Cable fault analysis system based on big data

Abstract

The invention relates to the technical field of cable fault detection, in particular to a cable fault analysis system based on big data, which comprises a disturbance feature extraction module, a time sequence hierarchy judgment module, a reflection characteristic analysis module, a topological section identification module and a fault path mapping module. According to the method, gradient inflection points are extracted by calculating voltage and current increment included angles at adjacent moments, spatial-temporal characteristics of disturbance signals are correlated, interference of noise to a starting point mark is reduced, time difference is fitted in combination with cable length, deviation over-threshold abnormal points are dynamically filtered, influences of synchronization errors on sorting are eliminated, and fault path mapping stability is improved; the method comprises the following steps: generating characteristic parameters by fusing amplitude values and wave crest intervals of main waves and reflection waves, quantifying amplitude attenuation and phase offset coupling, enhancing the multi-path superposed signal separation capability, judging the reflection characteristic similarity of a physical section based on a Manhattan distance and a support vector machine, and dynamically dividing a fault section in combination with line topology. The problem of fuzzy positioning in a branch parallel scene in a traditional method is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of cable fault detection, and in particular to a cable fault analysis system based on big data. Background Art

[0002] The field of cable fault detection technology encompasses the monitoring and diagnosis of the operating status of cable lines in power systems. The core of this technical field is the identification and location of internal or external cable faults through the collection and analysis of various physical parameters. The overall technology encompasses the collection and analysis of voltage and current signals, fault point location methods, cable insulation condition assessment, cable path identification, and characteristic parameter modeling. Common methods include pulse reflection, traveling wave, and impedance methods. In engineering practice, different detection technologies are selected based on the cable type and installation environment, combined with on-site testing equipment to enable fault analysis of complex systems such as underground, submarine, and high-voltage cables.

[0003] The cable fault analysis system, among other tasks, identifies faults during cable operation. By collecting transient current data from the cable and combining it with line length and structural parameters, it uses methods such as the time difference between traveling waves, the difference between currents before and after the fault, and waveform feature recognition to classify the fault type and calculate the fault location using a distance algorithm. The system typically uses synchronous sampling devices to acquire electrical signals from different nodes, performing comparative analysis using timestamps. It also uses frequency domain transformation to extract abnormal components from the fault signal and, combined with line structural models, deduce the specific location and attributes of the fault point, improving the accuracy and reliability of fault diagnosis.

[0004] Existing technologies rely on timestamp comparisons of synchronous sampling devices, but clock drift or communication delays in long-distance cable scenarios can easily lead to deviations in the calculation of traveling wave propagation time differences. For example, the time difference at the end node of a submarine cable deviates from the actual value due to signal transmission delays, reducing the accuracy of fault location. When the impedance method is used to evaluate the insulation status of a cable, the impact of changes in line distribution parameters on the impedance characteristics is not dynamically corrected. For example, underground cables may experience abnormal distributed capacitance due to soil moisture fluctuations, leading to misjudgment of insulation aging or partial discharge. The pulse reflection method is interfered with by the superposition of reflected waves in multi-branch cable systems, and the original waveform features are obscured. For example, urban underground pipeline cables cannot effectively identify the type of fault due to the aliasing of multi-branch reflected waves. The existing node hierarchical sorting lacks a dynamic elimination mechanism for abnormal data. When some nodes output abnormal time differences due to electromagnetic interference, the hierarchical sorting results deviate from the true path, increasing the risk of misjudgment of fault path mapping. Summary of the Invention

[0005] The purpose of the present invention is to solve the shortcomings of the prior art and to propose a cable fault analysis system based on big data.

[0006] In order to achieve the above objectives, the present invention adopts the following technical solutions: A cable fault analysis system based on big data includes:

[0007] A disturbance feature extraction module is used to obtain voltage and current time series data through cable monitoring nodes, calculate the directional angle between voltage increments and current increments at adjacent time points, extract the gradient change inflection point, generate a disturbance starting position marker, and pass the disturbance starting position marker to the time series level determination module;

[0008] A timing hierarchy determination module is configured to call the disturbance starting position marker, fit the time difference in combination with the cable length data, generate a node hierarchy ranking list, eliminate abnormal nodes whose time difference deviation exceeds three times the standard deviation, and pass the node hierarchy ranking list to the fault path mapping module;

[0009] A reflection characteristic analysis module is used to collect the amplitudes and peak intervals of the main wave and the reflected wave, calculate the amplitude attenuation coefficient and the phase offset, combine them into reflection characteristic parameters, and transmit the reflection characteristic parameters to the topology segment identification module;

[0010] A topological segment identification module is used to call the reflection characteristic parameters, calculate the Manhattan distance of the reflection characteristic parameters in the physical segment, determine the similarity through a support vector machine classification model, output a segment identification label, and pass the segment identification label to the fault path mapping module.

[0011] As a further solution of the present invention, the disturbance starting position mark is specifically the disturbance gradient inflection point position, the node hierarchy sorting list includes the node propagation order, the valid node set, and the abnormal node set, the reflection characteristic parameters include the amplitude attenuation coefficient and the phase offset, and the section identification label specifically refers to the cable topology physical section identification code.

[0012] As a further solution of the present invention, the disturbance feature extraction module includes:

[0013] The incremental vector calculation submodule monitors the voltage and current time series data of the cable nodes, extracts the voltage and current differences between adjacent time points, calculates the incremental vectors through the differential algorithm, and arranges the continuous time series incremental vectors in chronological order to generate an incremental vector sequence;

[0014] The gradient inflection point detection submodule calculates the angle between adjacent vector directions based on the incremental vector sequence using the direction cosine formula, applies a sliding window algorithm to perform local mean filtering on the angle sequence, and determines the inflection point position in combination with the angle change rate threshold to generate a gradient inflection point set.

[0015] The perturbation mark generation submodule calls the gradient inflection point set, selects the starting inflection point according to the difference in gradient direction before and after the inflection point using the extreme value positioning algorithm, maps the timestamp corresponding to the inflection point to the original time series data, and generates a perturbation starting position mark.

[0016] As a further solution of the present invention, the timing level determination module includes:

[0017] The time difference fitting submodule uses the node position coordinates recorded in the disturbance starting position mark, combines the conductor material parameters in the cable length data and the line segment length, calculates the signal transmission delay between adjacent nodes through linear interpolation, and generates a time difference set;

[0018] The hierarchical sorting construction submodule arranges the nodes from small to large according to the transmission delay based on the time difference set, and performs secondary sorting on the nodes with the same delay value using the cable length data to generate a hierarchical node sequence table;

[0019] The abnormal node elimination submodule calls the sorting weights in the hierarchical node sequence table, calculates the standard deviation of all node weight values, screens out nodes whose weight deviation exceeds three times the standard deviation threshold, removes them from the hierarchical node sequence table, and generates a node hierarchical sorting list.

[0020] As a further solution of the present invention, the reflection characteristic analysis module includes:

[0021] The waveform acquisition submodule collects synchronous measurement data of the main wave amplitude voltage peak value and the reflected wave amplitude voltage peak value, records the millisecond-level change in the time interval between the two peaks, integrates the main wave peak value, reflected wave peak value and time difference data into a waveform recording unit with a unified timestamp, and generates a waveform feature set;

[0022] The parameter calculation submodule extracts the main wave peak value and the reflected wave peak value in the waveform feature set, calculates the natural logarithm value of the amplitude ratio of the two using logarithmic operation, combines the peak time difference with the preset electromagnetic wave propagation speed constant, converts the product of the time difference and the speed into a phase angle offset radian value, and generates a characteristic parameter component;

[0023] The characteristic integration submodule calls the amplitude attenuation coefficient and the phase offset radian in the characteristic parameter components, aligns the two in time series, and then merges them into a double-column matrix to generate reflection characteristic parameters.

[0024] As a further solution of the present invention, the topology segment identification module includes:

[0025] The parameter processing submodule obtains the amplitude characteristics and phase difference characteristics of the reflection characteristic parameters, normalizes the characteristic sequence using the mean and standard deviation, compares the deviation degree of each data point in the characteristic sequence based on the parameter fluctuation threshold, selects the continuous segments with deviation values less than the threshold, and generates a standardized parameter set;

[0026] The distance calculation submodule extracts the amplitude difference sequence and phase difference difference sequence of adjacent physical segments based on the standardized parameter set, calculates the absolute deviation value of the difference sequence point by point according to the Manhattan distance formula, slides through all segment pairs using a fixed step window, accumulates the deviation values in the window to generate the difference metric between segments, and generates the Manhattan distance matrix;

[0027] The classification discrimination submodule calls the inter-segment difference measure value in the Manhattan distance matrix, inputs the pre-trained support vector machine classification model, calculates the geometric interval between the distance value and the classification hyperplane according to the model kernel function, determines whether the interval value is higher than the model decision boundary threshold, outputs the binary classification probability value and maps it to the segment attribution category, and generates the segment identification label.

[0028] As a further embodiment of the present invention, the system further comprises:

[0029] a fault path mapping module, configured to receive the node hierarchical sorting list and the section identification label, match the node hierarchical sorting list with the section identification label, screen a node set that meets the timing and reflection characteristics, iteratively calculate the path transition probability through a hidden Markov model, generate an actual propagation path topology map of the cable fault, and output the actual propagation path topology map to a monitoring terminal;

[0030] The actual propagation path topology diagram includes a fault propagation starting point, a propagation relay node, and a fault terminal node.

[0031] As a further solution of the present invention, the fault path mapping module includes:

[0032] The node matching submodule calls the node hierarchical sorting list and the timing threshold, extracts the node numbers in the segment identification label that meet the reflection characteristic constraint, compares the timing step in the node hierarchical sorting with the phase continuity of the segment identification label, selects the node combination whose timing deviation is less than the step threshold and whose phase continuity is higher than the phase threshold, and generates a candidate node set;

[0033] The probability iteration submodule loads the state transition matrix and observation probability matrix of the hidden Markov model based on the node transfer relationship in the candidate node set, calculates the initial state transition probability according to the fault propagation time sequence step, and iteratively updates the transition probability weights between nodes through the forward-backward algorithm to generate a transition probability matrix;

[0034] The path generation submodule calls the probability weights between nodes in the transfer probability matrix, selects the transfer path nodes layer by layer according to the maximum a posteriori probability criterion, superimposes the node hierarchy order and the probability weights to generate a path connection relationship, traverses all the hierarchical nodes to complete the path topology construction, and generates a propagation path topology map.

[0035] Compared with the prior art, the advantages and positive effects of the present invention are:

[0036] In the present invention, by calculating the angle between the voltage and current increment directions at adjacent time points and extracting the gradient change inflection point, the spatiotemporal characteristics of the transient disturbance signal are directly associated, reducing the interference of noise on the starting position mark. The time difference is fitted in combination with the cable length data, and abnormal nodes with deviations exceeding the threshold are dynamically filtered to eliminate the influence of sampling synchronization errors on hierarchical sorting, thereby improving the stability of fault path mapping. The reflection characteristic parameters are generated by integrating the amplitude and peak interval time of the main wave and the reflected wave, and the coupling relationship between amplitude attenuation and phase offset is quantified to enhance the signal separation capability in the scenario of multiple reflection paths superposition. Based on the Manhattan distance and support vector machine classification model to judge the similarity of reflection parameters within the physical section, the fault section is dynamically divided in combination with the line topology structure to solve the positioning ambiguity problem caused by cable branching or parallel connection in traditional methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 is a system flow chart of the present invention;

[0038] Figure 2 This is a flow chart of the disturbance feature extraction module of the present invention;

[0039] Figure 3 This is a flow chart of the timing level determination module of the present invention;

[0040] Figure 4 This is a flow chart of the reflection characteristic analysis module of the present invention;

[0041] Figure 5 This is a flow chart of the topology segment identification module of the present invention;

[0042] Figure 6 This is a flow chart of the fault path mapping module of the present invention. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0044] In the description of the present invention, it should be understood that the terms "length," "width," "up," "down," "front," "back," "left," "right," "vertical," "horizontal," "top," "bottom," "inside," "outside," and the like, indicating positions or relationships, are based on the positions or relationships shown in the accompanying drawings and are intended only to facilitate the description of the present invention and simplify the description. They do not indicate or imply that the devices or elements referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limiting the present invention. Furthermore, in the description of the present invention, "plurality" means two or more, unless otherwise expressly and specifically defined.

[0045] Example 1

[0046] See also Figure 1 , a cable fault analysis system based on big data includes:

[0047] The disturbance feature extraction module is used to obtain voltage and current time series data through the cable monitoring node, calculate the direction angle between the voltage increment and the current increment at adjacent time points, extract the gradient change inflection point, generate the disturbance starting position mark, and pass the disturbance starting position mark to the time series level determination module;

[0048] The gradient change inflection point is calculated by calculating the directional change rate of the voltage and current increment vectors at adjacent time points. When the directional change rate exceeds a preset threshold, it is determined to be an inflection point.

[0049] The timing hierarchy determination module is used to call the disturbance starting position marker, fit the time difference with the cable length data, generate a node hierarchy sorting list, eliminate abnormal nodes with time difference deviation exceeding three times the standard deviation, and pass the node hierarchy sorting list to the fault path mapping module;

[0050] The time difference deviation is calculated by the mean and standard deviation of the time difference of each node and the normal distribution assumption is used for anomaly detection.

[0051] The reflection characteristic analysis module is used to collect the amplitude and peak interval time of the main wave and the reflected wave, calculate the amplitude attenuation coefficient and phase offset, combine them into reflection characteristic parameters, and pass the reflection characteristic parameters to the topology section identification module;

[0052] The amplitude attenuation coefficient is defined as the ratio of the reflected wave amplitude to the main wave amplitude, and is used to quantify the energy loss during signal propagation.

[0053] The topological segment identification module is used to call the reflection characteristic parameters, calculate the Manhattan distance of the reflection parameters in the physical segment, identify the similarity through the support vector machine classification model, output the segment identification label, and pass the segment identification label to the fault path mapping module;

[0054] The support vector machine classification model uses the radial basis kernel function to distinguish the differences in reflectance characteristics of different segments through supervised learning.

[0055] The fault path mapping module is used to receive the node hierarchical sort list and section identification labels, match the hierarchical sort with the section labels, screen the node set that meets the timing and reflection characteristics, iteratively calculate the path transfer probability through the hidden Markov model, generate the actual propagation path topology map of the cable fault, and output the actual propagation path topology map to the monitoring terminal;

[0056] The path transition probability constructs a state transition matrix based on the temporal correlation between nodes and the similarity of reflection parameters, and the optimal path sequence is solved by the Viterbi algorithm.

[0057] The disturbance starting position mark is specifically the disturbance gradient inflection point position, the node hierarchy sorting list includes the node propagation order, the valid node set, and the abnormal node set, the reflection characteristic parameters include the amplitude attenuation coefficient and the phase offset, the section identification label specifically refers to the cable topology physical section identification code, and the actual propagation path topology diagram includes the fault propagation starting point, the propagation relay node, and the fault terminal node.

[0058] See also Figure 2 , the disturbance feature extraction module includes:

[0059] The incremental vector calculation submodule monitors the voltage and current time series data of the cable nodes, extracts the voltage and current differences between adjacent time points, calculates the incremental vectors through the differential algorithm, and arranges the continuous time series incremental vectors in chronological order to generate an incremental vector sequence;

[0060] Monitor the voltage and current time series data of the cable nodes. For node A, the voltage U is monitored at time t = 1ms. t；1 is 220.1V, current I t；1 is 10.5A, and the voltage U monitored at the next time point t = 2ms is t is 219.8V, current I t =10.8A, perform voltage difference calculation ΔU=U t -U t；1 =219.8-220.1=

[0061] -0.3V, perform current difference calculation ΔI=I t -I t；1 =10.8-10.5=0.3A, combine these two differences to form the increment vector at the current time point t=2ms Repeat the process of obtaining voltage and current values and calculating the difference at subsequent time points t = 3ms, 4ms, ..., T. For example, obtain U t<3ms =219.9V,It<3ms =10.7A, we get ΔU=219.9-219.8=0.1V, ΔI=10.7-10.8=-0.1A, and get the incremental vector The increment vector calculated at each time point Arrange them in the time sequence of t=2ms, 3ms, ..., T to form a sequence containing multiple incremental vectors. This sequence reflects the dynamic situation of the cable node voltage and current changing with time, and generates an incremental vector sequence.

[0062] The gradient inflection point detection submodule calculates the angle between adjacent vector directions based on the incremental vector sequence using the direction cosine formula, applies a sliding window algorithm to perform local mean filtering on the angle sequence, and determines the inflection point position in combination with the angle change rate threshold to generate a gradient inflection point set.

[0063] Based on the sequence of incremental vectors generated in the previous step, for example and Calculate the angle θ between the two adjacent vectors t<3ms , the direction cosine formula is used for calculation, which involves the dot product of two vectors and the product of their respective moduli. Specifically, the calculation vector Length of mold Calculating vectors Length of mold Calculates the dot product of two vectors Then calculate the cosine of the angle Due to the calculation accuracy or noise, the cosine value slightly exceeds the range of [-1,1] and is corrected to -1, then the angle θ t<3ms =

[0064] arccos(-1) = π radians. Repeat this angle calculation for all adjacent vector pairs in the incremental vector sequence to obtain an angle sequence [θ3,θ4,…,θ T ], set the sliding window size W = 5, perform local mean filtering on the angle sequence, for example, calculate the filtered angle at the i-th position It is necessary to take the W angle values from i to i+W-1 in the original angle sequence and calculate their arithmetic mean. For example, if the sequence is [3.14, 3.0, 2.8, 1.5, 1.6, 1.7, ...], and the first window covers [3.14, 3.0, 2.8, 1.5, 1.6], calculate the mean

[0065] Radians, slide the window one position to the right, calculate radians, continue this process to obtain the filtered angle sequence, and then calculate the rate of change of the filtered angle sequence, that is, the difference between adjacent filtered angles For example Radians / sampling point, set the angle change rate threshold Θ rate_th The threshold is set by referring to the statistical distribution of the angle change rate in the historical normal operation data. Usually, the statistical mean plus three times the standard deviation is taken. For example, if the standard deviation of the angle change rate is 0.1 radians / sampling point and the mean is 0.05 radians / sampling point according to the historical data analysis, the threshold can be set as Θ rate_th =0.05+3×0.1=0.35 radians / sampling point, the calculated angle change rate With threshold Θ rate_th For comparison, if The time point corresponding to position i is determined to be an inflection point, and all positions i that meet the conditions are collected to generate a gradient inflection point set.

[0066] Direction cosine formula calculates two vectors and The cosine of the angle θ between them, in represents the dot product of vectors, and Respectively represent the modulus of the vector, the dot product calculates the sum of the products of the corresponding elements, and the modulus calculates the square root of the sum of the squares of each element. The entire formula reflects the directional relationship between vectors (same direction, opposite direction, orthogonal) through the dot product, and then normalizes it through the modulus product to obtain the cosine value in the range of [-1,1] to determine the angle, which is used here to quantify the degree of directional change between consecutive incremental vectors. Using the aforementioned vector and θ t<3ms = arccos(-1) = π radians (or 180 degrees). Threshold Θ rate_th Set the angle change rate data for a period of time (for example, one week) under historical normal operating conditions and calculate its mean μ Δθ and standard deviation σ Δθ , for example, we can calculate μ Δθ =0.05 radians / sampling point, σ Δθ = 0.1 radians / sampling point. To effectively identify significant changes while avoiding misjudgment of noise, the 3σ principle is used to set the threshold, Θ rate_th =μ Δθ +3×σ Δθ =0.05+3×0.1=0.35 radians / sampling point.

[0067] Calculated angle change rate Less than the threshold Θ rate_th =0.35 radians / sampling point, indicating that the change at this point is not significant and is not determined to be an inflection point. Radians / sampling point, since 0.4>0.35, the point k is determined to be an inflection point and recorded in the gradient inflection point set.

[0068] The perturbation mark generation submodule calls the gradient inflection point set, selects the starting inflection point according to the difference in gradient direction before and after the inflection point using the extreme value positioning algorithm, maps the timestamp corresponding to the inflection point to the original time series data, and generates a perturbation starting position mark.

[0069] Call the gradient inflection point set generated in the previous step, which contains multiple time points identified as significant changes. For example, the set is {k1, k2, k3, ...}. For each inflection point k i , check the gradient direction difference value of the incremental vector sequence before and after, this difference value is calculated by the inflection point k i The previous paragraph (for example, k i -W to k i -1) The average direction of the vector and the inflection point k i The latter part (e.g. i to k i The average direction can be obtained by calculating the direction angle of the sum of the vectors in this segment, or by directly calculating the direction angle of the average incremental vector in this segment, for example, calculating W=5 vectors before the inflection point k2 The average vector Calculate its direction angle φ before Similarly, after calculating the inflection point k2, W = 5 vectors The average vector and its direction angle φ after , calculate the direction difference value Set the direction difference threshold Φ diff_th For example, based on experience or disturbance characteristic analysis, it is set to π / 2 radians (90 degrees), indicating that the direction has deflected significantly. The extreme value positioning algorithm is used to select the starting inflection point. This process is to find the first one that satisfies The inflection point k i , that is, find the inflection point with the smallest sequence number in the gradient inflection point set and the difference in the gradient direction before and after it is greater than the threshold. For example, in the inflection point set {k1, k2, k3}, we can calculate If Φ diff_th =1.57 (i.e., π / 2), then k2 is the first inflection point that meets the conditions, so k2 is selected as the starting inflection point, and the timestamp corresponding to this starting inflection point k2 is, for example Map it back to the original voltage and current time series data sequence, mark the time point 15ms as the starting moment of the disturbance event, and generate a disturbance starting position mark.

[0070] See also Figure 3 , the timing level determination module includes:

[0071] The time difference fitting submodule uses the node position coordinates recorded in the disturbance starting position mark, combines the conductor material parameters in the cable length data and the line segment length, calculates the signal transmission delay between adjacent nodes through linear interpolation, and generates a time difference set;

[0072] Call the node position coordinates recorded in the disturbance start position mark. For example, the mark shows that the disturbance starts at node B, and its position coordinates on the cable line are x B = 150m (from the starting point of the line), and at the same time obtain the position coordinates of the adjacent monitoring nodes, such as the coordinates of node A x A =50m, node C coordinate x C = 250m, combined with the conductor material parameters and line segment length information obtained from the cable length data file or database, for example, the cable model is YJV-10kV, its conductor material is copper, and the corresponding electromagnetic wave propagation speed v≈1.7×10 8 m / s (this is an approximate value, which can be accurately determined according to the cable parameter manual or measured and calibrated), and the line segment length L between node A and node B is obtained. AB =|x B -x A |=|150-50|=100m, obtain the line segment length L between node B and node C BC =|x C -x B |=|250-150|=100m, apply linear interpolation to calculate the signal transmission delay between adjacent nodes. Since the signal (disturbance wave) propagates along the cable, its speed is approximately constant, so the delay time Δt is proportional to the distance L, Δt=L / v, calculate the transmission delay Δt from node A to node B AB =L AB / v=100m / (1.7×10 8 m / s)≈0.588×10 ；6 s = 0.588 μs, calculate the transmission delay Δt from node B to node C BC =L BC / v=100m / (1.7×10 8 m / s)≈0.588μs, repeat this process for all adjacent monitoring node pairs in the system, for example, there is node D at x D =180m, then calculate the delay Δt from B to D BD =|180-150| / (1.7×10 8 )=30 / (1.7×10 8 )≈0.176μs, collect all the calculated delay time values to form a set, such as

[0073] {Δt AB:0.588μs,Δt BC :0.588μs,Δt BD :0.176μs,…}, generating a time difference set.

[0074] In this scenario, the linear interpolation method assumes that the signal propagates at a constant speed along the cable, and the propagation time is linearly related to the distance, t = L / v. This is itself a linear relationship. Therefore, "linear interpolation" here mainly refers to calculating the propagation time between two points based on the known node position (distance) and propagation speed. In essence, it uses a linear proportional relationship. If it is known that the disturbance occurs at point B (150m), then the time difference for the disturbance to propagate to point A (50m) is (150-50) / v = 100 / v, and the time difference for the disturbance to propagate to point D (180m) is (180-150) / v = 30 / v.

[0075] The hierarchical sorting construction submodule arranges the nodes from small to large according to the transmission delay based on the time difference set, and performs secondary sorting on the nodes with the same delay value using the cable length data to generate a hierarchical node sequence table;

[0076] Based on the time difference set generated in the previous step, this set contains the signal transmission delay between each adjacent node. For example, the set is {AB: 0.588μs, BC: 0.588μs, BD: 0.176μs, CE: 0.4μs, ...} (assuming it has been converted into a node pair form). Now it is necessary to hierarchically sort all the nodes participating in the monitoring (such as A, B, C, D, E). The sorting is based on the transmission delay of the signal from the disturbance source (assuming it is point B) to each node. The disturbance source B is taken as the 0th layer, and the total delay from B to other nodes is calculated. For directly adjacent nodes, the delay is the value in the time difference set, such as the delay from B to A is t BA =0.588μs, delay t from B to C BC =0.588μs, delay from B to D t BD = 0.176μs. For non-directly adjacent nodes, the delay on the path needs to be accumulated. For example, assuming E is connected to B through C, the delay from B to E is t BE =t BC +t CE =

[0077] 0.588μs+0.4μs=0.988μs, obtain the delay list of all nodes relative to the disturbance source B {B:0μs,D:0.176μs,A:0.588μs,C:0.588μs,E:0.988μs}, arrange the nodes from small to large according to the transmission delay, and obtain the preliminary sorting [B,D,A,C,E]. Check whether there are nodes with the same delay value. It is found that the delay of node A and node C is both 0.588μs. At this time, a secondary sorting is required, and the cable length data is retrieved to compare the physical cable path lengths of the two nodes to the disturbance source B. In this case, L BA =100m,

[0078] L BC =100m, and the length is also the same. If there are other configurations, for example, one is connected by an overhead line and the other is connected by an underground cable, even if the straight-line distance is the same, the actual cable length may be different. Or if the topology allows, the path segment attributes (such as model, laying method) from B to A and B to C are different, resulting in different equivalent lengths or priorities, then they are distinguished according to preset rules (such as giving priority to trunk line nodes, or sorting by node number). Assuming that the nodes are sorted in alphabetical order, A is ranked before C, and the final order is still [B, D, A, C, E]. Each node is assigned a weight based on this order, for example, B has a weight of 0, D has a weight of 1, A has a weight of 2, C has a weight of 3, and E has a weight of 4, to generate a hierarchical node sequence table.

[0079] The abnormal node elimination submodule calls the sorting weights in the hierarchical node sequence table, calculates the standard deviation of all node weight values, screens out nodes whose weight deviation exceeds three times the standard deviation threshold, removes them from the hierarchical node sequence table, and generates a node hierarchical sorting list.

[0080] Call the hierarchical node sequence list and its corresponding sorting weight generated in the previous step, for example, the list is

[0081] [B(0), D(1), A(2), C(3), E(4)], weight value is W=[0,1,2,3,4], calculate the standard deviation σ of the weight value of all nodes W , first calculate the average value of the weight μ W =(0+1+2+3+4) / 5=10 / 5=2, calculate the variance Calculating standard deviation Set a screening threshold, which is defined as three times the standard deviation, or 3σ W =3×1.414=4.242, check the weight value W of each node i Relative to the average value μ W Deviation of |W i -μ W|, and compare it with the three-standard deviation threshold to calculate the deviation of each node: |0-2|=2, |1-2|=1, |2-2|=0, |3-2|=1, |4-2|=2. Comparing these deviations with the threshold of 4.242, it is found that the deviation of all nodes (the maximum is 2) does not exceed 4.242, which means that in this example, no node's weight deviation reaches the level that needs to be eliminated. Suppose there is another node F, whose weight is calculated as W F =10, then μ W =(0+1+2+3+4+10) / 6=20 / 6≈3.33, 3σ W =3×3.25=9.75, the deviation of node F

[0082] |W F -μ W |=|10-3.33|=6.67, which does not exceed the threshold of 9.75. If the node F weight is 15, then μ W =(0+1+2+3+4+15) / 6=25 / 6≈4.17, Var(W)≈19.97,σ W ≈

[0083] 4.47,3σ W ≈13.41, the node F deviation |15-4.17|=10.83, which still does not exceed the threshold. This shows that the 3σ rule may not be sensitive under small sample sizes, or the threshold standard needs to be adjusted according to the actual data distribution (for example, changing to 2σ or a fixed threshold). Assume that the adjusted threshold is 2σ W ≈2×4.47=8.94, then the deviation of node F is 10.83>8.94, and node F is determined to be an abnormal node, and it is removed from the hierarchical node sequence list. In the original example, no node is removed, so the final node hierarchical sort list remains as

[0084] [B(0),D(1),A(2),C(3),E(4)], generates a hierarchical sorted list of nodes.

[0085] See also Figure 4 , the reflection characteristics analysis module includes:

[0086] The waveform acquisition submodule collects synchronous measurement data of the main wave amplitude voltage peak value and the reflected wave amplitude voltage peak value, records the millisecond-level change in the time interval between the two peaks, integrates the main wave peak value, reflected wave peak value and time difference data into a waveform recording unit with a unified timestamp, and generates a waveform feature set;

[0087] Collect synchronous measurement data of the main wave amplitude voltage peak and the reflected wave amplitude voltage peak. For example, in a disturbance event, the voltage peak U of the main wave (incident wave) is recorded at the monitoring point (such as the disturbance starting point B) by a high-frequency sampling device. main_peak =15.5kV. After a period of time, the reflected wave generated by the line end or impedance mismatch point is recorded, and its voltage peak value U refl_peak =8.2kV, and accurately record the time points of the two peaks, and calculate the time interval Δt between the two peaks. peak = 5.8μs. This time difference is recorded at the millisecond level (the original text here is at the millisecond level, but the example is at the microsecond level. We continue with the microsecond level example and assume that the original text refers to high-precision time measurement) or higher precision. These three data points—main wave peak value 15.5kV, reflected wave peak value 8.2kV, time difference 5.8μs—are integrated together and associated with a unified timestamp (for example, the disturbance start timestamp 15ms plus the time when the main wave peak reaches this point) to form a waveform recording unit. For example, the unit content is

[0088] {Timestamp:15.005ms,MainPeak:15.5kV,ReflPeak:8.2kV,TimeDiff:5.8μs}. Repeat this process for multiple sets of such main wave and reflected wave data captured during disturbance or periodic detection to generate a set containing multiple waveform recording units and generate a waveform feature set.

[0089] The parameter calculation submodule extracts the main wave peak value and the reflected wave peak value in the waveform feature set, calculates the natural logarithm value of the amplitude ratio of the two using logarithmic operation, combines the peak time difference with the preset electromagnetic wave propagation speed constant, converts the product of the time difference and the speed into a phase angle offset radian value, and generates a characteristic parameter component;

[0090] Extract the main wave peak value and reflected wave peak value data from the waveform feature set generated in the previous step. For example, take the recording unit data of the previous step: U main_peak =15.5kV and U refl_peak =8.2kV, perform logarithmic operation to calculate the natural logarithm value of the amplitude ratio of the two, that is, the amplitude attenuation coefficient α, first calculate the amplitude ratio Ratio = U refl_peak / U main_peak =8.2kV / 15.5kV≈0.529, then calculate its natural logarithm α=ln(Ratio)=ln(0.529)≈-0.637 (unit: Neper). This negative value indicates that the amplitude of the reflected wave has attenuated relative to the main wave. Then combine it with the recorded peak time difference Δt peak =5.8μs=5.8×10 ；6s and the preset electromagnetic wave propagation speed constant v=1.7×10 8 m / s (this value is determined by the cable type and insulation material and should be pre-calibrated or obtained from the manual). Convert the product of the time difference and the speed into a phase angle offset in radians Δφ to calculate the round-trip distance L of the signal. round_trip =v×Δt peak =(1.7×10 8 m / s)×(5.8×10 ；6 s) = 986m, this distance is the total distance the signal propagates from the measurement point to the reflection point and then returns to the measurement point, the one-way distance L single =L round_trip / 2=986 / 2=493m. The phase angle shift is usually related to the wavelength λ, which is λ=v / f, where f is the signal frequency. If the phase relationship at the fundamental frequency or a certain characteristic frequency is considered, assuming f=50Hz, then λ=(1.7×10 8 ) / 50=3.4×10 6 m, phase shift Δφ=(2π×L round_trip ) / λ(mod 2π), that is, Δφ=(2π×986) / (3.4×10 6 )(mod 2π)≈0.00182(mod 2π)≈0.00182 radians, or the phase offset can also be understood as a delay relative to a reference period, directly using the time difference Δt peak If the time delay itself is of interest as a characteristic, it can be converted into a phase angle Δφ = 2πfΔt corresponding to a specific frequency (such as f = 1MHz) peak =2π×(1×10 6 Hz)×(5.8×10 ；6 s)=11.6π radians, which is 1.6π radians after modulo 2π. Here, it is assumed that the calculated radian value is directly used, for example, Δφ=0.00182 radians, and the calculated amplitude attenuation coefficient α≈-0.637 and the phase offset radian value Δφ≈0.00182 are used as components of the characteristic parameter to generate the characteristic parameter component.

[0091] The logarithmic operation is specifically performed by calculating the natural logarithm of the amplitude ratio α=ln(U refl_peak / U main_peak ) is used to convert the multiplicative attenuation relationship into an additive one, which facilitates subsequent processing and analysis. It also makes the representation of the attenuation degree more consistent with physical perception (logarithmic scale). The result unit is neper (Np). The calculation example is as follows:

[0092] α=ln(8.2 / 15.5)=ln(0.529)≈-0.637Np.

[0093] Phase angle calculation specifically reflects the time (or space) delay of the reflected wave relative to the main wave through the phase angle offset Δφ, which is calculated as Δφ = (2π × round-trip distance) / wavelength (mod 2π), or directly relates the time difference and frequency Δφ = 2πfΔt (mod 2π). The distance-based calculation method is used here, and in the example, Δφ ≈ 0.00182 radians.

[0094] The characteristic integration submodule calls the amplitude attenuation coefficient and the phase offset radian in the characteristic parameter components, aligns the two in time series, and then merges them into a double-column matrix to generate reflection characteristic parameters.

[0095] Call the characteristic parameter components generated in the previous step, namely the amplitude attenuation coefficient α and the phase shift radian Δφ. Assume that by processing multiple waveform recording units, a series of amplitude attenuation coefficient sequences A=[α1,α2,…,α N ] and the phase shift radian sequence P = [Δφ1, Δφ2, ..., Δφ N ], for example, A = [-0.637, -0.650, -0.642, ...], P = [0.0018, 0.0019, 0.0017, ...] (the phase values are adjusted to be closer to the example values), ensuring that the two sequences are arranged in the same timestamp order, i.e., α i and Δφ i Corresponding to the measurement or calculation result at the i-th time point, these two one-dimensional sequences of length N are merged column by column to form an N×2 double-column matrix. Each row of the matrix represents the reflection characteristics of a time point. The first column is the amplitude attenuation coefficient, and the second column is the phase offset in radians. For example, the first three rows of the matrix M are: This matrix integrates information reflecting two dimensions, namely reflection intensity and delay time (phase), to generate reflection characteristic parameters.

[0096] See also Figure 5 , the topology segment identification module includes:

[0097] The parameter processing submodule obtains the amplitude characteristics and phase difference characteristics of the reflection characteristic parameters, normalizes the characteristic sequence using the mean and standard deviation, compares the deviation degree of each data point in the characteristic sequence based on the parameter fluctuation threshold, selects the continuous segments with deviation values less than the threshold, and generates a standardized parameter set;

[0098] Obtain the reflection characteristic parameter matrix M generated in the previous step. This matrix contains two columns of data: the amplitude attenuation coefficient (amplitude feature) and the phase offset radian (phase difference feature). Normalize each column of the feature sequence to eliminate the dimension and value range differences. Use mean and standard deviation for normalization (Z-score normalization). First, calculate the mean μ of the amplitude feature sequence A = [-0.637, -0.650, -0.642, ...]. A and standard deviation σ A , assuming that μ is calculated A =-0.645 and σ A =0.007, for each element in the sequence α i Perform normalization calculation A′ i =(α i -μ A ) / σ A For example, after the first element is standardized, A′1=(-0.637-(-0.645)) / 0.007=0.008 / 0.007≈1.14, and the second element A′2=(-0.650-(-0.645)) / 0.007=-0.005 / 0.007≈-0.71.

[0099] P=[0.0018,0.0019,0.0017,…] Similarly, calculate the mean μ P and standard deviation σ P , assuming μ P =0.0018 and σ P = 0.0001, calculate the standardized phase difference sequence P′ i =(P i -μ P ) / σ P , for example, P′1=(0.0018-0.0018) / 0.0001=0, P′2=(0.0019-0.0018) / 0.0001=1, P′3=(0.0017-0.0018) / 0.0001=-1, obtain the standardized feature sequences A′ and P′, and set a parameter fluctuation threshold T fluct , the threshold is used to define the stable range of the data. Its setting can be based on the statistical analysis of the fluctuation of the standardized parameters under normal operating conditions. For example, a value of 1.5 means that the deviation from the mean is allowed to be no more than 1.5 times the standard deviation. Comparing each data point A' in the standardized feature sequence A' and P' i ,P′ i The absolute value of |A′ i |,|P′ i |With threshold T fluct , filter out those |A′ i | <T fluctAnd |P′ i | <T fluct The segment composed of continuous data points, for example, if T fluct =1.5, then (A′1=1.14, P′1=0) meets the condition, (A′2=-0.71, P′2=1) meets the condition, (A′3, P′3) (assuming A′3=-0.43, P′3=-1) also meets the condition, if (A′4=

[0100] 2.0, P′4=0.5), then A′4 exceeds the threshold and the point does not meet the requirements. If the points that continuously meet the requirements constitute a time period, then this time period is regarded as a continuous segment. All the selected continuous segments and their corresponding standardized parameter data are collected to generate a standardized parameter set.

[0101] Threshold T fluct The setting method is to collect the reflection characteristic parameters of the cable under normal operating conditions, perform standardization, analyze the fluctuation range of the standardized parameter sequence, calculate its standard deviation (theoretically, the standard deviation after standardization is 1), and observe the data distribution. For example, set T fluct =1.5 means that the segment within the data fluctuation range of about 86.6% (for normal distribution) is accepted as the stable segment, which can be adjusted according to actual needs, such as T fluct =2 corresponds to approximately 95.4% of the data.

[0102] The distance calculation submodule extracts the amplitude difference sequence and phase difference difference sequence of adjacent physical segments based on the standardized parameter set, calculates the absolute deviation value of the difference sequence point by point according to the Manhattan distance formula, slides through all segment pairs using a fixed step window, accumulates the deviation values in the window to generate the difference metric between segments, and generates the Manhattan distance matrix;

[0103] Based on the standardized parameter set generated in the previous step, which contains the standardized amplitude feature A′ and phase difference feature P′ sequences of multiple continuous stable segments, the feature sequences corresponding to adjacent physical segments are extracted. For example, the feature sequence of segment j is The characteristic sequence of segment j+1 is Assuming that the two segments are adjacent in time, calculate the difference between them and extract the average value of the feature sequences of the two segments as a representative, for example, the average amplitude of segment j and average phase Average amplitude of segment j+1 and average phase Calculate the amplitude difference and phase difference The two differences are calculated according to the Manhattan distance formula to obtain the difference measure D between segments j and j+1. j,j:1 =

[0104] |ΔA′ j,j:1 |+|ΔP′ j,j:1 |, that is (The original description here seems ambiguous, is it the Manhattan distance of the difference sequence or the Manhattan distance between the segment representative values? The latter interpretation is more reasonable). Assuming that the average normalized value of segment j is Segment j+1 is Then D j,j:1 =|-0.5-1.0|+|0.8-0.5|=|-1.5|+

[0105] |0.3|=1.5+0.3=1.8. Another way to understand this is that the original text refers to calculating the difference sequence point by point within the two segments, and then calculating a certain cumulative value of the difference sequence. However, according to "extracting the amplitude difference sequence and phase difference difference sequence of adjacent physical segments" and then "calculating the absolute deviation value of the difference sequence point by point" and then "accumulating the deviation value within the window", this seems more complicated. For the time being, we will explain it by calculating the Manhattan distance of the segment representative value. Repeat this difference metric calculation for all identified adjacent segment pairs (j, j+1) and add all the calculated D j,k The values (where j, k represent the segment numbers) are organized into a matrix, and the elements (j, k) of the matrix represent the difference measure (distance) between segment j and segment k, generating a Manhattan distance matrix.

[0106] Manhattan distance formula explanation: For two points X = (x1, x2) and Y = (y1, y2), their Manhattan distance D M (X,Y)=|x1-y1|+|x2-y2|. This formula calculates the sum of the distances two points travel along the axis in the standard coordinate system. Here, it is used to quantify the comprehensive difference between the two segments in the standardized amplitude and phase feature dimensions. The greater the difference, the greater the distance. The calculation example is as follows:

[0107] Use the average normalized value of segments j and j+1 and Manhattan distance D j,j:1 =|1.0-(-0.5)|+|0.5-0.8|=|1.5|+|-0.3|=1.5+0.3=1.8.

[0108] The classification discrimination submodule calls the inter-segment difference measure value in the Manhattan distance matrix, inputs the pre-trained support vector machine classification model, calculates the geometric interval between the distance value and the classification hyperplane according to the model kernel function, determines whether the interval value is higher than the model decision boundary threshold, outputs the binary classification probability value and maps it to the segment attribution category, and generates the segment identification label.

[0109] Call the Manhattan distance matrix generated in the previous step, which stores the difference measurement value D between each segment j,k , these distance values are input as features into a pre-trained support vector machine (SVM) classification model. The model has been trained by learning a large number of segment pairs with known categories (for example, which segment pairs belong to the same fault type or the same cable segment). For the segment pair (j, k) that needs to be distinguished, its corresponding distance value D j,k =1.8 is input to the SVM model, and the model performs operations internally (involving kernel functions, such as linear kernels or Gaussian kernels, calculating the distance between the input point and the support vector, and combining weights), and finally calculates the input point D j,k The geometric interval (distance) to the classification hyperplane, assuming that the interval value calculated by the model is g = 0.6, and the model also has a preset decision boundary threshold T SVM , the threshold is determined during model training and is used to distinguish different categories, such as T SVM =0.5, compare the calculated interval value g with the decision boundary threshold T SVM , because g=0.6>T SVM = 0.5, the interval value is judged to be higher than the threshold, which usually means that the point (segment pair) clearly belongs to a certain class. The SVM model outputs a binary probability value (or confidence), such as P(class1|D j,k )=0.85 (the probability of belonging to category 1 is 85%), and this probability value is mapped to the segment category label. For example, if the probability is set to be greater than 0.7, it is determined to be category 1 (which may mean "the segments belong to the same fault area"), otherwise it is category 0 ("the segments belong to different areas or states"). Since 0.85>0.7, it is determined that segments j and k belong to the same category, and the identification label "category 1" is assigned to this pair of segments or one of the segments (depending on the specific task objectives). This process is repeated for all segment pairs that need to be classified to generate segment identification labels.

[0110] Decision boundary threshold T SVM The setting method is: This threshold is usually determined by optimizing the classification performance (such as accuracy, recall rate, F1 score) during the SVM model training and validation phase. For example, different thresholds are tested on the validation set, and the threshold that makes a certain evaluation indicator optimal is selected. Assume that T is determined by cross-validation. SVM =0.5 can achieve the best balance.

[0111] See also Figure 6 , the fault path mapping module includes:

[0112] The node matching submodule calls the node hierarchical sorting list and the timing threshold, extracts the node numbers in the segment identification label that meet the reflection characteristic constraint, compares the timing step in the node hierarchical sorting with the phase continuity of the segment identification label, selects the node combination whose timing deviation is less than the step threshold and whose phase continuity is higher than the phase threshold, and generates a candidate node set;

[0113] Call the node hierarchy sorting list generated in the previous step, such as [B(0), D(1), A(2), C(3), E(4)], and a preset timing threshold T time_step , which defines the maximum time interval allowed for two nodes to be considered adjacent on the fault propagation path, for example, T time_step = 1.0μs, and at the same time, the segment identification labels are called. These labels may divide some nodes into specific segments (for example, segment 1 contains nodes {B, D}, and segment 2 contains nodes {A, C}). The node numbers that meet the reflection characteristic constraints in the segment identification labels are extracted. This means that the selected nodes and the segments they constitute need to meet the stability or change pattern requirements obtained from the previous reflection characteristic analysis. The timing step lengths of adjacent nodes in the node hierarchy sorting list (that is, the calculated transmission delay between them) are compared with the timing threshold T time_step For example, nodes D(1) and A(2) are adjacent in the sorting, and the temporal relationship between them needs to be further confirmed. Assume that the disturbance actually reaches D at time t D =0.176μs, time to reach A t A = 0.588 μs (relative to source B), time difference t A -t D =0.412μs, this difference is less than the threshold T time_step =1.0μs, which satisfies the timing step constraint. At the same time, the phase continuity indicated by the segment identification label also needs to be considered, that is, whether the two nodes are located in the same or logically continuous segment, and whether their corresponding reflection characteristic parameters (especially phase) show continuity or conform to the expected change pattern. The phase continuity threshold T is set. phase_cont For example, it is required that the phase difference of adjacent nodes is less than 0.1 radians, or they belong to the same continuous segment determined by SVM. Assume that nodes D and A both meet the requirement of timing deviation less than 1.0 μs and their phase continuity is higher than the threshold T according to segment label and phase analysis. phase_cont , then the node pair (D, A) is considered as a possible connection, and this screening process is performed on the adjacent node pairs in all hierarchical sorted lists and the possible spanning node pairs (if supported by the segment identifier). All node combinations (node pairs or longer sequences) that meet the timing and phase constraints are collected to generate a candidate node set.

[0114] Timing threshold Ttime_step The setting method is: refer to the accuracy of signal measurement in the system, the expected fault propagation speed and the minimum distance between nodes, and set a reasonable threshold. For example, if the minimum inter-node delay is 0.1μs and the measurement error is 0.05μs, then T can be set. time_step The time interval is between 0.1 + 0.05 + margin = 0.2 μs and 1.0 μs, and a balance must be established between missed detection and false detection.

[0115] Phase threshold T phase_cont The setting method is: based on the statistical distribution of the phase change between consecutive nodes under normal operation and known fault modes, for example, taking the upper limit of the normal fluctuation range, or setting the expected phase jump amplitude according to the fault type, such as 0.1 radian.

[0116] The probability iteration submodule loads the state transition matrix and observation probability matrix of the hidden Markov model based on the node transfer relationship in the candidate node set, calculates the initial state transition probability according to the fault propagation time sequence step, and iteratively updates the transition probability weights between nodes through the forward-backward algorithm to generate a transition probability matrix;

[0117] Based on the candidate node set generated in the previous step, this set contains possible node transfer relationships that meet the timing and phase constraints. For example, the set contains node pairs such as (B, D), (D, A), (D, C), and (C, E). Load a pre-built hidden Markov model (HMM) state transfer matrix A and observation probability matrix B. The element a of the state transfer matrix A is ij represents the prior probability of transferring from node i to node j (for example, based on historical failure data statistics or network topology connection strength), and the element b of the observation probability matrix B is j (O t ) indicates that a certain feature O is observed at node j tThe probability of state transition is calculated based on the timing step of fault propagation (such as specific reflection waveform characteristics or delay time). For example, the reciprocal of the delay time can be used as a factor of the transition probability. The shorter the delay, the higher the initial transition probability. Alternatively, the adjacent relationship determined by the hierarchical sequence table can be directly used as the basis for the higher initial probability. The forward-backward algorithm (the algorithm is not described in detail here, only its application is explained) is used to combine the observed actual data (such as the disturbance arrival time and waveform characteristics of each node) to iteratively update the transition probability weights between nodes. The algorithm will calculate Given an observation sequence, the probability of the model being in each state (node) and the probability of a state transition occurring are calculated iteratively, and the probability values in the state transition matrix are continuously adjusted so that the model can best explain the observed data sequence. For example, the initial transition probability of (D, A) may be lower than that of (D, C) (if the physical distance between D and C is closer), but if the observed data shows that the disturbance reaches A first, the forward-backward algorithm will increase the transition probability of (D, A) and reduce the probability of (D, C). Ultimately, an updated probability matrix reflecting the most likely transition path in the current fault event is obtained, generating a transition probability matrix.

[0118] The path generation submodule calls the probability weights between nodes in the transfer probability matrix, selects the transfer path nodes layer by layer according to the maximum a posteriori probability criterion, superimposes the node hierarchy order and the probability weights to generate a path connection relationship, traverses all the hierarchical nodes to complete the path topology construction, and generates a propagation path topology map.

[0119] Call the transition probability matrix generated in the previous step. The matrix A′ contains the probability weight a′ that reflects the most likely transition relationship between nodes in the current fault event after iterative update. ij , the final fault propagation path is constructed according to the maximum a posteriori probability (MAP) criterion. Starting from the known disturbance starting node (e.g., node B, level 0), the node with the maximum transfer probability in the next level node is selected, that is, the node with the maximum transfer probability is found such that a′ B,j The largest node j (where j belongs to the node of level 1, such as D), assuming that in the updated transition probability matrix, a′ B,D = 0.9 is the largest, then D is selected as the second node on the path. The path is currently B->D. Next, starting from node D, find the node with the largest transfer probability among the level 2 nodes (such as A, C), that is, compare a′ D,A and a′ D,C , assuming a′ D,A=0.7 is the largest, then A is selected as the third node on the path, and the path is updated to B->D->A. This process continues, searching for the node with the highest probability of transferring from the current node (A, level 2) to the next level node (such as C, level 3, assuming that there is a possible transfer between A and C with the highest probability, or only C is available), such as a′ A,C =0.8 maximum, the path is B->D->A->C, and then from C (level 3) find the maximum probability transfer to the next level node (such as E, level 4), assuming a′ C,E =0.95 maximum, the path is B->D->A->C->E, the original node hierarchy order [B(0),D(1),A(2),C(3),E(4)] is combined with the calculated maximum probability transfer path B->D->A->C->E, where the probability weight a′ ij The connection relationship indicated by the hierarchical sorting is superimposed (or confirmed), and all hierarchical nodes are traversed until the path can no longer be extended (reaching the end node or there are no subsequent nodes that meet the conditions). The construction of the path topology is completed, and finally one or more (if there are branches with similar probabilities) most likely fault propagation paths are obtained, and a propagation path topology diagram is generated.

[0120] The above are merely preferred embodiments of the present invention and do not limit the present invention in any other form. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A cable fault analysis system based on big data, characterized in that: The system comprises: A disturbance feature extraction module is used to obtain voltage and current time series data through cable monitoring nodes, calculate the directional angle between voltage increments and current increments at adjacent time points, extract the gradient change inflection point, generate a disturbance starting position marker, and pass the disturbance starting position marker to the time series level determination module; A timing hierarchy determination module is configured to call the disturbance starting position marker, fit the time difference in combination with the cable length data, generate a node hierarchy ranking list, eliminate abnormal nodes whose time difference deviation exceeds three times the standard deviation, and pass the node hierarchy ranking list to the fault path mapping module; A reflection characteristic analysis module is used to collect the amplitudes and peak intervals of the main wave and the reflected wave, calculate the amplitude attenuation coefficient and the phase offset, combine them into reflection characteristic parameters, and transmit the reflection characteristic parameters to the topology segment identification module; A topological segment identification module is used to call the reflection characteristic parameters, calculate the Manhattan distance of the reflection characteristic parameters in the physical segment, determine the similarity through a support vector machine classification model, output a segment identification label, and pass the segment identification label to the fault path mapping module.

2. The cable fault analysis system based on big data according to claim 1, characterized in that: The disturbance starting position mark is specifically the disturbance gradient inflection point position, the node hierarchy sorting list includes the node propagation order, the valid node set, and the abnormal node set, the reflection characteristic parameters include the amplitude attenuation coefficient and the phase offset, and the section identification label specifically refers to the cable topology physical section identification code.

3. The cable fault analysis system based on big data according to claim 2, characterized in that: The disturbance feature extraction module includes: The incremental vector calculation submodule monitors the voltage and current time series data of the cable nodes, extracts the voltage and current differences between adjacent time points, calculates the incremental vectors through the differential algorithm, and arranges the continuous time series incremental vectors in chronological order to generate an incremental vector sequence; The gradient inflection point detection submodule calculates the angle between adjacent vector directions based on the incremental vector sequence using the direction cosine formula, applies a sliding window algorithm to perform local mean filtering on the angle sequence, and determines the inflection point position in combination with the angle change rate threshold to generate a gradient inflection point set. The perturbation mark generation submodule calls the gradient inflection point set, selects the starting inflection point according to the difference in gradient direction before and after the inflection point using the extreme value positioning algorithm, maps the timestamp corresponding to the inflection point to the original time series data, and generates a perturbation starting position mark.

4. The cable fault analysis system based on big data according to claim 3 is characterized in that: The timing level determination module includes: The time difference fitting submodule uses the node position coordinates recorded in the disturbance starting position mark, combines the conductor material parameters in the cable length data and the line segment length, calculates the signal transmission delay between adjacent nodes through linear interpolation, and generates a time difference set; The hierarchical sorting construction submodule arranges the nodes from small to large according to the transmission delay based on the time difference set, and performs secondary sorting on the nodes with the same delay value using the cable length data to generate a hierarchical node sequence table; The abnormal node elimination submodule calls the sorting weights in the hierarchical node sequence table, calculates the standard deviation of all node weight values, screens out nodes whose weight deviation exceeds three times the standard deviation threshold, removes them from the hierarchical node sequence table, and generates a node hierarchical sorting list.

5. The cable fault analysis system based on big data according to claim 4 is characterized in that: The reflection characteristic analysis module includes: The waveform acquisition submodule collects synchronous measurement data of the main wave amplitude voltage peak value and the reflected wave amplitude voltage peak value, records the millisecond-level change in the time interval between the two peaks, integrates the main wave peak value, reflected wave peak value and time difference data into a waveform recording unit with a unified timestamp, and generates a waveform feature set; The parameter calculation submodule extracts the main wave peak value and the reflected wave peak value in the waveform feature set, calculates the natural logarithm value of the amplitude ratio of the two using logarithmic operation, combines the peak time difference with the preset electromagnetic wave propagation speed constant, converts the product of the time difference and the speed into a phase angle offset radian value, and generates a characteristic parameter component; The characteristic integration submodule calls the amplitude attenuation coefficient and the phase offset radian in the characteristic parameter components, aligns the two in time series, and then merges them into a double-column matrix to generate reflection characteristic parameters.

6. The cable fault analysis system based on big data according to claim 5, characterized in that: The topology segment identification module includes: The parameter processing submodule obtains the amplitude characteristics and phase difference characteristics of the reflection characteristic parameters, normalizes the characteristic sequence using the mean and standard deviation, compares the deviation degree of each data point in the characteristic sequence based on the parameter fluctuation threshold, selects the continuous segments with deviation values less than the threshold, and generates a standardized parameter set; The distance calculation submodule extracts the amplitude difference sequence and phase difference difference sequence of adjacent physical segments based on the standardized parameter set, calculates the absolute deviation value of the difference sequence point by point according to the Manhattan distance formula, slides through all segment pairs using a fixed step window, accumulates the deviation values in the window to generate the difference metric between segments, and generates the Manhattan distance matrix; The classification discrimination submodule calls the inter-segment difference measure value in the Manhattan distance matrix, inputs the pre-trained support vector machine classification model, calculates the geometric interval between the distance value and the classification hyperplane according to the model kernel function, determines whether the interval value is higher than the model decision boundary threshold, outputs the binary classification probability value and maps it to the segment attribution category, and generates the segment identification label.

7. The cable fault analysis system based on big data according to claim 6, characterized in that: The system further comprises: a fault path mapping module, configured to receive the node hierarchical sorting list and the section identification label, match the node hierarchical sorting list with the section identification label, screen a node set that meets the timing and reflection characteristics, iteratively calculate the path transition probability through a hidden Markov model, generate an actual propagation path topology map of the cable fault, and output the actual propagation path topology map to a monitoring terminal; The actual propagation path topology diagram includes a fault propagation starting point, a propagation relay node, and a fault terminal node.

8. The cable fault analysis system based on big data according to claim 7, characterized in that: The fault path mapping module includes: The node matching submodule calls the node hierarchical sorting list and the timing threshold, extracts the node numbers in the segment identification label that meet the reflection characteristic constraint, compares the timing step in the node hierarchical sorting with the phase continuity of the segment identification label, selects the node combination whose timing deviation is less than the step threshold and whose phase continuity is higher than the phase threshold, and generates a candidate node set; The probability iteration submodule loads the state transition matrix and observation probability matrix of the hidden Markov model based on the node transfer relationship in the candidate node set, calculates the initial state transition probability according to the fault propagation time sequence step, and iteratively updates the transition probability weights between nodes through the forward-backward algorithm to generate a transition probability matrix; The path generation submodule calls the probability weights between nodes in the transfer probability matrix, selects the transfer path nodes layer by layer according to the maximum a posteriori probability criterion, superimposes the node hierarchy order and the probability weights to generate a path connection relationship, traverses all the hierarchical nodes to complete the path topology construction, and generates a propagation path topology map.

Images