A power distribution system safety state intelligent management method

Abstract

This invention discloses an intelligent management method for the safety status of power distribution systems, relating to the field of smart power operation and maintenance technology. The method includes: acquiring multi-dimensional operating parameters of the power distribution circuit and performing adaptive preprocessing; extracting time-series feature vectors based on load working cycles; analyzing the feature vectors using a hybrid decision model to obtain safety status and risk assessment scores; executing a graded response strategy according to the risk level; and entering the fault monitoring and recovery process after protection. This invention achieves accurate identification, rapid response, and adaptive management of safety hazards in power distribution systems through multi-dimensional perception, intelligent feature extraction, hybrid model decision-making, graded response, and self-learning optimization. This method transforms traditional passive protection into proactive early warning and intelligent protection, effectively reducing the risk of electrical fires and accidents, improving operation and maintenance efficiency and system reliability, and is applicable to various engineering and industrial power distribution scenarios.

Description

Technical Field

[0001] This invention relates to the fields of intelligent power operation and maintenance and industrial Internet of Things technology, specifically to an intelligent management method for the safety status of power distribution systems. Background Technology

[0002] With the deepening of smart city and smart construction site construction, higher requirements are being placed on the intelligent and safe operation and maintenance of power infrastructure. As a critical terminal node in power supply, the safe and stable operation of distribution boxes directly affects the safety of personnel and property and the progress of projects. Traditional distribution boxes mainly rely on mechanical circuit breakers, residual current devices (RCDs), and regular manual inspections. This approach has significant drawbacks: it cannot monitor the operating status in real time, cannot provide early warnings of potential hazards such as overload, leakage, poor contact, and abnormal temperature rise, and relies on manual troubleshooting after a fault occurs, resulting in slow response, low efficiency, and a lack of preventative maintenance capabilities supported by historical data.

[0003] To improve the safety of distribution boxes, existing technologies are mainly developing in two directions. The first is a remote monitoring solution based on the Internet of Things (IoT). This solution involves installing various independent sensors (such as current transformers and temperature sensors) and communication modules inside the distribution box to upload collected data such as current, voltage, and temperature to a cloud platform for centralized monitoring and over-limit alarms. While this solution achieves remote monitoring, it has significant problems: First, the sensing dimension is limited, typically monitoring only basic electrical parameters and lacking effective means to detect deeper hidden dangers such as leakage current, harmonics, and cable joint conditions. Second, the judgment logic is simple, usually only setting fixed thresholds for over-limit alarms, failing to identify complex and gradual fault modes (such as gradual temperature rise caused by poor contact), resulting in high false alarm and false negative rates. Third, the system response is lagging, with a long cycle from data upload and cloud analysis to command issuance, which cannot meet the millisecond-level rapid protection requirements for emergency faults such as short circuits.

[0004] The second approach is a local intelligent diagnostic solution based on edge computing. This solution integrates a computing unit into the terminal device to run diagnostic algorithms and achieve rapid local diagnosis. While this approach shortens response time, it still has shortcomings: First, terminal computing resources are limited, typically only able to run simple threshold comparisons or rule-based judgments, making it difficult to deploy complex intelligent analysis models, resulting in limited diagnostic accuracy and depth. Second, each terminal operates independently, preventing data and knowledge sharing and the ability to leverage collective experience for self-optimization; fault modes learned by one terminal cannot empower other terminals. Third, the coordination between perception, analysis, and execution is insufficient; typically, discrete sensors, processors, and actuators are used, resulting in low system integration, high cost, and weak reliability assurance mechanisms for protective actions.

[0005] Therefore, how to construct an intelligent monitoring and safety control method for engineering distribution boxes that integrates multi-dimensional accurate perception, embedded deep intelligent analysis, reliable and rapid protection execution, and system self-evolution capability has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] To address the aforementioned technical problems, a method for intelligent management of the safety status of a power distribution system is provided, the method comprising: A multi-dimensional set of operating parameters for the target power distribution circuit is acquired. During the acquisition process, a weak surface acoustic wave signal excited by the vibration of the power distribution circuit conductor and associated with a specific harmonic component of its power frequency current is simultaneously acquired, and this signal is used as an independent dimension of the multi-dimensional set of operating parameters. The multi-dimensional operating parameter set is adaptively preprocessed to eliminate noise interference and signal drift, and generate a standardized monitoring data stream. Based on the standardized monitoring data stream, a time-series feature vector characterizing the operational safety status of the target power distribution circuit is extracted. The extraction process of the time-series feature vector is segmented according to the load working cycle of the target power distribution circuit, and transient phase features, steady-state phase features, and end phase features are extracted for each working cycle segment. The time-series feature vector is input into a pre-configured safety status identification model for analysis. The safety status identification model is a hybrid decision model, which includes a rule engine submodule for processing deterministic rules and a lightweight artificial intelligence analysis submodule for processing complex correlation patterns. The analysis outputs the current safety status determination result and the corresponding risk assessment score of the target power distribution circuit. Based on the safety status determination result and the risk assessment score, a graded response strategy matching the risk level is executed. The graded response strategy includes at least one or more of the following: generating a warning log, triggering a local alarm prompt, and executing a circuit disconnection protection command. After the execution circuit cut-off protection command is issued, the fault monitoring and recovery process is initiated. The fault characteristics are periodically checked to see if they have been eliminated. Once the safety conditions are confirmed to be met, a prompt message indicating that power supply can be restored is generated.

[0007] By adopting the above technical solution, a multi-dimensional fusion perception and deep intelligent decision-making effect for potential hazards in power distribution systems is achieved. Its working principle involves introducing weak surface acoustic wave (SAW) signals as an independent perception dimension. When a conductor carries current, it generates microscopic mechanical vibrations due to electromagnetic forces and thermal effects. The vibration spectrum is closely related to current harmonics, contact states, and mechanical stress. While collecting traditional parameters such as electrical, thermal, and switching parameters, the system simultaneously captures this acoustic signal, thus fusing electrical and mechanical state information at the data source. Subsequently, this multi-dimensional raw data undergoes adaptive preprocessing to become a standardized data stream, which is then segmented into physically meaningful operational process segments based on the load cycle. The transient, steady-state, and termination features of each segment are extracted to form a time-series feature vector containing mechanical vibration characteristics. This feature vector is then input into a hybrid decision-making model composed of a rule engine and a lightweight AI model for comprehensive analysis. The rule engine quickly processes explicit rules, while the AI ​​model deeply mines complex correlation patterns, ultimately collaboratively outputting accurate judgment results including risk scores. Based on this result, the system executes a tiered response from early warning and alarm to power outage protection. After the protection action, the system does not passively wait, but actively performs periodic fault characteristic retests to provide a basis for decision-making for possible automatic recovery. The entire process constitutes a complete intelligent closed loop of "multi-dimensional perception - intelligent feature engineering - hybrid model decision-making - hierarchical reliable execution - proactive aftercare recovery".

[0008] Furthermore, the acquisition of the multi-dimensional operating parameter set of the target power distribution circuit includes: A non-contact composite sensing unit synchronously acquires the current signal of the target power distribution circuit, the leakage current characteristic signal derived from the current induction, and the cable connection temperature rise trend signal associated with the current signal. The non-contact composite sensing unit integrates a miniature flaw detection coil capable of actively emitting a specific frequency scanning magnetic field and receiving its modulation feedback. Microscopic defects on the conductor surface are indirectly assessed by analyzing the distortion characteristics of the feedback magnetic field; and / or, The temperature and humidity data inside the distribution box are collected by an independent wireless low-power sensor tag, which periodically broadcasts the collected data. The non-contact composite sensing unit and the sensing data of the wireless low-power sensing tag together constitute the multi-dimensional operating parameter set.

[0009] By adopting the above technical solution, a non-contact, integrated, and synchronous detection of electrical parameters, early physical defects, and temperature rise trends is achieved within a single sensing unit. The working principle lies in the fact that the non-contact composite sensing unit designed in this solution is not a simple stacking of sensors, but rather integrates a miniature flaw detection coil capable of actively emitting a scanning magnetic field. This coil emits a magnetic field of a specific frequency towards the conductor being tested. When the conductor has microscopic defects such as surface cracks or loose connections, its eddy current distribution changes, causing a characteristic distortion in the reflected magnetic field (i.e., modulation feedback). By analyzing this distortion characteristic, the unit can indirectly assess the mechanical integrity of the conductor's surface without contact or damage, achieving early defect detection. Simultaneously, the unit still collects the main current and leakage current characteristics through electromagnetic induction and senses the temperature rise trend through thermoelectric coupling. In this way, a single unit can simultaneously output three strongly correlated types of information: current, potential defect indication, and temperature rise. Meanwhile, a wireless low-power tag is responsible for collecting ambient temperature and humidity background values. These two types of data together constitute a raw dataset that is richer in dimensions and more interconnected in information than traditional methods, providing an unprecedented high-quality data foundation for subsequent in-depth analysis.

[0010] Furthermore, the adaptive preprocessing of the multi-dimensional operating parameter set includes: Adaptive filtering based on dynamic load characteristic baseline is performed on the current signal and the leakage current characteristic signal. The dynamic load characteristic baseline is established by learning the current waveform characteristics corresponding to the load type of the target power distribution circuit in real time. The filtering process is to remove high-frequency random fluctuation components that do not conform to the current dynamic load characteristic baseline. The cable connection temperature rise trend signal and the ambient temperature data are subjected to drift compensation based on a historical relationship model. The historical relationship model is used to describe the long-term relationship between sensor readings and ambient temperature, so as to dynamically correct the zero-point drift and sensitivity drift of the sensor. The drift compensation introduces a pressure compensation factor that is dynamically corrected based on recent ambient pressure fluctuation data. A debouncing check based on a dynamic time window is performed on the switch state sequence. The duration of the dynamic time window is negatively correlated with the current load current of the target power distribution circuit. A change in switch state is confirmed only when the switch state signal remains stable within the dynamic time window.

[0011] By adopting the above technical solution, the environmental adaptability and data fidelity of the preprocessing stage are significantly improved, particularly enhancing the scientific rigor and accuracy of temperature sensor drift compensation. Its working principle involves innovatively introducing a pressure compensation factor based on recent ambient pressure fluctuation data, in addition to classic filtering and anti-shake calibration in the preprocessing stage. Changes in ambient pressure affect the gas state and thermal convection conditions inside the sensor package, thus causing nonlinear and slow interference to the temperature sensor readings—an interference different from traditional temperature drift models. This method incorporates recent pressure fluctuation data as a dynamic correction factor when establishing a historical relationship model between sensor readings and ambient temperature. The system continuously monitors the trend of ambient pressure changes, and when a significant and sustained pressure change is detected, it dynamically adjusts the parameters of the drift compensation algorithm according to a pre-calibrated "pressure-sensor characteristic offset" relationship. This allows the drift compensation for temperature sensors, especially contact temperature sensors for cable connectors, to take into account not only time and temperature, but also another important environmental physical quantity. This enables the more accurate removal of environmental interference in scenarios such as changes in weather systems and altitude, restoring the true temperature rise signal determined solely by the electrical connection status, greatly improving the reliability of subsequent feature extraction and status recognition.

[0012] Further, the extraction of the time-series feature vector characterizing the operational safety status of the target power distribution circuit includes: The standardized monitoring data stream is automatically divided into continuous operation process segments based on the start and end points of the load cycle. For each of the aforementioned operating process segments, the following feature extraction operations are performed: the peak current surge and current rise time of the initial transient phase are extracted as the first feature subset; the effective current value, current fluctuation period, and average power of the steady-state operating phase are extracted as the second feature subset; and the current decay time constant of the final transient phase is extracted as the third feature subset. The rate of temperature change of the operation segment during the steady-state operation phase is calculated and correlated with the average current of the operation segment, which is used as the fourth feature subset. Record the action combination sequence of multiple switches in multiple consecutive operation process segments and the time interval between adjacent actions to form a fifth feature subset; From the weak surface acoustic wave signal, extract the energy entropy value change curve of the frequency component corresponding to the dominant vibration mode within the corresponding operation process segment, and use it as the sixth feature subset; The first feature subset, the second feature subset, the third feature subset, the fourth feature subset, the fifth feature subset, and the sixth feature subset are merged and encoded to form the time-series feature vector.

[0013] By adopting the above technical solution, a first-ever quantification of the mechanical vibration state of a conductor into analyzable features is achieved, deeply integrating it with electrical and thermal characteristics to construct a high-dimensional feature vector that comprehensively characterizes the "electric-thermal-mechanical" coupled operation state of a power distribution circuit. Its working principle involves a feature extraction step that, after segmenting and extracting conventional features such as current, temperature, and switching timing, specifically processes the weak surface acoustic wave (SAW) signal introduced in claim 1. The system performs time-frequency analysis on the SAW signal within each operational segment, identifying the dominant vibration mode and its corresponding frequency components for that time period. The key innovation lies in the fact that it does not simply analyze amplitude, but rather calculates the energy entropy value change curves of these dominant frequency components. Energy entropy values ​​quantify the concentration and dispersion of vibration energy; sudden changes or trends in entropy values ​​often correspond to changes in the mechanical state. For example, the initial loosening of a screw can lead to a redistribution of vibration energy across different frequencies, causing a change in entropy value. This entropy change curve is used as the sixth feature subset. It is then combined and encoded with the transient, steady-state, and decay features of the current, the temperature rise rate feature, and the switching action sequence feature. The resulting time-series feature vector simultaneously contains three types of information: electrical load characteristics, thermal evolution process, and mechanical vibration state. This provides AI models with unprecedented and highly discriminative feature inputs to identify complex hidden dangers such as "poor contact accompanied by mechanical loosening".

[0014] Further, the step of inputting the time-series feature vector into a pre-configured security state identification model for analysis includes: First, the time-series feature vector is input into the rule engine submodule. The rule engine submodule has a set of deterministic judgment rules based on electrical safety specifications and empirical thresholds. If the time-series feature vector triggers any rule in the set of deterministic judgment rules, the corresponding first safety state judgment result is directly output. If the time-series feature vector does not trigger any of the deterministic judgment rules, it is further input into the lightweight artificial intelligence analysis submodule. The lightweight artificial intelligence analysis submodule is a compressed and quantized time-series neural network model used to identify the complex fault modes and risk signs hidden in the time-series feature vector, and output the second safety state judgment result and risk probability value. The lightweight artificial intelligence analysis submodule embeds a dynamic attention regulation mechanism that simulates the short-term plasticity of biological neural synapses, so that when processing continuous time series, it can adaptively enhance the attention weight of features within the time window before and after the mutation. By combining the first safety status determination result, the second safety status determination result, and the risk probability value, a final safety status determination result is generated, and the risk assessment score is calculated based on a predefined risk scoring mapping table.

[0015] By adopting the above technical solution, a significant improvement in the lightweight AI model's ability to capture and diagnose patterns related to key abrupt changes in time series is achieved. The working principle involves embedding a dynamic attention regulation mechanism that simulates the short-term plasticity of biological neural synapses within the lightweight AI analysis submodule. The short-term plasticity of biological neural synapses allows their signal transmission intensity to dynamically adjust according to the frequency and pattern of preceding and following stimuli, thus prioritizing novel or important information flows. In this AI model, this mechanism is implemented as a learnable, dynamic attention weight allocation algorithm. When the model processes the input time-series feature vectors, this mechanism analyzes the change patterns of the features in real time. Once a drastic change in a feature dimension is detected within a short period (such as a sudden increase in current, a surge in temperature, or a dramatic change in vibrational entropy), the mechanism automatically increases the attention weight of all feature vectors within a short time window before and after the change. This means that the model not only focuses on the abrupt change itself but also examines the "precursor" features leading to the change and the "evolution" features after the change with higher "resolution." This dynamic focusing capability enables the model to learn the complete temporal logic of fault occurrence and development more effectively, significantly improving the recognition rate of early symptoms and the depth of fault root cause analysis, thereby achieving complex pattern analysis capabilities on resource-constrained edge devices that are closer to those of large cloud models.

[0016] Furthermore, the method also includes an optimization step for the security status identification model: Receive human feedback information for historical alarm events, including a label indicating that the alarm event was indeed a false alarm; When the manual feedback information is confirmed to be a false alarm, the historical time series feature vector corresponding to the alarm event is marked as a negative sample and added to the model retraining dataset; When the manual feedback information is confirmed to be a new type of fault, the historical time-series feature vector and its annotation information corresponding to the alarm event will be marked as positive samples and added to the model retraining dataset. The lightweight artificial intelligence analysis submodule is incrementally trained periodically using the model retraining dataset to optimize its model parameters. The incremental training process introduces a robust enhancement training step based on adversarial example generation. By injecting specific pattern noise simulating extreme electromagnetic interference in the field into the training samples, the discrimination stability of the model in a strong noise environment is improved.

[0017] By adopting the above technical solution, the robustness and generalization ability of the intelligent analysis model in complex and harsh industrial electromagnetic environments are significantly enhanced, reducing false alarms caused by unknown interference in the field. Its working principle involves introducing a robust enhancement training step based on adversarial example generation into the model optimization process, building upon conventional incremental training based on manual feedback. Specifically, when retraining the model using newly added positive and negative samples, the system simultaneously generates a batch of "adversarial examples." These adversarial examples are not random noise, but rather simulate extreme but reasonable electromagnetic interference patterns that may occur in engineering fields, such as specific spectral surges caused by the start-up and shutdown of large motors, high-frequency harmonic pulse groups generated by welding machines, or common-mode noise caused by intermittent grounding. These simulated interferences are "injected" into the feature vectors of the original training samples with specific intensity and patterns, generating new samples that appear severely contaminated, but whose labels remain unchanged. Subsequently, these "adversarial examples" are used to train the model along with the original samples. This process forces the model to learn to ignore these specific, deceptive interference patterns and focus on uncovering the true fault characteristics hidden beneath the noise. After multiple rounds of such adversarial training, the model's decision boundaries become smoother and more robust. Even when encountering similar strongly interfering scenarios not explicitly included in the training set, it can maintain high judgment accuracy, thereby greatly improving the system's reliability and user trust in real complex scenarios.

[0018] Furthermore, the step of implementing a graded response strategy matching the risk level based on the security status determination result and the risk assessment score includes: A multi-dimensional weighted risk scoring model is established, in which a basic risk score is assigned to each abnormal feature type, and a correlation weight coefficient is set for the correlation between different abnormal feature types. Based on the types and severity of abnormal features identified by the security status determination results, calculate the basic risk score for each abnormal feature; If multiple abnormal features are identified, the joint risk score of the superimposed abnormal features is calculated based on the correlation weight coefficient. The basic risk score or the combined risk score is compared with a preset risk level threshold range. When the total score falls into the first threshold range, it is determined to be a minor anomaly. An early warning log is generated and recorded in the local memory. At the same time, a continuous high-resolution acquisition mode for the weak surface acoustic wave signal is started to capture potential precursor signals of mechanical loosening. When the total score falls into the second threshold range, it is determined to be a moderate anomaly. A local audible and visual alarm device is triggered, and a remote alarm message is generated and pushed to the management platform. When the total score falls into the third threshold range or an emergency fault characteristic combination is identified, it is determined to be a serious danger level. The circuit cut-off protection command is immediately generated and issued.

[0019] By adopting the above technical solution, a more refined and forward-looking risk warning system is achieved, particularly with a unique early detection capability for latent mechanical faults. Its working principle lies in the fact that its graded response strategy does not rigidly rely solely on the final risk score, but organically combines response actions with in-depth diagnostic detection. When the system determines a minor anomaly, while performing routine early warning log recording, it initiates a continuous high-resolution acquisition mode for weak surface acoustic wave (SAW) signals. In the minor anomaly stage, electrical and temperature parameters may only show weak signs, but early loosening of mechanical connections may produce earlier and more specific changes in the vibration spectrum. In conventional monitoring mode, acoustic signals may use a lower sampling rate to save resources. However, once in this mode, the system will allocate more computing and storage resources to continuously acquire and finely analyze the SAW signals at a high sampling rate and high resolution, aiming to capture characteristic frequency drifts or modal excitations that may be overlooked in conventional mode, indicating mechanical loosening. This is equivalent to mobilizing "specialist inspection" methods to conduct in-depth screening of the most likely potential hazards at the initial stage of risk development. This risk-level-based, targeted enhanced monitoring strategy enables the system to detect potential mechanical problems before a fault causes significant electrical or thermal effects, truly achieving a leap from "post-fault protection" to "early warning of potential hazards."

[0020] Furthermore, the triggering of the local alarm notification includes: Control a multi-color light source to display according to a predefined encoding rule, wherein the encoding rule specifies the mapping relationship between different colors, different flashing frequencies and different flashing modes and different safety status judgment results and risk levels, and the flashing mode includes a specific spatial scanning flashing sequence that uses the persistence of vision of the human eye to dynamically draw a simple fault icon pattern in the air. And / or, control a sound-generating device to output prompts of different frequencies and rhythms according to predefined encoding rules, wherein the encoding rules of the prompts are semantically consistent with the encoding rules of the multi-color light source, jointly indicating the current security status.

[0021] By adopting the above technical solution, the information carrying capacity and intuitiveness of local human-computer interaction are greatly enriched and optimized, enabling on-site personnel to quickly and accurately obtain complex status information under noisy, distant, or poor visibility conditions. Its working principle lies in the fact that the light signal prompting part of the local alarm prompt breaks through the traditional simple flashing mode and adopts a specific spatial scanning flashing sequence that dynamically draws simple fault icon patterns in the air using the persistence of vision effect of the human eye. This solution typically requires arranging multiple LED beads in one-dimensional or two-dimensional space, and using a microcontroller to precisely control the on / off sequence of each bead. When the bead array scans at high speed according to the preset program, due to the persistence of vision, the observer will perceive a seemingly stable graphic that is suspended in the air; for example, a lightning bolt icon represents a short circuit, a flame icon represents overheating, and a wrench icon represents mechanical failure. This prompting method transforms the abstract "risk level" encoding into an intuitive "fault type" graphic, resulting in extremely high information transmission efficiency. Even from a distance, or in noisy environments where the alarm sound is difficult to hear, on-site personnel can instantly grasp the general nature of the malfunction with just a glance at the light source, enabling them to respond more prepared and efficiently. This innovative interaction method significantly lowers the barrier to information interpretation and improves emergency response efficiency.

[0022] Furthermore, the execution circuit disconnection protection command includes: A first control command is generated and sent to the smart circuit breaker through a first communication link. The first control command is used to drive the operating mechanism inside the smart circuit breaker to perform a tripping operation. At the same time or after the first control command is generated, a second control command is generated to turn on a high-speed solid-state switch connected in series in the shunt trip coil circuit of the smart circuit breaker. The second control command is used to generate magnetic force through the shunt trip coil to force the smart circuit breaker to perform a tripping operation. The main protection path corresponding to the first control command and the backup protection path corresponding to the second control command together constitute a dual-channel power failure protection execution mechanism. Furthermore, before the second control command is generated, a pre-excitation current pulse lower than its action threshold is injected into the shunt trip coil circuit to offset the action delay caused by the coil inductance.

[0023] By adopting the above technical solution, the operating speed of the backup protection channel is optimized to the extreme, further shortening the limit response time of power outage protection and gaining crucial milliseconds of time to suppress the development of the most dangerous short-circuit fault arc. Its working principle is as follows: When executing dual-channel power outage protection, before issuing the formal second control command through the backup path (i.e., driving the shunt trip coil), the system first injects a pre-excitation current pulse below its operating threshold into the shunt trip coil circuit. The shunt trip coil is a large inductive load. When a rated voltage is suddenly applied, it takes a certain amount of time for the coil current to rise from zero to the threshold sufficient to drive the trip mechanism, because inductance impedes current transients. Although this pre-excitation pulse is insufficient to actuate the mechanism, it can pre-establish a magnetic field in the coil that is close to but below the operating threshold. When the subsequent full-voltage second control command arrives, the coil current rises from the existing preset level, and the time required to reach the operating threshold is greatly shortened. This is equivalent to eliminating most of the inherent electrical delay caused by the coil inductance. When dealing with faults requiring extremely high-speed disconnection, such as metallic short circuits, these advance milliseconds to tens of milliseconds may mean whether the arc energy can be effectively limited and whether the equipment damage can be significantly reduced, thus raising the limit performance of the protection system to a new level at the physical level.

[0024] Further, whether the periodic detection fault characteristics have been eliminated includes: After the circuit cut-off protection command is triggered, a periodic timer is started. Every first preset time interval, an attempt is made to close a high-impedance detection circuit, and the non-contact composite sensing unit collects the tentative electrical signals in the detection circuit. Analyze the probing electrical signal to determine whether the original fault characteristics still exist. If the fault characteristics disappear and the line insulation parameters return to normal, it is determined that the safety conditions are initially met. At the same time, assess whether the energy spectrum of the weak surface acoustic wave signal during the probing period has returned to a stable background noise level. And / or, After the circuit cut-off protection command is triggered, the ambient temperature change is continuously monitored through the wireless low-power sensor tag. If the abnormal temperature rise trend associated with the fault has returned to normal, and combined with the judgment of the electrical signal, it is determined that the safety conditions are initially met.

[0025] By adopting the above technical solution, a multi-physical quantity cross-verification is achieved in post-fault recovery detection, particularly by utilizing the high sensitivity of vibration signals to eliminate the risk of misjudging fault elimination due to temporarily "normal" electrical parameters. Its working principle is that the periodic fault characteristic elimination detection not only relies on the recovery of trial electrical signals (such as insulation resistance and leakage current), but also innovatively uses whether the energy spectrum of the weak surface acoustic wave signal recovers to a stable background noise level during the trial as a key criterion. For some faults, especially poor mechanical contact or partial discharge, after power is cut off, their electrical characteristics (such as contact resistance) may temporarily appear normal due to thermal expansion and contraction, arc extinction, etc., but their mechanical vibration state (such as continuous micro-vibration caused by loosening, or partial discharge vibration at internal cracks in the insulation material) may still persist. Once power is restored, the hidden danger will quickly recur. During each trial power-on detection, the system simultaneously acquires surface acoustic wave signals with high precision and analyzes their energy spectrum. If the energy spectrum still shows characteristic frequency peaks or abnormal broadband energy related to the fault, the system will determine that the safety conditions are not met even if the electrical parameters are "compliant". This "electro-acoustic" dual-criteria mechanism greatly improves the reliability and safety of fault elimination determination and avoids repeated faults or the expansion of accidents due to the limitations of a single criterion.

[0026] Furthermore, after generating the prompt message allowing power restoration, the method further includes: The system will upload a message containing "Fault has been eliminated, request remote confirmation for reset" and a snapshot of relevant safety parameters to the remote management platform via a wireless communication network. Receive a remote reset confirmation command issued by the remote management platform; After verifying the legality of the remote reset confirmation command, a circuit breaker closing command is generated to restore power supply to the target power distribution circuit. At the same time as generating the closing command, a "soft wake-up" voltage pulse sequence of a specific frequency and lower than the device start-up threshold is pre-sent to the load side of the power distribution circuit. And / or, after generating a prompt message allowing power restoration, if no remote command is received within a second preset time interval, it automatically enters a waiting state and issues a prompt through a local audio-visual device to wait for manual on-site handling.

[0027] By adopting the above technical solution, a smoother and more flexible power restoration process is achieved, effectively suppressing the secondary inrush current that may be generated when a group of loads restarts simultaneously, and improving the success rate of power restoration and the overall stability of the system. Its working principle involves the system simultaneously sending a specific frequency "soft wake-up" voltage pulse sequence, below the equipment start-up threshold, to the load side of the distribution circuit at the instant of remote confirmation of reset and generation of the circuit breaker closing command. This pulse sequence is specially designed so that its voltage amplitude is insufficient to start inductive loads such as motors and compressors, but sufficient to charge the internal control circuits of loads such as control circuits and switching power supplies. Its frequency may differ from the power frequency, aiming to orderly "wake up" the soft starters, frequency converters, or intelligent controllers on the load side. When the main circuit power frequency voltage is officially restored, the control systems of most loads are ready and can start in an orderly manner according to their internal logic, rather than the inrush currents of all loads being superimposed at the same power frequency phase point. This avoids the huge inrush current that may occur at the moment of power restoration, which could trigger overcurrent protection again, leading to restoration failure. This "soft wake-up" mechanism is equivalent to giving the load side an orderly "pre-notification" and "pre-preparation" before the formal restoration of power supply. This transforms the restart process of large-scale loads from a disorderly shock into a relatively orderly and smooth process, thus protecting both the power grid and the distribution box itself.

[0028] Furthermore, the method also includes system self-learning and parameter adaptation steps: During the initial operation phase of the system, a self-learning period is set. During the self-learning period, the system only performs monitoring and recording, and does not execute the circuit disconnection protection command. Based on the historical data collected during the self-learning period, a personalized operating parameter baseline for the target power distribution circuit is established. The personalized operating parameter baseline includes the normal current range, temperature range, and load mode characteristics. Based on long-term operating data and alarm history, the health coefficient of the target power distribution circuit is dynamically calculated. The health coefficient is used to characterize the overall aging and performance change trend of the circuit. The calculation of the health coefficient incorporates the "mechanical fatigue" evaluation sub-coefficient based on the slow change of the spectrum structure of the weak surface acoustic wave signal. The judgment threshold used in the security status identification model, the weight coefficient in the risk assessment score calculation, and the health coefficient are associated so that the judgment threshold and the weight coefficient can be dynamically adjusted as the health coefficient changes.

[0029] By adopting the above technical solution, a first-ever quantitative assessment of the long-term aging and fatigue state of mechanical connections in power distribution circuits is achieved, integrating this assessment into overall health management and making predictive maintenance more comprehensive. Its working principle involves innovatively incorporating a "mechanical fatigue" assessment sub-coefficient based on the slow changes in the spectral structure of weak surface acoustic wave (SAW) signals when calculating the overall health coefficient of the circuit, in addition to electrical and thermal historical data. Under long-term electrothermal stress cycling, conductors, terminals, and fasteners undergo slow changes in their microstructure, such as metal fatigue, creep, and stress relaxation. These changes gradually alter their vibration characteristics, manifesting as a slow drift of the dominant frequency in the SAW signal spectrum, changes in the proportion of harmonic components, or the emergence of new frequency components. By monitoring and analyzing the spectral structure of acoustic signals over a long period (e.g., months or years), the system can quantify this slow trend and map it to a "mechanical fatigue" sub-coefficient. This sub-coefficient, along with other sub-coefficients calculated based on electrical parameter aging models and insulation performance degradation models, are weighted to form the final health coefficient. This expands health assessment beyond electrical insulation and contact resistance to include the integrity of the physical structure, enabling early warning of potential problems that are difficult to detect through purely electrical monitoring, such as "decreased busbar clamping force." It provides direct quantitative evidence for scheduling mechanical maintenance such as fastener inspections, achieving true full-state predictive health management.

[0030] Furthermore, the method also includes a collaborative working step with a management platform: The multi-dimensional operating parameter set, the time-series feature vector, the security status determination result, the risk assessment score, and the execution record of the graded response strategy are encapsulated into a data packet and uploaded to the management platform. The management platform maintains a digital twin for each target power distribution circuit. The digital twin synchronously updates the real-time status and historical data of the target power distribution circuit. Based on real-time data and physical models, the digital twin can deduce and visualize the virtual heat map and electromagnetic field distribution map of the invisible area inside the power distribution box in real time. It receives control policy update, model parameter update, or feature library update instructions issued by the management platform and updates its local configuration accordingly.

[0031] And / or, when local computing resources are insufficient or unknown feature patterns are encountered, the time series feature vector is uploaded to the management platform, requesting the management platform to perform cloud-based collaborative analysis, and the analysis results returned by the management platform are received to assist local decision-making.

[0032] By adopting the above technical solution, a panoramic, immersive monitoring experience that transcends physical limitations is provided to remote maintenance personnel, greatly improving the intuitiveness and accuracy of remote diagnosis. Its working principle lies in the fact that the digital twin, working in conjunction with the management platform, is not a simple data mirror, but a high-fidelity model with real-time physical simulation. Based on a multi-dimensional set of operating parameters uploaded from the field in real time, combined with a multi-physics coupling calculation model of electromagnetic fields and heat conduction, it can simulate and visualize virtual thermal maps and electromagnetic field distribution maps of invisible areas inside the distribution box. For example, the twin can display microscopic hotspots at busbar connections, temperature rise cloud maps in densely cabled areas, and magnetic field strength distribution near switch contacts. Maintenance personnel on the management platform can not only see numbers and alarms, but also intuitively "see" how the temperature is distributed inside the box, where the magnetic field is concentrated, and how hotspots are formed. When an anomaly occurs, this visualization helps maintenance personnel quickly understand the physical nature and scope of the fault, such as determining whether overheating is due to poor contact or heat dissipation obstruction, or whether magnetic field anomalies are due to short circuits or harmonics. This is equivalent to giving maintenance personnel "X-ray vision" and "thermal imaging vision," enabling remote diagnosis to no longer rely on experience and guesswork, but rather on the intuitive presentation of physical phenomena, thus allowing for faster and more accurate decision-making, and even guiding on-site personnel to perform precise repairs.

[0033] Furthermore, the method executes the following during configuration initialization: When the device is powered on for the first time, it enters the distribution network discovery mode and broadcasts the device identification information; Establish near-field wireless communication connection with nearby configuration terminals; Through the near-field wireless communication connection, the system receives power distribution circuit parameter configuration information sent by the configuration terminal. The parameter configuration information includes rated current, rated voltage, and circuit identifier. After receiving the basic parameter configuration, the system enters the load type self-identification stage. By analyzing the characteristics of the current waveform collected subsequently, it matches them with the pre-stored typical load feature library to automatically infer the main load type connected to the target power distribution circuit. The main load type is then used to optimize the feature extraction and safety status identification strategy. In the load type self-identification stage, the system also analyzes the voltage sag waveform characteristics on the grid side caused by the instantaneous connection of the load during the initial power-on stage to help identify the type of load start-up impact characteristics.

[0034] By adopting the above technical solution, a deep understanding of load characteristics is achieved during the system initialization phase, providing key prior knowledge for subsequent customized monitoring strategies and enhancing the intelligence level of self-configuration. Its working principle involves analyzing not only the steady-state current waveform of the device itself during the load type self-identification phase, but also innovatively utilizing the voltage sag waveform characteristics caused by the instantaneous connection of the load during the initial power-on phase. When a large load (such as a high-power motor or transformer) is connected to the grid, it draws a huge starting current from the grid, causing a momentary voltage drop near the connection point, i.e., a voltage sag. The waveform of this sag (sag depth, duration, recovery shape) is strongly correlated with the load's starting characteristics (such as direct start, star-delta start, soft start). During the configuration phase, this method deliberately monitors or obtains the grid-side voltage waveform at the instant the distribution box is connected from the upper-level system. By analyzing the characteristics of this sag waveform and comparing it with a database of typical load starting modes, it can help identify the type of load starting impact and its approximate power level. This information, combined with the steady-state characteristics subsequently identified from the device's current waveform, enables the system to make a more comprehensive and accurate judgment of the load type. For example, it can distinguish between "a high-power motor" and "a combination of multiple low-power motors" because their starting impact waveforms are drastically different. This in-depth identification provides crucial initial parameters for subsequently setting personalized overload protection curves, starting impact discrimination thresholds, and vibration and noise background baselines, allowing the system to "better understand" the load it is protecting.

[0035] Furthermore, the method employs a multi-mode heterogeneous fusion communication strategy at the communication transport layer, specifically including: During routine monitoring and low-risk data reporting, low-power wide-area network communication modules are used for intermittent data transmission to reduce power consumption. When the risk level is detected to reach or exceed the medium abnormality level, the high-speed mobile communication module is automatically activated to establish a high-bandwidth connection for uploading detailed fault characteristic data, waveform segments and receiving high-priority control commands. When the high-priority data is packaged, a unique "electromagnetic fingerprint" watermark generated based on the current electrical signal segment is embedded for data integrity verification and non-repudiation verification in the cloud. And / or, in scenarios where there is no public network communication signal, automatically switch to power line carrier communication mode to conduct network communication with other devices or gateways within the same local area network, and forward data through the gateway.

[0036] By adopting the above technical solution, the integrity and immutability of critical alarm data are ensured at the data communication level, providing a reliable data foundation for accident tracing, responsibility determination, and high-level safety applications. Its working principle involves embedding a unique "electromagnetic fingerprint" watermark, generated based on the current electrical signal segment, during the remote transmission of high-priority data (such as fault characteristics and protection action records). This "electromagnetic fingerprint" is not a simple timestamp or CRC checksum, but a feature code extracted from the original current or voltage waveform segments within a certain period before and after the alarm is triggered, using a specific algorithm to characterize the uniqueness of the event. Since each fault's electrical transient process is microscopically unique (affected by multiple factors such as fault point impedance, phase, and power grid conditions), the generated fingerprint also possesses strong uniqueness and unpredictability. This fingerprint is encrypted and embedded as a digital watermark into the uploaded data packet. After receiving the data, the cloud management platform can verify the integrity and validity of the watermark. This achieves three key benefits: first, integrity verification, ensuring data is not accidentally tampered with during transmission; second, non-repudiation verification, as fingerprints originate from unreplicable physical events on-site, proving that the data indeed comes from a specific device's authentic record at a specific moment, and cannot be forged or altered afterward; and third, data association, ensuring that waveform segments, alarm records, and action logs within massive datasets can be precisely linked to the same event through a unified "fingerprint." This provides robust technical support for scenarios involving security liability determination, insurance claims, or high reliability requirements.

[0037] The beneficial effects of this application are as follows: This invention synchronously collects multi-dimensional operating parameters of power distribution circuits and performs adaptive preprocessing to eliminate interference. Data is segmented based on the load cycle, and time-series feature vectors characterizing transient, steady-state, and termination phases are extracted. A hybrid decision model is used to analyze these feature vectors, outputting safety status and risk assessment scores. Graded responses are executed according to risk levels, ranging from early warning and alarm to dual-channel power outage protection. After protection, the system automatically monitors whether the fault has been eliminated and supports safe recovery. Furthermore, the system possesses self-learning capabilities, can establish personalized baselines, dynamically update thresholds, and collaborate with a cloud-based digital twin platform. This method achieves a leap from passive protection to proactive early warning, intelligent protection, and adaptive management, significantly improving the safety, reliability, and intelligent operation and maintenance level of power distribution systems. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating the steps of an intelligent management method for the safety status of a power distribution system according to this application. Detailed Implementation

[0039] The technical problem to be solved by this invention is how to overcome the shortcomings of traditional distribution box monitoring methods, such as single monitoring dimensions, simple judgment logic, slow response and inability to adaptively evolve, and to provide a method for intelligent management of the safety status of power distribution systems that can achieve multi-dimensional deep perception, edge intelligent analysis, hierarchical reliable protection and continuous self-optimization.

[0040] Reference Figure 1 The technical solution of the present invention will be described in detail below with reference to specific embodiments and illustrations. Those skilled in the art will understand that these descriptions are exemplary and not intended to limit the scope of the invention.

[0041] Example 1 In a core embodiment of the present invention, an intelligent management method for the safety status of a power distribution system is provided. The overall process of this method can be summarized as core steps including data acquisition, data preprocessing, feature extraction, intelligent analysis, hierarchical response, and recovery monitoring, and integrates advanced functions such as self-learning and collaborative management. The following provides a detailed description of each step.

[0042] S100: Obtain a multi-dimensional set of operating parameters for the target power distribution circuit, including a sequence of electrical parameters, a sequence of temperature parameters, a sequence of switch states, and a weak surface acoustic wave signal excited by the vibration of the power distribution circuit conductor.

[0043] The purpose of this step is to comprehensively perceive the operating status of the power distribution circuit. The acquired parameter set includes not only traditional electrical and temperature parameters, but also information reflecting the mechanical status.

[0044] S101: The current signal of the target power distribution circuit, the leakage current characteristic signal derived from the current induction, the cable connection temperature rise trend signal associated with the current signal, and the signal of the conductor surface micro-defects are simultaneously acquired by a non-contact composite sensing unit. The unit actively emits a specific frequency scanning magnetic field and receives its modulation feedback through a micro-flaw detection coil integrated within the unit.

[0045] Multiple key signals from the target power distribution circuit are simultaneously acquired through a non-contact composite sensing unit. This unit typically employs a snap-on or clamp-type structure, facilitating installation on the main circuit's busbar or cable without disconnecting the power supply. It first acquires the circuit's current signal using a high-precision Rogowski coil or current transformer. Simultaneously, by analyzing the three-phase current vector sum and / or zero-sequence current component, leakage characteristic signals characterizing leakage current can be derived. The unit also integrates a sensing element based on a piezoelectric-thermoelectric composite thin film, capable of converting the temperature rise caused by Joule heating from the current flowing through the conductor into a measurable voltage signal, thereby indirectly obtaining the temperature rise trend signal at the cable connection. This is a key innovation of this embodiment, achieving synchronous non-contact sensing of electrical and thermal signals.

[0046] Furthermore, the unit integrates a miniature flaw detection coil. This coil can actively emit a scanning magnetic field of a specific frequency, such as 1 kHz to 100 kHz, towards the conductor under test. When micro-cracks exist on the conductor surface, or when air gaps or corrosion exist at the crimping joint, the distribution of the eddy current field will be distorted, causing the reflected modulated feedback magnetic field to carry defect characteristic information. By demodulating and analyzing this feedback signal, the micro-defect condition of the conductor surface can be assessed at an early stage, enabling preventative maintenance. This signal is acquired as an independent dimension.

[0047] S102: A separate, low-power wireless sensor tag collects ambient temperature and humidity data inside the distribution box. This tag periodically broadcasts the collected data. The ambient temperature and humidity data inside the distribution box are collected using a separate, button battery-powered wireless low-power sensor tag. This tag can use Bluetooth Low Energy (BLE) or Zigbee protocols and periodically broadcasts its collected data, for example, every 10 seconds or every minute. The main control unit can receive this data as environmental background parameters.

[0048] S103: The sensing data from the non-contact composite sensing unit and the wireless low-power sensing tag are fused to form the multi-dimensional operating parameter set. The current signal, leakage characteristic signal, temperature rise trend signal, and conductor flaw detection signal from the non-contact composite sensing unit are timestamped and fused with the ambient temperature and humidity data from the wireless low-power sensing tag to form an information-rich, multi-dimensional operating parameter set. This parameter set lays a solid data foundation for subsequent in-depth analysis.

[0049] Step S110: Adaptive preprocessing is performed on the multi-dimensional operating parameter set to generate a standardized monitoring data stream.

[0050] The raw acquired signals usually contain noise and drift, requiring preprocessing to improve the accuracy of subsequent analysis. The preprocessing in this embodiment is adaptive, dynamically adjusting according to the actual situation on site.

[0051] S111: Perform adaptive filtering based on dynamic load characteristic baseline on the current signal and the leakage current characteristic signal, wherein the dynamic load characteristic baseline is established by learning the current waveform characteristics corresponding to the load type of the target power distribution circuit in real time.

[0052] Adaptive filtering based on a dynamic load characteristic baseline is performed on the current signal and leakage current characteristic signal. Conventional fixed cutoff frequency filtering may filter out useful load change information. In this embodiment, the system learns in real time the current waveform characteristics corresponding to the current load type, such as the sine wave of resistive load, the inductive hysteresis waveform of motor load, and the harmonic characteristics of switching power supply load, to establish a dynamic characteristic baseline under the current load. The filtering algorithm uses this as a reference, mainly filtering out high-frequency random fluctuations that do not conform to the current baseline characteristics, while retaining normal fluctuations and changes that conform to the load characteristics, thereby protecting effective information while denoising.

[0053] S112: Perform drift compensation on the cable connection temperature rise trend signal and the ambient temperature data based on a historical relationship model and recent ambient air pressure fluctuation data.

[0054] The system performs drift compensation based on a historical relationship model and recent ambient air pressure fluctuation data, dynamically correcting the temperature rise trend signal at cable connections and ambient temperature data. Temperature sensors, especially contact temperature sensing elements, experience zero-point drift and sensitivity drift over time. This embodiment not only establishes a historical statistical relationship model between sensor readings and ambient temperature but also innovatively introduces ambient air pressure as a correction factor. This is because changes in air pressure affect the heat exchange conditions inside the sensor package. The system continuously monitors the ambient air pressure, and when a significant and continuous change in air pressure is detected, such as during a typhoon or seasonal change, the parameters of the drift compensation model are dynamically adjusted according to a pre-calibrated air pressure-sensor characteristic offset curve, thereby obtaining a more accurate true temperature rise signal.

[0055] S113: Perform anti-jitter verification based on a dynamic time window on the switch state sequence. The duration of the dynamic time window is negatively correlated with the current load current of the target power distribution circuit.

[0056] Debounce verification based on a dynamic time window is performed on the switch state sequence. Switches, especially contactors, may experience contact bounce during opening and closing. In this embodiment, the debounce time window is not a fixed value, but rather negatively correlated with the current load current. The principle is that under high current loads, the arc generated when the switch operates is stronger, and the contact bounce time is usually shorter; while under low current or no-load conditions, the bounce may be more pronounced. Therefore, when a change in switch state is detected, if the current current is large, a shorter time window, such as 20 milliseconds, is used to verify whether the state is stable; if the current is small, a longer time window, such as 100 milliseconds, is used. This improves the accuracy and real-time performance of state judgment.

[0057] Step S120: Based on the standardized monitoring data stream, extract the time-series feature vector that characterizes the operational safety status of the target power distribution circuit.

[0058] The feature extraction in this embodiment is not a simple statistical calculation, but is closely integrated with the working physical process of electrical equipment.

[0059] The standardized monitoring data stream is automatically segmented into continuous operation segments based on the start and end points of the load cycle. The start and end points can be determined by the load current rising from zero to a certain threshold or falling from a certain value to near zero. Each segment represents a complete equipment start-up / shutdown or operating cycle.

[0060] For each operating segment, the peak current surge and current rise time during the initial transient phase are extracted as the first feature subset; the RMS current, current fluctuation period, and average power during the steady-state operation phase are extracted as the second feature subset; and the current decay time constant during the final transient phase is extracted as the third feature subset. These three subsets comprehensively describe the electrical dynamic characteristics of the load.

[0061] The rate of temperature change during the steady-state operation of this process segment is calculated and correlated with the average current of that segment, forming a fourth feature subset. This reflects the temperature rise efficiency under this load condition and is an important indicator for determining whether the contact resistance has increased.

[0062] Record multiple switches, such as the main circuit breaker and branch contactors, and the time intervals between the action combinations within multiple consecutive operating segments to form a fifth feature subset. For example, a valid motor starting sequence might be "circuit breaker closing -> 2-second delay -> contactor closing," while an abnormal sequence might be "contactor erroneously opening -> circuit breaker not tripping within 100 milliseconds." This reflects the health of the control logic.

[0063] From weak surface acoustic wave signals, the energy entropy value change curves of the frequency components corresponding to the dominant vibration mode within the corresponding operational segment are extracted and used as the sixth feature subset. Conductors generate mechanical vibrations when energized and heated, and their spectrum changes due to changes in mechanical states such as loosening or cracking. Energy entropy values ​​can quantify the concentration of vibration energy in the spectrum; sudden changes in entropy values ​​often indicate changes in mechanical states. This makes the feature vector contain coupled "electric-thermal-mechanical" information.

[0064] The first to sixth feature subsets mentioned above are merged and standardized to form a high-dimensional time-series feature vector with clear physical meaning, which is used for subsequent intelligent analysis.

[0065] Step S130: Input the time-series feature vector into the pre-configured security status identification model for analysis, and output the security status determination result and risk assessment score.

[0066] This embodiment employs a hybrid decision-making model that balances the determinism of rules with the deep mining capabilities of artificial intelligence.

[0067] Step S131: First, input the timing feature vector into the rule engine submodule. This module has a built-in set of deterministic judgment rules based on electrical safety standards, such as national standard GB, and engineering experience thresholds. For example, a rule might be "If the effective current value continuously exceeds 120% of the rated value for 5 seconds, it is judged as overload." If the feature vector triggers any such explicit rule, the corresponding first safety state judgment result, such as "overload," is directly output.

[0068] Step S132: If the feature vector does not trigger any deterministic rules, it indicates a potentially more complex and hidden problem. At this point, it is further input into a lightweight artificial intelligence analysis submodule. This module is a temporal neural network, such as a one-dimensional convolutional neural network or a gated recurrent unit network, that can run on resource-constrained embedded processors like the ARM Cortex-M7 after model compression and quantization. The key innovation of this embodiment is that the network embeds a dynamic attention regulation mechanism that simulates the short-term plasticity of biological neural synapses. This means that when processing feature sequences, the network can automatically increase the attention weight to data within the time window before and after a drastic change in the feature, thereby more sensitively capturing the precursors and evolution of faults. The module outputs a second safety state determination result and a risk probability value between 0 and 1, such as "Poor contact risk, probability 0.76".

[0069] Step S133: Combine the first safety status determination result, the second safety status determination result, and the risk probability value to generate the final safety status determination result. Simultaneously, based on a predefined risk scoring mapping table, the determination result and probability value are quantified into a specific risk assessment score, for example, 0-100 points, with higher scores representing greater risk.

[0070] Step S140: Based on the safety status determination result and risk assessment score, implement a tiered response strategy that matches the risk level. The response strategy is multi-layered and aims to achieve precise intervention.

[0071] A multi-dimensional weighted risk scoring model is established. This model assigns a base risk score to each abnormal feature type and sets correlation weight coefficients for the relationships between different abnormal feature types. The model assigns a base risk score to each abnormal feature type, such as overload, leakage, rapid temperature rise, and abnormal vibration; for example, overload has a base score of 20, and leakage has a base score of 30. Simultaneously, weight coefficients are set for the relationships between different abnormal features. For example, when "overload" and "rapid temperature rise" occur simultaneously, the correlation weight coefficient is 1.5, indicating a stronger risk superposition effect. The system calculates the base score based on the identified abnormal features and their severity. If multiple features occur simultaneously, a weighted joint risk score is calculated.

[0072] Based on the abnormal feature types and their severity identified by the safety status determination results, a basic risk score is calculated for each abnormal feature. The basic risk score or the combined risk score is compared with a preset risk level threshold range, and a graded response is executed according to the comparison result: if it falls into the first threshold range, it is determined to be a minor abnormality level, an early warning log is generated, and a continuous high-resolution acquisition mode for the weak surface acoustic wave signal is initiated; if it falls into the second threshold range, it is determined to be a moderate abnormality level, and the local audible and visual alarm device is triggered, and a remote alarm message is generated and pushed to the management platform; if it falls into the third threshold range or an emergency fault feature combination is identified, it is determined to be a severe danger level, and the circuit cut-off protection command is immediately generated and issued.

[0073] If multiple anomalous features are identified, a joint risk score is calculated based on the correlation weighting coefficients, resulting from the aggregation of these features. The calculated total risk score is then compared with three preset risk level threshold intervals, and a response is executed.

[0074] If the total score falls within the first threshold range, such as 0-30 points, it is judged as a minor anomaly. At this time, an early warning log is generated and recorded to local memory. Simultaneously, the system initiates a continuous high-resolution acquisition mode for weak surface acoustic wave signals. In normal mode, acoustic signals may be acquired at a lower sampling rate to save power. In the case of minor anomalies, the system switches to a high sampling rate and high-resolution spectrum analysis mode to capture more subtle vibration spectrum changes that may indicate mechanical loosening, enabling early and in-depth screening of potential hazards.

[0075] If the total score falls within the second threshold range, such as 31-70 points, it is judged as a moderate abnormality level. At this time, the local audible and visual alarm device is triggered, and a remote alarm message is generated and pushed to the cloud management platform to notify the operation and maintenance personnel to pay attention.

[0076] If the total score falls within the third threshold range, such as 71-100 points, or if an emergency fault characteristic combination such as "short-circuit current characteristics" or "rapid increase in leakage current" is directly identified, it is judged to be at a severe danger level. At this time, a circuit disconnection protection command is immediately generated and issued to prevent the accident from escalating.

[0077] In step S150, after executing the circuit disconnection protection command, the system enters the fault monitoring and recovery process. After the protection action, the system does not passively wait, but actively performs subsequent management.

[0078] After triggering the circuit cut-off protection command, a periodic timer is started. Every first preset time interval, an attempt is made to close a high-impedance detection circuit, and the non-contact composite sensing unit collects the tentative electrical signal in the detection circuit. For example, every 5 minutes, the system will attempt to temporarily close a detection circuit containing a high-resistance resistor by controlling a high-impedance relay, so that a weak tentative current is generated in the circuit.

[0079] The system analyzes the probing electrical signals to determine if the original fault characteristics still exist. If the fault characteristics disappear and the line insulation parameters return to normal, the safety conditions are preliminarily met. Simultaneously, the system assesses whether the energy spectrum of the weak surface acoustic wave signal during the probing period has returned to a stable background noise level. Electrical signals in this probing circuit are acquired using a non-contact composite sensing unit, and the original fault characteristics, such as short-circuit impedance and leakage current, are analyzed. If the electrical parameters return to normal, the safety conditions are preliminarily met. As an important supplementary assessment, the system evaluates whether the energy spectrum of the weak surface acoustic wave signal acquired during the probing period has returned to a stable background noise level. This prevents misjudgment that the fault has been eliminated due to temporarily normal electrical parameters but unstable mechanical hazards, such as unstable arcing points.

[0080] The system uploads a message containing "Fault resolved, requesting remote confirmation for reset" along with a snapshot of relevant security parameters to the remote management platform via a wireless communication network. The system then waits for and receives a remote reset confirmation command from the management platform. This command typically requires authentication and encryption to ensure security.

[0081] After verifying the validity of the remote reset confirmation command, a circuit breaker closing command is generated. At the instant the closing command is generated, this embodiment includes an innovative step: a "soft wake-up" voltage pulse sequence at a specific frequency, such as 25Hz or 75Hz, below the equipment start-up threshold, is pre-sent to the load side of the distribution circuit. This low-voltage, non-power frequency pulse is insufficient to start large loads such as motors, but it can charge the internal capacitors of the load-side controller and soft starter, preparing its control circuitry. When the actual power frequency voltage is restored, each load can start in an orderly manner, avoiding the huge inrush current impact caused by all loads starting simultaneously, thus improving the success rate of power restoration and system stability.

[0082] Example 2 In addition, the system establishes a closed-loop model optimization based on human feedback.

[0083] Maintenance personnel can mark historical alarm events on the remote management platform to confirm whether they are false alarms or a new type of fault.

[0084] If the alarm is confirmed to be a false alarm, the system will mark the historical time-series feature vector of the alarm as a "negative sample" and store it in the model retraining dataset. If the alarm is confirmed to be a new fault, it will be marked as a "positive sample" and stored.

[0085] The system periodically, such as weekly or monthly, incrementally trains the lightweight AI analysis submodule using this retraining dataset. During training, this embodiment introduces a robustness enhancement training step based on adversarial example generation. Specifically, the algorithm simulates typical strong electromagnetic interference patterns found in engineering sites, such as welding machine pulse bursts and large motor start-stop surges, digitally "injecting" these interference patterns into the feature vectors of the training samples to generate adversarial examples. The model is then trained using both the original and adversarial examples, forcing the model to learn to ignore these interferences and focus on real fault characteristics, thereby significantly improving the model's robustness and anti-interference ability in real-world, complex, and harsh electromagnetic environments.

[0086] In the initial phase after system installation and first commissioning, for example, a 72-hour self-learning period is set. During this period, the system only performs monitoring and data recording functions, does not execute any power outage protection commands, but will issue alarm prompts. Based on the historical data collected during this period, the system learns the current range of normal loads, typical temperature rise curves, and environmental background noise under this specific power distribution circuit, establishing a personalized operating parameter baseline and abandoning the "one-size-fits-all" fixed thresholds.

[0087] During long-term operation, the system dynamically calculates the health coefficient of the target power distribution circuit, ranging from 0.5 to 1.0, with 1.0 representing brand new. This coefficient is calculated not only based on the aging model of electrical insulation performance and the historical trend of contact resistance, but also incorporates a "mechanical fatigue" assessment sub-coefficient based on the slow changes in the spectral structure of weak surface acoustic wave signals. By analyzing the drift of dominant frequencies and changes in harmonic components in the vibration spectrum over months or even years, the mechanical aging and fatigue degree of conductor connections can be quantitatively assessed.

[0088] The various judgment thresholds used in the safety status identification model, as well as the weighting coefficients of the risk scoring model in step S141, are associated with this health coefficient. For example, when the health coefficient drops from 1.0 to 0.8, the overload alarm current threshold may be dynamically adjusted from 120% of the rated value to 115%, becoming more conservative; at the same time, the weighting coefficient of the "rapid temperature rise" feature may be increased because aging equipment is more prone to overheating. This allows the system's protection strategy to adaptively adjust according to the equipment's own state, achieving personalized predictive maintenance.

[0089] The local intelligent unit packages and uploads multi-dimensional operating parameter sets, time-series feature vectors, safety status judgment results, risk assessment scores, and all operation records to the cloud management platform via a communication module. The core of this platform is to create a high-fidelity digital twin for each physical distribution box. This twin not only displays data synchronously but also, based on uploaded real-time data and multi-physics simulation models, calculates and visualizes virtual heat maps and electromagnetic field distribution maps of areas invisible to the naked eye inside the distribution box. This gives maintenance personnel "X-ray vision," allowing them to intuitively see hotspot locations and areas of concentrated magnetic fields.

[0090] The management platform can also issue instructions to local units, such as updating the feature library, optimizing model parameters, and adjusting protection settings. In another scenario, when local edge devices encounter extremely complex unknown patterns that computing resources cannot handle, they can upload feature vectors, request the management platform's cloud-based big data model for collaborative analysis, and then send the results back to the local device to assist in decision-making, forming a cloud-edge collaborative intelligent system.

[0091] When the device is powered on for the first time, it enters the network discovery mode and broadcasts its device identifier via Bluetooth Low Energy.

[0092] Maintenance personnel use a dedicated mobile app to approach the device. After the app scans and discovers the device, it establishes a near-field wireless communication connection with it.

[0093] Through the app, maintenance personnel can easily set power distribution circuit parameters, such as rated current 630A and rated voltage 400V, and name the circuit, for example, "Tower Crane No. 1 Main Circuit". This configuration information is sent to the device via Bluetooth.

[0094] After receiving the basic configuration, the device enters the load type self-identification stage. It not only analyzes the steady-state current waveforms collected over the next few days, but also innovatively utilizes the voltage sag waveform characteristics on the upstream grid side caused by the instant the load connects to the grid during the initial power-up phase. By analyzing the depth, duration, and recovery shape of this sag, and matching it with a database of starting impacts for typical loads such as directly started motors, star-delta started motors, frequency converters, and lighting loads, it can help identify the approximate type and starting characteristics of the load. This provides valuable prior knowledge for subsequently setting personalized monitoring strategies.

[0095] S200: During routine monitoring and low-risk data reporting, a low-power wide area network communication module is used for intermittent data transmission.

[0096] When the system is under normal monitoring, with no abnormalities or only low-risk warnings, use low-power wide-area network communication modules, such as NB-IoT narrowband IoT, to perform intermittent, low-volume data reporting in order to save power consumption to the maximum extent and extend the working time of the device in passive power or battery-powered scenarios.

[0097] Once a risk level of moderate anomaly is detected, the system automatically activates a backup high-power, high-speed mobile communication module, such as a 4G LTE or 5G module. Through this high-bandwidth channel, detailed characteristic data and high-definition waveform clips before and after the fault can be rapidly uploaded to the cloud, and high-priority control commands that may be required can be received. A security innovation of this embodiment is that all high-priority data packets uploaded through this channel are embedded with a unique "electromagnetic fingerprint" watermark generated based on the unique electrical signal fragments of the current fault event during packaging. Since the electrical transients of each fault are unique, this watermark is unforgeable. The cloud platform can use it to verify whether data has been tampered with during transmission and to achieve a strong correlation between data and physical events, providing reliable technical evidence for accident tracing and liability determination.

[0098] Example 3 When a local alarm needs to be triggered, the system performs two coordinated actions.

[0099] On one hand, a multi-color light source is controlled, typically a light panel composed of multiple RGB LED beads arranged in a certain space, such as a 3x3 matrix, displaying information according to predefined encoding rules. These rules define the correspondence between color, flashing frequency, and flashing pattern and different faults. For example, a solid red light indicates a serious fault and power failure, a fast-flashing yellow light indicates an overload warning, and a slow-flashing blue light indicates a communication interruption. The ingenuity of this embodiment lies in the flashing pattern. The system can control the LED matrix to scan and illuminate row by row or column by column at extremely high frequencies, such as above 100Hz, according to a specific sequence. Due to the persistence of vision, the observer will see a stable, simple fault icon pattern that appears to float in the air. For example, a flashing flame graphic represents overheating, a lightning bolt graphic represents a short circuit risk, and a loose screw graphic represents abnormal mechanical vibration. This information presentation method is extremely intuitive, allowing on-site personnel to instantly understand the nature of the fault, even from a distance or in a noisy environment.

[0100] On the other hand, a sound-generating device, such as a piezoelectric buzzer or an active speaker, is controlled to output prompt sounds according to coding rules that coordinate with the semantics of the lights. For example, a sharp, rapid "beep beep" sound is paired with a red lightning bolt icon, while a low, slow "dudu" sound is paired with a yellow warning light. This sound-light co-coding greatly enriches the information capacity and recognizability of human-computer interaction.

[0101] When a serious danger is detected and a power outage is required, the system executes a dual-protection power outage action.

[0102] The first control command is generated, typically a digital signal or a specific communication message, and sent to the intelligent circuit breaker in the power distribution circuit via a first communication link, such as an RS485 bus or a CAN bus. This command drives the internal motor operating mechanism or permanent magnet mechanism of the intelligent circuit breaker to perform a tripping operation. This is the main protection path.

[0103] Simultaneously with the generation of the first control command, or after a very short delay, such as 1 millisecond, a second control command is generated. This command controls the conduction of a high-speed solid-state relay. This solid-state relay is connected in series in the drive circuit of the shunt trip coil of the smart circuit breaker. When the solid-state relay conducts, the rated voltage is applied to the shunt trip coil, generating a magnetic force that directly drives the trip mechanism, forcing the circuit breaker to trip. This is the backup hardware protection path. The key optimization of this embodiment is that before issuing the second control command to fully conduct the solid-state relay, the system first injects a brief pre-excitation current pulse with an amplitude lower than its operating threshold into the shunt trip coil circuit. Since the coil is an inductive load, the current cannot change abruptly; this pre-excitation pulse can establish a magnetic field close to the operating threshold in the coil in advance. When the subsequent full-voltage command arrives, the coil current rises from this preset value, thereby reaching the operating threshold more quickly and shortening the operating time of the entire backup protection path by several milliseconds, which is crucial for interrupting developing short-circuit arcs.

[0104] In summary, the intelligent management method for the safety status of power distribution systems provided by this invention constructs a complete intelligent closed loop from perception, analysis, decision-making, execution to optimization through non-contact composite sensing, multi-dimensional feature extraction, hybrid intelligent analysis models, hierarchical response and dual protection mechanisms, combined with self-learning, collaborative digital twins and secure communication strategies. This method elevates traditional passive and singular power distribution protection to proactive, multi-dimensional, adaptive, and evolvable intelligent safety management, significantly improving the safety, reliability, and operational efficiency of power distribution systems, and has promising prospects for industrial applications.

Claims

1. A power distribution system safety state intelligent management method, characterized in that, The method includes: A multi-dimensional set of operating parameters for the target power distribution circuit is acquired. During the acquisition process, a weak surface acoustic wave signal excited by the vibration of the power distribution circuit conductor and associated with a specific harmonic component of its power frequency current is simultaneously acquired, and this signal is used as an independent dimension of the multi-dimensional set of operating parameters. The multi-dimensional operating parameter set is adaptively preprocessed to eliminate noise interference and signal drift, and generate a standardized monitoring data stream. Based on the standardized monitoring data stream, a time-series feature vector characterizing the operational safety status of the target power distribution circuit is extracted. The extraction process of the time-series feature vector is segmented according to the load working cycle of the target power distribution circuit, and transient phase features, steady-state phase features, and end phase features are extracted for each working cycle segment. The time-series feature vector is input into a pre-configured safety status identification model for analysis. The safety status identification model is a hybrid decision model, which includes a rule engine submodule for processing deterministic rules and a lightweight artificial intelligence analysis submodule for processing complex correlation patterns. The analysis outputs the current safety status determination result and the corresponding risk assessment score of the target power distribution circuit. Based on the safety status determination result and the risk assessment score, a graded response strategy matching the risk level is executed. The graded response strategy includes at least one or more of the following: generating a warning log, triggering a local alarm prompt, and executing a circuit disconnection protection command. After the execution circuit cut-off protection command is issued, the fault monitoring and recovery process is initiated. The fault characteristics are periodically checked to see if they have been eliminated. Once the safety conditions are confirmed to be met, a prompt message indicating that power supply can be restored is generated.

2. The method of claim 1, wherein, The acquisition of the multi-dimensional operating parameter set of the target power distribution circuit includes: Synchronously acquire the current signal of the target power distribution circuit, the leakage current characteristic signal derived from current induction, and the cable connection temperature rise trend signal associated with the current signal; and / or, The wireless low-power sensor tag periodically broadcasts the collected data, which includes the ambient temperature and humidity inside the distribution box. Adaptive filtering based on dynamic load characteristic baseline is performed on the current signal and the leakage current characteristic signal. The dynamic load characteristic baseline is established by learning the current waveform characteristics corresponding to the load type of the target power distribution circuit in real time. The filtering process is to remove high-frequency random fluctuation components that do not conform to the current dynamic load characteristic baseline. The temperature rise trend signal of the cable connection and the ambient temperature data are subjected to drift compensation based on a historical relationship model to dynamically correct the zero drift and sensitivity drift of the sensor. The drift compensation introduces a pressure compensation factor that is dynamically corrected based on recent ambient air pressure fluctuation data. A debouncing check based on a dynamic time window is performed on the switch state sequence. The duration of the dynamic time window is negatively correlated with the current load current of the target power distribution circuit. A change in switch state is confirmed only when the switch state signal remains stable within the dynamic time window.

3. The method of claim 1, wherein, The extraction of the time-series feature vector characterizing the operational safety status of the target power distribution circuit includes: The standardized monitoring data stream is automatically divided into continuous operation process segments based on the start and end points of the load working cycle. For each of the aforementioned operating process segments, the following feature extraction operations are performed: the peak current surge and current rise time of the initial transient phase are extracted as the first feature subset; the effective current value, current fluctuation period, and average power of the steady-state operating phase are extracted as the second feature subset; and the current decay time constant of the final transient phase is extracted as the third feature subset. The rate of temperature change of the operation segment during the steady-state operation phase is calculated and correlated with the average current of the operation segment, which is used as the fourth feature subset. Record the action combination sequence of multiple switches in multiple consecutive operation process segments and the time interval between adjacent actions to form a fifth feature subset; From the weak surface acoustic wave signal, extract the energy entropy value change curve of the frequency component corresponding to the dominant vibration mode within the corresponding operation process segment, and use it as the sixth feature subset; The first feature subset, the second feature subset, the third feature subset, the fourth feature subset, the fifth feature subset, and the sixth feature subset are merged and encoded to form the time-series feature vector.

4. The method of claim 1, wherein, The step of inputting the time-series feature vector into a pre-configured security status identification model for analysis includes: First, the timing feature vector is input into a set of deterministic judgment rules based on electrical safety standards and empirical thresholds. If the timing feature vector triggers any rule in the set of deterministic judgment rules, the corresponding first safety state judgment result is directly output. If the time-series feature vector does not trigger any of the deterministic judgment rules, it is further input into the lightweight artificial intelligence analysis submodule. The lightweight artificial intelligence analysis submodule is a compressed and quantized time-series neural network model used to identify the complex fault modes and risk signs hidden in the time-series feature vector, and output the second safety state judgment result and risk probability value. The lightweight artificial intelligence analysis submodule embeds a dynamic attention regulation mechanism that simulates the short-term plasticity of biological neural synapses, so that when processing continuous time series, it can adaptively enhance the attention weight of features within the time window before and after the mutation. By combining the first safety status determination result, the second safety status determination result, and the risk probability value, a final safety status determination result is generated, and the risk assessment score is calculated based on a predefined risk scoring mapping table.

5. The method of claim 1, wherein, The step of implementing a graded response strategy matching the risk level based on the security status determination result and the risk assessment score includes: A multi-dimensional weighted risk scoring model is established, in which a basic risk score is assigned to each abnormal feature type, and a correlation weight coefficient is set for the correlation between different abnormal feature types. Based on the types and severity of abnormal features identified by the security status determination results, calculate the basic risk score for each abnormal feature; If multiple abnormal features are identified, the joint risk score of the superimposed abnormal features is calculated based on the correlation weight coefficient. The basic risk score or the combined risk score is compared with a preset risk level threshold range. When the total score falls into the first threshold range, it is determined to be a minor anomaly. An early warning log is generated and recorded in the local memory. At the same time, a continuous high-resolution acquisition mode for the weak surface acoustic wave signal is started to capture potential precursor signals of mechanical loosening. When the total score falls into the second threshold range, it is determined to be a moderate anomaly. A local audible and visual alarm device is triggered, and a remote alarm message is generated and pushed to the management platform. When the total score falls into the third threshold range or an emergency fault characteristic combination is identified, it is determined to be a serious danger level. The circuit cut-off protection command is immediately generated and issued.

6. The method of claim 1, wherein, The execution circuit cut-off protection command includes: Generate a first control command for driving the operating mechanism inside the smart circuit breaker to perform a tripping operation, and generate a second control command for generating magnetic force through the shunt trip coil to forcibly drive the smart circuit breaker to perform a tripping operation; The main protection path corresponding to the first control command and the backup protection path corresponding to the second control command together constitute a dual-channel power failure protection execution mechanism. Furthermore, before the second control command is generated, a pre-excitation current pulse lower than its action threshold is injected into the shunt trip coil circuit to offset the action delay caused by the coil inductance.

7. The method of claim 1, wherein, Whether the periodic detection fault characteristics have been eliminated includes: After the circuit cut-off protection command is triggered, a periodic timer is started. Every first preset time interval, an attempt is made to close a high-impedance detection circuit, and the non-contact composite sensing unit collects the tentative electrical signals in the detection circuit. Analyze the probing electrical signal to determine whether the original fault characteristics still exist. If the fault characteristics disappear and the line insulation parameters return to normal, it is determined that the safety conditions are initially met. At the same time, assess whether the energy spectrum of the weak surface acoustic wave signal during the probing period has returned to a stable background noise level. And / or, after triggering the circuit cut-off protection command, continuously monitor the ambient temperature change through the wireless low-power sensor tag. If the abnormal temperature rise trend associated with the fault has returned to normal, and combined with the judgment of the electrical signal, it is determined that the safety conditions are initially met.

8. The method of claim 1, wherein, The method also includes system self-learning and parameter adaptation steps: During the initial operation phase of the system, a self-learning period is set. During the self-learning period, the system only performs monitoring and recording, and does not execute the circuit disconnection protection command. Based on the historical data collected during the self-learning period, a personalized operating parameter baseline for the target power distribution circuit is established. The personalized operating parameter baseline includes the normal current range, temperature range, and load mode characteristics. Based on long-term operating data and alarm history, the health coefficient of the target power distribution circuit is dynamically calculated. The health coefficient is used to characterize the overall aging and performance change trend of the circuit. The calculation of the health coefficient incorporates the "mechanical fatigue" evaluation sub-coefficient based on the slow change of the spectrum structure of the weak surface acoustic wave signal. The judgment threshold used in the security status identification model, the weight coefficient in the risk assessment score calculation, and the health coefficient are associated so that the judgment threshold and the weight coefficient can be dynamically adjusted as the health coefficient changes.

9. The method of claim 1, wherein, The method also includes a collaborative working step with a management platform: The multi-dimensional operating parameter set, the time-series feature vector, the security status determination result, the risk assessment score, and the execution record of the graded response strategy are encapsulated into a data packet and uploaded to the management platform. The management platform maintains a digital twin for each target power distribution circuit. The digital twin synchronously updates the real-time status and historical data of the target power distribution circuit. Based on real-time data and physical models, the digital twin can deduce and visualize the virtual heat map and electromagnetic field distribution map of the invisible area inside the power distribution box in real time. It receives control policy update, model parameter update, or feature library update instructions issued by the management platform and updates its local configuration accordingly. And / or, when local computing resources are insufficient or unknown feature patterns are encountered, the time series feature vector is uploaded to the management platform, requesting the management platform to perform cloud-based collaborative analysis, and the analysis results returned by the management platform are received to assist local decision-making.

10. The method according to claim 1, characterized in that, The method executes the following during configuration initialization: When the device is powered on for the first time, it enters the distribution network discovery mode and broadcasts the device identification information; Establish near-field wireless communication connection with nearby configuration terminals; Through the near-field wireless communication connection, the system receives power distribution circuit parameter configuration information sent by the configuration terminal. The parameter configuration information includes rated current, rated voltage, and circuit identifier. After receiving the basic parameter configuration, the system enters the load type self-identification stage. By analyzing the characteristics of the current waveform collected subsequently, it matches them with the pre-stored typical load feature library to automatically infer the main load type connected to the target power distribution circuit. The main load type is then used to optimize the feature extraction and safety status identification strategy. In the load type self-identification stage, the system also analyzes the voltage sag waveform characteristics on the grid side caused by the instantaneous connection of the load during the initial power-on stage to help identify the type of load start-up impact characteristics.

Images