A method and system for diagnosing faults in electrical wire and cable

Abstract

The application relates to the technical field of pre-diagnosis and health management, and discloses a power cable fault diagnosis method and system. The method comprises the following steps: acquiring cable voltage and current signals, extracting mutation and gradient features after denoising and segmentation, and generating an abnormal distribution atlas; scale-decomposing a mutation region and extracting key feature points, clustering and smoothing to construct a maximum value chain; correcting time coordinate offset, complementing missing coordinates to obtain a complete time coordinate set; comparing signal strength with a fault threshold to determine a high-resistance or low-resistance fault type; and combining cable medium parameters to determine signal propagation speed, and converting the time coordinates into specific spatial positions of faults. The method can realize accurate fault diagnosis and type identification of power cables, and meets the safety and forward-looking requirements of cable operation and maintenance.

Description

Technical Field

[0001] This invention relates to the field of pre-diagnosis and health management technology, and in particular to a method and system for diagnosing faults in wires and cables. Background Technology

[0002] Currently, in the fields of pre-diagnosis and health management technology, with the continuous expansion of power system scale and the increasing complexity of cable application scenarios, their safe and stable operation is directly related to the reliability of the power system and the order of social production and life. As a key link in preventing the expansion of faults and avoiding power outages, wire and cable fault diagnosis needs to achieve early identification of hidden dangers by accurately capturing abnormal signal characteristics. The integration of multi-scale signal analysis and signal reconstruction technology has become the core technical direction for improving the accuracy and reliability of fault diagnosis.

[0003] Existing fault diagnosis methods for wires and cables in the industry mainly rely on single techniques such as traveling wave method, impedance method, or time-domain reflection method. For example, they determine the fault location by detecting the time difference of reflected signals or identify the fault type by using simple electrical parameter tests. However, this approach is clearly insufficient in complex operating environments. Due to long-distance cable transmission, signal attenuation is severe, waveform distortion occurs, and interference from noise and multipath reflections makes it difficult to extract clear fault anomaly features. Furthermore, it has low sensitivity to high-resistance grounding faults and intermittent faults, and is prone to misjudgment due to feature masking. At the same time, there is a lack of correction mechanisms for signal breakpoints and offsets, making it difficult to accurately locate fault locations, especially in cable lines with complex media distribution.

[0004] In summary, existing technologies are insufficient for accurate fault diagnosis and type identification of wires and cables, and cannot meet the safety and forward-looking requirements for cable operation and maintenance. Summary of the Invention

[0005] This invention provides a method and system for diagnosing faults in wires and cables, enabling accurate fault diagnosis and type identification of wires and cables, and meeting the safety and forward-looking requirements for cable operation and maintenance.

[0006] In a first aspect, to solve the above-mentioned technical problems, the present invention provides a method for diagnosing faults in wires and cables, comprising:

[0007] Acquire the voltage and current signals of the cable;

[0008] The voltage and current signals are denoised and segmented over time. The abrupt changes and gradient features of each time segment are extracted. The degree of anomaly of the signal is calculated based on the abrupt changes and gradient features, and an anomaly distribution map is generated.

[0009] The abrupt change regions in the anomalous distribution map are decomposed by scale and the modulus maxima at different scales are captured. Key feature points are extracted based on the scale evolution law of the modulus maxima to obtain a set of feature points for scale evolution.

[0010] Clustering and smoothing are performed on the feature point set to remove pseudo-extreme points and fill in the breakpoints, resulting in a continuous modulus maxima chain.

[0011] Based on the time domain position of each modulus maximum point in the modulus maximum chain, the time coordinates of the fault candidates are extracted, and the time coordinates are associated with the pre-acquired correction model to correct the time coordinate offset and obtain the corrected coordinate distribution.

[0012] The coordinate distribution is matched with the preset fault feature topology, and the missing coordinates are supplemented by interpolation to obtain a complete set of fault time coordinates.

[0013] The signal strength corresponding to the time coordinate set is compared with the preset high-impedance fault threshold and low-impedance fault threshold respectively. If it exceeds the high-impedance fault threshold, it is determined to be a high-impedance fault; if it is below the low-impedance fault threshold, it is determined to be a low-impedance fault, and the fault type result is obtained.

[0014] Based on the fault type result and the pre-acquired cable medium parameters, the propagation speed of the signal in the cable is determined. The propagation speed is then calculated with the coordinate set to convert the time coordinates into spatial positions, thereby obtaining the specific location of the cable fault.

[0015] Secondly, the present invention provides a wire and cable fault diagnosis system, comprising:

[0016] The signal acquisition module is used to acquire the voltage and current signals of the cable.

[0017] The anomaly map module is used to denoise and segment the voltage and current signals over time, extract the abrupt changes and gradient features of each time segment, calculate the degree of anomaly of the signal based on the abrupt changes and gradient features, and generate an anomaly distribution map.

[0018] The feature point extraction module is used to perform scale decomposition on the mutation regions in the abnormal distribution map and capture the modulus maxima at different scales. Based on the scale evolution law of the modulus maxima, key feature points are extracted to obtain a set of feature points of scale evolution.

[0019] The chain construction module is used to cluster and smooth the feature point set, remove pseudo-extreme points and fill in the breakpoints to obtain a continuous modulus maxima chain;

[0020] The coordinate correction module is used to extract the time coordinates of the fault candidates based on the time domain position of each of the modulus maxima points in the modulus maxima chain, associate the time coordinates with the pre-acquired correction model, correct the time coordinate offset, and obtain the corrected coordinate distribution.

[0021] The coordinate completion module is used to perform feature matching between the coordinate distribution and the preset fault feature topology, and to fill in the missing coordinates through interpolation to obtain a complete set of time coordinates of the fault.

[0022] The fault classification module is used to compare the signal strength corresponding to the time coordinate set with the preset high-impedance fault threshold and low-impedance fault threshold respectively. If the signal strength exceeds the high-impedance fault threshold, it is determined to be a high-impedance fault; if the signal strength is below the low-impedance fault threshold, it is determined to be a low-impedance fault, and the fault type result is obtained.

[0023] The location determination module is used to determine the propagation speed of the signal in the cable based on the fault type result and the pre-acquired cable medium parameters, perform calculations on the propagation speed and the coordinate set, convert the time coordinates into spatial positions, and obtain the specific location of the cable fault.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] (1) This invention obtains the voltage and current signals of the cable, filters out high-frequency interference by threshold noise reduction, uses a fixed-duration sliding window to segment, extracts the abrupt change features and gradient features of each segment, and generates a signal anomaly distribution map. This breaks through the limitations of traditional single diagnostic techniques in dealing with signal attenuation and distortion over long distances, explores the dynamic features of signal anomalies in cable faults, eliminates noise and multipath reflection interference, provides high-precision basic data support for fault diagnosis, effectively improves the feature capture rate of weak and intermittent faults, and solves the problem of missed judgment caused by the masking of fault features.

[0026] (2) This invention performs multi-scale decomposition of abrupt change regions in the distribution map, captures modulus maxima at different scales, combines Lee's index to screen key feature points, removes pseudo-extreme points and fills in discontinuities through clustering smoothing, and constructs a continuous modulus maxima chain. This breaks through the limitation that single-scale analysis cannot capture multi-scale fault features, accurately captures the scale evolution law of fault features, provides multi-dimensional basis for fault judgment, significantly improves the identification accuracy of complex faults such as high resistance and low resistance, and makes up for the defects of misjudgment caused by pseudo-extreme point interference in the existing technology.

[0027] (3) This invention corrects the time coordinate offset by a correction model, completes the missing coordinates by feature matching, distinguishes the fault type by combining signal strength, and converts the time coordinate into spatial location by using cable medium parameters. It breaks through the limitations of traditional diagnosis that lacks coordinate correction and accurate type identification, and provides maintenance personnel with dual basis for fault type and accurate location. It solves the problems of large fault location deviation and misjudgment of type, takes into account both diagnostic accuracy and maintenance response efficiency, and meets the safety and forward-looking needs of complex cable line maintenance. Attached Figure Description

[0028] Figure 1 This is a schematic flowchart of a wire and cable fault diagnosis method provided in the first embodiment of the present invention;

[0029] Figure 2 This is a schematic diagram of a wire and cable fault diagnosis system provided in the second embodiment of the present invention. Detailed Implementation

[0030] 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 skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Reference Figure 1 The first embodiment of the present invention provides a method for diagnosing faults in electric wires and cables, comprising the following steps:

[0032] S101, acquire the voltage and current signals of the cable;

[0033] S102, the voltage signal and current signal are denoised and segmented over time, the abrupt changes and gradient features of each time segment are extracted, the degree of anomaly of the signal is calculated based on the abrupt changes and gradient features, and an anomaly distribution map is generated.

[0034] S103, perform scale decomposition on the mutation regions in the abnormal distribution map and capture the modulus maxima at different scales, extract key feature points according to the scale evolution law of the modulus maxima, and obtain a set of feature points of scale evolution.

[0035] S104, perform clustering and smoothing on the feature point set, remove pseudo-extreme points and fill in the breakpoints to obtain a continuous modulus maxima chain;

[0036] S105, based on the time domain position corresponding to each of the modulus maxima points in the modulus maxima chain, extract the time coordinates of the fault candidates, associate the time coordinates with the pre-acquired correction model, correct the time coordinate offset, and obtain the corrected coordinate distribution;

[0037] S106, perform feature matching between the coordinate distribution and the preset fault feature topology, and supplement the missing coordinates by interpolation to obtain a complete set of time coordinates of the fault.

[0038] S107, compare the signal strength corresponding to the time coordinate set with the preset high-impedance fault threshold and low-impedance fault threshold respectively. If it exceeds the high-impedance fault threshold, it is determined to be a high-impedance fault; if it is lower than the low-impedance fault threshold, it is determined to be a low-impedance fault, and the fault type result is obtained.

[0039] S108. Based on the fault type result and the pre-acquired cable medium parameters, determine the propagation speed of the signal in the cable, perform calculations on the propagation speed and the coordinate set, convert the time coordinates into spatial positions, and obtain the specific location of the cable fault.

[0040] In step S101, acquiring the voltage and current signals of the cable includes:

[0041] Voltage and current signals during cable operation are collected by voltage and current sensors deployed on the cable lines.

[0042] It should be noted that, firstly, capacitive voltage divider sensors are used for voltage sensors, and Rogowski coil sensors are used for current sensors. Both types of sensors are suitable for the high-voltage operating environment of power cables. Sensor deployment follows the principles of uniform distribution and increased density at key nodes, with a typical deployment interval of 50-200 meters. Additional sensors are installed at cable joints, branches, and terminations—areas prone to failure—to ensure no blind spots in signal acquisition. The sampling frequency is 20kHz, based on the cable fault signal frequency range of 0-10kHz, satisfying the Nyquist sampling theorem. The voltage sensor range covers 0-120% of the rated voltage, and the current sensor range is 0-1000A, both with an accuracy of 0.2 class, accurately capturing signal changes under normal and fault conditions. For example, after deploying sensors in this manner on a 110kV cable line, stable signals of the 110kV line voltage and 200A phase current were successfully acquired during normal operation, and instantaneous waveforms of voltage drops and current surges were also captured during faults.

[0043] During data acquisition, the sensors simultaneously collect cable line voltage and phase current signals. The acquisition timing covers all operating conditions, including normal operation, load fluctuations, and equipment start-up and shutdown, to avoid missing intermittent fault signals. The acquired data is transmitted to the data processing unit via optical fiber, with transmission delay controlled within 10ms to ensure time-domain information synchronization. For example, when a cable fails due to insulation damage, the sensor captures the sudden change in voltage amplitude from 110kV to 105kV and current from 200A to 800A within 1ms of the fault occurring, providing complete raw data for subsequent fault diagnosis.

[0044] In step S102, the denoising and time segmentation of the voltage and current signals, the extraction of abrupt changes and gradient features in each time segment, the calculation of the anomaly degree of the signal based on the abrupt changes and gradient features, and the generation of an anomaly distribution map include:

[0045] The voltage and current signals are subjected to threshold denoising to filter out high-frequency noise and interference components, resulting in denoised signals.

[0046] The denoised signal is segmented using a fixed-duration sliding time window to obtain several signal segments.

[0047] Calculate the waveform energy, amplitude variance, and gradient of each signal segment, and extract the abrupt change features and gradient features of each segment;

[0048] Based on the mutation characteristics and gradient characteristics, the degree of abnormality of each segment is determined, and an abnormality distribution map is generated.

[0049] It should be noted that, firstly, when performing threshold denoising on the voltage and current signals to filter out high-frequency noise and interference components, wavelet threshold denoising technology is used to obtain the denoised signal. The db4 wavelet basis function is selected, and the decomposition level is set to 5 levels. The threshold calculation adopts a heuristic thresholding method, and the formula is as follows: ,in Indicates the threshold. The noise standard deviation is estimated using the decomposed high-frequency coefficients. The signal length is used. During the learning process, the number of decomposition layers is adjusted according to the signal sampling frequency of 20kHz to ensure effective separation of high-frequency noise. After five consecutive rounds of denoising, the signal-to-noise ratio is stabilized at over 15dB, at which point parameter fine-tuning is stopped. For example, after processing, the waveform smoothness of a voltage signal containing high-frequency electromagnetic interference is significantly improved, and the originally superimposed jagged noise is effectively filtered out.

[0050] It is worth noting that, The estimation method is as follows: take the high-frequency detail coefficients of the first layer of wavelet decomposition, calculate the median of their absolute values, and then divide them by the constant 0.6745.

[0051] Next, when segmenting the denoised signal into several signal segments using a fixed-length sliding time window, the window duration is set based on the typical duration of the cable fault signal, with a base duration of 0.1 seconds. This can be adjusted to 0.2 seconds for high-voltage cables and shortened to 0.05 seconds for low-voltage cables. The window step size is set to 50% of the window duration to ensure overlapping segments without omissions. For example, if the denoised current signal lasts 10 seconds, segmenting it with a 0.1-second window and a 0.05-second step size yields 199 signal segments, completely covering the entire time domain signal.

[0052] Subsequently, the waveform energy, amplitude variance, and gradient of each signal segment are calculated. When extracting the abrupt change and gradient features of each segment, the waveform energy is V²·s for voltage signals and A²·s for current signals, calculated by integrating the sum of squares of the signal segments. The amplitude variance is V² for voltage signals and A² for current signals, representing the squared mean of the deviations between the signal amplitude and the mean within each segment. The gradient is V / s for voltage signals and A / s for current signals, representing the mean of the absolute values ​​of the amplitude differences between adjacent segments. The abrupt change feature is quantized by multiplying the normalized waveform energy and amplitude variance, while the gradient feature uses the calculated result of the normalized gradient.

[0053] Due to the significant differences in the dimensions and orders of magnitude of the original features, min-max normalization is required for the three types of parameters. The upper and lower bounds of the normalization are not fixed constants, but are adaptively set based on the statistical distribution of features under normal and fault conditions of the cable over the past year: the 5% to 95th percentiles of historical samples are used as the upper and lower bounds for each type of feature, with a 10% margin reserved at each bound to cover abnormal peak values, thus avoiding normalization distortion caused by different voltage levels / current ranges. Waveform energy, amplitude variance, and gradient are all statistically bounded for voltage and current signals respectively, and then mapped to the [0,1] interval for subsequent feature fusion calculations. For example, in the statistics of a certain 110kV cable line over the past year, the 5% to 95th percentile range of voltage waveform energy is approximately 0 to 8.0 × 10⁻⁶. 6 V²·s, with a margin, the upper bound is taken as 8.8 × 10⁻⁶. 6 V²·s; If the energy of a certain segmented voltage waveform is 6.0 × 10⁻⁶ V²·s; 6 V²·s, after normalization, is approximately 6.0 × 10⁻⁶. 6 / 8.8×10 6 ≈0.68. If the normalized variance of the amplitude and the gradient of change in the same segment are 0.62 and 0.60 respectively, then the abrupt change feature = 0.68 × 0.62 ≈ 0.42, and the gradient feature = 0.60, indicating that there is a significant abrupt change in this segment.

[0054] It should be noted that the anomaly level is calculated using a weighted summation, where the weight for the abrupt change feature is set to 0.7, the weight for the gradient feature is set to 0.3, and the sum of all weights is 1. This weight combination is based on the characteristics of the fault signal. The abrupt change feature comprehensively reflects the amplitude of signal intensity changes, while the gradient feature reflects the rate of change. After normalization, the two are of the same order of magnitude, allowing for coordinated quantification of the anomaly level. For example, for a voltage signal segment with an abrupt change feature of 0.64 and a gradient feature of 0.56, the calculated anomaly level is 0.64 × 0.7 + 0.56 × 0.3 ≈ 0.61, intuitively reflecting the anomaly level of this segment and avoiding the problem of gradient features being ignored due to differences in the order of magnitude of the original data.

[0055] Finally, when generating the distribution spectrum of signal anomaly severity based on the correspondence between time and anomaly severity, the matplotlib visualization tool was used. The horizontal axis represents time (in seconds), and the vertical axis represents the anomaly severity. Data points were plotted sequentially in segmented order and connected to form a curve. The peak position in the spectrum corresponds to the moment of signal abrupt change, and the peak height corresponds to the magnitude of the anomaly severity. For example, the generated distribution spectrum shows an anomaly severity peak of 0.85 at 3.2 seconds, clearly indicating a signal abrupt change related to a cable fault at that moment.

[0056] In step S103, the abrupt change regions in the abnormal distribution map are scaled and the modulus maxima at different scales are captured. Key feature points are extracted based on the scale evolution of these modulus maxima to obtain a set of scale-evolved feature points, including:

[0057] Based on the anomaly distribution map, the time domain range of the mutation region is determined, and discrete signals within the time domain range are extracted to obtain the mutation region sequence;

[0058] The mutation region sequence is decomposed into a multi-scale coefficient matrix.

[0059] The multi-scale coefficient matrix is ​​traversed to filter out the modulus maxima points, and the modulus maxima evolution curve is constructed.

[0060] The decay characteristics of the modulus maxima evolution curve are calculated, the Lee index is estimated, and key feature points are selected based on the Lee index to obtain the feature point set of scale evolution.

[0061] It should be noted that, firstly, when determining the time domain range of the abrupt change region based on the distribution map, and extracting discrete signals within this range to obtain the abrupt change region sequence, the criterion for identifying the abrupt change region is that the degree of anomaly exceeds a preset abrupt change identification threshold. This threshold is set based on the statistical distribution of anomalies during normal cable operation over the past year. The basic threshold is 1.5 times the maximum anomaly level during normal operation; for high-voltage cables, this can be increased to 1.8 times, and for low-voltage cables, it can be decreased to 1.3 times. By traversing the distribution map, the time period continuously exceeding this threshold is determined as the time domain range of the abrupt change region. Discrete data points within this range are extracted according to the signal sampling frequency to form the abrupt change region sequence. For example, if the anomaly level from 3.2 seconds to 3.5 seconds in the distribution map all exceeds the abrupt change identification threshold, 6000 discrete signal points in this interval are extracted at a sampling frequency of 20kHz to form the abrupt change region sequence.

[0062] Next, when performing multi-scale decomposition on the abrupt change region sequence to obtain the multi-scale coefficient matrix, the db4 wavelet transform technique is employed. Specific parameters are set as follows: 5 decomposition levels, db4 wavelet basis function, symmetric extension for boundary processing, and sampling frequency adapted to the 20kHz original signal. The decomposition process is performed progressively from low to high scales. Each level of decomposition yields low-frequency approximation coefficients and high-frequency detail coefficients for the corresponding scale. Finally, the coefficients from all scales are integrated to form the multi-scale coefficient matrix. For example, after performing 5-level db4 wavelet decomposition on the abrupt change region sequence from 3.2 seconds to 3.5 seconds, the coefficient sequences at different scales are resampled / linearly interpolated to align to the same length, resulting in a 6×6000 matrix containing one row of low-frequency coefficients and five rows of high-frequency coefficients.

[0063] Subsequently, when traversing the multi-scale coefficient matrix to screen for modulo maxima and constructing the modulo maxima evolution curve, the screening of modulo maxima uses a 3-point neighborhood comparison method. That is, for each element in the coefficient matrix, if its value is greater than its left and right adjacent elements, it is determined to be a modulo maxima. Following the order of scales from 1 to 5, the modulo maxima at each scale are aligned according to their temporal location. The maxima at each temporal location at different scales are extracted and connected in scale order to form the modulo maxima evolution curve. For example, at the temporal location of 3.3 seconds, the local maxima corresponding to scales 1 to 5 are 8.2, 7.5, 6.8, 5.3, and 4.1, respectively. These values ​​are sorted by scale and connected to form the modulo maxima evolution segment at that location. All segments are integrated into a complete evolution curve.

[0064] Finally, the decay characteristics of the modulus maxima evolution curve are calculated, and the Lippi exponent is estimated. When selecting key feature points based on the Lippi exponent to obtain the feature point set for scale evolution, the decay characteristics are obtained by calculating the slope of the curve in different scale intervals. The slope is the ratio of the logarithm of the modulus maxima to the logarithm of the scale. The Lippi exponent α is estimated based on the decay slope m, where m is the fitting slope of the logarithm of the modulus maxima and the logarithm of the scale. When using the db4 wavelet basis function, the Lippi exponent satisfies α = m - 1 / 2; the smaller the α value, the stronger the signal anomaly. The selection threshold for key feature points is set based on historical fault data. The basic threshold is -0.4, which can be lowered to -0.6 for high-resistivity fault scenarios and raised to -0.2 for low-resistivity fault scenarios. Feature points with Lippi exponents less than the threshold are selected to form a set. For example, if the decay slope m of a certain modulus maxima evolution curve is −0.1, then the anomaly index α = m - 1 / 2 = -0.6, which is less than the basic threshold of -0.4. This feature point is included in the feature point set for scale evolution.

[0065] In step S104, the clustering and smoothing of the feature point set, the removal of spurious extrema and the filling of discontinuities, to obtain a continuous modulus maxima chain, includes:

[0066] Calculate the spatial distance between each point in the feature point set, and cluster them according to the spatial distance to obtain the modulus maxima cluster;

[0067] Based on the temporal proximity and amplitude propagation law of feature points within each cluster, cross-scale association is performed to generate a path with a modulus maxima.

[0068] Based on the attenuation law of the modulus maxima path, pseudo-maxima points that do not conform to the law are eliminated to obtain the signal skeleton containing the discontinuities;

[0069] The signal skeleton is interpolated and fitted to fill in the breakpoints and form a reconstruction evolution line;

[0070] The reconstructed evolution line is smoothed to suppress high-frequency jitter, resulting in a continuous modulus maxima chain.

[0071] It should be noted that, firstly, when calculating the spatial distance between points in the feature point set and clustering them by distance to obtain the modulus maxima cluster, the spatial distance is calculated using Euclidean distance. The feature dimension of the Euclidean distance includes the temporal position t of the feature point and the magnitude A(t) of the modulus maxima corresponding to that temporal position, thus forming a two-dimensional feature vector x=[t,A(t)]. A(t) is directly taken from the magnitude of the modulus maxima at that temporal position of the multi-scale coefficient matrix, used to characterize the intensity of local abrupt changes at that point. To avoid distance bias caused by the different dimensions of the temporal domain and magnitude, t and A(t) are standardized (e.g., using Z-score) before clustering, and the standardized results are then mapped to [0,1], and the Euclidean distance is calculated in the normalized feature space.

[0072] The K-means++ clustering algorithm was selected, and the number of clusters K (ranging from 3 to 8) was determined using the elbow rule. Initial cluster centers were selected using the K-means++ algorithm to avoid local optima. The number of iterations was set to 100, and the convergence condition was that the change in the sum of squared errors within each cluster was less than 0.001. This clustering method can group feature points with similar temporal location and amplitude characteristics into one class, highlighting the concentrated distribution characteristics of fault signals. For example, a feature point set containing 15 modulus maxima points was clustered into 3 clusters after K-means++, one of which contained feature points with a temporal range of 3.2-3.5 seconds and an amplitude of 5-8, corresponding to the same fault abrupt change region.

[0073] Next, cross-scale association is performed based on the temporal proximity and amplitude propagation law of feature points within each cluster. When generating the modulus maxima path, cross-scale association is performed in the order of "small scale → large scale". For each feature point within a cluster, it is first sorted by scale from smallest to largest, and then feature points of adjacent scales are matched based on the Euclidean distance of temporal location (threshold set to 0.01 seconds). At the same time, it is verified whether the amplitude conforms to the law of "monotonically decaying with increasing scale".

[0074] It is worth noting that the verification amplitude rule is as follows: if the amplitudes of adjacent scales are Am and Am+1 respectively, then when Am+1≤Am⋅(1+ε), the attenuation constraint is considered to be satisfied, where ε is a tolerance coefficient (e.g., 0.05~0.1, used to allow slight fluctuations caused by noise); if Am+1>Am⋅(1+ε) occurs in multiple consecutive scales, then the matching is considered unreliable and cross-scale association is not established.

[0075] The association weights are allocated as follows: temporal location similarity 0.6, amplitude attenuation compliance 0.4, with the sum of all weights being 1 to ensure the accuracy of the association. The quantification rule for temporal location similarity is as follows: based on the time difference between adjacent scale feature points, with a temporal threshold set to 0.01 seconds, time differences between 0 and 0.01 seconds are linearly mapped to a score of 0 to 1. A time difference of 0 receives 1 point, a time difference of 0.01 seconds receives 0 points, and a time difference exceeding 0.01 seconds is directly considered unrelated and will not participate in subsequent score calculations. The quantification rule for amplitude attenuation compliance is to calculate the amplitude ratio of adjacent scale feature points. If the ratio is between 0 and 1 (compliant with the attenuation rule), 1 point is awarded; if the ratio is between 1 and 1.05 (slight fluctuations within the tolerance range), 0.5 points are awarded; if the ratio exceeds 1.05 (significantly increasing in the opposite direction, not compliant with the attenuation rule), 0 points are awarded. The comprehensive association score is obtained by multiplying the temporal location similarity score by 0.6 and the amplitude attenuation compliance score by 0.4, and then summing them. Under the premise of meeting the temporal threshold, the adjacent scale feature point with the highest comprehensive score is selected to establish cross-scale association, thus taking into account both positional continuity and consistency of attenuation patterns. For example, within a cluster, a small-scale (scale 1) feature point has a temporal similarity of 3.3 seconds and an amplitude of 7.2, while the adjacent large-scale (scale 2) feature point has a temporal similarity of 3.305 seconds and an amplitude of 6.8. The time difference of 0.005 seconds corresponds to a temporal location similarity score of 0.5, and the amplitude ratio of 0.944 corresponds to an amplitude attenuation compliance score of 1. The comprehensive association score = 0.5 × 0.6 + 1 × 0.4 = 0.7, which meets the association requirements. Cross-scale association is established, and a preliminary modulus maxima path is generated.

[0076] Secondly, after eliminating spurious extrema based on the attenuation law of the modulus maxima path to obtain the signal skeleton containing discontinuities, the attenuation law is determined by the exponential attenuation function of the fitted path, as shown in the formula: ,in, This represents the signal amplitude of the feature point at the corresponding scale. This represents the signal amplitude at the initial scale of the path. The attenuation coefficient is used to characterize the rate attenuation of amplitude as the scale increases. The natural constant is represented by a value of approximately 2.718. This indicates the wavelet scale parameter or scale level number corresponding to the feature point (from small to large, corresponding to fine scale to coarse scale), used to characterize the change in the magnitude of the modulus maxima at different scales.

[0077] It should be noted that, The attenuation coefficient represents the rate at which the amplitude decays with increasing scale; it is obtained by performing an exponential decay fit on the scale and amplitude data of the path to the modulus maxima, for example, by taking the natural logarithm of both sides of the formula. Then, a linear regression is performed using the least squares method, and the negative of the regression slope is the attenuation coefficient. .

[0078] It is worth noting that the threshold for identifying false extreme points was obtained from statistical analysis of historical fault data over the past year. Specifically, an amplitude decay curve was first fitted to each modulus maximum path, and the deviation between the actual amplitude and the fitted amplitude at each scale was taken as the fitting error. After statistically analyzing the error levels of historical effective paths, 1.5 times their typical error was used as the base threshold. In high-noise scenarios, to avoid false rejection, the threshold was widened to 1.8 times, while in low-noise scenarios, to enhance the rejection effect, the threshold was tightened to 1.2 times.

[0079] If the deviation between the actual amplitude of a feature point and the fitted value exceeds the threshold, it is identified as a pseudo-extreme point and removed. For example, if the fitted decay function of a certain path is "Am=8×e^(-0.2×Scale)", and the actual amplitude of a certain feature point is 3.2 higher than the fitted value, exceeding the basic threshold of 2.5, it is identified as a pseudo-extreme point and removed. The remaining feature points constitute the signal skeleton containing the breakpoints.

[0080] Subsequently, when interpolating and fitting the signal skeleton to fill the gaps and form the reconstructed evolution line, a cubic spline interpolation algorithm is used. The specific process is as follows: two effective feature points before and after the gap are selected as interpolation nodes; a cubic spline function is constructed (satisfying the continuity of the first and second derivatives at the nodes); interpolation points are generated at 0.005-second time intervals to fill the gaps. This interpolation method maintains the smoothness of the path and the consistency of the attenuation trend. For example, if the signal skeleton has a gap in the time domain from 3.35 to 3.4 seconds, three virtual points are generated through interpolation of the feature points before and after the gap, with amplitudes of 5.2, 4.9, and 4.6, respectively, forming a continuous reconstructed evolution line after filling.

[0081] Finally, to smooth the reconstructed evolution line and suppress high-frequency jitter, a Gaussian filtering technique is used to obtain a continuous chain of modulo maxima. The filter kernel size is set to 5×1, and the standard deviation σ=1.2. The filtering process first performs sliding window convolution, and then performs amplitude renormalization to ensure that the amplitude range of the smoothed path is consistent with the original evolution line, while removing high-frequency jitter (frequency higher than 100Hz). For example, the reconstructed evolution line has slight jitter in the time domain of 3.4-3.45 seconds. After Gaussian filtering, the jitter is significantly suppressed, and the evolution line exhibits a smooth exponential decay trend, ultimately forming a continuous chain of modulo maxima.

[0082] In step S105, the step of extracting the time coordinates of fault candidates based on the time domain positions corresponding to each of the modulus maxima points in the modulus maxima chain, associating the time coordinates with a pre-acquired correction model, correcting the time coordinate offset, and obtaining the corrected coordinate distribution includes:

[0083] Extract the time-domain position corresponding to each of the modulus maxima points in the modulus maxima chain to determine the time coordinates of the fault candidate;

[0084] The time coordinates of the fault candidates are correlated with the pre-acquired correction model to identify the signal interference range;

[0085] Calculate the Lee's exponential deviation of the time coordinate within the signal interference interval. If the Lee's exponential deviation exceeds a preset deviation threshold, it is marked as an offset point.

[0086] Based on the correction model, a distortion-free reference waveform under ideal propagation conditions is simulated and generated. The time offset of the offset point is calculated based on the reference waveform. The time offset is then superimposed on the time coordinate of the offset point to obtain the corrected feature point sequence.

[0087] The feature point sequence is compared with a pre-acquired normal time-domain distribution template, and a time-domain consistency check and correction are performed to obtain the corrected coordinate distribution.

[0088] It should be noted that, firstly, when extracting the time-domain positions corresponding to each modulomaxima point in the modulomaxima chain to determine the time coordinates of fault candidates, the continuous modulomaxima chain is traversed in chronological order, and the timestamp corresponding to each modulomaxima point is recorded. These timestamps are the time coordinates of the fault candidates. This extraction process relies on the time synchronization mechanism during signal acquisition to ensure the accuracy of the time-domain positions. For example, if the modulomaxima chain contains three significant peak points with corresponding timestamps of 3.2 seconds, 3.3 seconds, and 3.4 seconds, these three time points are determined as the time coordinates of the fault candidates.

[0089] Subsequently, the time coordinates of the fault candidates were correlated with the pre-acquired correction model. When identifying signal interference intervals, the multipath propagation correction model was selected. The training set consisted of a 1:1 mix of multipath interference data and normal operating condition data from cable operation over the past year, with a total sample size of 50,000, covering scenarios under different cable types, laying environments, and meteorological conditions. The six dimensions of the input features were: signal propagation time (measured round-trip time of the signal between the two ends of the cable, in μs), path loss (calculated based on transmission distance and dielectric attenuation coefficient, in dB), dielectric constant (measured parameter of the cable insulation layer, dimensionless), signal amplitude fluctuation coefficient (standard deviation of the amplitude difference between adjacent sampling points, dimensionless), frequency offset (difference between the actual signal and the reference frequency, in Hz), and noise power spectral density (noise energy distribution in the signal frequency domain, in W / Hz). Each sample was labeled manually as either "interference condition" or "normal condition," corresponding to the interference probability in the output layer (interference condition was labeled as 1, normal condition as 0).

[0090] It's worth noting that the model employs a backpropagation (BP) neural network architecture, with an input layer of 6 dimensions, 2 hidden layers (32 neurons each), and an output layer of 1 dimension (outputting perturbation probabilities). A segmented activation function is used: both hidden layers use the ReLU activation function, which effectively alleviates the vanishing gradient problem and improves the training efficiency of deep networks; the output layer uses the Sigmoid activation function, mapping the output value to the 0-1 range, directly corresponding to the perturbation probability. The optimizer used is the Adam optimizer, with an initial learning rate of 0.01, decreasing by 0.1 every 20 epochs until reaching 0.001. The loss function is the mean squared error, and the loss value is monitored in real-time during training. Iteration stops when the loss fluctuation is less than 0.001 for 10 consecutive epochs (maximum training epochs 100).

[0091] Next, during the correlation process, the cross-correlation coefficient between the signal corresponding to the candidate coordinates and the theoretical signal output by the model is calculated. Continuous time periods with a cross-correlation coefficient below 0.6 are identified as signal interference intervals. For example, the cross-correlation coefficient for candidate coordinates from 3.2 seconds to 3.4 seconds is 0.52, which is below 0.6, and this interval is identified as a signal interference interval. Then, the Lippi exponent deviation of the time coordinates within the signal interference interval is calculated. If it exceeds a preset deviation threshold, it is marked as an offset point, and the deviation is estimated using the decay characteristics of the modulus maximum evolution curve. The deviation is the difference between the actual Lippi exponent and the theoretical Lippi exponent output by the Lippi exponent prediction model.

[0092] The Lee's index prediction model is an independently trained BP neural network model. The training set contains signal data from nearly one year of normal and fault conditions, with a total sample size of 60,000, including 27,000 samples from normal conditions and 33,000 samples from fault conditions, covering multiple scenarios such as high-voltage / low-voltage cables, high-resistance / low-resistance faults, and different laying environments (direct burial / overhead / conduit). The sample labels are Lee's indexes calculated through measured attenuation characteristics of modulus-maximum evolution curves, ensuring label accuracy. The six dimensions of the input features are: peak signal amplitude (V for voltage signals, A for current signals), instantaneous rate of change of signal (mean of amplitude difference between adjacent sampling points, V / s for voltage signals, A / s for current signals), harmonic distortion rate (THD, the ratio of total harmonic content to fundamental frequency content, %), signal smoothness (amplitude variance within a 5-point moving window, dimensionless), spectral centroid (energy-weighted average frequency in the signal frequency domain, Hz), and noise percentage (the ratio of noise energy to total signal energy, %). All input features are normalized by min-max and mapped to the [0,1] interval. The upper and lower bounds of the normalization are set based on the 5%-95% quantile of the training set to reserve margin to avoid the influence of outliers.

[0093] It should be noted that the model is configured with an input layer of 6 dimensions, 2 hidden layers (32 neurons per layer), and an output layer of 1 dimension. The activation function is segmented: both hidden layers use ReLU activation to alleviate the vanishing gradient problem and improve training efficiency; the output layer uses a linear activation function because the Lich King exponent is a continuous value (ranging from -1 to 1), and linear activation can directly output a prediction result within the true range. The optimizer is the Adam optimizer, with an initial learning rate of 0.01, decaying by 0.1 every 20 epochs until it reaches 0.001. The loss function is the mean squared error, used to minimize the deviation between the predicted and measured Lich King exponents. During training, 20% of the samples are used as a validation set, and the loss on the validation set is monitored in real time. Iteration stops when the loss fluctuation is less than 0.001 for 10 consecutive epochs, with a maximum training duration of 120 epochs to avoid overfitting. It is worth noting that the deviation threshold is set based on the statistical distribution of the Liebrion index under normal operation and fault conditions over the past year. The basic threshold is 0.15, which can be increased to 0.2 for high-voltage cable scenarios (where multipath interference is more severe) and decreased to 0.1 for low-voltage cable scenarios. This threshold has been verified in multiple scenarios and can accurately distinguish between normal fluctuations and offset interference. For example, if the actual Liebrion index of a certain time coordinate within the interference range is 0.8 and the theoretical Liebrion index is 0.6, the deviation is 0.2, exceeding the basic threshold, and this coordinate is marked as an offset point.

[0094] Based on a pre-acquired correction model, a distortion-free reference waveform under ideal propagation conditions is simulated and generated. The time offset of each offset point is calculated from the reference waveform. This offset is then inversely superimposed onto the time coordinate of the offset point to obtain the corrected feature point sequence. The reference waveform in the correction model is generated based on signal statistics from normal cable operation over the past year, extracting typical double-exponential pulse waveform characteristics: the amplitude is taken as the statistical average of the normal signal peak value (5V), the rising edge is taken as the average value (0.1ms), and the falling edge is taken as the average value (0.2ms), ensuring the waveform conforms to the signal propagation law under interference-free conditions. The time offset is the time-domain difference between the actual signal peak value corresponding to the offset point and the reference waveform peak value. Inverse superposition involves subtracting this offset from the corrected coordinates. For example, if the time coordinate of the offset point is 3.3 seconds and the time corresponding to the reference waveform peak value is 3.27 seconds, the calculated time offset is 0.03 seconds. After inverse superposition, the corrected coordinate is 3.3 - 0.03 = 3.27 seconds. All offset points are corrected to form the feature point sequence.

[0095] Finally, the feature point sequence is compared with the pre-acquired normal time-domain distribution template to perform a time-domain consistency test. When the corrected coordinate distribution is obtained, the normal time-domain distribution template is the statistical distribution of the time-domain intervals of the modulus maxima during the cable's normal operation over the past year (mean 0.05 seconds, variance 0.001 seconds). The consistency test uses a chi-square test, calculating the chi-square statistic of the time intervals of the feature point sequence and the template distribution. If the statistic is less than 3.84 (corresponding to a significance level of 0.05), the test is considered passed. The feature point sequence that passes the test is the corrected coordinate distribution. For example, if the corrected feature point sequence time intervals are 0.048 seconds and 0.052 seconds, the chi-square statistic is 1.2, which is less than 3.84, and the test passes; this sequence is the corrected coordinate distribution.

[0096] In step S106, the step of performing feature matching between the coordinate distribution and the preset fault feature topology, and supplementing the missing coordinates through interpolation to obtain a complete set of fault time coordinates, includes:

[0097] The coordinate distribution is mapped to a preset fault feature topology, and the local matching similarity is calculated.

[0098] If the local matching similarity is lower than a preset similarity threshold, then the signal attenuation characteristics are extracted based on the signal corresponding to the coordinate distribution to identify the feature break interval;

[0099] Extract the edge feature anchor points before and after the feature break interval, and construct a nonlinear trend constraint function based on the temporal location and anomaly degree of the edge feature anchor points;

[0100] Interpolation is performed based on the trend constraint function to generate virtual coordinate points;

[0101] By integrating the virtual coordinate points with the coordinate distribution, a complete set of time coordinates for the fault is obtained.

[0102] It should be noted that, firstly, the corrected coordinate distribution is mapped to a preset fault feature topology. When calculating local matching similarity, over 50,000 measured data points of high-resistance and low-resistance faults from the past year are collected, categorized by fault type, covering different cable models, laying environments, and fault severity. Then, the core features of each sample are extracted, including the core coordinate interval, coordinate density per unit time, coordinate interval variance, attenuation slope, and fluctuation coefficient, forming a 6-dimensional feature vector and standardizing it. Next, K-means++ clustering is used (4 classes for high-resistance and 3 classes for low-resistance), and the mean within each cluster is taken as the typical feature vector. A cubic polynomial is fitted to obtain the typical attenuation curve. Finally, the typical feature vectors and attenuation curves are integrated to construct a topological unit containing a "coordinate distribution layer" and an "attenuation trend layer," classifying and integrating them to form a complete topology. Local matching similarity is calculated using a cosine similarity algorithm. First, the corrected coordinate distribution and the coordinate sequence of the topology are converted into 128-dimensional feature vectors (generated through average pooling), and then the cosine similarity between the vectors is calculated, with a value ranging from 0 to 1. For example, the corrected coordinate distribution is concentrated in the range of 3.2-3.5 seconds. The coordinate pattern of the preset high-resistivity fault topology is also in this range and the decay trend is similar. The calculated local matching similarity is 0.68.

[0103] If the local matching similarity is lower than the preset similarity threshold, the signal attenuation characteristics are extracted based on the signal corresponding to the corrected coordinate distribution to identify the characteristic break interval. The similarity threshold is set based on statistical data of normal matching over the past year. The lowest matching similarity of all complete fault coordinate sets is statistically analyzed and used as the base threshold of 0.75. For high-voltage cables (where signal attenuation is more significant), this threshold can be increased to 0.8, and for low-voltage cables, it can be decreased to 0.7. This threshold has been verified in multiple scenarios and can accurately determine whether the coordinate distribution is complete. For example, if the local matching similarity is 0.68, which is lower than the base threshold of 0.75, analysis of the corresponding signal attenuation curve reveals that the signal strength drops sharply from 7.5 to 6.8 in the 3.35-3.4 second interval, with no coordinate points, thus identifying it as a characteristic break interval.

[0104] Next, edge feature anchor points before and after the feature break interval are extracted. When constructing the nonlinear trend constraint function based on their temporal location and anomaly degree, the edge feature anchor points are two consecutive valid coordinate points before and after the break interval. Their temporal location (accurate to microseconds) and anomaly degree values ​​are extracted. The nonlinear trend constraint function adopts the form of an exponential decay function, and the fitting formula is as follows: ,in The predicted anomaly level output by the function represents the degree of anomaly. This represents the degree of anomaly at the anchor point before the fracture section. The natural constant is represented by a value of approximately 2.718. The attenuation coefficient is used to characterize the rate at which the degree of anomaly decays over time. This represents the position of the function's argument in the time domain. This represents the temporal position of the anchor point before the fracture interval. The attenuation coefficient k is calculated by fitting the temporal difference between the anchor points before and after the fracture interval with the difference in anomaly degree, ensuring that the function fits the actual change trend of the anchor points. For example, if the temporal position of the anchor point before the fracture interval is 3.3 seconds with an anomaly degree of 8.2, and the temporal position is 3.35 seconds with an anomaly degree of 7.5, and the temporal position of the anchor point after the fracture interval is 3.4 seconds with an anomaly degree of 6.8, and the temporal position is 3.45 seconds with an anomaly degree of 6.1, the fitting calculation yields k=0.8, and the final constraint function is S=8.2×e^(-0.8×(t-3.3)).

[0105] It should be further explained that the calculation method for Ac is as follows: Based on the time-domain position (ti) and anomaly degree (Adi) of the four anchor points, the time-domain difference Δt = ti - t0 of each point is first calculated with t0 as the reference; then, the natural logarithm of the anomaly degree of each anchor point is taken to obtain ln(Adi / Ad); finally, the linear equation ln(Adi / Ad) = -Ac × Δt is fitted using the least squares method, and the absolute value of the slope of the equation is the attenuation coefficient Ac, ensuring that the function closely matches the actual changing trend of the anchor points. For example, after fitting the data with the given anchor points, Ac = 0.8.

[0106] Subsequently, interpolation is performed based on the trend constraint function. When generating virtual coordinate points, the interpolation step size is set to an integer multiple of the original signal sampling period to ensure that the time domain of the virtual point is consistent with the original sampling time axis and to facilitate subsequent indexing. During the interpolation process, each time domain position is traversed point by point within the break interval according to the stated step size, and the time domain position is substituted into the trend constraint function to directly calculate the corresponding anomaly degree, thereby generating the virtual coordinate point.

[0107] Finally, when integrating the virtual coordinate points with the corrected coordinate distribution to obtain the complete time coordinate set of the fault, the virtual coordinate points are inserted into the corrected coordinate sequence in temporal order to ensure the temporal continuity and consistent decay trend of the entire set. Duplicate coordinate points are removed after insertion (if any), ultimately forming a complete coordinate set covering the entire fault time period. For example, if the corrected coordinate sequence jumps directly from 3.35 seconds to 3.4 seconds, inserting 5 virtual points results in a continuous distribution from 3.3 seconds to 3.45 seconds, with the degree of anomaly at each coordinate point conforming to an exponential decay law, thus obtaining the complete time coordinate set of the fault point.

[0108] In step S107, the signal strength corresponding to the time coordinate set is compared with a preset high-impedance fault threshold and a low-impedance fault threshold, respectively. If the signal strength exceeds the high-impedance fault threshold, it is determined to be a high-impedance fault; if it is below the low-impedance fault threshold, it is determined to be a low-impedance fault, thus obtaining a fault type result, including:

[0109] Based on the time coordinate set index of the original signal, the voltage amplitude and current amplitude of the original signal at the corresponding time are extracted to construct a signal strength sequence;

[0110] Feature extraction is performed on the signal intensity sequence. Based on the pre-acquired fault identification sensitivity coefficient, the sensitivity of each feature to fault identification is determined. Weights are assigned to each feature according to the sensitivity and the features are weighted and fused to generate a weighted feature vector.

[0111] The weighted feature vector is compared with the preset high-resistance fault threshold and low-resistance fault threshold, respectively.

[0112] If the weighted feature vector exceeds the high-resistance fault threshold, it is marked as a high-resistance fault candidate point; if it is below the low-resistance fault threshold, it is marked as a low-resistance fault candidate point.

[0113] The waveform distortion features of the high-resistance fault candidate points and the low-resistance fault candidate points are extracted and compared with the pre-acquired standard fault features to confirm the fault type and obtain the fault type result.

[0114] It should be noted that, firstly, when constructing a signal strength sequence by indexing the original signal based on the time coordinate set and extracting the voltage and current amplitudes at the corresponding moments, the voltage and current amplitudes corresponding to each coordinate point are precisely indexed by matching the timestamp of the time coordinate with the timestamp of the original acquired signal, and arranged in chronological order to form a signal strength sequence. This sequence completely preserves the signal strength changes during the fault occurrence period, providing basic data for subsequent feature extraction. For example, if the time coordinate set contains three points at 3.2 seconds, 3.3 seconds, and 3.4 seconds, after indexing the original signal, the extracted voltage amplitudes are 220V, 180V, and 150V, and the current amplitudes are 150A, 120A, and 100A, respectively, and the constructed signal strength sequence is [(220,150),(180,120),(150,100)].

[0115] When extracting features from the signal strength sequence, assigning weights based on the fault identification sensitivity coefficient, and generating a weighted feature vector through weighted fusion, the feature extraction selects three core features: peak abrupt change, amplitude change rate, and harmonic distortion rate.

[0116] It should be noted that the calculation and synthesis characterization of peak mutations are as follows: First, the voltage peak mutation ΔU and the current peak mutation ΔI are calculated separately. ΔU is the maximum difference (in V) of voltage amplitude between adjacent time points or within a 5ms short window, and ΔI is the maximum difference (in A) of current amplitude in the corresponding interval. Then, both are normalized by min-max (with upper and lower bounds set based on the 5%-95% quantile of historical fault data) and mapped to the [0,1] interval to eliminate dimensional differences. Finally, the sum is weighted according to the weights of 0.5 for voltage and 0.5 for current to obtain the comprehensive peak mutation characteristics, ensuring that the influence of the two is balanced.

[0117] The comprehensive calculation of the amplitude change rate involves calculating the voltage amplitude change rate R_U (ΔU / reference voltage, dimensionless) and the current amplitude change rate R_I (ΔI / reference current, dimensionless) separately. The reference voltage / current is taken as the average value of stable operation before the fault. After the same normalization process as above, the comprehensive amplitude change rate is obtained by weighting them with a weight of 0.5:0.5, which intuitively reflects the speed of signal strength change.

[0118] The quantification and synthesis of harmonic distortion rate (THD) is carried out by calculating voltage THD_U and current THD_I (the ratio of total non-fundamental harmonic content to fundamental content, in %) according to the national standard method. After normalization, the THD is obtained by weighting with a weight of 0.5:0.5, which quantifies the degree of signal spectrum distortion and helps to distinguish fault types.

[0119] The fault identification sensitivity coefficients were obtained by first collecting over 60,000 labeled fault samples from the past year (28,000 high-resistivity and 32,000 low-resistivity). Next, a random forest model was used to evaluate feature importance and calculate the contribution of each feature to the identification accuracy of different fault types. Subsequently, statistics showed that peak mutations contributed the most to high-resistivity faults (sensitivity coefficient 0.4), amplitude change rate contributed significantly to low-resistivity faults (sensitivity coefficient 0.3), and THD assisted in differentiation (sensitivity coefficient 0.3). 4) The stability of the coefficients was verified through 5-fold cross-validation to ensure scientific weight allocation.

[0120] The weighted feature vector calculation is based on the three comprehensive features mentioned above: peak mutation, amplitude change rate, and THD. After normalization of each feature, the features are weighted and fused according to the sensitivity coefficient. For example, after normalization, the peak mutation of a certain sequence is 0.9, the amplitude change rate is 0.6, and the THD is 0.7. The weighted feature vector is 0.9×0.4+0.6×0.3+0.7×0.3=0.75.

[0121] When comparing the weighted feature vector with the preset high-resistance fault threshold and low-resistance fault threshold, the high-resistance fault threshold and low-resistance fault threshold are set based on fault data statistics from the past year. The minimum value of the weighted feature vector of all high-resistance fault cases is used as the basic threshold for high-resistance faults (0.8); the maximum value of the weighted feature vector of all low-resistance fault cases is used as the basic threshold for low-resistance faults (0.3). In high-voltage cable scenarios (where high-resistance faults are frequent), the high-resistance threshold can be lowered to 0.75, and in low-voltage cable scenarios (where low-resistance faults are more common), the low-resistance threshold can be raised to 0.35. This threshold has been verified in multiple scenarios and can effectively distinguish different types of faults. For example, if the weighted feature vector is 0.85, which exceeds the basic threshold for high-resistance faults (0.8), it needs to be further judged as a candidate for high-resistance faults; if the vector is 0.25, which is lower than the basic threshold for low-resistance faults (0.3), it is judged as a candidate for low-resistance faults.

[0122] If the weighted feature vector exceeds the high-impedance fault threshold, it is marked as a high-impedance fault candidate point; if it is below the low-impedance fault threshold, it is marked as a low-impedance fault candidate point. This marking is done point by point according to the comparison results. Multiple candidate points may exist in the same time coordinate set, and their corresponding time-domain positions and feature vector values ​​must be recorded separately. For example, the weighted feature vectors of the signal strength sequence are 0.85, 0.82, and 0.78, where the first two exceed the high-impedance threshold of 0.8 and are marked as high-impedance fault candidate points; the other sequence vectors are 0.28, 0.25, and 0.32, where the first two are below the low-impedance threshold of 0.3 and are marked as low-impedance fault candidate points.

[0123] Finally, waveform distortion features of high-resistance and low-resistance fault candidate points are extracted and compared with pre-acquired standard fault features to confirm the fault type, obtaining the fault type result. Waveform distortion features include waveform asymmetry, peak offset, and rising edge steepness. The standard fault feature library is constructed based on real fault cases from the past year and contains typical distortion feature vectors (128 dimensions) for high-resistance and low-resistance faults. The comparison uses the Euclidean distance algorithm to calculate the distance between the distortion features of the candidate point and the standard features; the smaller the distance, the higher the matching degree. If the distance between a high-resistance candidate point and the standard high-resistance feature is less than 0.2, a high-resistance fault is confirmed; similarly, if the distance between a low-resistance candidate point and the standard low-resistance feature is less than 0.2, a low-resistance fault is confirmed. For example, if the Euclidean distance between the waveform distortion feature of a high-resistance candidate point and the standard high-resistance feature is 0.15, which is less than 0.2, the fault type is ultimately confirmed as a high-resistance fault.

[0124] It should be noted that the waveform asymmetry is calculated by taking the signal period corresponding to the candidate point, calculating the area in both positive and negative half-cycles, and obtaining the voltage / current asymmetry (half-cycle area ratio) separately. After normalization, the values ​​are weighted and summed at a ratio of 0.5:0.5. The peak offset is calculated by comparing the time domain difference between the peak value of the fault waveform and the peak value of the normal waveform, calculating the voltage / current offset separately, and then weighting and fusing them after normalization. The rise edge steepness is calculated by taking the range of 10%-90% before the peak value, calculating the rise slope of the voltage / current, and then weighting it after normalization to obtain a comprehensive value. These three values ​​are concatenated according to a fixed dimension, and after dimensionality enhancement, they are converted into a 128-dimensional feature vector, which is then included in the standard fault feature library.

[0125] In step S108, determining the signal propagation speed in the cable based on the fault type result and pre-acquired cable medium parameters, performing calculations on the propagation speed and the coordinate set to convert the time coordinates into spatial location, and obtaining the specific location of the cable fault includes:

[0126] Based on the fault type results, the corresponding cable dielectric constant is determined, and the theoretical signal propagation speed is calculated in combination with the pre-acquired cable dielectric parameters.

[0127] The theoretical propagation speed of the signal is calculated using the set of time coordinates to obtain a preliminary linear distance;

[0128] Based on the preliminary linear distance and the pre-acquired cable impedance distribution data, a path segmentation model containing cable impedance variation characteristics is constructed, and the time delay of the secondary pulse reflection peak is extracted.

[0129] The initial linear distance is corrected using the time delay to obtain the one-dimensional fault location;

[0130] The one-dimensional fault location is mapped to a spatial coordinate system to obtain the specific location of the cable fault.

[0131] It should be noted that, firstly, the corresponding cable dielectric constant is determined based on the fault type. When calculating the theoretical signal propagation speed using pre-obtained cable dielectric parameters, different fault types correspond to fixed dielectric constants: high-resistance faults correspond to a dielectric constant of 2.3, and low-resistance faults correspond to 1.8. This correspondence is based on dielectric characteristic tests under different fault conditions over the past year. Cable dielectric parameters include relative permeability, conductor resistivity, and insulation thickness. The theoretical signal propagation speed is calculated using the lossy dielectric electromagnetic wave phase velocity formula derived from Maxwell's equations. The formula is: ×K, where: For the theoretical propagation speed of the signal, The speed of light in a vacuum (3 × 10⁻⁶) 8 m / s), The relative permittivity of the cable is dimensionless. Where is the relative permeability of the cable (dimensionless), and K is the loss correction coefficient (calculated by combining conductor resistivity and insulation thickness, with a value ranging from 0.95 to 1.0). This formula integrates the loss effects of dielectric constant and dielectric parameters, adapting to the propagation characteristics of lossy media in cables.

[0132] It should be noted that the specific calculation method for K is as follows: ,in For conductor resistivity, For the insulation layer thickness, international units are used in the calculation and dimensionless normalization is performed.

[0133] For example, the fault type is a high-resistance fault, with a corresponding relative permittivity εᵣ=2.3, and dielectric parameters including relative permeability μᵣ=1, conductor resistivity ρ=1.72e-8Ω·m, and insulation layer thickness d=10mm=0.01m. First, calculate K using the corrected loss correction factor formula: K≈0.996. Substituting this into the propagation velocity formula: first calculate √(2.3×1)=1.516, then use 3×10... 8 Dividing m / s by 1.516 gives 1.978 × 10⁻⁶ m / s. 8 The value is calculated as m / s, and then multiplied by a correction factor of 0.996 to obtain the theoretical propagation speed of the signal, which is approximately 2.04e8 m / s.

[0134] Subsequently, when calculating the preliminary linear distance by combining the theoretical signal propagation speed with the time coordinate set, the calculation logic is the conversion of the round-trip distance of the signal, i.e., the calculation expression is L=(V×T)÷2. Here, L represents the preliminary linear distance, V represents the theoretical signal propagation speed, and T represents the round-trip propagation delay corresponding to the core fault time coordinate. Dividing by 2 is because the signal propagates from the detection end to the fault point and then reflects back to the detection end, requiring conversion to a one-way distance. For example, if the theoretical signal propagation speed is 2.05e8 meters per second, and the round-trip propagation delay corresponding to the core fault time with the highest degree of anomaly in the complete fault time coordinate set is 3.1 microseconds, substituting into the expression, we get L=(2.05e8×3.1e-6)÷2≈318 meters, thus obtaining the preliminary linear distance between the fault point and the detection end.

[0135] Next, based on the preliminary linear distance and the pre-acquired cable impedance distribution data, a path segmentation model incorporating cable impedance variation characteristics is constructed. When extracting the time delay of the secondary pulse reflection peak from the passively acquired traveling wave / reflection signal, the path segmentation model is constructed using the K-means++ clustering algorithm. The specific parameters are: the number of clusters is determined by the number of impedance abrupt change points (3-5), the initial cluster centers are selected using the K-means++ algorithm, the iteration is performed 50 times, and the convergence condition is that the change in the sum of squared errors within the cluster is less than 0.001.

[0136] It should be noted that the extraction of secondary pulse reflection peaks is based on the acquired fault reflection waveforms without additional active pulse injection. Using the primary reflection peak as a benchmark, and combining it with the possible reflection intervals given by the path segmentation model, peak detection and matching are performed within the corresponding time window to identify subsequent secondary reflection peaks, and their time difference relative to the primary peak is recorded as the time delay. The path segmentation model is used to limit the peak search to the time range corresponding to the impedance abrupt change segment, thereby reducing false detections caused by noise peaks or irrelevant reflections.

[0137] For example, with an initial linear distance of 320 meters, impedance distribution data shows that there are obvious impedance abrupt changes at 200 meters and 300 meters, and the clustering is divided into 3 segments; in the passively acquired reflection waveform, the secondary reflection peak is detected within the time window of the corresponding segment, with a time delay of 0.08 microseconds between it and the main peak, using the main peak as a reference.

[0138] When using time delay to correct the initial linear distance to obtain the one-dimensional fault location, the correction logic is to calculate the corresponding additional propagation distance using the time delay and subtract this additional distance from the initial linear distance (due to the fault location after the impedance abrupt change point corresponding to the secondary reflection). For example, a time delay of 0.08 microseconds corresponds to an additional distance of (2.05e8 × 0.08e-6) ÷ 2 ≈ 8.2 meters. The corrected one-dimensional fault location is 320 - 8.2 ≈ 311.8 meters, which is closer to the actual fault location.

[0139] Finally, when mapping the one-dimensional fault location to a spatial coordinate system to obtain the specific location of the cable fault, the spatial coordinate system uses a mileage coordinate system specifically for cable lines, calibrated in conjunction with GPS positioning data from the laying process. During mapping, the geographic information ledger of the cable line is referenced, converting the one-dimensional mileage location into specific latitude and longitude or line station number. For example, a one-dimensional fault location of 311.8 meters corresponds to cable line mileage station number K0+311.8. After mapping through the geographic information ledger, the specific location is 116.35°E, 39.92°N, accurately pointing to the fault point of the underground cable, facilitating location and troubleshooting by maintenance personnel.

[0140] In summary, this invention discloses a method for diagnosing faults in wires and cables. By acquiring cable voltage and current signals, denoising and segmenting them, abrupt changes and gradient features are extracted to generate an anomaly distribution map. The method then decomposes abrupt change regions and extracts key feature points, using clustering and smoothing to construct a modulus maxima chain. Time coordinate offsets are corrected, and missing coordinates are filled in to obtain a complete time coordinate set. Signal strength is compared with fault thresholds to determine whether the fault type is high-resistance or low-resistance. Finally, the signal propagation speed is determined by combining cable dielectric parameters, converting the time coordinates into the specific spatial location of the fault. This method enables accurate fault diagnosis and type identification of wires and cables, meeting the safety and forward-looking requirements for cable operation and maintenance.

[0141] Reference Figure 2 The second embodiment of the present invention provides a wire and cable fault diagnosis system, comprising:

[0142] The signal acquisition module is used to acquire the voltage and current signals of the cable.

[0143] The anomaly map module is used to denoise and segment the voltage and current signals over time, extract the abrupt changes and gradient features of each time segment, calculate the degree of anomaly of the signal based on the abrupt changes and gradient features, and generate an anomaly distribution map.

[0144] The feature point extraction module is used to perform scale decomposition on the mutation regions in the abnormal distribution map and capture the modulus maxima at different scales. Based on the scale evolution law of the modulus maxima, key feature points are extracted to obtain a set of feature points of scale evolution.

[0145] The chain construction module is used to cluster and smooth the feature point set, remove pseudo-extreme points and fill in the breakpoints to obtain a continuous modulus maxima chain;

[0146] The coordinate correction module is used to extract the time coordinates of the fault candidates based on the time domain position of each of the modulus maxima points in the modulus maxima chain, associate the time coordinates with the pre-acquired correction model, correct the time coordinate offset, and obtain the corrected coordinate distribution.

[0147] The coordinate completion module is used to perform feature matching between the coordinate distribution and the preset fault feature topology, and to fill in the missing coordinates through interpolation to obtain a complete set of time coordinates of the fault.

[0148] The fault classification module is used to compare the signal strength corresponding to the time coordinate set with the preset high-impedance fault threshold and low-impedance fault threshold respectively. If the signal strength exceeds the high-impedance fault threshold, it is determined to be a high-impedance fault; if the signal strength is below the low-impedance fault threshold, it is determined to be a low-impedance fault, and the fault type result is obtained.

[0149] The location determination module is used to determine the propagation speed of the signal in the cable based on the fault type result and the pre-acquired cable medium parameters, perform calculations on the propagation speed and the coordinate set, convert the time coordinates into spatial positions, and obtain the specific location of the cable fault.

[0150] It should be noted that the wire and cable fault diagnosis system provided in this embodiment of the invention is used to execute all the process steps of the wire and cable fault diagnosis method in the above embodiment. The working principle and beneficial effect of the two are one-to-one, so they will not be described again.

[0151] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0152] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for diagnosing faults in electric wires and cables, characterized in that, include: Acquire the voltage and current signals of the cable; The voltage and current signals are denoised and segmented over time. The abrupt changes and gradient features of each time segment are extracted. The degree of anomaly of the signal is calculated based on the abrupt changes and gradient features, and an anomaly distribution map is generated. The abrupt change regions in the anomalous distribution map are decomposed by scale and the modulus maxima at different scales are captured. Key feature points are extracted based on the scale evolution law of the modulus maxima to obtain a set of feature points for scale evolution. Clustering and smoothing are performed on the feature point set to remove pseudo-extreme points and fill in the breakpoints, resulting in a continuous modulus maxima chain. Based on the time domain position of each modulus maximum point in the modulus maximum chain, the time coordinates of the fault candidates are extracted, and the time coordinates are associated with the pre-acquired correction model to correct the time coordinate offset and obtain the corrected coordinate distribution. The coordinate distribution is matched with the preset fault feature topology, and the missing coordinates are supplemented by interpolation to obtain a complete set of fault time coordinates. The signal strength corresponding to the time coordinate set is compared with the preset high-impedance fault threshold and low-impedance fault threshold respectively. If it exceeds the high-impedance fault threshold, it is determined to be a high-impedance fault; if it is below the low-impedance fault threshold, it is determined to be a low-impedance fault, and the fault type result is obtained. Based on the fault type result and the pre-acquired cable medium parameters, the propagation speed of the signal in the cable is determined. The propagation speed is then calculated with the coordinate set to convert the time coordinates into spatial positions, thereby obtaining the specific location of the cable fault.

2. The method for diagnosing wire and cable faults according to claim 1, characterized in that, The acquisition of the voltage and current signals of the cable includes: Voltage and current signals during cable operation are collected by voltage and current sensors deployed on the cable lines.

3. The method for diagnosing wire and cable faults according to claim 1, characterized in that, The process of denoising and time-segmenting the voltage and current signals, extracting abrupt changes and gradient features from each time segment, calculating the degree of signal anomaly based on the abrupt changes and gradient features, and generating an anomaly distribution map includes: The voltage and current signals are subjected to threshold denoising to filter out high-frequency noise and interference components, resulting in denoised signals. The denoised signal is segmented using a fixed-duration sliding time window to obtain several signal segments. Calculate the waveform energy, amplitude variance, and gradient of each signal segment, and extract the abrupt change features and gradient features of each segment; Based on the mutation characteristics and gradient characteristics, the degree of abnormality of each segment is determined, and an abnormality distribution map is generated.

4. The method for diagnosing wire and cable faults according to claim 1, characterized in that, The process involves scale decomposing the abrupt change regions in the abnormal distribution map and capturing modulus maxima at different scales. Key feature points are then extracted based on the scale evolution of these modulus maxima, resulting in a set of scale-evolved feature points, including: Based on the anomaly distribution map, the time domain range of the mutation region is determined, and discrete signals within the time domain range are extracted to obtain the mutation region sequence; The mutation region sequence is decomposed into a multi-scale coefficient matrix. The multi-scale coefficient matrix is ​​traversed to filter out the modulus maxima points, and the modulus maxima evolution curve is constructed. The decay characteristics of the modulus maxima evolution curve are calculated, the Lee index is estimated, and key feature points are selected based on the Lee index to obtain the feature point set of scale evolution.

5. The method for diagnosing wire and cable faults according to claim 1, characterized in that, The process of clustering and smoothing the feature point set, removing spurious extrema and filling in breakpoints to obtain a continuous modulo maxima chain includes: Calculate the spatial distance between each point in the feature point set, and cluster them according to the spatial distance to obtain the modulus maxima cluster; Based on the temporal proximity and amplitude propagation law of feature points within each cluster, cross-scale association is performed to generate a path with a modulus maxima. Based on the attenuation law of the modulus maxima path, pseudo-maxima points that do not conform to the law are eliminated to obtain the signal skeleton containing the discontinuities; The signal skeleton is interpolated and fitted to fill in the breakpoints and form a reconstruction evolution line; The reconstructed evolution line is smoothed to suppress high-frequency jitter, resulting in a continuous modulus maxima chain.

6. The method for diagnosing wire and cable faults according to claim 1, characterized in that, The step of extracting the time coordinates of fault candidates based on the time domain positions corresponding to each modulus maximum point in the modulus maximum chain, associating the time coordinates with a pre-acquired correction model, correcting the time coordinate offset, and obtaining the corrected coordinate distribution includes: Extract the time-domain position corresponding to each of the modulus maxima points in the modulus maxima chain to determine the time coordinates of the fault candidate; The time coordinates of the fault candidates are correlated with the pre-acquired correction model to identify the signal interference range; Calculate the Lee's exponential deviation of the time coordinate within the signal interference interval. If the Lee's exponential deviation exceeds a preset deviation threshold, it is marked as an offset point. Based on the correction model, a distortion-free reference waveform under ideal propagation conditions is simulated and generated. The time offset of the offset point is calculated based on the reference waveform. The time offset is then superimposed on the time coordinate of the offset point to obtain the corrected feature point sequence. The feature point sequence is compared with a pre-acquired normal time-domain distribution template, and a time-domain consistency check and correction are performed to obtain the corrected coordinate distribution.

7. The method for diagnosing wire and cable faults according to claim 1, characterized in that, The step of performing feature matching between the coordinate distribution and the preset fault feature topology, and supplementing missing coordinates through interpolation to obtain a complete set of fault time coordinates includes: The coordinate distribution is mapped to a preset fault feature topology, and the local matching similarity is calculated. If the local matching similarity is lower than a preset similarity threshold, then the signal attenuation characteristics are extracted based on the signal corresponding to the coordinate distribution to identify the feature break interval; Extract the edge feature anchor points before and after the feature break interval, and construct a nonlinear trend constraint function based on the temporal location and anomaly degree of the edge feature anchor points; Interpolation is performed based on the trend constraint function to generate virtual coordinate points; By integrating the virtual coordinate points with the coordinate distribution, a complete set of time coordinates for the fault is obtained.

8. The method for diagnosing wire and cable faults according to claim 1, characterized in that, The process involves comparing the signal strength corresponding to the time coordinate set with preset high-impedance fault thresholds and low-impedance fault thresholds. If the signal strength exceeds the high-impedance fault threshold, it is determined to be a high-impedance fault; if it falls below the low-impedance fault threshold, it is determined to be a low-impedance fault, thus obtaining the fault type result, including: Based on the time coordinate set index of the original signal, the voltage amplitude and current amplitude of the original signal at the corresponding time are extracted to construct a signal strength sequence; Feature extraction is performed on the signal intensity sequence. Based on the pre-acquired fault identification sensitivity coefficient, the sensitivity of each feature to fault identification is determined. Weights are assigned to each feature according to the sensitivity and the features are weighted and fused to generate a weighted feature vector. The weighted feature vector is compared with the preset high-resistance fault threshold and low-resistance fault threshold, respectively. If the weighted feature vector exceeds the high-resistance fault threshold, it is marked as a high-resistance fault candidate point; if it is below the low-resistance fault threshold, it is marked as a low-resistance fault candidate point. The waveform distortion features of the high-resistance fault candidate points and the low-resistance fault candidate points are extracted and compared with the pre-acquired standard fault features to confirm the fault type and obtain the fault type result.

9. The method for diagnosing wire and cable faults according to claim 1, characterized in that, The step of determining the signal propagation speed in the cable based on the fault type result and pre-acquired cable medium parameters, performing calculations on the propagation speed and the coordinate set to convert the time coordinates into spatial positions, and obtaining the specific location of the cable fault includes: Based on the fault type results, the corresponding cable dielectric constant is determined, and the theoretical signal propagation speed is calculated in combination with the pre-acquired cable dielectric parameters. The theoretical propagation speed of the signal is calculated using the set of time coordinates to obtain a preliminary linear distance; Based on the preliminary linear distance and the pre-acquired cable impedance distribution data, a path segmentation model containing cable impedance variation characteristics is constructed, and the time delay of the secondary pulse reflection peak is extracted. The initial linear distance is corrected using the time delay to obtain the one-dimensional fault location; The one-dimensional fault location is mapped to a spatial coordinate system to obtain the specific location of the cable fault.

10. A fault diagnosis system for electric wires and cables, characterized in that, include: The signal acquisition module is used to acquire the voltage and current signals of the cable. The anomaly map module is used to denoise and segment the voltage and current signals over time, extract the abrupt changes and gradient features of each time segment, calculate the degree of anomaly of the signal based on the abrupt changes and gradient features, and generate an anomaly distribution map. The feature point extraction module is used to perform scale decomposition on the mutation regions in the abnormal distribution map and capture the modulus maxima at different scales. Based on the scale evolution law of the modulus maxima, key feature points are extracted to obtain a set of feature points of scale evolution. The chain construction module is used to cluster and smooth the feature point set, remove pseudo-extreme points and fill in the breakpoints to obtain a continuous modulus maxima chain; The coordinate correction module is used to extract the time coordinates of the fault candidates based on the time domain position of each of the modulus maxima points in the modulus maxima chain, associate the time coordinates with the pre-acquired correction model, correct the time coordinate offset, and obtain the corrected coordinate distribution. The coordinate completion module is used to perform feature matching between the coordinate distribution and the preset fault feature topology, and to fill in the missing coordinates through interpolation to obtain a complete set of time coordinates of the fault. The fault classification module is used to compare the signal strength corresponding to the time coordinate set with the preset high-impedance fault threshold and low-impedance fault threshold respectively. If the signal strength exceeds the high-impedance fault threshold, it is determined to be a high-impedance fault; if the signal strength is below the low-impedance fault threshold, it is determined to be a low-impedance fault, and the fault type result is obtained. The location determination module is used to determine the propagation speed of the signal in the cable based on the fault type result and the pre-acquired cable medium parameters, perform calculations on the propagation speed and the coordinate set, convert the time coordinates into spatial positions, and obtain the specific location of the cable fault.

Images