense the alternating magnetic field generated by hysteresis / eddy current losses, directly correlated with the insulation aging fault mechanism, giving the acquired magnetic field energy characteristics clear fault diagnostic significance and providing key data for the construction of multi-physics energy fingerprints.

[0029] In summary, the above technical solutions effectively solve the problems of unstable energy conversion efficiency and insufficient signal reliability caused by the lack of specific implementation details of the sensor, providing continuous and stable energy support for the self-powered system, while ensuring the reliable construction of multi-physics energy fingerprints, and significantly improving the accuracy of fault feature capture and early warning.

[0030] In practical applications, a micro-thermal radiation sensor is a device capable of sensing thermal radiation signals from an object's surface and converting them into measurable electrical signals. It can be implemented using an uncooled infrared focal plane array fabricated using MEMS technology. An uncooled infrared focal plane array can be understood as an infrared sensor that operates without an additional cooling system, achieving high-sensitivity infrared signal acquisition through principles such as thermopile, pyroelectricity, or microbolometers. The sampling frequency refers to the number of times the sensor samples the target signal per unit time. Its setting needs to comprehensively consider the dynamic characteristics of fault temperature rise and system power consumption limitations; typically, a fixed-frequency sampling can be achieved using a timer interrupt trigger. The SPI interface is a synchronous serial communication protocol characterized by high data transmission rates and low pin occupancy, suitable for data interaction needs in low-power scenarios. The selection of the thermal radiation energy band is mainly based on atmospheric window theory, and a bandpass filter combined with an infrared detector can be used to achieve energy acquisition in a specific band.

[0031] Specifically, this solution utilizes the structural characteristics of the distribution box by deploying an uncooled infrared focal plane detector on its surface, directly capturing external thermal radiation signals and effectively avoiding the influence of internal airflow and electromagnetic interference. A 1Hz sampling frequency ensures timely response to rapidly changing temperature rise events, guaranteeing no early fault transient characteristics are missed. The use of an SPI interface for communication with the MCU reduces system power consumption while maintaining data transmission efficiency, supporting continuous monitoring requirements. The 8-14μm band is selected for thermal radiation energy detection, leveraging its good penetration within the atmospheric window to reduce environmental interference with the detection results. For identifying abnormal temperature rise zones, both area and temperature gradient thresholds are set simultaneously, filtering out false signals caused by tiny local hotspots while accurately capturing the rapid energy dissipation characteristics unique to faults, thus achieving precise identification of fault-related temperature rises. The entire solution is closely integrated with the multiphysics energy fingerprinting process, and through refined sensor configuration and decision logic design, significantly improves the timeliness and accuracy of fault warnings.

[0032] This application further proposes an integrated supercapacitor charging and discharging management circuit for a self-powered edge computing node. When there is sufficient external energy input, the capacitor is charged, and when there is no input, the capacitor is discharged to maintain the operation of the MCU.

[0033] The supercapacitor charge / discharge management circuit refers to a circuit structure capable of efficiently charging and stably discharging supercapacitors. It can be implemented using a DC-DC converter-based charge / discharge management circuit, aiming to adapt to the intermittent characteristics of multi-physics energy harvesting and ensure continuous system operation. A supercapacitor is an energy storage element with high energy density and long cycle life. It can be implemented using double-layer capacitors or pseudocapacitors, with the purpose of storing externally harvested energy and releasing it when needed to maintain system operation.

[0034] Specifically, this solution establishes a self-powered mechanism for edge computing nodes by integrating a supercapacitor charge / discharge management circuit. During distribution box monitoring, when external energy input is sufficient—for example, energy fluctuations such as vibrations and magnetic fields generated during poor contact or insulation aging are effectively collected—the supercapacitor charge / discharge management circuit converts this energy into electrical energy and stores it in the supercapacitor. This design fully utilizes the characteristics of multi-physics energy acquisition, avoiding the capacity limitations of traditional battery power. When the distribution box is in a stable state or during periods of low energy, i.e., when there is no external energy input, the supercapacitor begins to discharge, relying on its slow-release capability to provide power to the microcontroller unit, thereby ensuring that the microcontroller unit can continuously execute local fault probability calculation and early warning tasks. The entire solution, through closed-loop management of energy storage and release, ensures that edge computing nodes can still operate stably under conditions of discontinuous energy input, fundamentally overcoming the constraint of power supply bottlenecks on system reliability. Furthermore, this self-powered mechanism, combined with the multi-physics energy acquisition in the aforementioned distribution box fault early warning method, enables the system to achieve maintenance-free long-term monitoring in remote areas, solving the problem of sensor power supply relying on external power sources or batteries.

[0035] This application further proposes a specific implementation of the CNN-LSTM model in the above scheme: the CNN layer uses 3×3 convolutional kernels, 32 kernels in total, with a stride of 1 and same padding, to extract spatial features; the LSTM layer has an input dimension of 32, 64 hidden nodes, and an output dimension of 1, trained using the Adam optimizer with a loss function of binary cross-entropy, and 100 local training cycles. Specifically, the 3×3 convolutional kernel refers to a filter that performs a sliding window operation on two-dimensional data, with a width and height of 3 pixels or time step units, which can effectively capture local spatial correlations. In practical applications, the convolutional layer interface provided by deep learning frameworks such as TensorFlow or PyTorch can be used. Setting the number of convolutional kernels to 32 ensures feature extraction capability while strictly controlling the model size, adapting to the limited memory resources of edge nodes. The coordinated setting of a stride of 1 and same padding maintains the spatial resolution consistency of the feature maps, eliminates the additional upsampling operations required due to size changes, simplifies the data processing flow, and improves inference speed.

[0036] The LSTM layer's input dimension of 32 matches the CNN layer's output dimension, enabling seamless transfer of temporal features. The hidden layer's node count is limited to 64, providing sufficient temporal dependency modeling capability to capture fault evolution patterns, while the moderate node size prevents model bloat. The output dimension is set to 1, directly generating fault probability values ​​and simplifying the output layer structure. The Adam optimizer is an adaptive learning rate optimization algorithm that adjusts the learning rate of each parameter by calculating the first and second moment estimates of the gradient, accelerating the model convergence process. The binary cross-entropy loss function closely aligns with the binary classification nature of fault warning, optimizing the accuracy of the probability output.

[0037] In detail, this solution effectively addresses the core issue of balancing model efficiency and energy consumption in resource-constrained edge device scenarios by precisely configuring the structure and training parameters of the CNN-LSTM model. Small-sized convolutional kernels significantly reduce the number of parameters and computational complexity, enabling efficient execution of spatial feature extraction on low-power MCUs. Setting the number of convolutional kernels to 32 ensures feature extraction capabilities while strictly controlling the model size, preventing increased storage pressure due to feature channel redundancy. A stride of 1 and the use of "same" padding maintain the spatial resolution consistency of the feature maps, simplifying data processing and improving inference speed. The matching design of the LSTM layer input dimension of 32 and the CNN layer output dimension achieves seamless transfer of temporal features. Limiting the number of hidden layer nodes to 64 provides sufficient temporal dependency modeling capabilities to capture fault evolution patterns, ensuring stable operation under supercapacitor power constraints. Setting the output dimension to 1 directly generates fault probability values, allowing the model to accurately focus on binary classification early warning requirements. The adaptive learning rate mechanism of the Adam optimizer accelerates the model convergence process, reducing the number of iterations required to achieve stable performance. The loss function chosen is binary cross-entropy, which closely aligns with the binary classification nature of fault warning. The local training cycle is fixed at 100 rounds, ensuring that the model fully learns the feature patterns while suppressing the risk of overfitting and extending the continuous monitoring time of the self-powered system.

[0038] Through the above technical solutions, the model can achieve efficient fault probability prediction under the condition of limited edge device resources, meet the low power consumption operation requirements of self-powered edge nodes, and at the same time ensure prediction accuracy and response speed.

[0039] This application further proposes that in the federated learning framework, differential privacy technology is used when edge nodes upload gradients, and Gaussian noise is added to prevent the leakage of original data; after the gradients are aggregated in the regional cloud, the knowledge of the regional model is transferred to the global model through model distillation, thereby improving the generalization ability in small sample scenarios.

[0040] In practical applications, differential privacy technology refers to the technique of introducing random noise during data processing to protect individual data privacy. This can be achieved through Laplace or Gaussian mechanisms, aiming to ensure that the original information in the gradient data cannot be reverse-derived. Gaussian noise is a random perturbation that follows a normal distribution and can be generated using predefined noise scale and distribution parameters, aiming to balance privacy protection with the effectiveness of model training. Model distillation can be understood as a technique that compresses the knowledge of a complex model into a simplified model. This can be achieved through soft labeling guidance, temperature regulation, etc., aiming to improve the generalization performance of the model in small sample scenarios.

[0041] Specifically, in the federated learning framework, data collected by edge nodes is used to generate gradient information after local model training. Before being uploaded, this gradient information is enhanced with Gaussian noise using differential privacy techniques, effectively blocking potential data leakage paths. The regional cloud receives the noisy gradients uploaded from each edge node, aggregates them, and uses model distillation to transfer knowledge from the regional models to the global model. This process guides the global model's learning through soft-label information output by the regional models, avoiding the pitfalls of small-sample training based solely on raw data, enabling the global model to extract more universal feature representations. Furthermore, this scheme combines the advantages of differential privacy techniques and model distillation, not only addressing the privacy risks associated with data dispersion across regional distribution boxes but also significantly enhancing the system's adaptability and predictive robustness in data-scarce scenarios, thus specifically overcoming the shortcomings of traditional methods in privacy protection and generalization capabilities.

[0042] The present invention further provides that, in the dynamic threshold adaptive mechanism, the cumulative energy loss This is the sum of the absolute values ​​of the key energy features retained after PCA dimensionality reduction; Fitted through accelerated life testing Here, represents the dielectric constant of the XLPE insulating material. Specifically, the cumulative energy loss Q refers to the quantitative index of energy loss calculated based on the compressed and screened 8-dimensional principal features from the multiphysics energy fingerprint. This can be achieved by weighted summation or direct summation of the principal component features after PCA dimensionality reduction, aiming to avoid redundant interference from the original 15-dimensional features and ensure that the quantification of cumulative energy loss is closely related to the actual fault energy flow. The fatigue limit of the contact material. This refers to a parameter reflecting the mechanical wear process of contact materials, obtained through fitting experimental data. It can be achieved using stress-strain data analysis methods from accelerated life testing or fatigue crack propagation rate testing methods. The aim is to overcome theoretical estimation biases and ensure that the calculated aging coefficient α accurately reflects the aging state of the equipment during long-term operation. Dielectric constant of the insulating layer. This refers to the fixed parameter basis for XLPE, a commonly used material for insulation layers in distribution boxes. It can be achieved through material database queries or laboratory dielectric performance testing methods. The purpose is to ensure that the threshold calculation accurately captures the impact of insulation aging on the magnetic-thermal coupling effect.

[0043] In detail, the above technical solution solves the threshold inaccuracy problem caused by fuzzy parameter definitions by precisely defining the core parameters of the dynamic threshold adaptive mechanism. The cumulative energy loss Q is calculated based on the 8-dimensional principal features retained after PCA dimensionality reduction. These principal features originate from multi-physical field energy fingerprints such as resistive loss, hysteresis loss, and eddy current loss, and can comprehensively reflect energy anomalies during equipment operation. (Contact material fatigue limit) By fitting accelerated life testing data with the material performance degradation patterns under actual operating conditions, the calculated aging factor α more closely reflects the true degree of aging. (Insulating layer dielectric constant) The fixed parameters specified for XLPE material are matched with the actual material properties of the distribution box insulation layer, thereby improving the targeting and accuracy of threshold calculation. Through the precise definition of the above parameters, the dynamic threshold adaptive mechanism can deeply integrate the physical characteristics of the equipment with historical aging data, significantly improving the personalized accuracy of the warning threshold, and effectively reducing the probability of false alarms or missed alarms, especially in long-term operation.

[0044] The graded early warning system in this invention includes: like This triggers an emergency alert: the edge device calls the camera to detect personnel, and if no one is present, it immediately disconnects the shunt trip unit and sends an "Immediate Repair" work order. like This triggers a medium-level warning: the cooling fan or dehumidifier (10W power) is activated, and a "check within 48 hours" work order is pushed out. like This triggers a low-risk warning, calculates the remaining lifetime L=L0·exp(−β·cumulative energy loss) based on the Weibull distribution model, and generates a remaining lifetime prediction report, where L0=36 months and β=0.01.

[0045] Specifically, tiered early warning refers to a technical approach that divides early warning responses into multiple levels based on the difference between the local fault probability and a dynamic threshold. This can be achieved by setting different threshold ranges, aiming to dynamically match early warning actions with the severity of the fault, thereby improving the scientific rigor and adaptability of operational decisions. In practical applications, emergency early warnings can be implemented by judging... Does it exceed and Triggered by the sum of values, this design can capture significant risks exceeding the threshold, ensuring timely and necessary measures are taken in high-risk scenarios. Intermediate warnings can be based on... exist and Triggering within a range of 0.1 aims to delay fault development and ensure power supply continuity. Low-risk early warning can be based on... Below The conditions are triggered, and the remaining lifetime is calculated using the Weibull distribution model combined with parameters to guide proactive maintenance.

[0046] In detail, the above technical solution achieves refined classification of early warning response by quantifying the difference between the local fault probability and the dynamic threshold. For emergency early warning scenarios, when... Exceed When the sum is 0.1, the edge device uses a camera to detect people, ensuring that the power-off operation is only performed when no one is present, thus avoiding safety hazards caused by accidental power outages. For intermediate warnings, when... In and When the power supply is increased by 0.1, the cooling fan or dehumidifier with a power of 10W is activated to slow down the development of the fault through environmental control. Simultaneously, a work order for inspection within 48 hours is pushed out, achieving appropriate mitigation measures while ensuring power supply continuity. In low-risk warnings, when... Below At that time, based on the Weibull distribution model combined with L0 and β parameters, the remaining lifetime is calculated, and a prediction report is generated to guide operation and maintenance planning, transforming passive response into proactive maintenance. Overall, this mechanism solves the problem of inaccurate response caused by single threshold warnings through precise risk gradient division, and adapts to the comprehensive needs of safety, continuity and resource optimization in actual operation and maintenance.

[0047] Furthermore, the aforementioned graded early warning mechanism is combined with a dynamic threshold adaptive mechanism, by introducing the aging coefficient α and the fatigue limit of the contact material. Dielectric constant of insulating layer Including key parameters such as cumulative energy loss Q, this design further improves the accuracy and adaptability of early warning. This design not only considers individual differences in equipment aging but also achieves accurate assessment of fault risks through quantitative analysis, providing strong support for the full lifecycle management of distribution boxes.

[0048] In another embodiment, this application also discloses a distribution box fault early warning system for implementing the aforementioned method, comprising: Multiphysics sensing module: piezoelectric energy harvester, electromagnetic induction coil, micro-thermal radiation sensor and matching energy harvesting circuit; Self-powered edge computing module: low-power MCU, MPPT energy management chip, supercapacitor and local inference model; Federated learning collaboration module: communication interfaces and model aggregation algorithms for edge computing, regional cloud, and global cloud; Fault tracing module: Bayesian network inference engine and maintenance work order generation unit; Active protection module: camera, cooling fan, dehumidifier and circuit breaker shunt trip unit.

[0049] The core innovation of this application lies in combining a multi-physics sensing module with a self-powered edge computing module, while introducing a federated learning collaboration module and a fault tracing module. This systematically solves the problems of failing to capture early hidden faults due to the lack of energy characteristics, difficulties in powering sensors, and cross-regional data privacy conflicts in power distribution box fault early warning, achieving the effects of accurate early warning, long-term maintenance-free monitoring, and distributed privacy protection.

[0050] Specifically, this system integrates five modules working collaboratively to provide a comprehensive solution to the problems of missing energy characteristics, power supply dependence, and data privacy barriers in power distribution box fault early warning. The multi-physics sensing module collects multi-dimensional physical field data, including resistance loss vibration energy, hysteresis / eddy current loss alternating magnetic field, and thermal radiation energy, through a piezoelectric energy harvester, electromagnetic induction coil, and micro-thermal radiation sensor. It also utilizes a matching energy harvesting circuit to capture energy, solving the problem of traditional methods failing to capture early, latent faults such as poor contact and insulation aging due to their single data dimension, thus providing a comprehensive feature basis for fault early warning. The self-powered edge computing module integrates a low-power MCU, an MPPT energy management chip, and a supercapacitor. It uses the energy collected by the sensing module for charge and discharge management, storing electrical energy when external energy is sufficient and maintaining system operation when there is no input. Combined with a local inference model, it achieves on-site data processing, effectively overcoming the problem of sensors in remote areas relying on external power supply and ensuring long-term maintenance-free monitoring. The federated learning collaboration module, through communication interfaces and model aggregation algorithms between the edge nodes, regional clouds, and the global cloud, enables edge nodes to upload only model gradients rather than raw data. After gradient aggregation, the regional cloud collaborates with the global cloud to generate differentiated models, avoiding the privacy risks associated with centralized cross-regional data transmission and adapting to the aging differences of various distribution boxes. The fault tracing module uses a Bayesian network inference engine to analyze energy fingerprint sources, accurately locating the root cause component of the fault. It outputs priority-based maintenance instructions through a maintenance work order generation unit, overcoming the shortcomings of traditional methods in tracing the root cause of faults. The active protection module integrates cameras, cooling fans, dehumidifiers, and circuit breaker shunt trip units, automatically triggering corresponding protective actions based on the warning level. For example, it disconnects the circuit or activates environmental control equipment when no one is present, forming a closed-loop protection chain from monitoring to response. The modules work closely together: the multiphysics sensing module provides raw energy data input, the self-powered edge computing module ensures continuous system operation and preprocesses data, the federated learning collaboration module optimizes model performance while protecting privacy, the fault tracing module deeply analyzes the causes of faults, and the proactive protection module executes intervention measures in real time, together achieving accurate early warning of faults, distributed processing with privacy and security, and self-sustaining proactive protection.

[0051] Furthermore, in practical applications, multiphysics energy fingerprinting can be understood as a set of energy features extracted from multiple physical fields, with the primary purpose of capturing energy dissipation anomalies in early latent faults. For example, similar functionality can be achieved by deploying other types of energy harvesting devices, such as photoelectric converters or thermoelectric generators. For electromagnetic induction coils, different magnetic field strengths and frequency ranges can be adapted by changing the winding method or number of turns, for example, by using flat coils or multi-layer winding structures. Micro-thermal radiation sensors can be implemented by selecting detectors in different wavelength bands, such as short-wave infrared detectors or long-wave infrared detectors, to adapt to different thermal radiation characteristics.

[0052] Furthermore, the 15-dimensional energy feature extraction and compression process is primarily aimed at focusing on key fault information and eliminating redundant data. Specifically, feature extraction can be achieved through other signal processing methods, such as wavelet transform or Fourier transform, to analyze the time-frequency characteristics of the signal. Principal component analysis can be replaced by other dimensionality reduction methods, such as linear discriminant analysis or independent component analysis, to achieve feature compression. The core of the federated learning collaborative early warning mechanism lies in avoiding data leakage through gradient aggregation rather than raw data transmission. As a preferred implementation, the lightweight model at the edge can be implemented using other machine learning algorithms, such as support vector machines or random forests. The model aggregation algorithm for the regional cloud can also employ other distributed optimization methods, such as asynchronous stochastic gradient descent or distributed alternating direction multiplier method. In the calculation formula for dynamically adjusting the threshold, the aging coefficient α and the fatigue limit of the contact material are used. Dielectric constant of insulating layer The specific calculation method for cumulative energy loss Q can be adjusted according to the actual application scenario.

[0053] In summary, through the above technical solutions, this application achieves accurate detection of early latent faults in distribution boxes, long-term maintenance-free monitoring, and distributed privacy protection, forming a closed loop of full life cycle management, which significantly improves the accuracy of fault early warning and the reliability of the system.

[0054] In another embodiment, this application also discloses a storage medium storing a computer program, which, when executed by a processor, implements the above-described distribution box fault early warning method.

[0055] As used in the specification and claims, certain terms refer to specific components. Those skilled in the art will 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" 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 margin of error, those skilled in the art can solve the technical problem and substantially achieve the technical effect within a certain margin of error.

[0056] 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 early warning of faults in a distribution box, characterized in that, Includes the following steps: S1. Multi-physics field energy fingerprint acquisition: A piezoelectric energy harvester, an electromagnetic induction coil, and a micro-thermal radiation sensor are deployed in the distribution box to collect the resistance loss vibration energy when there is poor contact, the hysteresis / eddy current loss alternating magnetic field when the insulation is aging, and the thermal radiation energy in the abnormal temperature rise zone, respectively, to construct a multi-physics field energy fingerprint that includes resistance loss, hysteresis loss, eddy current loss, sound wave energy, and thermal radiation energy. S2. 15-Dimensional Energy Feature Extraction and Compression: Time-frequency domain analysis of the energy fingerprint is performed to extract and compress the following 15-dimensional energy features. Resistive characteristics of resistance loss: proportion of high-frequency component energy, variance of resistance loss energy fluctuation, attenuation rate of fundamental frequency component energy, and correlation between resistance and vibration energy. Hysteresis loss related characteristics: hysteresis loss spectrum kurtosis, high-frequency magnetic energy ratio, hysteresis loop area change rate, magnetic energy-temperature rise coupling coefficient; Eddy current loss characteristics: eddy current loss energy abrupt change rate, vibration-eddy current energy delay, eddy current frequency bandwidth; Sound wave energy related characteristics: Mel frequency cepstral coefficients MFCC-2, sound wave energy pulse interval; Characteristics related to thermal radiation energy: temperature rise gradient, spatial distribution entropy of thermal radiation energy; Principal component analysis was used to retain the first 8 principal features. S3, Federated Learning Collaboration Warning: At the edge, a lightweight CNN-LSTM model is trained based on compressed features to output the local fault probability. ; The regional cloud aggregates the model gradients of edge nodes using the FedAvg algorithm to update the regional model parameters. Global cloud platform generates differentiated early warning thresholds based on distribution box type. ; Calculate the dynamic adjustment threshold at the edge ,in ;in α =0.05 is the aging coefficient. The fatigue limit of the contact material. The dielectric constant of the insulating layer, The accumulated energy loss is calculated using the retained features; S4. Full lifecycle fault tracing: Locate the root cause of the fault by inferring the energy fingerprint source through Bayesian network and generate a work order with maintenance priority.

2. The method according to claim 1, characterized in that: The piezoelectric energy harvester uses PZT-5A piezoelectric ceramic sheets, which are attached to the busbar joints and circuit breaker contacts. The vibration mechanical energy is converted into electrical energy through a rectifier bridge, with an output voltage of 3.3V. The electromagnetic induction coil is made of 200 turns of 0.1mm diameter enameled wire and is installed in the gap between the insulator and the enclosure to sense the alternating magnetic field generated by hysteresis / eddy current loss.

3. The method according to claim 1, characterized in that: The micro-thermal radiation sensor is an uncooled infrared focal plane detector, deployed on the surface of the enclosure, with a sampling frequency of 1Hz. It communicates with the MCU via an SPI interface, detects thermal radiation energy in the 8-14μm band, and locates abnormal temperature rise areas. The abnormal temperature rise area is determined by: area ≥ 2cm² and temperature gradient > 2℃ / s.

4. The method according to claim 1, characterized in that: The self-powered edge computing node integrates a supercapacitor charging and discharging management circuit. When there is sufficient external energy input, the capacitor is charged, and when there is no input, the capacitor is discharged to maintain the operation of the MCU.

5. The method according to claim 1, characterized in that: In the CNN-LSTM model, the CNN layer uses 3×3 convolutional kernels, with 32 kernels, a stride of 1, and is padded with the same kernel to extract spatial features; the LSTM layer has an input dimension of 32, 64 hidden layer nodes, and an output dimension of 1. It is trained using the Adam optimizer, with a loss function of binary cross-entropy, and a local training cycle of 100 rounds.

6. The method according to claim 1, characterized in that: In the federated learning framework, differential privacy technology is used to add Gaussian noise when the edge nodes upload gradients to prevent the leakage of the original data; after the gradients are aggregated in the regional cloud, the knowledge of the regional model is transferred to the global model through model distillation to improve the generalization ability in small sample scenarios.

7. The method according to claim 1, characterized in that: In the aforementioned dynamic threshold adaptive mechanism, the cumulative energy loss This is the sum of the absolute values ​​of the key energy features retained after PCA dimensionality reduction; Fitted through accelerated life testing, is the dielectric constant of XLPE insulating material.

8. The method according to claim 1, characterized in that: The tiered early warning system includes: like This triggers an emergency alert: the edge device calls the camera to detect personnel, and if no one is present, it immediately disconnects the shunt trip unit and sends an "Immediate Repair" work order. like This triggers a medium-level warning: the cooling fan or dehumidifier (10W power) is activated, and a "check within 48 hours" work order is pushed out; like This triggers a low-risk warning, and the remaining lifetime is calculated based on the Weibull distribution model. L = L 0∙··exp(− β • Cumulative energy loss), generating a remaining lifespan prediction report, in which L 0 = 36 months β =0.

01.

9. A distribution box fault early warning system implementing the method of any one of claims 1-8, characterized in that, include: Multiphysics sensing module: piezoelectric energy harvester, electromagnetic induction coil, micro-thermal radiation sensor and matching energy harvesting circuit; Self-powered edge computing module: low-power MCU, MPPT energy management chip, supercapacitor and local inference model; Federated learning collaboration module: communication interfaces and model aggregation algorithms for edge computing, regional cloud, and global cloud; Fault tracing module: Bayesian network inference engine and maintenance work order generation unit; Active protection module: camera, cooling fan, dehumidifier and circuit breaker shunt trip unit.

10. A storage medium, characterized in that, The device contains a computer program that, when executed by a processor, implements the method described in any one of claims 1-8.

Images