A high-sensitivity detection system for series arc faults in a cable branch box

Abstract

This invention discloses a highly sensitive detection system for series fault arcs in cable branch boxes, relating to the field of remote condition monitoring and fault diagnosis technology for power distribution networks. Its key technical features include a distributed synchronous acquisition and composite excitation injection module, a topology inverse filtering and phase dispersion compensation module, an anti-pseudo-chaotic dynamic feature extraction and confidence assessment module, and a confidence-gated health assessment and prediction module, which are connected in series in the order of signal perception, physical reconstruction, feature purification, and health prediction. Through the collaborative work of distributed acquisition, topology inverse filtering compensation, pseudo-chaotic feature isolation, and confidence-gated prediction, this system restores the true arc waveform at the physical layer and isolates pseudo-chaotic interference at the algorithm layer, solving the problems of multi-branch signal distortion and transient interference contamination, thus achieving highly sensitive and reliable detection of series fault arcs.

Description

Technical Field

[0001] This invention relates to the field of remote condition monitoring and fault diagnosis technology for power distribution networks, specifically a highly sensitive detection system for series fault arcing in cable branch boxes. Background Technology

[0002] Cable branch boxes are critical nodes in power distribution networks that connect distribution transformers with multiple user branches. Due to thermal expansion and contraction, fretting wear, and moisture corrosion during long-term operation, the cable joints inside are prone to increased contact resistance. This can lead to intermittent microsecond-level series fault arcs near the current zero crossing, causing the joint temperature to rise continuously and insulation to deteriorate rapidly. In severe cases, this can even cause fires. Therefore, highly sensitive detection of series fault arcs is crucial for ensuring power distribution safety.

[0003] Existing arc detection technologies for series faults in cable branch boxes mainly employ fixed-band filtering or time-domain characteristic thresholding methods. These methods extract current distortion or high-frequency components by installing a single detection point on the incoming line side to determine the arc. Some solutions also incorporate zero-crossing voltage detection.

[0004] The aforementioned existing technologies still have some problems in the actual operation of complex industrial scenarios with mixed loads such as photovoltaics, charging piles, variable frequency motors, and water pumps: First, in the multi-branch parallel topology, the parasitic capacitance to ground of each branch cable and the stray inductance of the busbar form a time-varying frequency response network. The high-frequency characteristic signal generated by the arc undergoes frequency-selective attenuation when it propagates to the detection point on the incoming line side, accompanied by nonlinear phase winding and time delay dispersion, which makes it impossible for a single detection point to obtain the complete arc waveform, forming a physical detection blind zone; Second, the steep current impact generated by normal load transients such as water pump start-up or charging pile switching is similar to the early arc in the time and frequency domain, which will mislead feature extraction and generate a large number of misjudged features, causing pseudo-chaos. As a result, feature extraction based on a fixed window cannot distinguish between the real arc and interference, and low-quality data is input into the subsequent health model, causing long-term prediction pollution, resulting in the drift of prediction model parameters and the accumulation of prediction errors until failure. Summary of the Invention

[0005] To achieve the above objectives, the present invention provides the following technical solution: A high-sensitivity detection system for series fault arc in a cable branch box includes: a distributed synchronous acquisition and composite excitation injection module, used to inject a broadband detection signal into the bus side of the cable branch box to obtain raw data with a high signal-to-noise ratio, and synchronously acquire the voltage and current of each branch, outputting a synchronous data frame containing the current sequence, voltage sequence, and a copy of the detection signal; a topology inverse filtering and phase dispersion compensation module, used to receive the synchronous data frame, construct an inverse filter by identifying the dynamic admittance matrix, perform frequency domain inverse filtering on the current sequence to eliminate amplitude attenuation caused by parallel resonance of multiple branches, and perform phase calibration and time delay dispersion correction, outputting the restored waveform at the arc source end; and an anti-pseudo-chaotic dynamic feature extraction and confidence assessment module, used to receive the restored waveform, extract the transient impact geometric fingerprint of the waveform and calculate the maximum Lyapunov exponent, generating a dynamic The confidence weights and weighted chaos are adaptively determined by a dynamic gating adjustment mechanism to determine the analysis window length. Within the analysis window, the arc features of series faults are extracted, and the arc feature vector and corresponding dynamic confidence weights are output. The confidence-gated health assessment and prediction module receives the arc feature vector and dynamic confidence weights, integrates slowly varying sensor data, and uses a graph neural network to model the spatial correlation of monitoring nodes to output the node health status probability distribution. After gating the health status probability distribution with dynamic confidence weights, it drives a hidden Markov model for state tracking. Based on the state tracking results, the health index is calculated, and a Wiener process model gated with dynamic confidence weights is used to predict the remaining fault time and confidence interval, outputting hierarchical operation and maintenance instructions to realize remote health status monitoring and predictive maintenance of cable branch boxes.

[0006] Furthermore, in the distributed synchronous acquisition and composite excitation injection module, the broadband detection signal is a pseudo-random binary sequence encoded signal, which is injected in a manner that repeats once per power frequency cycle, with a spectrum covering 10 kHz to 200 kHz and an amplitude not greater than one-thousandth of the line rated current; the synchronous acquisition adopts a precise time protocol to achieve multi-node clock synchronization.

[0007] Furthermore, in the topological inverse filtering and phase dispersion compensation module, a recursive least squares algorithm with a forgetting factor is used to identify the dynamic admittance matrix online. The topological inverse filtering and phase dispersion compensation module has a built-in anti-saturation mechanism. When the amplitude of the input signal exceeds twice the preset safety threshold, the covariance matrix inside the recursive least squares algorithm is reset to the identity matrix and the integration stage is frozen. The module is then switched to open-loop prediction mode to prevent parameter estimation from diverging under large signal interference.

[0008] Furthermore, in the topological inverse filtering and phase dispersion compensation module, the inverse filter transfer function is obtained by inverting the admittance matrix at each frequency point. The current sequence is then multiplied point by point with the inverse filter transfer function after undergoing a fast Fourier transform, and then subjected to an inverse fast Fourier transform to obtain the decoupled current signal, thereby eliminating the amplitude attenuation caused by multi-branch parallel resonance.

[0009] Furthermore, in the topology inverse filtering and phase dispersion compensation module, the theoretical phase offset is calculated based on the physical length of the cable and the signal propagation speed. The decoupled current signal is subjected to reverse phase rotation through a full-pass phase equalization filter composed of cascaded second-order full-pass sections for phase calibration. The group delay of each frequency component is calculated, and non-uniform resampling is performed using cubic spline interpolation to complete the time delay dispersion correction and output the restored waveform.

[0010] Furthermore, in the anti-spurious chaotic dynamic feature extraction and confidence evaluation module, the transient impact geometric fingerprint includes rising edge steepness, half-wave zero-crossing offset, and third harmonic phase jump angle; the transient impact geometric fingerprint is input into a pre-trained gradient boosting tree classifier, which outputs event category labels and original confidence, and the maximum Lyapunov exponent is calculated using the Wolf algorithm within a fixed initial window.

[0011] Furthermore, the dynamic confidence weight is calculated based on the event category label and the original confidence, and the weighted chaos is equal to the product of the dynamic confidence weight and the maximum Lyapunov exponent. The dynamic gating adjustment mechanism adopts a dual asymmetric threshold method to divide the weighted chaos into a low chaos region, a fuzzy region, and a high chaos region. In the fuzzy region, a proportional-integral controller is used to dynamically adjust the analysis window length. Combined with the chaos and confidence coupling decision table, the analysis window length and output data label are corrected according to the relationship between the weighted chaos interval and the dynamic confidence weight to avoid misjudgment of the analysis window due to pseudo-chaos.

[0012] Furthermore, in the anti-spurious chaos dynamic feature extraction and confidence assessment module, the arc feature vector is a four-dimensional feature vector containing the intermittent arcing frequency, half-wave anomaly proportion, high-frequency energy growth rate, and arc resistance dynamic change trend. According to the dynamic confidence weight, the corresponding four-dimensional feature vector is divided and marked as a high-quality data stream and an isolated data packet. When the weighted chaos degree is in the high chaos region and the dynamic confidence weight is less than or equal to the preset low confidence trigger threshold, a covariance reset request signal is sent to the confidence gating health assessment and prediction module. The isolated data packet is only used as background reference data and does not participate in the subsequent health assessment and prediction model core parameter update.

[0013] Furthermore, in the confidence-gated health assessment and prediction module, slowly varying sensor data is input into a long short-term memory network encoder to obtain a slowly varying embedding vector, which is then concatenated with the statistics of the arc feature vector according to a fixed fusion period to form a fused feature vector. The graph neural network constructs a topology containing six electrical vertices based on the fused feature vector and performs message passing and state updates, outputting the node health state probability distribution. The hidden Markov model uses the node health state probability distribution as the original observation likelihood and multiplies the original observation likelihood with the corresponding dynamic confidence weight to obtain the corrected observation likelihood, which is used to drive state tracking.

[0014] Furthermore, the Wiener process model estimates the drift rate of the health index online using a one-dimensional Kalman filter; the confidence-gated health assessment and prediction module maintains a continuous counter for low-confidence data packets, and automatically resets the posterior error covariance of the Kalman filter to its initial value when it continuously receives covariance reset request signals and the count exceeds a preset continuous low-confidence count threshold; the confidence interval for the remaining fault time is obtained by calculating the overall fluctuation variance based on the random fluctuation intensity of the health index, the expected value of the remaining fault time, and the posterior error covariance; the hierarchical operation and maintenance instructions are generated based on the numerical range of the health index.

[0015] This invention provides a highly sensitive detection system for series fault arc in cable branch boxes, which has the following advantages: 1. This invention utilizes the coordinated efforts of four modules: distributed synchronous acquisition and composite excitation injection, topological inverse filtering and phase dispersion compensation, anti-pseudo-chaotic dynamic feature extraction and confidence assessment, and confidence-gated health assessment and prediction. It leverages active detection signals and inverse filtering networks to compensate for frequency-selective attenuation, phase entanglement, and time-delay dispersion caused by multi-branch topologies, restoring the true waveform at the arc source and eliminating detection blind spots caused by physical propagation distortion. Simultaneously, it generates dynamic confidence weights based on transient impact geometric fingerprints and chaos degree calculations. Through a dynamic gating adjustment mechanism and a chaos degree confidence-coupled decision table, it quantitatively isolates pseudo-chaotic interference caused by load switching, separating high-quality data streams from isolated data packets, thus preventing interference signals from contaminating subsequent models at the source. Therefore, it simultaneously solves the signal distortion problem and pseudo-chaotic interference problem under multi-branch topologies, achieving high signal-to-noise ratio and high robustness in series fault arc detection.

[0016] 2. This invention achieves a cross-scale closed loop for cable branch boxes, from transient arc detection to long-term health management and fault prediction. Furthermore, it uses the dynamic confidence weights of front-end feature extraction as gating factors to drive the state tracking of the back-end Hidden Markov Model. The prediction parameters for health status and remaining fault time are updated only when high-confidence data arrives. A Kalman filter prediction error covariance gating reset mechanism cuts off the error propagation link of low-quality data. Simultaneously, a graph neural network models the spatial association of electrical nodes within the branch box, outputting the node health status probability distribution. Combined with the Wiener process model, it outputs a health index and remaining fault time confidence interval, ultimately generating hierarchical operation and maintenance instructions, seamlessly connecting transient detection with long-term health prediction. This achieves cross-scale health management from microsecond-level arc detection to monthly-level remaining life prediction, solving the functional fragmentation problem of traditional protection devices that only output the presence or absence of arc Boolean values ​​and cannot predict health trends.

[0017] 3. This invention constructs a complete technical chain of propagation distortion compensation, pseudo-chaotic immunity, and confidence-gated prediction. This enables the system to maintain a high detection signal-to-noise ratio, reduce health index prediction errors, improve the detection sensitivity of series fault arcs and the reliability of health prediction, and reduce unnecessary power outages and maintenance costs. This provides reliable technical support for the intelligent operation and maintenance of distribution branch boxes. Attached Figure Description

[0018] Figure 1 This is a system framework diagram of Embodiment 1 of the present invention.

[0019] Figure 2 This is a flowchart of the anti-spurious chaotic dynamic feature extraction and confidence evaluation process of the present invention.

[0020] Figure 3 This is a flowchart of the confidence-gated health assessment and prediction process of the present invention.

[0021] Figure 4 This is a flowchart of the method in Embodiment 2 of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0023] Example 1: Please see Figures 1 to 3This embodiment provides a highly sensitive detection system for series fault arcing in cable branch boxes, comprising the following four core modules, which are connected in series in the order of signal sensing, physical reconstruction, feature purification, and health prediction to form a complete technology chain: Module 1: Distributed Synchronous Acquisition and Composite Excitation Injection Module; responsible for injecting safe broadband detection signals on the bus side and synchronously acquiring the raw voltage and current data of each branch with high precision, providing a reliable data source for the entire system; Module 2: Topology Inverse Filtering and Phase Dispersion Compensation Module; Receives raw voltage and current data from Module 1, constructs an inverse filter by online identification of the network dynamic admittance matrix, and compensates for phase distortion and time delay broadening during signal propagation, ultimately restoring the true waveform at the arc occurrence point, thus solving the signal distortion and detection blind zone problems caused by multi-branch topology from a physical perspective; Module 3: Anti-pseudo-chaotic dynamic feature extraction and confidence assessment module; Receives the waveform restored by Module 2, first extracts the geometric fingerprint of the impact event and combines it with the chaos degree calculation to generate dynamic confidence weights, then adaptively adjusts the analysis window through an innovative dynamic gating mechanism, and finally extracts high-quality arc feature sequences, effectively identifying and isolating pseudo-chaotic phenomena caused by strong interference such as load switching, and preventing them from contaminating subsequent models; Module 4: Confidence-Gated Health Assessment and Prediction Module; Receives the feature sequence with confidence labels output from Module 3, integrates slowly varying signals such as temperature and partial discharge, uses confidence weights to control the weights of health status updates, and performs spatial correlation analysis and status tracking through graph neural networks and hidden Markov models. Finally, it outputs health index, remaining time-of-failure prediction, and hierarchical operation and maintenance instructions, solving the problem of the separation between detection and prediction functions, and realizing remote health status monitoring and predictive maintenance of cable branch boxes based on high-confidence data.

[0024] The system works as follows: Module 1 is deployed on the inlet side of the branch box and at each outlet branch end. It actively injects broadband detection signals and simultaneously collects the original waveforms of voltage and current of multiple branches, providing a time-aligned material data basis for the entire system. Module 2 receives data from Module 1, compensates for frequency selective attenuation caused by multi-branch parallel topology through dynamic admittance identification and inverse filtering network, and eliminates nonlinear phase winding and time delay dispersion by combining all-pass phase equalization and non-uniform resampling technology, outputting a high-fidelity arc source end restored waveform. Module 3 receives the restored waveform, performs impact geometric fingerprint extraction and chaos degree calculation in parallel, and generates dynamic confidence weights and weighted chaos degree; it introduces a dual asymmetric threshold method and a chaos degree confidence coupled decision table to realize the dynamic adaptive adjustment of the analysis window, and splits the data into high-quality data streams and isolated data packets; Module 4 receives the high-quality feature sequence, slowly varying sensor embedding vector, and dynamic confidence weights output by Module 3. It performs spatial correlation modeling through graph neural networks, uses confidence weight-gated hidden Markov model for state tracking, and employs Wiener process combined with error covariance gating reset mechanism for remaining lifetime prediction. Finally, it outputs health index and hierarchical operation and maintenance decisions.

[0025] This system eliminates signal propagation distortion and detection blind spots in multi-branch parallel topologies by coordinating amplitude and phase distortion compensation through Module 1 and Module 2. Module 3 quantitatively identifies and isolates pseudo-chaotic interference through a dynamic gating adjustment mechanism and a coupled decision table, solving the problems of feature collapse and sliding window misjudgment caused by non-stationary phase transitions of high-resistivity micro-arcs. Modules 3 and 4 cut off the pollution link from low-quality data to the health model through confidence data diversion and Kalman filter error covariance gating reset mechanism, solving the industry problem of the disconnect between transient detection and long-term health prediction functions.

[0026] Furthermore, for Module 1: Distributed Synchronous Acquisition and Composite Excitation Injection Module, its working process includes 4 stages.

[0027] Step 1.1: Generating a broadband detection signal; A low-power signal injection unit is installed on the bus side of the cable branch box. The microcontroller inside the unit calls the preset pseudo-random binary sequence encoded data through a lookup table, and generates an active detection signal through a sixteen-bit digital-to-analog converter circuit. The active detection signal is output in a manner that repeats once per power frequency cycle, with an amplitude not exceeding one-thousandth of the rated current of the line, and a frequency range covering 10kHz to 200kHz, exhibiting good autocorrelation characteristics.

[0028] Step 1.2: Forming a composite excitation; the active detection signal and the transient current naturally generated by the load of each branch during operation, such as motor starting or charging pile connection, together form a composite excitation for subsequent impedance identification.

[0029] Step 1.3: Perform synchronous sampling; High-frequency current transformers and voltage transformers with a synchronization accuracy ≤1 microsecond are configured at the incoming and outgoing ends of each branch, with a sampling frequency ≥200kHz. Multi-node clock synchronization is achieved using GPS or the IEEE 1588 precise time protocol. The installation locations of these transformers are called detection points, which are physically fixed by Module 1 during the system deployment and commissioning phase. Voltage transformers sense voltage changes at their locations in real time and convert the sensing results into discrete digital signals. The set of digital signals sampled by all voltage transformers at the same time, arranged in chronological order, constitutes the voltage sequence. High-frequency current transformers sense current changes at their locations in real time and convert the sensing results into discrete digital signals. The set of digital signals sampled by all high-frequency current transformers at the same time, arranged in chronological order, constitutes the current sequence.

[0030] Step 1.4: Complete data packaging; encapsulate the current sequence, voltage sequence and active injection signal copies collected from each branch into a synchronization data frame according to a unified time tag.

[0031] Furthermore, for Module 2: Topological Inverse Filtering and Phase Dispersion Compensation Module, its working process includes 4 steps.

[0032] Step 2.1: Identify the dynamic admittance matrix using an anti-saturation adaptive observer.

[0033] A dynamic equivalent circuit model is established for each branch, consisting of equivalent resistance, equivalent inductance, and equivalent capacitance to ground connected in series. The equivalent resistance, equivalent inductance, and equivalent capacitance to ground are estimated in real time by a recursive least squares algorithm to characterize the dynamic impedance characteristics of the branch during operation.

[0034] A recursive least squares algorithm with a forgetting factor is used to estimate the parameters of the synchronization voltage sequence, current sequence, and active detection signal replicas. The forgetting factor ranges from 0.95 to 0.995, with a default value of 0.98. This value is obtained through cross-validation calibration using historical long-term aging test data. It is used to make the weight of earlier sampled data decay exponentially, enabling the algorithm to track parameter time-varying events on the order of 0.5 seconds, while suppressing white noise. The forgetting factor directly participates in the update calculation of the gain matrix, determining the proportion of historical data retained in the parameter estimation. The algorithm execution update interval is set to 0.01 seconds. This interval is determined based on the power frequency cycle duration of the power distribution system and the computing power of the digital signal processor, specifying that a parameter estimation calculation is performed every 0.01 seconds using the latest sampling window data.

[0035] The equivalent parameters of each branch in the cable distribution box change slowly with temperature, current, and operating time, making it a time-varying system. Recursive least squares can track parameter drift in real time with low computational cost, making it suitable for embedded implementation. However, when the input signal amplitude is too large, the covariance matrix of the recursive least squares algorithm tends to zero, causing the gain to disappear and the parameters to fail to update, i.e., the estimator is locked. When the signal suddenly returns to normal, the locked estimator cannot track quickly. Therefore, when the input exceeds twice the safety threshold, the covariance is reset and the update is frozen, switching to open-loop prediction. The system restarts after the signal returns to normal, ensuring the robustness of the estimator.

[0036] Specifically, the recursive least squares algorithm calculates the prediction error and the gain matrix. The prediction error is equal to the difference between the actual sampled value at the current moment and the theoretical prediction value calculated based on the estimated parameters at the previous moment. The gain matrix is ​​obtained by matrix inversion and multiplication of the real-time updated covariance matrix and the current input vector. The covariance matrix is ​​used to characterize the degree of uncertainty of the parameter estimation, and its initial value is set as the identity matrix. The current parameter estimate of the algorithm is equal to the parameter estimate at the previous moment plus the product of the gain matrix and the prediction error. This operation directly outputs the real-time updated values ​​of the equivalent resistance, equivalent inductance, and equivalent capacitance to ground.

[0037] The recursive least squares algorithm incorporates an anti-saturation mechanism. The preset safety threshold is set at 5 times the rated current of the branch box, a value determined according to the power equipment overload protection standard. The power equipment overload protection standard typically uses 2 times the rated current as the upper limit for short-term overload; exceeding this value triggers the anti-saturation mode. Therefore, when the input signal amplitude exceeds twice the preset safety threshold, the algorithm determines that the system is in a state of drastic load change or surge impact. At this time, the algorithm automatically forces the covariance matrix to reset to an identity matrix, freezes the integration stage, and switches to open-loop prediction mode. The integration stage is the cumulative update process of parameter estimation in the recursive least squares algorithm. The open-loop prediction mode is a working state where parameter updates stop, and only the original fixed parameters are used for output. Both together constitute the anti-saturation mechanism to prevent parameter estimation divergence. The reset covariance matrix is ​​directly used as the initial input for calculating the gain matrix at the next time step.

[0038] The algorithm outputs updated equivalent resistance, equivalent inductance, and equivalent capacitance to ground every 0.01 seconds. Based on these three parameters, it calculates the admittance value corresponding to each operating frequency point. The admittance value is equal to the complex combination of the reciprocal of the resistance and the reciprocal of the reactance. The admittance values ​​of all branches at all operating frequency points are arranged in a matrix format and combined to generate a time-varying admittance matrix for the entire network. This matrix completely records the variation of network impedance with time and frequency.

[0039] Step 2.2: Construction of the topological inverse filter network and frequency domain decoupling.

[0040] The inverse admittance value at each frequency point in the admittance matrix is ​​obtained by calculating the reciprocal of the admittance value. The inverse admittance values ​​of all frequency points are combined in frequency order to form the inverse filter transfer function. This transfer function is used to accurately cancel the frequency selective attenuation caused by multi-branch parallel resonance in the frequency domain. The amplitude-frequency curve of the inverse filter transfer function is a mirror image of the amplitude-frequency curve of the original network admittance.

[0041] The original current sequence of the detection points collected by Module 1 is subjected to Fast Fourier Transform (FFT) to convert it into a frequency domain complex signal. The frequency domain complex signal is then multiplied point-by-point by the inverse filter transfer function at the corresponding frequency points. The multiplied frequency domain signal is then subjected to Inverse Fast Fourier Transform (IFFT) to convert it back to a time domain signal. This signal eliminates the amplitude attenuation caused by multi-branch parallel resonance and is called the decoupled current signal.

[0042] Step 2.3: Perform phase frequency response modeling and phase calibration.

[0043] Theoretical phase offset is calculated based on the physical length of the cable, the signal propagation speed, and the current frequency component to be processed. For a given frequency, the propagation distance of the signal from the arc point to the detection point is equal to the physical length of the branch cable. The signal propagation speed is determined by the dielectric constant and permeability of the cable insulation layer. The dielectric constant is obtained from the cable material specification, and the permeability is an internationally recognized physical constant. For commonly used cross-linked polyethylene cables, the dielectric constant is approximately 2.3, and the permeability is approximately 1. The theoretical phase offset is equal to negative twice pi multiplied by the current operating frequency multiplied by the propagation distance, and then divided by the signal propagation speed. A negative result indicates that the signal phase lags during propagation in the line.

[0044] Phase equalization compensation: A full-pass phase equalization filter is designed, which consists of four cascaded second-order full-pass sections, with a total order of 8. The center frequency and quality factor of each second-order full-pass section are obtained by fitting the theoretical phase offset curve. The filter applies a reverse phase rotation to the time-domain decoupled signal, which is equal in value but opposite in direction to the theoretical phase offset, thereby compensating for the phase lag caused by propagation. Since the amplitude-frequency characteristic of the filter is always 1, the signal amplitude remains unchanged during the compensation process. The reverse phase rotation operation directly cancels the phase winding caused by propagation. The compensated signal is called the phase correction signal.

[0045] Step 2.4: Perform time delay dispersion correction.

[0046] Calculate group delay: For a phase-corrected signal, calculate the derivative of its phase with respect to frequency to obtain the group delay of each frequency component. The group delay characterizes the time difference of different frequency components propagating in the line and is the direct cause of the stretching or compression of the waveform in the time domain. The group delay curve is obtained by performing sliding differential operation on the phase sequence using a digital differentiator.

[0047] Non-uniform resampling alignment: The entire working frequency band is divided into 128 equal bands. The number of equal bands is set according to the uniformity requirements of the frequency band energy distribution measured by the spectrum analyzer to ensure that the phase change in each sub-band is approximately linear. The theoretical arrival time of each frequency band component is calculated based on the group delay. The cubic spline interpolation method is used to perform non-uniform resampling of the signal waveform according to the deviation between the theoretical arrival time of each frequency band and the unified reference time axis. The number of interpolation nodes is set to at least four sampling points per frequency band to ensure waveform continuity and smoothness. The timestamp after resampling is directly aligned to the initial synchronization clock reference acquired by module one.

[0048] After the above processing, the distortion of the signal in the three dimensions of amplitude, phase and time delay is completely corrected, and the restored true time-domain waveform of the arc source is output. This waveform eliminates the time stretching and compression effects and fully preserves the zero-rest characteristics of the weak arc and the microsecond-level impedance step details.

[0049] Furthermore, for Module 3: Anti-spurious chaotic dynamic feature extraction and confidence assessment module, its working process includes 7 steps.

[0050] Step 3.1: Multi-resolution morphological filtering preprocessing.

[0051] For the restored real time-domain waveform of the arc source, opening and closing operations are performed sequentially. Opening is performed using a flat structuring element of length 5. The opening operation involves taking the minimum value within the structuring element's coverage area and performing an erosion operation, followed by taking the maximum value and performing an expansion operation. The length of 5 is obtained by calibrating the frequency response characteristics to filter out high-frequency glitches above 10kHz and is used to generate a preliminary smooth waveform. Then, closing is performed using a structuring element of length 3. The closing operation involves first expanding and then eroding. The length of 3 is obtained by calibrating the time-domain characteristics to fill in the microsecond-level waveform discontinuities. After filtering, a denoised waveform with edge preservation is obtained.

[0052] Step 3.2: Transient impact geometric fingerprint extraction.

[0053] Three geometric fingerprints are extracted from the filtered waveform and combined into a three-dimensional impact geometric fingerprint vector: Fingerprint 1 is the rising edge steepness. Locate the peak value of the positive half-wave of the current in each power frequency cycle in the waveform, find the moment when the current amplitude reaches 10% of the peak value and the moment when it reaches 90% of the peak value, calculate the time difference between the two moments, and the reciprocal of the time difference is the rising edge steepness. Fingerprint 2 is the half-wave zero-crossing offset. The zero-crossing time of the smooth waveform current and the zero-crossing time of the synchronously acquired voltage are calculated separately using linear interpolation. The absolute value of the difference between the two times is the half-wave zero-crossing offset. The calculation is performed once for each half-wave, and the average value of the calculation results of all half-waves within the window is taken.

[0054] The fingerprint is the third harmonic phase jump angle. Zero-crossing detection is performed on the smooth waveform to extract the fundamental phase. At the same time, a second-order Butterworth digital filter with a center frequency of 150Hz and a bandwidth of 20Hz is used to extract the third harmonic component. Hilbert transform is used to obtain the instantaneous phase of the third harmonic. The difference between the third harmonic phase of the current half-wave and three times the fundamental phase is calculated. Then, the difference of the previous half-wave is subtracted. If the absolute value exceeds 180 degrees, the angle is folded by subtracting the value from 360 degrees to obtain the third harmonic phase jump angle.

[0055] Section 3.3: Lightweight Gradient Boosting Tree Event Classification.

[0056] The three-dimensional impact geometric fingerprint vector is input into a pre-trained lightweight gradient boosting tree classifier, which consists of 100 decision trees with a maximum depth of 3. During the training phase, the Gini coefficient is used as the node splitting criterion and the learning rate is set to 0.1. The classifier passes the input vector through all decision trees in sequence, accumulates the output values ​​of the leaf nodes, and then transforms them through a logistic function to output the normal load switching probability and the probability of the real arc event. The probability of a real electric arc event is used as the original confidence score output. If the probability of a real electric arc event is greater than the arc confidence score threshold of 0.7, it is judged as a real electric arc event and the category label is assigned a value of 1. If the probability of a real electric arc event is ≤0.7, it is judged as a normal load switching event and the category label is assigned a value of 0. The arc confidence score threshold of 0.7 is determined by the optimal segmentation point of the working characteristic curve of 100,000 subjects.

[0057] Step 3.4: Calculation of the maximum Lyapunov index.

[0058] For the restored arc source end real time domain waveform, an initial fixed waveform window with a length of 2 milliseconds is selected, corresponding to 400 sampling points. This length is determined based on the shortest arc microsecond step duration of the power distribution system. The Wolf algorithm is used for phase space reconstruction, with the embedding dimension set to 5 and the time delay set to 4 sampling points, i.e., 20 microseconds.

[0059] In the reconstructed phase space, the first phase point is selected as the reference point, and the nearest neighbor point is searched and the initial distance is recorded. The process is advanced with a time delay as the step size, and the new distance between the two points after evolution is calculated. When the new distance exceeds 10 times the initial distance, a new neighbor point that meets the condition that the directional angle is less than 45 degrees is searched near the reference point and replaced.

[0060] The natural logarithm of the distance ratios of all evolution steps is accumulated and divided by the total evolution time to obtain the maximum Lyapunov exponent, which characterizes the degree of waveform chaos.

[0061] Step 3.5: Dynamic confidence weight and weighted chaos generation.

[0062] Based on the category label, dynamic confidence weights are calculated in two cases.

[0063] If the category label is 1, the dynamic confidence weight is equal to the original confidence plus the compensation coefficient 0.1. Then the calculation result is compared with the upper limit of dynamic confidence 0.95, and the smaller value is taken. The compensation coefficient 0.1 is set based on the lowest confidence of the load switching event to ensure that the isolated data packet still has a very low weight to participate in the environment reference. The upper limit of 0.95 is used to prevent overfitting and retain 5% uncertainty.

[0064] If the category label is 0, the matching degree is calculated first. The matching degree is the maximum value of the similarity between the current rising edge steepness and the typical steepness values ​​in the pre-stored typical load impact template library. The similarity is calculated using the Gaussian kernel function. The template library is extracted from historical normal switching data by clustering. The template library contains typical features of common loads such as motor starting and charging pile access. The dynamic confidence weight is equal to 0.1 plus 0.2 multiplied by 1 minus the difference of the matching degree, ensuring that the weight range of load switching events is limited to between 0.1 and 0.3.

[0065] The weighted degree of chaos is equal to the dynamic confidence weight multiplied by the maximum Lyapunov exponent.

[0066] Section 3.6: Weighted Chaos Dynamics Gating Regulation Mechanism.

[0067] The analysis window is adaptively controlled using weighted chaos degree and dynamic confidence weights.

[0068] First, the interval is divided using a dual asymmetric threshold method: the weighted chaos degree is compared with a preset chaos threshold. The low chaos threshold is 0.3, which is determined by the 90th percentile of the upper bound of the weighted chaos degree of a large amount of normal load switching data; the high chaos threshold is 0.7, which is determined by the 10th percentile of the lower bound of the weighted chaos degree of the actual arc evolution stage. If the weighted chaos degree is <0.3, it is determined to be a low chaos region, and the sliding window length is locked at 200 milliseconds; If the weighted chaos degree is >0.7, it is determined to be a high chaos region, and the sliding window length is locked at 10 milliseconds; If 0.3 ≤ weighted chaos degree ≤ 0.7, it is determined to be a fuzzy region. The proportional-integral controller (PIC) is activated to dynamically adjust the window length. The input error of the PIC is equal to the weighted chaos degree minus the target chaos degree of 0.5, where the target chaos degree is the equilibrium point of the fuzzy region, achieving a trade-off between time and frequency resolution. The controller outputs a window length change step size, which is equal to the proportional coefficient 0.1 multiplied by the input error plus the integral coefficient 0.02 multiplied by the cumulative value of the input error over time. The proportional coefficient is tuned using the Ziegler-Nichols critical proportionality method to ensure no overshoot in the critical response. The integral coefficient eliminates steady-state error, and the integral time constant is 200 milliseconds, matching the power frequency cycle. The cumulative value is calculated using an accumulator, with each accumulation being the input error multiplied by a 5-microsecond sampling interval. The new window length is equal to the old window length plus the change step size, and the result is forcibly limited to the range of 10 milliseconds to 200 milliseconds. 10 milliseconds and 200 milliseconds correspond to half a power frequency cycle and 10 power frequency cycles, respectively.

[0069] Secondly, a decision table coupled with chaos level and confidence level is introduced for cross-validation: a two-dimensional cross-validation rule is established: The preset low confidence trigger threshold is 0.7, which is consistent with the judgment threshold of the gradient boosting tree classifier in step 3.3, ensuring the consistency of the judgment criteria between the feature extraction step and the quality control step. When the weighted chaos degree is in the low chaos region, if the dynamic confidence weight is >0.7, the window length is kept at 200 milliseconds and the data is marked as high-confidence stationary data; if the dynamic confidence weight is ≤0.7, the window length is kept at 200 milliseconds and the data is marked as low-confidence stationary data. When the weighted chaos degree is in the high chaos region, if the dynamic confidence weight is >0.7, the locking window length is 10 milliseconds, and it is marked as high-confidence transient data; if the dynamic confidence weight is ≤0.7, the locking window length is 10 milliseconds, and a covariance reset request signal is sent to module four at the same time. When the weighted chaos is in the fuzzy region, if the dynamic confidence weight is >0.7, dynamic sliding window is allowed, and the proportional and integral coefficients of the proportional-integral controller are multiplied by a gain boost factor of 1.2; if the dynamic confidence weight is ≤0.7, dynamic sliding window is allowed, but the proportional and integral coefficients of the proportional-integral controller are multiplied by a gain decay factor of 0.6. The gain boost factor is set based on the principle that it enhances the controller response at high confidence levels, but does not exceed the system stability margin; the gain decay factor is set based on the principle that it suppresses control actions at low confidence levels to avoid misadjustment.

[0070] It should be noted that a dynamic confidence weight > 0.7 indicates that the current data is highly believed by the classifier to be a real arc event. At this time, the system should respond quickly to changes in chaos and accelerate window adjustment in order to capture the rapid evolution of arc features in a timely manner. Multiplying by 1.2 can increase the controller gain by 20%, accelerate the integral action of the error, and make the window length converge towards the target chaos level of 0.5 more quickly. Choosing 1.2 instead of a higher multiple is to avoid overshoot and oscillation.

[0071] A dynamic confidence weight of ≤0.7 indicates that the current data is more likely to be in a state of disturbance or mixture. In this case, the window should be adjusted carefully to prevent the window from jittering due to pseudo-chaos. Multiplying by 0.6 reduces the controller gain by 40%, weakens the integral accumulation effect, and greatly slows down the change in window length. This conservative strategy ensures that the window remains relatively stable during disturbances and avoids passing incorrect time scale parameters to the subsequent health model.

[0072] The gain attenuation factor of 0.6 and the gain enhancement factor of 1.2 are not strictly inversely related, but the attenuation factor of 0.6 reduces the closed-loop response speed by about half, while the enhancement factor of 1.2 increases it by about 20%. This asymmetric design reflects the engineering logic of the system trusting reliable data and doubting unreliable data. Simulation verification shows that this combination has the smallest window length fluctuation variance in the simulation test.

[0073] Step 3.7: Adaptive four-dimensional feature extraction within the window.

[0074] Within the final determined window length, four arc diagnostic features are calculated for the restored waveform segment: Feature 1: Intermittent arcing frequency; within the window, search for a range of 10 microseconds before and after each half-wave zero-crossing point. If the current amplitude within this range is less than 30% of the half-wave peak value and the duration is between 2 and 50 microseconds, it is determined as a zero-rest arcing event; count the total number of all zero-rest arcing events within the window to obtain the intermittent arcing frequency, which reflects the arcing frequency of the electric arc. Feature 2: Half-wave anomalous proportion; calculate the total number of half-waves within the window, the total number of half-waves is equal to the window time length divided by 10 milliseconds; detect each half-wave, if a flat shoulder appears near the zero crossing and the amplitude is less than 30% of the rated value, it is judged as an anomalous half-wave; the half-wave anomalous proportion is equal to the number of anomalous half-waves divided by the total number of half-waves, reflecting the degree of waveform distortion caused by the electric arc. Feature 3: High-frequency energy growth rate; The signal within the window is passed through an eighth-order Butterworth digital bandpass filter with a passband range of 10kHz to 100kHz; The current window energy is obtained by multiplying the sum of the squares of the amplitudes of all sampled points of the filtered signal by the 5-microsecond sampling interval; The energy value of the previous window is obtained, the difference between the current window energy and the previous window energy is calculated, and then divided by the previous window energy value to obtain the high-frequency energy growth rate; If it is the first window started by the system, the growth rate is set to 0; The high-frequency energy growth rate reflects the trend of the intensity change of the high-frequency components generated by the electric arc; Feature 4: Dynamic trend of arc resistance; within the window, sampling points with current amplitude greater than 10% of the rated current and far from the zero-crossing point are selected, and the instantaneous resistance at each point is calculated. The instantaneous resistance is equal to the voltage value at that point divided by the current value at that point. The selected instantaneous resistance sequence is linearly fitted with the corresponding time sequence using the least squares method to calculate the slope. The slope is the dynamic trend of arc resistance, reflecting the evolution direction of arc impedance. The slope calculation method is as follows: first, calculate the average value of all time and the average value of all resistances; then, calculate the difference between each time value and the time average value, multiply it by the difference between each resistance value and the resistance average value, and sum them to obtain the numerator; calculate the square of the difference between each time value and the time average value and sum them to obtain the denominator; divide the numerator by the denominator to obtain the slope. The unit of slope is ohms per second.

[0075] The above four features are combined into a four-dimensional feature vector, namely the electric arc feature vector.

[0076] After the above processing, the output includes the arc feature vector, dynamic confidence weight, and region label. Branch routing is then executed based on the dynamic confidence weight. Branch 1: When the dynamic confidence weight is greater than the arc confidence judgment threshold of 0.7, it is judged as the real arc evolution stage. The data in this branch is marked as high-quality data stream, the window adaptive adjustment record is completely preserved, and it is transmitted to module 4 to participate in the core update of the health trajectory and the state tracking of the hidden Markov model. Branch 2: When the dynamic confidence weight is ≤0.7, it is determined to be a normal load switching or a strong interference transient, and the data in this branch is marked as an isolated data packet; a 0.3 attenuation coefficient is applied to the weighted chaos degree, that is, the weighted chaos degree used for subsequent sliding window calculation is taken as 0.3 times the original value, in order to suppress severe window jitter; the isolated data packet is transmitted to Module 4 only as background environment reference data and does not participate in the parameter learning for health index calculation and remaining failure time prediction.

[0077] Furthermore, for Module 4: Confidence-Gated Health Assessment and Prediction Module, its working process includes 6 steps.

[0078] Step 4.1: Slowly changing state quantity acquisition and long short-term memory network coding.

[0079] Collect three types of slowly changing state variables: Terminal surface temperature: The temperature sequence is obtained by an infrared thermal imager. The infrared thermal imager is deployed inside the branch box and aimed at each cable joint. The sampling period is set to 1 second, which is calibrated based on the thermal conduction response time constant. Accumulated partial discharge energy: Acquired by an ultrasonic sensor to obtain a partial discharge sequence. The ultrasonic sensor is attached to the inner wall of the chamber, and the sampling period is set to 0.1 seconds. This period is calibrated based on the characteristics of partial discharge pulse accumulation. Temperature and humidity inside the chamber: collected by temperature and humidity sensors to obtain temperature and humidity sequences. The temperature and humidity sensors are installed at the ventilation opening on the top of the chamber, and the sampling period is set to 10 seconds. This period is calibrated based on the slow change law of the environment and climate.

[0080] The three types of slow variable sequences are input into the Long Short-Term Memory network encoder. The encoder architecture consists of two layers, each containing 64 hidden units. The number of units is determined by grid search and cross-validation to balance representation ability and computational cost.

[0081] The temperature sequence is input into the encoder using the 10 most recent sampling points, and the output is an embedding vector with a dimension of 128; the partial discharge sequence is input into the encoder using the 100 most recent sampling points, and the output is an embedding vector with a dimension of 128; the temperature and humidity sequence is input into the fully connected mapping layer using the two values ​​at the current time, and the output is an embedding vector with a dimension of 128; the three embedding vectors are concatenated end to end according to the feature dimension to obtain a slowly varying feature embedding vector with a total dimension of 384.

[0082] Step 4.2: Asynchronous fusion of feature layers.

[0083] A fusion time point is set every 0.5 seconds. This 0.5-second period is derived by combining the microsecond-level arc event response speed with the hardware computing power load balance.

[0084] At each fusion time point, all arc feature vectors output by Module 3 within that time period are collected. The mean vector, variance vector, and median vector of these vectors are calculated and concatenated to form a fast feature statistical vector with a dimension of 12. The fast feature statistical vector is concatenated with the slow-varying feature embedding vector at the current time point to obtain a fusion feature vector with a dimension of 396. At the same time, the maximum value among all dynamic confidence weights within that time period is extracted as the fusion dynamic confidence weight for that fusion point.

[0085] Step 4.3: Spatial correlation modeling using graph neural networks: Using the physical topology of the cable branch box as a blueprint, graph neural networks are used to mine the spatial correlation between the health status of different monitoring nodes.

[0086] Construct the graph topology: Define six vertices, corresponding to the incoming line, four branch connectors, and the bus node, respectively; based on the electrical connection relationship, establish undirected edges between directly connected vertices, with the incoming line connected to the bus node, and the bus node connected to the four branch connectors respectively; the initial weight of each edge is equal to the reciprocal of the connecting cable length plus 0.1, and then normalize the weights of all edges originating from each vertex so that their sum is 1; the cable length is measured and recorded using a laser rangefinder during the system installation phase.

[0087] The fused feature vector of each node is input into the perceptron, which consists of two layers. The first layer outputs 128 dimensions, and the second layer outputs 64 dimensions. The activation function is a linear rectified function. The output of the perceptron is used as the initial hidden state of the node.

[0088] A three-layer message passing iteration is performed. In each layer, each node collects the hidden state of all directly connected neighboring nodes in the previous layer, multiplies each neighbor's hidden state by a 64x64 learnable weight matrix, sums the results, and then passes them through a linear rectified activation function to obtain an aggregated message. The node's current hidden state is updated through a gated loop unit, combining its own hidden state in the previous layer with the aggregated message. The update mechanism includes an update gate and a reset gate, which are used to control the proportion of historical information retained.

[0089] After three iterations, the final hidden state of each node is input into the fully connected layer, and five nodes are output, corresponding to five health states. The five health states are then converted into probability distributions of the five health states by the Softmax function, and their sum is 1. The five states are normal, abnormally high contact resistance, intermittent micro-arc, stable arc, and destructive arc.

[0090] Through the above topological mapping, initialization of hidden states, multi-layer iterative computation and output mapping, a graph neural network with spatial correlation modeling capabilities was constructed and run.

[0091] Step 4.4: Confidence-gated Hidden Markov Model State Tracking.

[0092] An independent Hidden Markov Model is established for each node, with its hidden states being the five healthy states mentioned above. The state transition matrix is ​​a 5x5 square matrix, and the matrix elements represent the transition probabilities between states. This matrix is ​​obtained through pre-training and statistical analysis of accelerated aging test data of cable joints. The diagonal elements are dominant to reflect the gradual nature of state evolution.

[0093] The graph neural network outputs a probability distribution of five health states for each node. The probability value corresponding to each state in the probability distribution is the original observation likelihood of that state. The value of the original observation likelihood ranges from 0 to 1. The larger the value, the better the observation matches the hypothetical state.

[0094] The fusion confidence weight is used as a gating factor to scale the original observation likelihood. A high weight makes the observation have a strong influence on the state estimation, while a low weight weakens the influence of the observation. For each state, the corrected observation likelihood is equal to the original observation likelihood value multiplied by the current fusion dynamic confidence weight.

[0095] When the fusion dynamic confidence weight is close to 0.95, the corrected observation likelihood is almost equal to the original observation likelihood, and the current graph neural network observation results play a dominant role in the posterior update of the hidden Markov model's state. When the fusion confidence weight is close to 0.1, the corrected observation likelihood is significantly compressed, and the contribution of the current observation to the state update is significantly diluted. The state estimation of the hidden Markov model will mainly rely on the prior evolution law represented by the pre-trained state transition matrix.

[0096] Update the hidden Markov model using the modified observation likelihood: The posterior probabilities of the states are recursively calculated using the forward algorithm. At each time step, the posterior probabilities of each hidden state at the previous time step are multiplied by the elements of the corresponding column of the state transition matrix and summed to obtain the prior probabilities of each state at the current time step. The prior probabilities are multiplied by the values ​​of the corresponding states in the corrected observation likelihood to obtain the unnormalized joint probability. The joint probability of all states is normalized so that the sum of the probabilities of each state equals 1, thus obtaining the posterior probabilities of each hidden state at the current time step. The most likely health state sequence is decoded using the Viterbi algorithm. At each time step, the algorithm calculates the maximum cumulative probability of each state by multiplying the maximum cumulative probability of the previous time step by the state transition probability, and then multiplying by the corrected observation likelihood. The algorithm synchronously records the optimal predecessor state pointer at each step. At the end of the sequence, the state with the highest cumulative probability is selected, and the algorithm backtracks along the recorded optimal predecessor state pointer to output the complete health state evolution trajectory. This trajectory sequence is stored in local non-volatile memory for subsequent offline optimization of the hidden Markov state transition matrix parameters and provides a visual traceability basis for the equipment degradation process for the operation and maintenance platform.

[0097] Step 4.5: Health index and remaining time-to-failure calculation and error covariance gating.

[0098] The health index is calculated using a weighted summation method. The health index equals 100 multiplied by the probability of a normal state, plus 75 multiplied by the probability of an abnormally high contact resistance state, plus 50 multiplied by the probability of an intermittent micro-arc state, and plus 25 multiplied by the probability of a stable arc state. The probability of a destructive arc state has a weight of 0, meaning it does not contribute to the health index. The weighting coefficients 100, 75, 50, and 25 are determined comprehensively based on fault severity classification, remaining lifespan percentage, and engineering experience, as detailed below: Normal state, weight 100: This corresponds to a health index of 100, indicating that the equipment is in the early stages of its designed lifespan with no signs of deterioration; the weight is set to the maximum value to intuitively reflect the ideal state. Abnormally high contact resistance, weight 75: This state is an early warning stage where the circuit is detectable but no arc has yet occurred; usually, the increased resistance is accompanied by slight heating, but it does not affect the insulation integrity; a weight of 75 indicates that the health has decreased by 25%, and the remaining lifespan is approximately 70%-80% of the normal state; this value is based on the typical deduction range for the condition of attention in the power equipment condition evaluation standard. Intermittent micro-arc, weight 50: Zero-rest characteristics and microsecond-level arcing have appeared, but the arc energy is low and has not yet caused carbonization of the insulation surface; this state is the critical node for the arc to transition from poor contact to stable combustion, and the health status is halved; a weight of 50 corresponds to approximately 50% of the remaining life, which is consistent with the engineering judgment for this type of fault. Stable arc, weight 25: The arc continues to burn with high energy density, which will produce obvious high-frequency noise and local high temperature; at this time, the insulation has suffered irreversible damage, and the health status is only 25%, so power outage and maintenance must be arranged immediately; this weight reflects the severity of the fault. Destructive arc, weight 0: The arc has ignited the insulating material or caused the metal to melt, and the equipment is about to fail or has already failed; at this time, the health index is 0, and the corresponding remaining lifespan is 0.

[0099] Predicting Remaining Failure Time: The Wiener process model is used to describe the stochastic degradation process of the health index. Based on the critical point of sudden temperature rise in the joint during accelerated aging test, the failure threshold of the health index is set to 20. The expected value of the remaining failure time is equal to the difference between the current health index and 20, and then divided by the absolute value of the drift rate of the health index decline.

[0100] Drift rate was estimated online using historical health index sequences, employing a one-dimensional Kalman filter. The state variable was the drift rate, and the process noise variance was set to 10. -6 The value is derived based on the variance calibration of the long-term natural fluctuations of the health index; the variance of the observation equation is set to 0.1, which is derived based on the systematic error calibration of the health index calculation process; the time step interval is set to 0.5 seconds; the filter recursion includes a prediction step and an update step. In the prediction step, the prior state estimate is equal to the posterior estimate at the previous time step, and the prediction error covariance is equal to the posterior error covariance at the previous time step plus the process noise variance; in the update step, the Kalman gain is equal to the prediction error covariance divided by the sum of the prediction error covariance and the observation noise variance, the posterior state estimate is equal to the prior estimate plus the Kalman gain multiplied by the observation residual, and the posterior error covariance is equal to one minus the Kalman gain multiplied by the prediction error covariance.

[0101] A Kalman filter prediction error covariance gating mechanism based on confidence weights is introduced. When estimating the drift rate online, the contribution of each historical health index data point is weighted by the dynamic confidence weight at its corresponding moment. A continuous low-confidence data packet counter is maintained, with an initial value of 0. When module 3 sends a covariance reset request signal, the counter is incremented by 1. When the accumulated value of the counter exceeds the preset continuous low-confidence counting threshold of 3, the posterior error covariance of the Kalman filter is automatically reset to the initial covariance matrix.

[0102] After the covariance is reset, the Kalman gain is temporarily increased. Its working logic is: when the system suspects that the previous estimate has been contaminated, the covariance is reset to increase the confidence in subsequent new data.

[0103] Setting the continuous low confidence count threshold to 3 means that the system allows for brief interference or data loss, usually within 1.5 seconds. However, if low-quality data is reported for three consecutive fusion cycles, totaling 1.5 seconds, it is considered a persistent data contamination risk and the protection mechanism needs to be triggered. This threshold setting achieves a balance between false alarm rate and response speed.

[0104] The diagonal elements of the initial covariance matrix are 10 times the square of the initial drift rate estimate, and the initial drift rate is the historical average of the same type of branch bin. The reset operation directly cuts off the error propagation link of old and low-quality data. After the reset, the counter is cleared to ensure that the Kalman gain increases rapidly when subsequent high-confidence data arrives, and the model parameters converge quickly to the true degradation trend, thereby ensuring the stability of long-term prediction.

[0105] Simultaneously, the 90% confidence interval for the remaining time to failure is calculated to quantify the uncertainty of the prediction. The overall variance of the remaining time to failure is equal to the square of the random fluctuation intensity of the health index multiplied by the expected value of the remaining time to failure, plus the square of the expected value of the remaining time to failure multiplied by the estimated uncertainty of the drift rate. The square root of the overall variance to obtain the standard deviation is then taken. The standard deviation is multiplied by the quantile value corresponding to the 90% probability in the standard normal distribution to obtain the correction amount. The lower limit of the confidence interval is obtained by subtracting the correction amount from the expected value of the remaining time to failure, and the upper limit of the confidence interval is obtained by adding the correction amount to the expected value of the remaining time to failure. When the lower limit is less than 0, it is taken as 0, thus forming a complete 90% confidence interval.

[0106] Among them, the random fluctuation intensity of the health index is the diffusion coefficient of the Wiener process model, which is obtained by offline calibration through historical long-term operating data, and the typical initial value is set to 0.5; the estimation uncertainty of the drift rate is synchronously output by the Kalman filter during each online update, and the posterior error covariance output by the Kalman filter is directly used as the quantitative index of the drift rate estimation uncertainty.

[0107] Step 4.6: Tiered Operation and Maintenance Decision Output: Automatically generate executable operation and maintenance instructions based on quantified health indices.

[0108] Based on the power equipment operation and maintenance procedures and on-site fault statistics calibration, the health index grading thresholds are determined to be 80, 60, and 40, corresponding to different response levels: Health index > 80: Determined to be healthy, outputting a command that no intervention is required, and the system can enter an energy-saving mode that reduces the main frequency sampling frequency; If the health index is between 60 and 80, it is considered a cause for concern. The system will issue a recommendation for a recent check-up and prompt the user to take infrared temperature measurements. The system will maintain high-frequency sampling and automatically generate a trend report every 24 hours. If the health index is 40 < health index ≤ 60, it is considered abnormal. The system will output a maintenance instruction and prompt that a power outage inspection is required. The system will automatically lower the arc confidence judgment threshold of module 3 from 0.7 to 0.5 to improve the detection sensitivity of weak arcs. If the health index is ≤40, it is judged as dangerous, and an immediate power-off command is output. At this time, the highest level alarm is issued, and a trip signal is directly output to the local circuit breaker protection device through hard wiring to execute local power-off protection. At the same time, the original waveforms of key time periods before and after the fault are recorded for accident tracing and analysis.

[0109] After the above processing, the output includes the continuous health index value of each node, the expected value of the remaining fault time and its 90% confidence interval, the hierarchical operation and maintenance instruction text, and the alarm signal and trip protection signal.

[0110] All results are uploaded to the operation and maintenance center monitoring platform in real time via a remote communication interface for centralized monitoring and scheduling, and are displayed on the local display panel of the branch box. The historical sequence of health index and the confidence weight sequence are bidirectionally transmitted back to Module 3 and Module 2 for offline optimization of gradient boosting tree classifier weights, hidden Markov transition matrix and Wiener process drift parameters. The system automatically switches between four operating branches according to the health index interval, corresponding to energy-saving monitoring mode, high-frequency reporting mode, high-sensitivity maintenance mode and emergency protection mode, respectively, to realize a complete closed loop from transient feature identification to long-term health management.

[0111] Taking a cable distribution box deployed in a coastal industrial park with high humidity and heat as an example, the working process of this system is as follows: First, a broadband detection signal is injected into the bus side through module one, and a composite excitation is formed with the transient current of each branch load. The voltage and current sequences of the incoming line and the four outgoing branches are collected synchronously at 200 kHz to provide high signal-to-noise ratio raw data for subsequent analysis. After receiving this data, Module 2 identifies the time-varying dynamic admittance matrix online and constructs an inverse filter. It performs frequency domain inverse filtering on the current sequence to eliminate the frequency-selective amplitude attenuation caused by multi-branch parallel resonance. Then, it compensates for phase winding and group delay dispersion through full-pass phase equalization and non-uniform resampling to restore the true waveform at the arc source end. This solves the detection blind zone problem caused by signal propagation distortion from a physical perspective. Module 3 extracts transient impact geometric fingerprints such as rising edge steepness and zero-crossing offset from the restored waveform, and calculates the maximum Lyapunov exponent. It then integrates and generates dynamic confidence weights and weighted chaos, and uses a dual asymmetric threshold method and a chaos confidence coupling decision table to achieve dynamic adaptive adjustment of the analysis window. Within the window, it extracts four-dimensional features such as intermittent arcing frequency, half-wave anomaly proportion, high-frequency energy growth rate, and arc resistance change trend. Based on the weights, the data is split into high-quality data streams and isolated data packets, thereby effectively isolating pseudo-chaotic interference caused by load switching and preventing low-quality data from contaminating subsequent models. Module 4 integrates high-quality features with slowly varying sensor data such as infrared temperature and ultrasonic partial discharge. It uses a graph neural network to establish the spatial association of the six electrical nodes in the branch box, outputs the health status probability distribution, and then uses dynamic confidence weights as gating factors to correct the observation likelihood of the hidden Markov model to drive state tracking. At the same time, it uses a Wiener process with Kalman filtering covariance gating to predict the remaining fault time and its 90% confidence interval. Finally, it outputs the node health index and four-level hierarchical operation and maintenance instructions: no intervention required, inspection recommended, maintenance scheduled, or immediate power outage.

[0112] In this way, the entire process achieves closed-loop detection and intelligent operation and maintenance, from microsecond-level distortion compensation and pseudo-chaotic immunity to long-term health prediction.

[0113] Example 2: Please see Figure 4 Based on Embodiment 1, this embodiment also provides a highly sensitive method for detecting series fault arcs in cable branch boxes, comprising the following steps: Step 1: Synchronous Acquisition and Excitation Injection: Inject broadband detection signals into the bus side of the cable branch box, synchronously acquire voltage and current signals of each branch, and generate a synchronous data frame containing current sequence, voltage sequence and a copy of the detection signal; Step 2: Topology Inverse Filtering and Amplitude-Phase Correction: Based on the online identification of the dynamic admittance matrix of the synchronous data frame, an inverse filter is constructed and frequency domain inverse filtering is completed. Phase calibration and time delay dispersion correction are performed on the signal, and the restored waveform at the arc source end is output. Step 3: Pseudo-chaos suppression and arc feature extraction: Extract transient impact geometric fingerprints from the restored waveform, calculate the maximum Lyapunov exponent, generate dynamic confidence weights and weighted chaos degree, determine the analysis window length through dynamic gating adjustment, extract the arc features of series faults and output the arc feature vector and corresponding dynamic confidence weights; Step 4: Confidence-gated health assessment and prediction: Integrate arc feature vectors, dynamic confidence weights, and slowly varying sensor data, use graph neural networks to complete spatial correlation modeling of monitoring nodes, use dynamic confidence weight-gated hidden Markov models to achieve state tracking, calculate health indices and predict remaining fault time and confidence intervals, and generate hierarchical operation and maintenance instructions.

[0114] Furthermore, in step 1, a pseudo-random binary sequence is used as a broadband detection signal, which is repeatedly injected with the power frequency cycle as the period, and the signal amplitude is no greater than one-thousandth of the rated current of the line; a precise time protocol is used to realize the clock synchronization of multiple nodes, with a synchronization accuracy of no more than 1 microsecond and a sampling frequency of no less than 200kHz, to ensure that the time reference of each electrical signal is unified.

[0115] In step 2, a recursive least squares algorithm with a forgetting factor is used to identify the equivalent electrical parameters of each branch online, and a time-varying admittance matrix of the entire network is constructed. When the signal amplitude exceeds twice the preset safety threshold, the covariance matrix is ​​automatically reset and the integral element is frozen, and the open-loop prediction mode is switched to prevent parameter divergence. The inverse filter transfer function is obtained by inverting the admittance matrix at each frequency point, and decoupling is achieved through fast Fourier transform and inverse transform. The phase offset is calculated based on the physical length and propagation speed of the cable, and the phase distortion is compensated by an all-pass phase equalization filter. The group delay is calculated and non-uniform resampling is completed by cubic spline interpolation to eliminate time delay dispersion.

[0116] In step 3, the rising edge steepness, half-wave zero-crossing offset, and third harmonic phase jump angle are extracted to form a geometric fingerprint. The fingerprint is input into a pre-trained gradient boosting tree classifier to obtain class labels and original confidence scores. Within a fixed initial window, the Wolf algorithm is used to calculate the maximum Lyapunov exponent. Dynamic confidence weights are calculated based on the class labels and original confidence scores. The dynamic confidence weights are multiplied by the maximum Lyapunov exponent to obtain the weighted chaos degree. The dual asymmetric threshold method is used to divide the chaotic interval. The analysis window length is corrected by combining the chaos degree and confidence degree coupling decision table. Within the window, the frequency of intermittent arcing, the proportion of half-wave anomalies, the high-frequency energy growth rate, and the dynamic change trend of arc resistance are extracted to form a four-dimensional arc feature vector. Based on the dynamic confidence weights, the feature vector is marked as a high-quality data stream or isolated data packet. In high-chaos, low-confidence scenarios, a covariance reset request signal is sent.

[0117] In step 4, the slowly varying sensor data is input into the long short-term memory network encoder to obtain the slowly varying embedding vector. This vector is then concatenated with the arc feature vector statistics according to a fixed fusion period to obtain the fused feature vector. Based on the fused feature vector, a graph neural network structure containing six electrical vertices is constructed. Three layers of message passing and state updates are performed, and the health state probability distribution of each node is output. The health state probability distribution is used as the original observation likelihood and multiplied by the dynamic confidence weight to obtain the corrected observation likelihood, which drives the hidden Markov model to complete state tracking. A one-dimensional Kalman filter is used to estimate the drift rate online. When the continuously received covariance reset request signal exceeds the threshold, the posterior error covariance of the Kalman filter is automatically reset. The confidence interval is calculated based on the health index fluctuation intensity, the expected value of the remaining fault time, and the posterior error covariance. Four-level hierarchical operation and maintenance instructions are generated according to the health index value range. When the health index is lower than the danger threshold, power outage protection is triggered.

[0118] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0119] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0120] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A highly sensitive detection system for series fault arcing in a cable branch box, characterized in that: include: The distributed synchronous acquisition and composite excitation injection module injects broadband detection signals into the bus side of the cable branch box, synchronously acquires the voltage and current of each branch, and outputs a synchronous data frame containing the current sequence, voltage sequence and a copy of the detection signal. The topology inverse filtering and phase dispersion compensation module receives synchronous data frames, constructs an inverse filter by identifying the dynamic admittance matrix, performs frequency domain inverse filtering on the current sequence, performs phase calibration and time delay dispersion correction, and outputs the restored waveform at the arc source end. The anti-spurious chaos dynamic feature extraction and confidence assessment module receives the restored waveform, extracts the transient impact geometric fingerprint of the waveform and calculates the maximum Lyapunov exponent, generates dynamic confidence weights and weighted chaos degree, adaptively determines the analysis window length through a dynamic gating adjustment mechanism, extracts the arc features of the series fault within the analysis window, and outputs the arc feature vector and the corresponding dynamic confidence weights. The confidence-gated health assessment and prediction module receives arc feature vectors and dynamic confidence weights, and integrates slowly varying sensor data. It uses a graph neural network to model the spatial correlation of monitoring nodes to output the probability distribution of node health status. After gating the probability distribution of health status with dynamic confidence weights, it drives a hidden Markov model to track the state. Based on the state tracking results, it calculates the health index and uses a Wiener process model gated with dynamic confidence weights to predict the remaining fault time and confidence interval, and outputs hierarchical operation and maintenance instructions.

2. The high-sensitivity detection system for series fault arc in a cable branch box according to claim 1, characterized in that: In the distributed synchronous acquisition and composite excitation injection module, the broadband detection signal is a pseudo-random binary sequence encoded signal, which is injected in a manner that repeats once per power frequency cycle, with an amplitude not exceeding one-thousandth of the line's rated current; the synchronous acquisition uses a precise time protocol to achieve multi-node clock synchronization.

3. The high-sensitivity detection system for series fault arc in a cable branch box according to claim 1, characterized in that: In the topological inverse filtering and phase dispersion compensation module, a recursive least squares algorithm with a forgetting factor is used to identify the dynamic admittance matrix online. The topological inverse filtering and phase dispersion compensation module has a built-in anti-saturation mechanism. When the amplitude of the input signal exceeds twice the preset safety threshold, the covariance matrix inside the recursive least squares algorithm is reset to the identity matrix and the integration stage is frozen, switching to open-loop prediction mode.

4. The high-sensitivity detection system for series fault arc in a cable branch box according to claim 3, characterized in that: In the topological inverse filtering and phase dispersion compensation module, the inverse filter transfer function is obtained by inverting the admittance matrix at each frequency point. The current sequence is then multiplied point by point with the inverse filter transfer function after being subjected to a fast Fourier transform, and finally subjected to an inverse fast Fourier transform to obtain the decoupled current signal.

5. The high-sensitivity detection system for series fault arc in a cable branch box according to claim 4, characterized in that: In the topology inverse filtering and phase dispersion compensation module, the theoretical phase offset is calculated based on the physical length of the cable and the signal propagation speed. The decoupled current signal is subjected to reverse phase rotation through a full-pass phase equalization filter composed of cascaded second-order full-pass sections for phase calibration. The group delay of each frequency component is calculated, and non-uniform resampling is performed using cubic spline interpolation to complete the time delay dispersion correction and output the restored waveform.

6. The high-sensitivity detection system for series fault arc in a cable branch box according to claim 1, characterized in that: In the anti-spurious chaotic dynamic feature extraction and confidence evaluation module, the transient impact geometric fingerprint includes rising edge steepness, half-wave zero-crossing offset and third harmonic phase jump angle; the transient impact geometric fingerprint is input into a pre-trained gradient boosting tree classifier, which outputs event category labels and original confidence, and the Wolf algorithm is used to calculate the maximum Lyapunov exponent within a fixed initial window.

7. The high-sensitivity detection system for series fault arc in a cable branch box according to claim 6, characterized in that: The dynamic confidence weight is calculated based on the event category label and the original confidence. The weighted chaos is equal to the product of the dynamic confidence weight and the maximum Lyapunov exponent. The dynamic gating adjustment mechanism adopts a dual asymmetric threshold method to divide the weighted chaos into a low chaos region, a fuzzy region, and a high chaos region. In the fuzzy region, a proportional-integral controller is used to dynamically adjust the analysis window length. Combined with the chaos and confidence coupling decision table, the analysis window length and output data label are corrected according to the relationship between the weighted chaos interval and the dynamic confidence weight.

8. The high-sensitivity detection system for series fault arc in a cable branch box according to claim 1, characterized in that: In the anti-spurious chaos dynamic feature extraction and confidence assessment module, the arc feature vector is a four-dimensional feature vector containing the frequency of intermittent arcing, the proportion of half-wave anomalies, the high-frequency energy growth rate, and the dynamic change trend of arc resistance. According to the dynamic confidence weight, the corresponding four-dimensional feature vector is divided and marked as a high-quality data stream and an isolated data packet. When the weighted chaos degree is in the high chaos region and the dynamic confidence weight is less than or equal to the preset low confidence trigger threshold, a covariance reset request signal is sent to the confidence gating health assessment and prediction module.

9. The high-sensitivity detection system for series fault arc in a cable branch box according to claim 1, characterized in that: In the confidence-gated health assessment and prediction module, slowly varying sensor data is input into a long short-term memory network encoder to obtain a slowly varying embedding vector. This vector is then concatenated with the statistics of the arc feature vector according to a fixed fusion period to form a fused feature vector. The graph neural network constructs a topology containing six electrical vertices based on the fused feature vector and performs message passing and state updates, outputting a node health state probability distribution. The hidden Markov model uses the node health state probability distribution as the original observation likelihood and multiplies the original observation likelihood with the corresponding dynamic confidence weight to obtain a corrected observation likelihood, which is used to drive state tracking.

10. A high-sensitivity detection system for series fault arc in a cable branch box according to claim 1, characterized in that: The Wiener process model estimates the drift rate of the health index online using a one-dimensional Kalman filter; the confidence-gated health assessment and prediction module maintains a continuous counter for low-confidence data packets, and automatically resets the posterior error covariance of the Kalman filter to its initial value when it continuously receives covariance reset request signals and the count exceeds a preset continuous low-confidence count threshold; the confidence interval of the remaining fault time is obtained by calculating the overall fluctuation variance based on the random fluctuation intensity of the health index, the expected value of the remaining fault time, and the posterior error covariance; the hierarchical operation and maintenance instructions are generated based on the numerical range of the health index.

Images