An arc fault pre-judgment method and system for a fault self-recovery type intelligent integrated distribution box

Abstract

The application discloses a kind of arc fault pre-judgment method and system of fault self-healing type intelligent comprehensive distribution box, the method utilizes the current transformer and voltage transformer of existing distribution box to collect each loop electrical signal, the electrical signal is extracted with sliding window feature to calculate normalized abnormality index, abnormality index is compared with dynamic baseline to output abnormality mark and feature vector;Multi-parameter correlation analysis is carried out in combination with historical operation data to evaluate fault risk level;When reaching early warning level, part of the load of target loop is transferred to adjacent redundant loop, response change is continuously observed and feedback correction dynamic baseline is fed back, so that feature extraction, risk assessment and load regulation form closed-loop linkage;According to response change, determine the nature of the fault and execute corresponding control strategy.The application realizes early prediction and preventive elimination of arc fault through cross-coupling closed-loop linkage mechanism without special arc light sensor, eliminates sensor space blind area and environmental light false alarm problem.

Description

Technical Field

[0001] This invention relates to intelligent protection technology for power distribution systems, specifically a method and system for predicting arc faults in a self-healing intelligent integrated distribution box. Background Technology

[0002] In medium and low voltage power distribution systems, arc flash faults have received widespread attention due to their intense energy release and destructive power. Existing arc flash protection schemes rely on deploying a dedicated arc flash sensor array inside the switchgear to capture the light radiation signal generated by the fault, and then a dedicated main control unit performs signal processing and logic judgment.

[0003] However, this approach faces the inherent drawback of incomplete sensor spatial coverage in practical applications. The internal structure of the switch cabinet is complex and contains obstructions and blind spots. No matter how many probes are added, it is difficult to achieve full coverage of all potential fault points. Furthermore, the arc light sensor is essentially a non-selective light intensity measurement device. While improving the sensitivity to detect weak initial arc light, it inevitably increases the sensitivity to ambient light interference, which may trigger false alarms due to daytime cabinet door opening maintenance, flashes generated by adjacent cabinet operations, or even changes in the lighting of the power distribution room.

[0004] Furthermore, in the existing process, the information flow between each processing step strictly follows a unidirectional sequence. The abundant electrical parameters already collected by the current transformer are only used for secondary verification after a fault occurs rather than for pre-fault prediction. Historical fault data is completely detached from the online decision-making process and cannot dynamically optimize the detection threshold. Key fault precursor information such as the temporal evolution trend of electrical parameters and thermal accumulation status is fragmented in each step and fails to form effective information fusion. Moreover, the overall design belongs to a passive response mode, that is, waiting for the arc to actually occur before detection and removal, completely ignoring the usable early warning information contained in the gradual evolution of arc faults from insulation degradation to partial discharge to breakdown discharge.

[0005] Current technology considers arc detection to be an optical-physical process that requires photoelectric conversion hardware, but fails to recognize that multi-dimensional time-series correlation analysis and closed-loop verification of existing electrical quantities in the distribution box can achieve earlier and more reliable fault prediction than direct optical measurement. Summary of the Invention

[0006] The technical problem to be solved by the present invention is the dilemma of sensor blind spots and false alarms in arc fault prediction. That is, the arc sensor has physical blind spots and is easily affected by ambient light interference, which can cause false alarms. The existing solutions lack interaction between the processing steps, and cannot achieve early prediction and preventive elimination of faults.

[0007] Therefore, the present invention provides a method for predicting arc faults in a self-healing intelligent integrated distribution box.

[0008] To address the aforementioned technical problems, this invention provides a method for predicting arcing faults in a self-healing intelligent integrated distribution box, comprising: The existing current transformers and voltage transformers in the distribution box are used to collect electrical signals of each circuit. The electrical signals are then subjected to sliding window feature extraction to calculate a normalized anomaly index. The anomaly index is compared with a dynamic baseline. When the deviation reaches a preset ratio, an anomaly marker and the corresponding feature vector are output. Based on the anomaly markers and feature vectors, multi-parameter correlation analysis is performed using historical operational data of the target circuit to assess the fault risk level; wherein, the fault risk level includes attention level, early warning level, and emergency level; When the fault risk level reaches the warning level, part of the load of the target circuit is transferred to the adjacent redundant circuit according to priority to reduce insulation stress. The electrical characteristic response of the target circuit after the load transfer is continuously observed, and the response change is fed back to the dynamic baseline for correction, so that feature extraction, risk assessment and load regulation form a closed loop linkage. The nature of the fault is determined based on the response change, and the corresponding control strategy is executed: if the abnormal characteristics decrease as the load decreases, the current load allocation is maintained; if the abnormal characteristics do not improve or worsen, the fault risk level is raised to the emergency level and protection action is triggered.

[0009] The present invention also provides a fault self-healing intelligent integrated distribution box arc fault prediction system, including an electrical feature extraction unit, a risk assessment unit, a load regulation and verification unit, and a fault determination and execution unit. Each unit undertakes the functions of feature extraction and anomaly calculation, risk level assessment, preventive load transfer and closed-loop verification, fault nature determination and control strategy execution, etc. The load regulation and verification unit feeds back the response changes to the electrical feature extraction unit to correct the dynamic baseline, so that the units form a closed-loop linkage.

[0010] In summary, the present invention has the following beneficial effects: By establishing a coupled closed-loop linkage between the electrical feature extraction step and the preventive load control step, the problem of arc fault detection requiring direct optical measurement using a dedicated arc sensor array and main control unit is overcome. By utilizing the existing current transformer and voltage transformer signals in the distribution box, fault prediction can be achieved earlier than optical detection, fundamentally eliminating the problems of sensor blind spots and ambient light false alarms.

[0011] The load control steps in the closed-loop linkage have both fault prevention and active verification functions. By "applying load disturbance and observing electrical response", the correctness of fault risk judgment is actively confirmed, avoiding the misjudgment risk of the traditional single threshold comparison method. At the same time, it preventively reduces insulation stress, thereby eliminating potential fault hazards before arcing occurs.

[0012] The dynamic baseline adaptive correction mechanism formed by reverse coupling enables the system to continuously learn. Each closed-loop interaction not only confirms or eliminates the current fault risk, but also performs an online calibration of the anomaly detection model. The fault identification accuracy of the system continues to improve with the accumulation of running time.

[0013] By eliminating the need for dedicated arc light sensor arrays and supporting fiber optic transmission systems, the risks of hardware failures such as optical component aging, fiber breakage, and probe contamination are eliminated. This significantly improves long-term operational reliability and maintainability while reducing system hardware costs and installation complexity.

[0014] This invention forms a hierarchical collaborative relationship with the existing overcurrent protection and short-circuit protection of the distribution box as a prediction layer. When the prediction system determines that the situation is emergency, it triggers protection action to cooperate with conventional protection, ensuring that even if the prediction function is abnormal, conventional protection still serves as a backup. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the overall process and cross-coupling closed-loop linkage structure of the arc fault prediction method in Embodiment 1 of the present invention.

[0016] Figure 2 This is a schematic diagram of the sliding window feature extraction process in Embodiment 2 of the present invention.

[0017] Figure 3 This is a schematic diagram of the fusion logic for calculating the normalized anomaly index in Embodiment 3 of the present invention.

[0018] Figure 4 This is a schematic diagram of the dynamic baseline initialization and update mechanism in Embodiment 4 of the present invention.

[0019] Figure 5 This is a schematic diagram of the risk level determination logic for multi-parameter correlation analysis in Embodiment 5 of the present invention.

[0020] Figure 6 This is a schematic diagram of the load transfer priority sequence and execution flow in Embodiment Six of the present invention.

[0021] Figure 7 This is a schematic diagram of the electrical characteristic response observation process after load transfer in Embodiment 7 of the present invention.

[0022] Figure 8 This is a schematic diagram of the dynamic baseline feedback correction and event classification marking logic in Embodiment 8 of the present invention.

[0023] Figure 9 This is a schematic diagram of the module structure and inter-unit information interaction of the arc fault prediction system in Embodiment 10 of the present invention.

[0024] Figure 10This is a schematic diagram of the adaptive baseline evolution and knowledge accumulation mechanism in Embodiment 9 of the present invention.

[0025] Figure 11 This is a schematic diagram showing the comparison results of the verification experiment in Embodiment 11 of the present invention. Detailed Implementation

[0026] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0027] Example 1: Overall process and cross-coupled closed-loop linkage structure of arc fault prediction method like Figure 1 As shown, this embodiment provides a method for predicting arcing faults in a self-healing intelligent integrated distribution box. This method establishes a closed-loop linkage between three core steps: electrical feature extraction, fault risk assessment, and preventative load control. This replaces the passive optical detection function of traditional arcing sensor arrays, enabling early prediction and preventative elimination of arcing faults. This method is positioned as the prediction layer in the distribution box protection system, forming a hierarchical collaborative relationship with the existing overcurrent and short-circuit protection of the distribution box. Even if this prediction method fails due to abnormal operating conditions, conventional overcurrent and short-circuit protection can still serve as backup to clear the fault. The overall method includes the following steps:

[0028] S1 uses the existing current transformers and voltage transformers in the distribution box to collect electrical signals from each circuit, performs sliding window feature extraction on the electrical signals to calculate the normalized anomaly index, compares the anomaly index with the dynamic baseline, and outputs an anomaly mark and the corresponding feature vector when the deviation reaches a preset ratio.

[0029] It should be noted that the core technical logic of this step lies in transforming arc fault detection from a direct photoelectric conversion measurement of optical radiation signals to a correlation inference based on the temporal characteristics of electrical parameters. The essence of an arc fault is a gas discharge phenomenon following insulation breakdown. Before the arc occurs, the breakdown point inevitably undergoes a gradual process of local insulation degradation, micro-discharge, and heat accumulation. These precursory processes leave identifiable characteristic imprints in the electrical circuit, including high-frequency pulse components in the current waveform, transient dips in the voltage waveform, and abnormal increases in the amplitude of specific harmonics in the harmonic spectrum. This electrical characteristic information already exists in the data collected by the existing current transformers and voltage transformers in the distribution box and can be acquired without additional optical sensors.

[0030] The sliding window analysis method segments continuous electrical signals, extracts multidimensional features within each time window, and synthesizes them into a normalized anomaly index. This index reflects the degree of deviation of the circuit insulation health status from the normal operating level within the current time period. For example, for a power distribution system with a power frequency of 50Hz, the sampling frequency is preferably selected to be 200 times the fundamental frequency, i.e., 10kHz, to ensure that the high-frequency pulse characteristics generated by partial discharge can be captured; the sliding time window length is preferably 10 power frequency cycles, i.e., 200ms. This length can obtain statistically stable feature values ​​without obscuring rapidly developing abnormal signals.

[0031] The dynamic baseline is the core difference between this step and the traditional fixed threshold judgment. It is continuously corrected by the feedback results of step S3, so that the anomaly judgment criteria can be dynamically adjusted according to changes in the system's operating status and the accumulation of knowledge. The preset ratio for anomaly judgment is preferably set to 3 times the baseline fluctuation range, i.e., k=3. This value refers to the 3σ criterion in statistical process control, which corresponds to a normal fluctuation false trigger probability of about 0.27% under the assumption of normal distribution, and can achieve a reasonable balance between detection sensitivity and false alarm suppression.

[0032] When the anomaly degree deviates from the dynamic baseline by a preset proportion, the system considers the circuit to have signs of insulation degradation and outputs an anomaly marker and a feature vector containing feature values ​​of each dimension for use in subsequent steps. The coupling relationship of this step in the closed-loop structure is as follows: receiving feedback information from step S3 to correct the dynamic baseline, while simultaneously outputting the anomaly degree and feature vector to step S2 in a positive direction.

[0033] S2, based on anomaly markers and feature vectors, combined with historical operating data of the target circuit, performs multi-parameter correlation analysis to assess the fault risk level; among which, the fault risk level includes attention level, early warning level, and emergency level.

[0034] It should be noted that the function of this step is to transform the raw anomaly detection results output from step S1 into a risk level judgment with clear operational significance. A high anomaly value alone does not necessarily mean there is a real fault risk; it may originate from non-fault factors such as load fluctuations, changes in ambient temperature, or normal equipment aging. Therefore, this step introduces multi-parameter correlation analysis, comprehensively considering multiple dimensions such as the absolute value of the anomaly degree, the concentration of anomaly features in the frequency domain, the duration of the anomaly state, and the slope of the anomaly degree trend over time, combined with the historical operating data and load characteristics of the circuit, to make a comprehensive judgment.

[0035] The three risk levels correspond to different follow-up strategies: the "Attention Level" triggers an enhanced monitoring mode, which shortens the feature extraction time window to obtain more refined monitoring data; the "Early Warning Level" triggers preventative load regulation in step S3; and the "Emergency Level" directly triggers the circuit breaker protection to disconnect the faulty circuit. The emergency level triggering complements the existing overcurrent protection in the distribution box, as the overcurrent protection is based on the absolute value of the current exceeding the limit.

[0036] The emergency-level triggering in this method is based on a sharp deterioration in the anomaly trend. Both provide protection from different dimensions, and triggering by either one can independently execute a tripping action. This tiered response mechanism avoids applying the same response strength to all abnormal signals.

[0037] S3, when the fault risk level reaches the warning level, a portion of the load of the target circuit is transferred to the adjacent redundant circuit according to priority to reduce insulation stress. The electrical characteristic response changes of the target circuit after the load transfer are continuously observed, and the response changes are fed back to the dynamic baseline for correction, so that feature extraction, risk assessment and load regulation form a closed loop linkage.

[0038] It should be noted that this step is part of the entire cross-coupling closed-loop linkage process, and has two functions: preventive fault elimination and proactive risk verification.

[0039] 1) Preventive fault elimination: By transferring part of the load of the target circuit to an adjacent redundant circuit, the current thermal effect and insulation electric field stress of the target circuit are reduced, thereby preventing potential faults before the arc is generated.

[0040] 2) Proactive risk verification: Load regulation is equivalent to applying a controllable disturbance signal to the power circuit. The response characteristics of the circuit after the disturbance can reflect the true health status of the insulation. This proactive detection strategy of "applying excitation and observing the response" essentially replaces the function of traditional arc flash sensors that "passively wait for the fault to occur and then detect the light signal," and provides deeper diagnostic information.

[0041] This step feeds back the electrical response characteristics after load regulation to step S1 to correct the dynamic baseline of anomalies. This reverse coupling enables the system to have adaptive learning capabilities. If the abnormal characteristics subside after the load is reduced, it is confirmed that the anomaly originates from reversible insulation stress caused by heat accumulation, and the dynamic baseline is adjusted accordingly to reflect the current normal operating state of the circuit. If the abnormal characteristics persist or even worsen, the judgment threshold of the dynamic baseline is lowered to increase the sensitivity to subsequent anomalies in the circuit. Through this feedback correction, the system's feature discrimination accuracy continuously improves with the increase of operating time.

[0042] 3) Execution constraints: This step only initiates load transfer when adjacent circuits have sufficient capacity margin. When the load rate of all adjacent circuits exceeds the safety threshold and cannot accept additional load, the system skips the load transfer operation and directly enters enhanced monitoring mode, awaiting manual intervention from maintenance personnel. Furthermore, the load switching operation itself may generate transient overvoltages and inrush currents. Therefore, each switching operation is performed near the zero-crossing point of the load current to minimize transient impacts, and subsequent operations are only performed after the system reaches steady state following the switching.

[0043] S4. Determine the nature of the fault based on the response change and execute the corresponding control strategy: If the abnormal characteristics decrease as the load decreases, maintain the current load distribution; if the abnormal characteristics do not improve or worsen, raise the fault risk level to the emergency level and trigger the protection action.

[0044] It should be noted that this step determines the nature of the fault and decides on the subsequent control strategy based on the changes in electrical characteristic response observed in step S3. The logical basis for determining the nature of the fault is that thermally accumulated insulation stress is reversible. When the external stress decreases, the performance of the insulation material can gradually recover, and the corresponding abnormal electrical characteristics will also subside. However, irreversible insulation damage, such as carbonization, cracks, or penetrating defects in the insulation material, will not repair itself due to a reduction in load. Its abnormal electrical characteristics will persist or even worsen with further insulation deterioration.

[0045] Based on this physical mechanism, if the abnormal characteristics decrease as expected, the current load distribution scheme will be maintained until maintenance personnel intervene for repair; if the abnormal characteristics do not improve or even worsen, the risk level will be immediately upgraded to the emergency level and protective actions such as circuit breaker tripping will be performed to avoid the actual occurrence of arc flash faults.

[0046] It is important to note that this method is designed as an early prediction and preventative measure for arc faults, not as a sole protective layer to replace all other protective functions. For sudden insulation breakdown scenarios without gradual warning signs, the existing overcurrent and short-circuit protection in the distribution box will continue to operate independently as backup measures, without interference. The core value of this method lies in identifying and eliminating most arc faults with gradual development characteristics in their early warning stages, thereby significantly reducing the actual probability of arc fault occurrence.

[0047] In summary, steps S1 to S4 form a closed-loop structure of S1→S2→S3→S1, with the reverse coupling between steps S1 and S3 constituting the core adaptive mechanism of the system. Each closed-loop interaction simultaneously completes two tasks: confirming or eliminating the current fault risk, and performing an online calibration of the anomaly detection model. This spiral-upward mechanism of "detection-verification-correction" enables the system's discrimination accuracy to continuously improve over time, which is impossible to achieve with passive detection schemes based on fixed hardware.

[0048] Example 2: Processing flow for sliding window feature extraction like Figure 2 As shown, this embodiment corresponds to the following content: "The sliding window feature extraction of the electrical signal includes: continuously sampling the current and voltage waveforms of each circuit at a preset sampling frequency; segmenting the sampled data into units of sliding time windows; performing frequency domain decomposition on the current waveform within each time window and extracting the energy proportion of high-frequency pulse components as the first feature; performing transient analysis on the voltage waveform within each time window and extracting the frequency and amplitude of voltage transient dips as the second feature; and performing harmonic analysis on the electrical signal within each time window and extracting the growth rate of the amplitude of a specific harmonic relative to the baseline period as the third feature."

[0049] The sliding window feature extraction step transforms continuous electrical signals into quantifiable indicators of impending faults. The selection of the preset sampling frequency needs to balance the ability to capture high-frequency transient signals with the computational load of data processing. For a power distribution system with a power frequency of 50Hz, the high-frequency pulse components generated by partial discharge are typically distributed in the frequency range of hundreds to thousands of Hz. According to the Nyquist sampling theorem, the sampling frequency needs to be at least twice the target frequency. A preferred sampling frequency is 10kHz, which is 200 times the fundamental frequency. This frequency can effectively capture the characteristics of partial discharge pulses distributed below 5kHz, while the data volume is within the computational capabilities of existing embedded processors. Within each sampling clock cycle, one sliding time window (200ms) contains 2000 sampling points, which is sufficient to support the accuracy requirements of frequency domain decomposition and harmonic analysis.

[0050] The preferred sliding time window length is 200ms, which is equivalent to 10 complete cycles of a 50Hz power frequency system. This window length is chosen based on two criteria: first, including 10 complete power frequency cycles ensures that the frequency domain decomposition yields complete fundamental and harmonic information, avoiding spectral leakage; second, a 200ms window length corresponds to a feature update rate of 5 windows / second, enabling the accumulation of 100 continuous feature samples within 20 seconds for trend analysis, meeting the tracking requirements for slowly developing anomalies. The preferred sliding step size between adjacent windows is 50% of the window length, i.e., 100ms. This means that there is 50% data overlap between adjacent windows, improving the temporal resolution of the features to 10 windows / second without increasing the total computational load.

[0051] The first characteristic quantity is the energy proportion of the high-frequency pulse component. When local degradation occurs in the insulating material, minute partial discharges may occur under the influence of an electric field. These discharge events manifest as high-frequency pulses superimposed on the current waveform. By performing a fast Fourier transform frequency domain decomposition on the current waveform, the ratio of the energy in the frequency band above the set cutoff frequency to the total signal energy can be calculated, thus obtaining a quantitative indicator reflecting the degree of partial discharge activity.

[0052] The cutoff frequency is preferably set to 10 times the fundamental frequency, i.e., 500Hz. The energy components above this frequency mainly come from the superposition of non-power frequency sinusoidal signals, including partial discharge pulses, switching operation transients, etc.

[0053] The specific calculation method is as follows: Perform a Fast Fourier Transform on 2000 current sampling points within the window to obtain the spectrum, and calculate the sum of the power spectral densities of all frequency components above 500Hz, denoted as P. hf The sum of the power spectral densities across the entire frequency band is denoted as P. total Then the original value of the first characteristic quantity is V1=P hf / P total Under normal operating conditions, V1 is usually maintained at a very low level; when localized insulation deterioration occurs and persistent partial discharge activity occurs, V1 will show a significant and continuous upward trend.

[0054] The second characteristic quantity is the frequency and amplitude of voltage transient sags. When intermittent breakdown or flashover occurs at a certain point in the insulation of a circuit, the breakdown will cause a brief transient sag in the voltage of that circuit.

[0055] Transient dips are detected by setting a threshold for amplitude drop in the voltage waveform. The specific method is as follows: At each sampling point, the deviation between the instantaneous voltage value and its theoretical sinusoidal envelope is calculated. When this deviation exceeds 5% of the nominal voltage amplitude (i.e., the detection threshold is set to 5% of the nominal amplitude), it is recorded as the start of a dip event. A dip event ends when the deviation returns to below the threshold. The cumulative number N of dip events is counted within each sliding window. sag And the maximum indentation amplitude D among all indentation events max (Expressed as a percentage of nominal voltage amplitude). Since the second characteristic quantity contains both frequency and amplitude dimensions, it needs to be synthesized into a single value for subsequent normalization. The synthesis method is to multiply the number of indentations by the maximum indentation amplitude to obtain the original synthesized value of the second characteristic quantity:

[0056] ; Where V2 is the comprehensive value of voltage transient sag, in units of times per second; N sag This is the cumulative number of indentation events within the current window, expressed in times; D max This represents the percentage of the maximum indentation amplitude within the current window relative to the nominal voltage amplitude, expressed as a percentage. The physical meaning of this product form is: only high-frequency but small-amplitude indentations (such as normal load fluctuations) will not produce a high composite value; only large-amplitude but extremely low-frequency single indentations (such as external transient interference) will not produce a sustained high composite value. Only when both frequency and amplitude are high does it indicate the existence of sustained intermittent breakdown activity. Under normal operating conditions, N... sag It is usually 0 or a very small number, and V2 is close to 0.

[0057] The third characteristic is the growth rate of a specific harmonic amplitude relative to the baseline period. Nonlinear electrical behavior caused by insulation degradation introduces harmonic components of specific orders into the circuit. The specific harmonics of concern are mainly the 3rd, 5th, and 7th harmonics, as these odd harmonics are most closely related to nonlinear electrical characteristics. The growth rate of each harmonic is calculated as follows:

[0058] The amplitude A of the nth harmonic within the current window is extracted using Fast Fourier Transform. n The corresponding subharmonic amplitude A in the normal operating baseline period recorded during the system initialization phase. n0 By comparison, the growth rate Gn=(A) is calculated. n -A n0 ) / A n0 Since the growth rates of three subharmonics are monitored simultaneously, they need to be combined into a single value for subsequent normalization. The combination method is to take the maximum value among the three growth rates as the original combined value of the third characteristic:

[0059] ; Wherein, V3 is the comprehensive value of the growth rate of a specific harmonic, dimensionless; G3, G5, and G7 are the growth rates of the 3rd, 5th, and 7th harmonics, respectively, dimensionless. The basis for taking the maximum value is that different types of insulation degradation mechanisms may produce the most significant response on different harmonics, and taking the maximum value can capture the changes of the most sensitive harmonics without being diluted by the normal fluctuations of other harmonics.

[0060] The three features characterize the precursory information of insulation degradation from different physical dimensions, complementing each other and jointly constituting a multi-dimensional design for the insulation health status of the circuit. Compared with the single light intensity measurement of traditional arc light sensors, this multi-dimensional feature extraction method has the advantages of richer information dimensions and is not limited by spatial blind spots. Insulation degradation at any location will leave an identifiable trace in the overall electrical parameters of the circuit.

[0061] Example 3: Fusion Logic for Calculating Normalized Outlier Index like Figure 3 As shown, this embodiment corresponds to the following content: "The calculation of the normalized anomaly index includes: normalizing the first, second, and third feature quantities respectively to eliminate dimensional differences; assigning adaptive weighting coefficients to each normalized feature quantity based on its sensitivity to insulation degradation; fusing the weighted feature quantities to obtain the normalized anomaly index characterizing the overall insulation health status within the current time window; and periodically updating the adaptive weighting coefficients based on the historical correction records of the dynamic baseline."

[0062] The purpose of normalization is to eliminate dimensional and numerical scale differences among the various characteristic quantities. The first characteristic quantity, V1, is a dimensionless energy ratio, typically ranging from 0 to 0.1. The second characteristic quantity, V2, is the product of frequency and amplitude percentage, expressed in times per percent (%), and its value can range from 0 to tens depending on the insulation condition. The third characteristic quantity, V3, is a dimensionless growth rate ratio, ranging from 0 to several times depending on the insulation condition. Since the numerical ranges and physical dimensions of these three quantities are different, a direct weighted sum has no physical meaning.

[0063] The normalization process employs a min-max normalization method, mapping the original comprehensive values ​​of each feature to a unified interval between 0 and 1. Specifically, for each feature, the lower bound of the normal operating range obtained during the system initialization phase is used as the minimum value V. min The maximum value V is taken as a preset multiple (preferably 3 times) of the upper limit of this range. max This linearly maps the current feature value to the interval between 0 and 1.

[0064] The normalization formula is: ; Among them, f i V is the normalized value of the i-th feature, dimensionless, and ranging from 0 to 1; i V represents the original composite value of the i-th feature in the current window; min,i V is the normal operating lower bound obtained statistically during the initialization phase for the i-th feature; max ,i is the normalized upper bound of the i-th feature, which is taken as 3 times the normal operating upper bound during the initialization phase. When the calculation result f i When f < 0, take 0; when f i When the value is greater than 1, it is set to 1. V is defined as 3 times the upper bound of the normal range during the initialization phase. max The basis for this is that under normal operating conditions, the characteristic values ​​are distributed within the normal range, and 3 times the upper limit corresponds to a severe abnormal state. Values ​​exceeding this limit are considered to be full-scale abnormalities.

[0065] The initial allocation of adaptive weighting coefficients is based on the theoretical sensitivity differences of each characteristic quantity at different stages of insulation degradation. The preferred initial weighting configuration during the initial stage of system operation is: w1=0.4, w2=0.3, w3=0.3. In this initial configuration, the first characteristic quantity, namely the proportion of high-frequency pulse energy, has the most direct and sensitive response to insulation degradation because partial discharge is the most characteristic electrical accompanying phenomenon of insulation degradation, and therefore has the highest initial weight. The second and third characteristic quantities provide supplementary information from the perspectives of voltage transients and harmonic components, respectively, and have equal initial weights.

[0066] The mathematical expression for fusion computing is: ; Where A is the normalized anomaly index, dimensionless, ranging from 0 to 1; f1, f2, and f3 are the first, second, and third normalized features, dimensionless, each ranging from 0 to 1; w1, w2, and w3 are adaptive weight coefficients, dimensionless, satisfying w1 + w2 + w3 = 1 and any weight not less than 0.1. Since all normalized features and weight coefficients are within the range of 0 to 1 and the sum of the weights is 1, the value of the fusion result A is strictly limited to between 0 and 1. The closer the A value is to 0, the healthier the insulation state; the closer it is to 1, the more severe the degradation symptoms.

[0067] The weighting coefficients are periodically updated based on the historical correction records of the dynamic baseline. This indicates that after the system verifies the authenticity of a certain type of anomaly through load disturbance in each closed-loop interaction, it records the actual performance of each feature quantity in that verification and adjusts the weight allocation accordingly. The update cycle is preferably set to 30 days or the condition that is met first among 10 completed closed-loop interactions.

[0068] The update method is as follows: Calculate the discriminative contribution of each feature in all labeled events within the past update cycle, appropriately increase the weight of features with higher discriminative contributions, and ensure that no weight is lower than 0.1 to avoid completely ignoring information of a certain dimension. The specific calculation method of discriminative contribution will be explained in detail in Example 9.

[0069] Example 4: Dynamic Baseline Initialization and Update Mechanism like Figure 4 As shown, this embodiment corresponds to the following content: "The comparison of the anomaly index with the dynamic baseline further includes: when the system is initially put into operation, initializing the dynamic baseline with the statistical mean and standard deviation of the anomaly index over multiple consecutive sliding windows under fault-free operation; the dynamic baseline includes two parameters: baseline mean and baseline fluctuation range, and the preset ratio is a preset multiple by which the anomaly index deviates from the baseline mean by the baseline fluctuation range; after receiving feedback correction from the load control steps, the dynamic baseline updates the baseline mean and the baseline fluctuation range using an exponentially weighted moving average method."

[0070] The dynamic baseline is one of the core technical features that distinguishes this invention from the fixed threshold judgment method in traditional arc flash protection devices. In traditional solutions, the arc flash criterion threshold is a fixed value or is only adjusted in simple increments based on day and night, which cannot adapt to the continuous changes in the circuit's operating state, nor can it utilize new information accumulated during operation for optimization. The dynamic baseline in this invention achieves continuous updating and adaptive optimization through a closed-loop feedback mechanism.

[0071] The initialization of the dynamic baseline is performed during the initial system commissioning. Since there is no historical operational data available at this time, the system is allowed to run continuously for an initialization period, preferably 24 hours, under confirmed fault-free operation conditions, covering the complete diurnal load fluctuation cycle of the power distribution system. During this 24-hour initialization period, the system continuously calculates anomaly indicators with a window length of 200ms and a sliding step size of 100ms, accumulating approximately 864,000 anomaly samples. The statistical mean μ0 and statistical standard deviation σ0 of these samples are calculated, with μ0 used as the initial value for the baseline mean parameter μ and σ0 used as the initial value for the baseline fluctuation range parameter σ.

[0072] The triggering condition for exception detection is: ; Where A is the normalized anomaly index for the current time window, dimensionless; μ is the baseline mean parameter of the dynamic baseline, dimensionless; σ is the baseline fluctuation range parameter of the dynamic baseline, dimensionless; and k is a preset multiple, dimensionless. k=3 is preferred, derived from the 3σ criterion in statistical process control: under the assumption of normal distribution, the probability of a signal deviating from the mean by more than 3 standard deviations is only 0.27%, meaning that under normal operating conditions, on average only 1 out of every 370 windows may produce a false trigger. Calculated with a 200ms window and a 100ms step size, an average of approximately 1 false trigger candidate is generated every 37 seconds. However, since subsequent multi-parameter correlation analysis in step S2 performs a secondary judgment (which must simultaneously meet conditions such as duration thresholds), this frequency will not lead to actual false alarms.

[0073] To prevent the dynamic baseline from shrinking excessively when the system is running stably, thus causing the system to become overly sensitive, a lower limit value σ for the baseline fluctuation range needs to be set. min Preferred σ min The standard deviation σ0 during the initialization phase is set to 50%, meaning σ must never fall below half of the initial fluctuation range. When the closed-loop feedback correction attempts to reduce σ to σ0... min The system will clamp σ at σ when the following conditions are met. min The purpose of this lower limit setting is to ensure that even if the equipment operates extremely smoothly, reducing the normal fluctuation range, the system still maintains a minimum tolerance for abnormal signals, avoiding alarms for any tiny fluctuations due to excessively small σ values.

[0074] After receiving feedback corrections from load regulation steps, the dynamic baseline is updated using an exponentially weighted moving average method. ; Where, μ new The updated baseline mean is dimensionless; μ old The baseline mean before the update is dimensionless; A feedback The steady-state value of the anomaly degree fed back from the load control steps is dimensionless; α is the exponential weighting coefficient, also dimensionless. α is preferably 0.15. The reason for this selection is that α = 0.15 means that after each update, new data contributes 15% of the weight, historical data contributes 85%, and after approximately 4 to 5 closed-loop interactions, the cumulative contribution of old data drops below 50%. The system's memory half-life is approximately 4 to 5 closed-loop interactions.

[0075] For a typical power distribution system, if 1 to 2 warning-level events occur per month, the system needs approximately 4 to 6 months to update the baseline data. This rate matches the typical timescale of insulation aging in power distribution equipment. If the value of α is too large (e.g., >0.5), a single feedback will have too great an impact on the baseline, potentially causing drastic baseline jumps due to individual abnormal feedback values. If the value of α is too small (e.g., <0.05), the baseline update will be too slow, making it difficult to track normal drift in equipment status. The baseline fluctuation range parameter σ is updated using a similar formula, and its α value can share the same setpoint as the baseline mean.

[0076] This closed-loop feedback-based dynamic update mechanism enables the dynamic baseline to continuously self-calibrate during system operation. Each execution of preventative load regulation and observation of its electrical response provides new calibration information for the dynamic baseline, gradually bringing it closer to the precise setting of the actual insulation state of the circuit. This is the core adaptive capability resulting from the reverse coupling of step S3 to step S1.

[0077] Example 5: Risk Level Determination Logic of Multi-parameter Correlation Analysis like Figure 5 As shown, this embodiment corresponds to the following content: "Perform multi-parameter correlation analysis to assess the fault risk level, including: obtaining the numerical value of the anomaly index, the frequency domain concentration of the abnormal features in the feature vector, the duration of the anomaly, and the trend slope of the anomaly over time; inputting the numerical value, the frequency domain concentration, the duration of the anomaly, and the trend slope into a multi-parameter correlation analysis model for comprehensive evaluation; when the anomaly index exceeds the baseline fluctuation range but the trend slope is stable or decreasing, it is determined to be at the attention level; when the anomaly index continues to rise and the frequency domain concentration is higher than a preset concentration threshold, it is determined to be at the warning level; when the anomaly index rises sharply or the absolute value exceeds the safety limit, it is determined to be at the emergency level."

[0078] The core of multi-parameter correlation analysis is to avoid simple threshold judgments based on a single indicator value, and instead comprehensively assess the severity and urgency of fault risks from multiple dimensions. This multi-dimensional assessment method can significantly reduce the risk of misjudgment caused by fluctuations in a single indicator, while improving the accuracy of identifying real fault risks.

[0079] The definitions and calculation methods for the four evaluation dimensions are as follows: 1) The value of the anomaly index is the current value of the normalized anomaly index A output in step S1. The value range is from 0 to 1, reflecting the intensity level of the current insulation deterioration symptoms.

[0080] 2) Frequency domain concentration describes the degree of concentration of anomalous features in the frequency domain. Specifically, it is calculated as follows: In the frequency domain decomposition results of step S1, the high-frequency band above 500Hz is divided into several equal-width frequency bands (each band's bandwidth is 5 times the frequency resolution, i.e., 25Hz). The ratio of the energy in the single frequency band with the highest energy concentration to the total energy of the high-frequency band is calculated. This ratio is denoted as C, and its value ranges from 0 to 1.

[0081] A higher C-value indicates that the abnormal energy is more concentrated in a specific frequency band and has a more definite physical source. A preferred preset concentration threshold is 0.6, meaning that when more than 60% of the abnormal energy is concentrated in a narrow frequency band, it is considered an anomaly with a definite physical source. This threshold is based on the fact that the frequency domain concentration of random noise and broadband interference is typically below 0.3, while the frequency domain concentration of partial discharge signals with a deterministic physical source is typically above 0.6.

[0082] 3) The duration of the anomaly is denoted as T, and its calculation starts from the first triggering of the anomaly marker (i.e., The timer continues from the moment the anomaly is first established until the anomaly flag is removed or the risk level assessment is completed. Brief anomalies (lasting less than 10 windows, i.e., 1 second) may originate from external transient disturbances or load surges; the system automatically ignores these and does not include them in the risk assessment process. Only anomalies whose duration exceeds the set minimum threshold T are considered valid. min Anomalies that last for at least 60 seconds are considered to be of evaluable significance.

[0083] 4) The trend slope, denoted as S, is obtained by performing a least-squares linear fit on the anomaly index sequence over the most recent 100 consecutive time windows (i.e., 10 seconds). The fitting result is the slope S in A(t) = S·t + b, with units of 1 / window. A positive slope indicates that the anomaly is worsening, while a negative slope indicates that the anomaly is mitigating. The preferred stability threshold is S. th =0.001 / window, absolute value less than S th The slope is considered to be stationary.

[0084] The determination of the three risk levels is based on the following conditional logic: 1) The trigger condition for the attention level is: the anomaly degree meets the trigger condition ( (The condition is met) and the duration of the anomaly T exceeds the minimum threshold T. min Meanwhile, the trend slope S satisfies S≤S th (That is, the slope is flat or decreasing). This can be represented by the following logical expression:

[0085] Triggering exception AND T>T min AND S≤S th → Attention Level This status indicates that although the anomaly exists, it is not continuously worsening. The system switches to enhanced monitoring mode, shortening the sliding window step size to 50% of the normal value to obtain more refined status tracking information.

[0086] 2) The triggering conditions for the early warning level are: the anomaly degree meets the triggering conditions and the anomaly duration T exceeds the minimum threshold T. min Meanwhile, the trend slope S is greater than the stationarity threshold S. th (i.e., the slope is positive and continuously increasing), and the frequency domain concentration C is higher than the preset concentration threshold C. th In logical terms, it can be represented as:

[0087] Triggering exception AND T>T min AND S>S th AND C>C th → Warning level This state indicates that an insulation degradation process with a clear physical source is accelerating, triggering preventative load regulation in step S3.

[0088] Additionally, for anomalies that meet the triggering condition and T>T min S>S th But C≤C th The situation indicates that although the anomaly is increasing, there is a lack of a clear and concentrated physical source. The system is treated as a matter of concern and maintains enhanced monitoring mode. The situation will be reassessed after the frequency domain concentration or trend further evolves.

[0089] The emergency level is triggered by one of two conditions: one is that the trend slope S increases sharply within a short period of time (S exceeds 5·S). th That is, if the absolute value of the slope exceeds 5 times the normal alarm slope, it indicates that the anomaly is deteriorating rapidly; secondly, if the absolute value of the anomaly index A exceeds the safety limit A. max (Option A) max =0.85, meaning the normalized anomaly rate reaches 85% of full scale, indicating that the insulation condition is close to the danger boundary. This can be expressed logically as:

[0090] (S>5·S) th OR (A>A) max → Emergency Level Emergency protection directly triggers rapid protection actions such as circuit breaker tripping, without requiring load control verification.

[0091] The execution order of the above decision-making logic is as follows: the system first checks the emergency level conditions; if they are met, emergency protection is immediately executed, and other decisions are skipped; if the emergency level conditions are not met, the warning level conditions are checked; if the warning level conditions are also not met, the attention level conditions are checked. This descending priority decision-making order ensures that the most urgent risks receive the fastest response.

[0092] Example 6: Load Transfer Priority Sequence and Execution Flow like Figure 6 As shown, this embodiment corresponds to the following content: "Transferring a portion of the load from the target circuit to adjacent redundant circuits according to priority includes: establishing a load transfer priority sequence based on the importance level of the loads connected to the target circuit, with lower priority loads transferred first; querying the current load rate of adjacent redundant circuits and selecting circuits with sufficient margin as load transfer targets; performing load switching operations sequentially according to the priority sequence, waiting for a preset stabilization time after each switch before performing the next switch, until the current of the target circuit drops below a preset safety margin."

[0093] Load transfer is a specific operational means to achieve preventative fault elimination. Its core objective is to reduce the current-induced thermal effect and insulation electric field stress of the target circuit without affecting the power supply reliability of the distribution system. The thermal effect of current is proportional to the square of the current; reducing the circuit current can effectively alleviate the thermal load on conductors and insulation materials, and slow down the insulation degradation process caused by heat accumulation.

[0094] This method is applicable to fault-healing intelligent integrated distribution boxes. Its physical topology requires that each outgoing circuit within the distribution box has an electrical connection for load transfer. Typical implementations include inter-circuit tie switches, dual-power switching devices, or backup busbar tie circuit breakers, allowing some load branches on a single circuit to be transferred to adjacent circuits via a switching switch. This premise is a fundamental structural feature distinguishing fault-healing distribution boxes from ordinary distribution boxes. In distribution boxes lacking inter-circuit tie channels, steps S1 to S2 (feature extraction and risk assessment) of this method can still be executed normally and provide early warning information; however, the load transfer operation in step S3 cannot be executed automatically and requires manual intervention by maintenance personnel.

[0095] The load transfer priority sequence is established based on the importance level classification of the loads connected to the target circuit. In a power distribution system, different electrical equipment has different requirements for power supply continuity, and can be divided into multiple levels according to importance. Low-priority loads are usually auxiliary equipment, non-critical lighting, and other electrical equipment that can be temporarily interrupted or transferred; transferring them out first will not affect the critical functions of the system. High-priority loads are usually critical production equipment, safety systems, and other electrical equipment that should not be frequently switched; they are only transferred if the safety margin requirements cannot be met even after all low-priority loads have been transferred out. The priority sequence is set by maintenance personnel based on the actual list of electrical equipment during the initial system configuration and stored in the system parameter database.

[0096] The selection of adjacent redundant circuits requires querying their current load rates to ensure sufficient capacity margin after accepting additional loads. The load rate is calculated as the ratio of the current actual current of the circuit to its rated capacity current. The adjacent circuit with the lowest load rate is selected as the priority for accepting the transferred load. The expected load rate of the circuit after accepting the transferred load must not exceed the set upper limit (preferably 80% of the rated capacity) to avoid the circuit itself entering the overload risk zone. When the load rates of all adjacent circuits exceed the upper limit, the system determines that there are no available circuits to accept the transferred load. At this time, the load transfer operation is skipped, the status is marked as transfer blocked, and the system remains in enhanced monitoring mode, awaiting manual intervention from maintenance personnel.

[0097] The step-by-step switching strategy is a key detail in the load transfer process. Only one load unit is switched at a time, and after switching, a preset settling time is allowed for the system to enter a new steady state. The preset settling time is preferably 30 seconds. This value is chosen because the electromechanical transient time constant of a typical inductive load in a power distribution system is usually within 10 seconds, and a 30-second waiting time ensures the system fully enters a steady state. After the waiting period, the current in the current loop is assessed to see if it has decreased below a preset safety margin. If not, the next load unit is switched. The preset safety margin is preferably 70% of the rated capacity; that is, when the target loop current decreases to below 70% of the rated current, the safety requirement is met, and load transfer is stopped.

[0098] This step-by-step execution method has two functional characteristics: 1) Avoid switching a large amount of load at once, which can cause severe transient processes in the system and generate electrical disturbances that may be misjudged as fault characteristics; 2) It provides progressive load change information for the observation in step S1, which is beneficial for accurately analyzing the correspondence between the degree of load reduction and the change in anomaly, and improving the information quality of closed-loop verification. In addition, each switching operation is performed near the zero-crossing point of the load current, which utilizes the zero-current condition at the current zero crossing to minimize the arc energy and transient impact during the switching process.

[0099] Example 7: Observation Procedure for Electrical Characteristic Response After Load Transfer like Figure 7 As shown, this embodiment corresponds to the following content: "Continuously observe the changes in the electrical characteristic response of the target circuit after load transfer, including: starting an observation window after load transfer is completed, continuously calculating the anomaly index of the target circuit within the observation window in the same manner as feature extraction; recording the trajectory of the anomaly index over time within the observation window, and calculating the anomaly decay rate or growth rate; and outputting the trajectory and the decay rate or growth rate as the changes in the electrical characteristic response."

[0100] Among them, electrical characteristic response observation is a key execution link in realizing the "active detection" function in the cross-coupled closed-loop linkage structure. Its technical principle is similar to the "applying a known input and observing the system output" method in system identification theory: the load transfer operation is equivalent to applying a known and controllable disturbance (current reduction) to the power circuit. By observing how the electrical characteristics of the circuit respond to this disturbance, the true state of the insulation system inside the circuit can be inferred.

[0101] The observation window is initiated after the load transfer is completed, starting at the moment after the last load switching operation has been completed and a preset stabilization time (30 seconds) has elapsed. The preferred duration of the observation window is 30 minutes. This value is chosen because the thermal response time constant of insulating materials after current reduction is typically on the order of minutes to ten minutes. A 30-minute observation period can cover 2 to 3 thermal time constants, sufficient to observe the complete process of the thermal effect anomaly index decaying from its initial high value to its steady-state value. The duration of the observation window can also be adaptively adjusted based on whether the anomaly change curve has entered a steady state during actual observation. That is, if the anomaly change amplitude is less than 0.01 within 5 consecutive minutes, the observation window can be terminated early.

[0102] Within the observation window, the system continuously calculates the anomaly index of the target loop in the same way as the feature extraction in step S1 (same sampling frequency 10kHz, same window length 200ms, same feature extraction algorithm and normalization parameters), ensuring that the response observation data and the data when the anomaly marker is triggered are completely consistent in processing method and are comparable.

[0103] The anomaly rate change trajectory is recorded, with time on the horizontal axis and anomaly index values ​​on the vertical axis, forming a curve describing the evolution of anomaly rate over time. Key feature parameters are extracted from this curve: the decay rate describes the speed at which the anomaly rate decreases, and it is calculated by performing an exponential decay fit on the anomaly index sequence within the observation window. The fitting formula is:

[0104] ; Where A(t) is the anomaly index value at time t, which is dimensionless; A0 is the initial anomaly value at the start of the observation window, which is dimensionless; A ss The constant is the steady-state value after decay, dimensionless; τ is the decay time constant in minutes; t is the elapsed time from the start of the observation window in minutes. The decay rate is quantified by the reciprocal of the time constant τ, 1 / τ, with a larger 1 / τ indicating faster decay. The growth rate describes the speed at which the anomaly index increases. When the anomaly index does not show a decay trend within the observation window, it is quantified by obtaining the slope Sobs through linear regression of the anomaly index sequence. A positive slope indicates that the anomaly index continues to grow after the load decreases.

[0105] The decay rate and growth rate, as quantitative outputs of response changes, provide a basis for subsequent fault nature determination. A faster decay rate (τ < 10 minutes) means that the anomalous source is highly sensitive to load changes, supporting the determination of heat accumulation type faults; a sustained growth rate (positive Sobs) means that the anomalous source is independent of the load level or even deteriorates autonomously over time, supporting the determination of irreversible insulation damage type faults.

[0106] A complete record of the change trajectory can also indicate more complex response patterns, such as a "bouncing" pattern that first briefly decays and then rises again. In this pattern, the decay phase reflects the fading of thermal effects, while the subsequent rise phase reflects the continued development of underlying insulation damage. The system labels this composite pattern as an insulation damage type to ensure safe-side handling.

[0107] Example 8: Dynamic Baseline Feedback Correction and Event Classification Marking Logic like Figure 8 As shown, this embodiment corresponds to the following content: "The feedback of the response change to the dynamic baseline for correction includes: if the response change indicates that the anomaly index shows a decreasing trend after the load decreases, then the current closed-loop interaction is marked as a heat accumulation type event, and the baseline mean of the dynamic baseline is updated with the decreased steady-state value of the anomaly; if the response change indicates that the anomaly index does not improve or shows an increasing trend after the load decreases, then the current closed-loop interaction is marked as an insulation damage type event, and the baseline fluctuation range of the dynamic baseline is lowered to improve the sensitivity of subsequent detection; the marking results of each closed-loop interaction and the corresponding baseline correction amount are stored as learning samples."

[0108] The dynamic baseline feedback correction is a specific implementation method for the reverse coupling from step S3 to step S1, and it also reflects the adaptive learning capability in the entire cross-coupled closed-loop linkage structure. The result of each closed-loop interaction is transformed into calibration information for the anomaly detection model, enabling the system to continuously accumulate experience during operation.

[0109] The criterion for identifying heat accumulation events is that the anomaly index shows a clear decay trend within the observation window. Specifically, the criterion is the steady-state value A of the anomaly index at the end of the observation window. ss The anomaly is below 70% of the initial anomaly value A0 at the start of the observation window (i.e., the attenuation exceeds 30%), and the exponential fitting coefficient of determination R for the attenuation process is... 2 A value >0.8 indicates that the decay curve conforms to the exponential decay model.

[0110] When the above conditions are met, this event is labeled as heat accumulation type. The physical mechanism of heat accumulation is as follows: when the loop current continues to operate at a high level, the temperature of the conductor and insulating materials rises due to the Joule heating effect and dielectric loss. The increase in temperature leads to a decrease in the dielectric properties of the insulating material and a reduction in the partial discharge threshold, which is manifested in the gradual increase of the anomaly index in electrical parameters. When the load is transferred and the current decreases, the thermal effect weakens, the temperature of the insulating material gradually drops, its dielectric properties recover, and the anomaly index correspondingly decays to a lower level. The steady-state value A after decay. ss This reflects the "healthy baseline" level of the circuit after load reduction. When using it to update the baseline mean parameter μ of the dynamic baseline, an exponentially weighted moving average formula is used for smoothing the update. .

[0111] The criteria for classifying an insulation damage event are that the anomaly index does not show a decreasing trend or shows an increasing trend within the observation window. Specifically, the criteria are: the anomaly value at the end of the observation window is not less than 90% of the initial value (i.e., the decrease is less than 10%), or the linear regression slope (Sobs) of the anomaly within the observation window is positive. If either of these conditions is met, the event is classified as an insulation damage event.

[0112] The physical mechanism of irreversible insulation damage is that the insulating material has undergone permanent structural changes such as carbonized channels, dendritic discharge traces, or mechanical cracks. Once these damages are formed, they will not disappear on their own due to a reduction in load, and may even continue to deteriorate as the damage points continue to develop under the influence of an electric field.

[0113] For such events, the system needs to reduce the baseline fluctuation range parameter σ of the dynamic baseline. The preferred reduction is 80% of the current σ value, i.e., updating σ to 0.8·σ, but it must not be lower than σ. min The purpose of tightening the anomaly detection threshold is to make the system more alert to any minor anomalies that subsequently occur in the circuit, because once irreversible insulation damage is confirmed, the probability of the circuit developing into a real fault increases significantly.

[0114] For the intermediate case where the attenuation is between 10% and 30%, and for the case where the attenuation exceeds 30% but the exponential fitting coefficient of determination R is still high... 2 For cases ≤0.8 (where the decay trajectory does not conform to the exponential model), the system marks them as undetermined events and does not correct the dynamic baseline, but still records them as learning samples for subsequent statistical analysis in the adaptive evolution step.

[0115] The storage of learning samples archives complete information from each closed-loop interaction, including: the feature vector (normalized values ​​of f1, f2, and f3) at the time of triggering the anomaly, the anomaly degree value A at the time of triggering, the frequency domain concentration C and trend slope S at the time of triggering, the specific operation record of load transfer (which loads were transferred, and the current values ​​before and after the transfer), the complete anomaly degree change trajectory within the observation window, the fitting parameters (τ or Sobs), the event type label, and the specific correction amount applied to the dynamic baseline. These learning samples provide the data foundation for subsequent adaptive baseline evolution steps.

[0116] Example 9: Adaptive Baseline Evolution and Knowledge Accumulation Mechanism like Figure 10 As shown, this embodiment corresponds to the following content: "It also includes an adaptive baseline evolution step: during long-term system operation, all stored learning samples are statistically analyzed at preset intervals to identify the characteristic distribution patterns of thermal accumulation events and insulation damage events; the judgment boundaries of the attention level, early warning level, and emergency level in the multi-parameter correlation analysis model are adjusted according to the characteristic distribution patterns; and the weight allocation of each feature quantity in the adaptive weight coefficient is optimized based on the characteristic differences of different types of events in the accumulated learning samples, so that the system's ability to distinguish different types of fault precursors continuously improves with operating time."

[0117] The adaptive baseline evolution step is an optimization step in the entire cross-coupled closed-loop linkage structure. It operates on a longer timescale than a single closed-loop interaction, utilizing the learning samples accumulated from multiple closed-loop interactions for statistical analysis and model optimization. The dynamic baseline correction in Example 8 is online event calibration, while the adaptive evolution in Example 9 is batch offline optimization. Together, they constitute the system's continuous learning capability mechanism.

[0118] The preset statistical period is preferably set to 30 days, or it can be triggered when the cumulative number of new learning samples reaches 10, whichever comes first. This setting ensures that the system updates the model in a timely manner when there are sufficient samples, while avoiding prolonged periods of inactivity during stable operation with few events. Each time a statistical period is triggered, the system performs a full statistical analysis on all historical learning samples (not limited to new samples in the current period).

[0119] The statistical analysis includes the statistical distribution characteristics of heat accumulation events and insulation damage events across various feature dimensions, specifically the mean and standard deviation of the first feature quantity f1, the second feature quantity f2, and the third feature quantity f3 when the two types of events are triggered. The contribution of each feature quantity to event type classification is quantified by calculating the class discrimination ratio of the two types of events across each feature dimension. The discrimination ratio is calculated using the Fisher discriminant ratio method, where the Fisher discriminant ratio F for the i-th feature quantity is... i Defined as:

[0120] ; Among them, F i Fisher's discrimination ratio for the i-th feature is dimensionless. The larger the value, the stronger the ability of the feature to distinguish between the two types of events. The mean of the normalized values ​​of the i-th feature quantity in all samples of thermal accumulation events is dimensionless. The mean of the normalized values ​​of the i-th feature quantity in all insulation damage event samples is dimensionless. denoted as the variance of the normalized value of the i-th feature quantity in the thermal accumulation event sample, which is dimensionless; F represents the variance of the normalized value of the i-th feature in the insulation damage event sample, which is dimensionless. The physical meaning of the Fisher discriminant ratio is: the numerator reflects the difference in mean between the two classes of events along this feature dimension (inter-class distance), and the denominator reflects the dispersion of each class of events along this dimension (intra-class dispersion). i The larger the value, the more effectively the feature can distinguish between the two types of events.

[0121] The specific method for adjusting the risk level determination boundary based on the characteristic distribution law is as follows: when statistical analysis reveals that the mean of the frequency domain concentration of insulation damage events is significantly higher than that of heat accumulation events, the boundary between the two can be used as the preset concentration threshold C. th The basis for updating.

[0122] The specific update method is to take the arithmetic mean of the frequency domain concentrations of the two types of events as the new C. th For example, if the mean frequency domain concentration of accumulated insulation damage samples is 0.75 and the mean frequency domain concentration of thermal accumulation samples is 0.42, then the updated C th =(0.75+0.42) / 2=0.585. Similarly, the stationarity threshold S of the trend slope... th The slope can also be adjusted accordingly based on the statistical distribution of the slope before the two types of events are triggered.

[0123] The method for optimizing the adaptive weight coefficients is as follows: weights are redistributed based on the Fisher discriminant ratios of the three features. The new weights are proportional to the Fisher discriminant ratios, and the calculation method is as follows:

[0124] ; in, Let w1 be the dimensionless weight coefficient after updating the i-th feature. This formula automatically satisfies the constraint w1 + w2 + w3 = 1. If any calculated weight is lower than 0.1, it is clamped to 0.1, and the remaining weights are scaled proportionally to maintain a sum of 1. This ensures that even if a feature has low discriminative power in the current sample set, it will not be completely ignored, preserving its potential to play a role in different types of fault scenarios in the future.

[0125] This step endows the entire system with the technical characteristic of becoming more accurate with use: as the running time increases and the number of closed-loop interactions increases, the system accumulates richer learning samples, and its ability to identify and classify fault precursors becomes more and more accurate. This is the core advantage brought about by the cross-coupled closed-loop linkage structure, and traditional detection schemes based on fixed hardware parameters do not possess this evolutionary capability.

[0126] Example 10: Module Structure and Inter-unit Information Interaction of an Arc Fault Prediction System like Figure 9 As shown, this embodiment corresponds to the following content: "A fault-self-healing intelligent integrated distribution box arc fault prediction system, which adopts a fault-self-healing intelligent integrated distribution box arc fault prediction method, includes: an electrical feature extraction unit, which uses existing current transformers and voltage transformers in the distribution box to collect electrical signals of each circuit, performs sliding window feature extraction on the electrical signals to calculate a normalized anomaly index, compares the anomaly index with a dynamic baseline, and outputs an anomaly marker and corresponding feature vector when the deviation reaches a preset proportion; the transmission bandwidth of the current transformer and voltage transformer should not be less than 5kHz; and a risk assessment unit, which is used to, based on the anomaly marker and the feature vector, combine the target..." The historical operating data of the circuit is subjected to multi-parameter correlation analysis to assess the fault risk level; a load regulation and verification unit is used to transfer part of the load of the target circuit to an adjacent redundant circuit according to priority when the fault risk level reaches the warning level, continuously observe the changes in the electrical characteristic response of the target circuit after the load transfer, and feed the response changes back to the electrical characteristic extraction unit to correct the dynamic baseline, so that the electrical characteristic extraction unit, the risk assessment unit and the load regulation and verification unit form a closed-loop linkage; a fault determination and execution unit is used to determine the fault nature based on the response changes and execute the corresponding control strategy.

[0127] This system consists of four functional units, which communicate and collaborate with each other through defined data interfaces, forming a cross-coupled closed-loop linkage architecture.

[0128] The electrical feature extraction unit is the system's sensing front end. Its hardware foundation consists of existing current and voltage transformers in the distribution box, and an embedded signal processing board. This board's functions include: performing continuous analog-to-digital conversion on the current and voltage signals of each circuit at a 10kHz sampling frequency; dividing the circuit into sliding windows with a 200ms window length and a 100ms step size; and extracting three feature quantities within each window using a fast Fourier transform and transient detection algorithm, based on V²=N. sag ·D max The original comprehensive value of each feature quantity is calculated using V3=max(G3,G5,G7). After min-max normalization, f1, f2, and f3 are obtained. Weighted fusion is then performed to obtain the anomaly index A, which is compared in real-time with the dynamic baseline parameters stored in onboard memory. This unit also maintains dynamic baseline parameter data (μ and σ) and receives feedback from the load control and verification unit to correct and update the baseline. The output interface of this unit transmits anomaly markers and feature vectors (containing f1, f2, f3, and frequency domain concentration C) to the risk assessment unit, while the input interface receives response change data (A) from the load control and verification unit. ss (or Sobs). No dedicated arc light sensor or optical detection hardware is required.

[0129] The risk assessment unit is the system's decision-making center. It receives anomaly markers and feature vectors from the electrical feature extraction unit, combines them with stored historical operational data to perform multi-parameter correlation analysis, and outputs a three-level risk level judgment. This unit maintains the parameters of the multi-parameter correlation analysis model, including the frequency domain concentration threshold C. th Trend slope stability threshold S th Safety Limit A max These parameters undergo periodic optimization during the adaptive baseline evolution process. The decision logic strictly follows the conditional expressions defined in Example 5, with a priority order of emergency level → early warning level → concern level descending checks. The output of this unit triggers different subsequent actions based on the risk level: concern level feedback to the electrical feature extraction unit requires shortening the sampling window step size; early warning level triggers the load regulation and verification unit to initiate load transfer; emergency level triggers the fault determination and execution unit to directly execute protection actions.

[0130] The load control and verification unit is the execution and verification hub of the system, and also the implementation carrier of reverse coupling in the closed-loop structure. After receiving the early warning trigger signal from the risk assessment unit, this unit queries the load priority database and the load rate information of adjacent circuits, formulates and executes a preventive load transfer plan, and after the transfer is completed, initiates a 30-minute observation window to continuously monitor the electrical characteristic response of the target circuit. The key feature of this unit is its bidirectional data interaction capability: downward, it controls the load switching device (tethering switch or contactor) to perform physical operations through the communication interface, and upward, it transmits the observed response change data (including the decay time constant τ or growth slope Sobs, steady-state value A) to the system. ss (etc.) Feedback is sent to the electrical feature extraction unit for dynamic baseline correction. This reverse information flow is the technical essence of the cross-coupling structure, enabling the entire system to transform from open-loop passive detection to closed-loop active verification and adaptive learning. When all adjacent loops are unable to handle the load, the unit outputs a "transfer blocked" flag and notifies the operation and maintenance system to request manual intervention.

[0131] The fault diagnosis and execution unit is the final decision-making and execution endpoint of the system. Based on the response change data output by the load regulation and verification unit, it determines the nature of the fault and executes corresponding control strategies. The judgment criterion is: the attenuation exceeds 30% and R... 2 An attenuation value >0.8 indicates a thermal accumulation event, while an attenuation of less than 10% or a positive observation slope indicates an insulation damage event. For thermal accumulation events, the unit issues a command to maintain the current load distribution and sends maintenance suggestions to the upper-level maintenance system via the communication interface. For insulation damage events, the unit issues an emergency protection command to trigger the circuit breaker to trip and disconnect the faulty circuit. The unit's trip output signal is logically ORed with the existing overcurrent protection trip signal in the distribution box and connected to the circuit breaker's trip coil. This means that either signal trigger can independently complete the tripping action, ensuring the independence of this predictive system from conventional protection and their mutual backup relationship. The unit is also responsible for recording complete information for each fault determination and storing it as a learning sample, providing annotation information for the system's long-term learning.

[0132] The information flow between the four units forms a complete closed-loop path: electrical feature extraction unit → risk assessment unit → load control and verification unit → electrical feature extraction unit. The reverse information flow from the load control and verification unit to the electrical feature extraction unit constitutes the core adaptive mechanism of the system.

[0133] Example 11: Verification Experiment like Figure 11 As shown in Example 11, to verify the effectiveness of the method proposed in this invention, the following comparative verification experiment was designed.

[0134] 1) Verification Scenario and Comparison Scheme The verification scenario selected a typical 10kV switchgear distribution system, containing 8 outgoing circuits interconnected by tie switches, capable of load transfer, with an operating load rate of 60% to 80% of the rated capacity. During a 180-day continuous operation test, three typical fault simulation scenarios were set up: a thermal accumulation insulation stress scenario, simulating cyclic aging of insulation temperature by applying periodic overloads of 110% to 120% of the rated current to a specific circuit (each lasting 4 hours before returning to normal); an insulation damage scenario, simulating irreversible insulation degradation by implanting air gap defects of known size at cable joints; and an environmental interference scenario, testing false alarms by simulating cabinet door opening operation lights, adjacent operation flashes, and changes in distribution room lighting.

[0135] A total of 15 simulated fault scenarios were set up during the test, including 8 thermal accumulation type and 7 insulation damage type. The key technical indicators were defined as: early warning hit rate (number of fault scenarios in which the system correctly issued an early warning / total number of fault scenarios), false alarm count (cumulative number of times the system incorrectly issued an early warning or protection action when there was no actual fault or fault precursor), functional coverage (percentage of fault scenario types that the system could detect), and early warning lead time (the time interval from the system first issuing an early warning to the simulated fault condition reaching the breakdown threshold).

[0136] Two comparative schemes were set up and run in parallel under the same test conditions. Scheme A is the existing technical solution, which adopts a linear process of traditional arc light sensor array with dedicated main control unit. It includes 8 point arc light probes covering each cabinet space, fiber optic transmission to main control unit for fixed threshold judgment, and supplemented by current over-limit joint criterion for secondary verification. This scheme is installed and debugged according to the manufacturer's preferred standard.

[0137] Scheme B is the scheme of this invention, which adopts a method of cross-coupling closed-loop linkage between electrical feature extraction, risk assessment and preventive load control, and utilizes the existing current transformer and voltage transformer signals of each circuit without using arc sensor hardware.

[0138] The complete parameter configuration used in Scheme B is as follows: sampling frequency 10kHz; sliding window length 200ms (10 power frequency cycles); sliding step size 100ms; high-frequency pulse cutoff frequency 500Hz; voltage descent detection threshold is 5% of the nominal amplitude; the second feature quantity synthesis method is V2=N sag ·D max The specific harmonics of interest are the 3rd, 5th, and 7th harmonics; the third characteristic quantity is synthesized using V3=max(G3,G5,G7); the normalization method is minimum-maximum normalization, with the maximum value being 3 times the upper limit of the normal range; the initial weighting coefficients are w1=0.4, w2=0.3, w3=0.3; the lower limit of the weight is 0.1; the preset multiple is k=3; the exponential weighting coefficient is α=0.15; and the lower limit of the baseline fluctuation range is σ.min =0.5·σ0; Preset concentration threshold C th =0.6; Safety limit A max =0.85; Trend slope plateau threshold S th =0.001 / window; Emergency slope threshold 5·s th =0.005 / window; minimum duration threshold T for anomalies min =60 seconds; Maximum load rate of the receiving circuit: 80%; Load switching stabilization waiting time: 30 seconds; Safety margin of the target circuit: 70% of the rated current; Response observation window: 30 minutes; Thermal accumulation type judgment attenuation threshold: 30% and R 2 >0.8; Insulation damage type judgment attenuation is less than the threshold of 10% or the observation slope is positive; σ down-adjustment coefficient is 0.8; Adaptive evolution period is 30 days or 10 new samples are accumulated; Weight update method is Fisher discriminant ratio normalization allocation; Initialization time is 24 hours.

[0139] result: Option A only has the function of passive detection and rapid removal after the arc fault occurs. Its function is positioned as a rapid response after the fact. It cannot issue an early warning before the arc occurs, nor can it prevent the elimination of potential fault hazards by adjusting the operation mode.

[0140] Solution B, through a cross-coupled closed-loop linkage structure, adds three capabilities to the functionality of Solution A: The first is the ability to identify and classify early signs of faults, enabling the system to detect abnormal trends in the early stages of insulation degradation.

[0141] The second is the ability to prevent load regulation and fault elimination, enabling the system to eliminate potential faults before arcing occurs by actively adjusting its operating mode.

[0142] The third is the proactive verification capability of fault characteristics based on closed-loop feedback, which enables the system to confirm the correctness of risk judgment through the strategy of "applying load disturbance and observing response", and to distinguish between heat accumulation type reversible anomalies and insulation damage type irreversible anomalies.

[0143] It should be noted that the rapid post-event response function of Solution A and the pre-event prediction and prevention function of Solution B are at different technical levels. Solution A detects the optical signal after the arc has occurred, while Solution B detects the electrical precursor signals before the arc has occurred. They address different physical phenomena at different stages, therefore their performance indicators should not be directly compared on the same level. Solution B eliminates the arc sensor while retaining the existing overcurrent protection in the distribution box as a backup.

[0144] Comparison of effects: (1) Early warning capability: Option A is designed for post-event detection and does not have pre-event warning capabilities; therefore, it lacks a pre-warning time indicator.

[0145] Solution B provides an average warning lead time of 3.8 hours for thermal accumulation faults (average across 8 scenarios, ranging from 1.2 to 7.5 hours) and an average warning lead time of 28.6 hours for insulation damage faults (average across 7 scenarios, ranging from 8.3 to 52 hours). Warning lead time is a new capability dimension of Solution B, enabling the system to issue warnings several hours to tens of hours before fault conditions reach the breakdown threshold, providing maintenance personnel with ample time to intervene.

[0146] (2) Detection function coverage: The effective optical sensing coverage of the eight point probes in Scheme A within the cabinet is related to the sensor layout. In this test configuration, spatial occlusion analysis confirmed that two simulated fault points were located at the edge of the probe's field of view. In 15 simulated fault scenarios, Scheme A was able to detect normally in 13 scenarios (the remaining two were due to the arc not yet being generated and the fault location being in the edge area of ​​the probe), with a functional coverage rate of 86.7%.

[0147] Solution B is based on the overall electrical signal analysis of the circuit, and is not limited by the sensor viewpoint or spatial obstruction. It successfully triggered the anomaly marker in all 15 simulated fault scenarios, with a functional coverage rate of 100%.

[0148] (3) Early warning hit rate: In 13 out of 15 simulated fault scenarios, Scheme B successfully issued warnings before the fault conditions reached the breakdown threshold, achieving a warning hit rate of 86.7%. The two scenarios that failed to issue warnings were both insulation damage scenarios, with the simulated defect development speed set to a fast mode (the time interval from implantation to the breakdown threshold was less than 2 hours). At this time, the system was still in the early operational phase after baseline establishment (within the first 45 days), and the adaptive weights and decision boundaries had not yet been fully optimized. In the 11 scenarios after day 45, the warning hit rate was 100% (all 11 scenarios were successful).

[0149] It should be noted that the total sample size of the 15 scenarios is limited, and the 86.7% hit rate represents the observation results in this test rather than the theoretical limit of the method's performance.

[0150] (4) False alarm situations: During the 180-day test, Scheme A generated 12 false alarms. Among them, 7 false alarms were caused by strong light triggering the arc light criterion during cabinet door opening maintenance (these false alarms were partially filtered in the current over-limit joint criterion verification process of the main control unit, but there were still 7 false alarms that were simultaneously met due to load fluctuations during maintenance operations), 3 false alarms were caused by instantaneous flashes generated by adjacent operations, and 2 false alarms were caused by sensor sensitivity drift over time. The average daily false alarm rate was 0.067 times / day.

[0151] It should be noted that the cabinet door opening and maintenance operations were relatively frequent in this test scenario (2 to 3 times per week on average), and some maintenance operations were accompanied by load switching, which is a relatively harsh operating environment for the arc flash protection system. Solution B generated 4 false alarms during the 180-day test period. Of these, 3 false alarms occurred during the early self-learning phase in the first 15 days (during which the dynamic baseline was not yet stable, which is a normal process of system self-learning); 1 false alarm occurred between day 15 and day 45 (this false alarm triggered the load control step, and after closed-loop verification, it was confirmed to be a transient anomaly caused by normal load fluctuations, and the system automatically corrected the dynamic baseline); and there were zero false alarms after day 45. The average daily false alarm rate for the entire cycle was 0.022 times / day, a 67% reduction compared to Solution A.

[0152] (5) Adaptability: The detection performance of Scheme A remained unchanged during the 180-day test, with the judgment threshold set at the factory default value, and no automatic adjustments were made based on actual operation. The warning hit rate of Scheme B showed an increasing trend over time: 50% in the first 45 days (system learning period) (2 hits out of 4 scenarios), and 100% from day 45 to day 180 (system maturity period) (all 11 scenarios hit). The weighting coefficients evolved from the initial w1=0.4, w2=0.3, w3=0.3 at the end of 180 days to w1=0.45, w2=0.22, w3=0.33, reflecting that in this test scenario, the first characteristic quantity (high-frequency pulse energy) and the third characteristic quantity (harmonic growth) contributed more to the characterization of insulation degradation than the second characteristic quantity (voltage dip). The frequency domain concentration threshold was adjusted from the initial 0.6 to 0.55, reflecting that the system appropriately relaxed the trigger sensitivity of the warning level based on the actual event distribution. The aforementioned weight evolution and threshold adjustment are both automatically completed by the adaptive baseline evolution step through Fisher discriminant ratio calculation.

[0153] (6) Preventive effect: In 8 out of 13 successful early warning scenarios, Option B successfully eliminated heat accumulation-type fault hazards through preventative load regulation, preventing the insulation condition from deteriorating further to breakdown. In these 8 scenarios, the anomaly index decayed to an average of 45% of its initial value after load regulation, and the average decay time constant τ was 6.2 minutes, all correctly identified as heat accumulation-type events. In the remaining 5 successful early warning scenarios involving insulation damage, the system correctly identified irreversible damage characteristics through closed-loop verification (the anomaly index did not decay to below 90% of its initial value within a 30-minute observation window after load reduction) and triggered emergency protection actions. Option A lacks preventative regulation capabilities; protection clearance for all fault scenarios requires waiting for the actual occurrence of an arc before execution.

[0154] (7) Hardware costs: Option A requires eight arc flash detectors, a matching fiber optic transmission channel, a dedicated arc flash protection control unit, and installation and commissioning costs. Option B utilizes only the existing current transformer and voltage transformer signals in the distribution box, adding one embedded signal processing board and corresponding software. The additional hardware cost of Option B is approximately 20% of the dedicated hardware cost of Option A, resulting in an overall hardware cost reduction of approximately 80%.

[0155] (8) System availability: During the 180-day test, Option A required one cleaning and maintenance operation due to dust accumulation on the surface of one probe causing a decrease in sensitivity, and another inspection was required due to a loose fiber optic connection. The total maintenance downtime was approximately 3 hours, with an availability rate of 99.93%. Option B had no hardware failure events during the 180-day test, with zero system downtime and an availability rate of 100%.

[0156] Comparative verification experiments show that the present invention has significant advantages in terms of functional integrity, early warning accuracy, and long-term reliability.

[0157] Solution B achieves the pre-emptive prediction function that Solution A lacks, with an average early warning time of hours (3.8 hours for thermal accumulation type and 28.6 hours for insulation damage type), providing ample intervention time window for operation and maintenance; the functional coverage rate increases from 86.7% in Solution A to 100%, eliminating sensor blind spots; the false alarm rate decreases from an average of 0.067 times per day in Solution A to an average of 0.022 times per day in Solution B, a reduction of 67%; the system achieves a 100% early warning hit rate and zero false alarms during the mature operation period (after day 45); and preventive load control successfully eliminated potential faults in 8 thermal accumulation scenarios.

[0158] The core reason mentioned above lies in the cross-coupling closed-loop linkage mechanism adopted in this invention: the reverse coupling between the electrical feature extraction step and the preventive load control step enables the system to have active verification and adaptive learning capabilities, transforming the traditional passive optical detection into active electrical feature reasoning and closed-loop confirmation. At the same time, through the accumulation of experience from each closed-loop interaction, the system performance is continuously optimized, realizing the self-evolutionary characteristic that the detection capability continuously enhances over time.

[0159] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A method for predicting arcing faults in a self-healing intelligent integrated distribution box, characterized in that, include: The existing current transformers and voltage transformers in the distribution box are used to collect electrical signals of each circuit. The electrical signals are then subjected to sliding window feature extraction to calculate a normalized anomaly index. The anomaly index is compared with a dynamic baseline. When the deviation reaches a preset ratio, an anomaly marker and the corresponding feature vector are output. Based on the anomaly markers and feature vectors, multi-parameter correlation analysis is performed using historical operational data of the target circuit to assess the fault risk level; wherein, the fault risk level includes attention level, early warning level, and emergency level; When the fault risk level reaches the warning level, part of the load of the target circuit is transferred to the adjacent redundant circuit according to priority to reduce insulation stress. The electrical characteristic response of the target circuit after the load transfer is continuously observed, and the response change is fed back to the dynamic baseline for correction, so that feature extraction, risk assessment and load regulation form a closed loop linkage. The nature of the fault is determined based on the response change and the corresponding control strategy is executed: if the abnormal characteristics decrease as the load decreases, the current load allocation is maintained; if the abnormal characteristics do not improve or worsen, the fault risk level is raised to the emergency level and protection action is triggered. Sliding window feature extraction of the electrical signal includes: The current and voltage waveforms of each circuit are continuously sampled at a preset sampling frequency, and the sampled data are segmented in units of sliding time windows. The current waveform within each time window is decomposed in the frequency domain, and the energy ratio of the high-frequency pulse component is extracted as the first feature quantity. Transient analysis is performed on the voltage waveform within each time window, and the frequency and amplitude of voltage transient dips are extracted as the second characteristic quantity. Harmonic analysis is performed on the electrical signal within each time window, and the growth rate of the amplitude of a specific harmonic relative to the baseline period is extracted as the third characteristic quantity. Calculate the normalized outlier index, including: The first feature quantity, the second feature quantity, and the third feature quantity are respectively normalized to eliminate dimensional differences; Based on the sensitivity of each characteristic quantity to insulation degradation, an adaptive weighting coefficient is assigned to each normalized characteristic quantity. The weighted features are fused together to obtain the normalized anomaly index that characterizes the overall insulation health status within the current time window. The adaptive weighting coefficients are periodically updated based on the historical correction records of the dynamic baseline.

2. The method for predicting arc faults in a self-healing intelligent integrated distribution box according to claim 1, characterized in that, The anomaly index is compared with the dynamic baseline, and this process includes: When the system is first put into operation, the dynamic baseline is initialized with the statistical mean and standard deviation of the anomaly index of multiple consecutive sliding windows under fault-free operation. The dynamic baseline includes two parameters: the baseline mean and the baseline fluctuation range. The preset ratio is the number of times that the anomaly index deviates from the baseline mean by the baseline fluctuation range. After receiving feedback corrections from the load control steps, the dynamic baseline updates the baseline mean and the baseline fluctuation range using an exponentially weighted moving average method.

3. The method for predicting arc faults in a self-healing intelligent integrated distribution box according to claim 2, characterized in that, Perform multi-parameter correlation analysis to assess the failure risk level, including: Obtain the numerical value of the anomaly index, the frequency domain concentration of the abnormal features in the feature vector, the duration of the anomaly, and the slope of the anomaly index changing over time; The numerical value, the frequency domain concentration, the duration of the anomaly, and the trend slope are input into a multi-parameter correlation analysis model for comprehensive evaluation. When the anomaly index exceeds the baseline fluctuation range but the trend slope is stable or decreasing, it is determined to be at the attention level; when the anomaly index continues to rise and the frequency domain concentration is higher than the preset concentration threshold, it is determined to be at the warning level; when the anomaly index rises sharply or the absolute value exceeds the safety limit, it is determined to be at the emergency level.

4. The method for predicting arc faults in a self-healing intelligent integrated distribution box according to claim 3, characterized in that, Transferring a portion of the load of the target circuit to adjacent redundant circuits according to priority includes: Establish a load transfer priority sequence based on the importance level of the loads connected to the target circuit, with lower priority loads being transferred first. Query the current load rate of adjacent redundant loops and select the loop with sufficient margin as the load transfer target; The load switching operation is performed sequentially according to the priority sequence. After each switch, a preset stabilization time is waited before the next switch is performed, until the current of the target circuit is reduced to below the preset safety margin.

5. The method for predicting arc faults in a self-healing intelligent integrated distribution box according to claim 4, characterized in that, Continuously observe the changes in the electrical characteristic response of the target circuit after load transfer, including: After the load transfer is completed, an observation window is started, and within the observation window, the anomaly index of the target loop is continuously calculated in the same way as feature extraction. Record the trajectory of the anomaly index over time within the observation window, and calculate the anomaly decay rate or growth rate. The change trajectory and the decay rate or growth rate are output as the electrical characteristic response change.

6. The method for predicting arc faults in a self-healing intelligent integrated distribution box according to claim 5, characterized in that, Feeding the response change back to the dynamic baseline for correction includes: If the response change indicates that the anomaly index shows a decaying trend after the load decreases, then this closed-loop interaction is marked as a heat accumulation event, and the baseline mean of the dynamic baseline is updated with the decayed steady-state value of the anomaly index. If the response change indicates that the anomaly index does not improve or shows an upward trend after the load is reduced, then this closed-loop interaction is marked as an insulation damage event, and the baseline fluctuation range of the dynamic baseline is lowered to improve the subsequent detection sensitivity. The labeling results and corresponding baseline corrections for each closed-loop interaction are stored as learning samples.

7. The method for predicting arc faults in a self-healing intelligent integrated distribution box according to claim 6, characterized in that, It also includes an adaptive baseline evolution step: During the long-term operation of the system, all stored learning samples are statistically analyzed at preset intervals to identify the characteristic distribution patterns of thermal accumulation events and insulation damage events. The decision boundaries for attention level, early warning level, and emergency level in the multi-parameter correlation analysis model are adjusted according to the aforementioned characteristic distribution pattern. Based on the feature differences of different types of events in the accumulated learning samples, the weight allocation of each feature quantity in the adaptive weight coefficient is optimized, so that the system's ability to distinguish different types of fault precursors continues to improve with running time.

8. A fault-self-healing intelligent integrated distribution box arc fault prediction system, applicable to the fault-self-healing intelligent integrated distribution box arc fault prediction method according to any one of claims 1 to 6, characterized in that, include: An electrical feature extraction unit is used to collect electrical signals of each circuit using the existing current transformers and voltage transformers in the distribution box, perform sliding window feature extraction on the electrical signals to calculate a normalized anomaly index, compare the anomaly index with a dynamic baseline, and output an anomaly marker and corresponding feature vector when the deviation reaches a preset ratio. A risk assessment unit is used to perform multi-parameter correlation analysis based on the anomaly marker and the feature vector, combined with the historical operating data of the target circuit, to assess the fault risk level. The load control and verification unit is used to transfer part of the load of the target circuit to an adjacent redundant circuit according to priority when the fault risk level reaches the warning level, continuously observe the change of electrical characteristic response of the target circuit after the load transfer, and feed the response change back to the electrical characteristic extraction unit to correct the dynamic baseline, so that the electrical characteristic extraction unit, the risk assessment unit and the load control and verification unit form a closed-loop linkage. The fault determination and execution unit is used to determine the nature of the fault based on the response change and execute the corresponding control strategy.

Images