Cable branch box fault intelligent classification system based on time domain reflection data

Abstract

The present application belongs to the technical field of cable time domain reflection detection and fault identification, and particularly relates to a cable branch box fault intelligent classification system based on time domain reflection data. The system establishes a branch topology according to the main trunk section, branch section, connection node and end position of the cable branch box, and constructs a normal reflection event dictionary using historical normal time domain reflection data. The system performs topology constraint decomposition on the time domain reflection waveform to be measured, obtains normal topology reflection components and abnormal residual reflection components, and constructs a reflection event graph for candidate echo events and their path adjacency, overlap and residual continuity. The system outputs fault categories such as poor contact, local insulation degradation, open circuit, low resistance short circuit and damp impedance gradient and corresponding branch position intervals through a graph attention classification model. The system can distinguish between normal branch echoes and fault abnormal echoes, and reduce the interference of multiple branch echo superposition on the classification result.

Description

Technical Field

[0001] This invention belongs to the field of cable time-domain reflection detection and fault identification technology, specifically involving an intelligent classification system for cable branch box faults based on time-domain reflection data. Background Technology

[0002] Time-domain reflectometry (TD-RS) is a commonly used fault location and condition identification method in cable electrical measurements. Its basic process involves injecting a step or pulse signal into the cable under test from the test end, acquiring the reflected waveform from the location of impedance discontinuity, and determining the location and type of the anomaly based on the arrival time, polarity, amplitude, and waveform width of the reflected wave. In scenarios with a single cable or a relatively simple line structure, conventional systems typically perform DC bias removal, smoothing filtering, amplitude normalization, time zero-point calibration, and noise thresholding on the acquired time-domain reflectometry data. The location of the reflection peak is then determined using a fixed threshold, peak search, or manual point selection, and the fault distance is calculated using propagation speed and round-trip time. For typical faults such as open circuits and low-impedance short circuits, conventional solutions use empirical classification based on the positive / negative change in reflection polarity, the magnitude of the reflection peak amplitude, and the steepness of the peak shape. If the reflection peak is weak, multiple sampling averaging, local window magnification, and manual verification are used to assist in the judgment. The data processing link of this type of scheme mainly consists of waveform preprocessing, peak location, distance conversion and fault type comparison. The detection object is usually abstracted as a continuous propagation path, and normal joints or end reflections are only subtracted as known background. It is suitable for cable detection objects with few impedance changes, clear echo propagation paths and weak normal reflection background.

[0003] Time-domain reflection detection in cable branch boxes is more complex than in single-path cables. Branch boxes typically contain a main section, multiple branch sections, connection nodes, and end points. Each node itself generates discontinuous echoes with normal impedance, and overlapping reflections with similar arrival times occur between branch paths. The most common approach involves acquiring a reference waveform under normal conditions after the branch box is put into operation or undergoes maintenance. This reference waveform is then saved as a manual or threshold template. During subsequent detection, the waveform to be tested is compared with the reference waveform using amplitude difference, correlation matching, or local peak comparison. When the difference in a certain time interval exceeds a set range, the corresponding time interval is marked as abnormal. The branch location of the abnormality is then calculated using a branch length table. Some systems pre-divide multiple detection windows according to known branch lengths, extracting statistical features such as peak amplitude, peak width, local energy, polarity changes, and arrival time offset within each window. Multiple historical fault samples are input into a supervised classification model, which outputs categories such as poor contact, insulation degradation, open circuit, or low-resistance short circuit. Although such schemes introduce template matching or classification models, normal branch echoes are usually still regarded as background curves, window features or differential benchmarks, and are not decomposed into reflection events with topological constraints.

[0004] The main technical problem with the aforementioned conventional solutions is that normal branch reflections and fault-related abnormal reflections in cable branch boxes coaxially superimpose in the time domain waveform. Existing processing methods lack a mechanism for unified constraint separation of branch topology, normal reflection events, and abnormal residual reflections. This makes it difficult for classification models to determine whether a particular abnormal peak originates from amplitude drift in the echo from a normal connection node or from new reflections generated at a fault point. The technical reason for this problem is that the propagation times of different branch paths within the branch box may be similar, and echoes from connection nodes, end positions, and fault locations may fall within the same sampling resolution window. Conventional peak detection can only observe the mixed local morphology and cannot confirm the topological origin of each reflection component. When the reference waveform changes slightly with the operating state, the template differential results will also mix the natural fluctuations at normal impedance discontinuities with the actual abnormal reflections. Furthermore, the reflection amplitudes of faults such as poor contact, local insulation degradation, and gradual impedance changes due to moisture may not necessarily form independent spikes, but may instead manifest as morphological shifts, broadening, or residual energy accumulation near normal reflection segments. If only fixed template difference or ordinary feature classification is used, the fault classification results are easily affected by the similarity of branch lengths, echo shielding and peak mixing, making it difficult to form a fault category judgment consistent with the actual propagation path of the cable branch box. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent classification system for cable branch box faults based on time-domain reflection data, which can effectively solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The intelligent classification system for cable branch box faults based on time-domain reflectometry data includes: a topology modeling component, which generates branch topology based on the main trunk, branch segments, connection nodes and end positions of the cable branch box; A normal reflection dictionary construction component is used to form a set of normal reflection events corresponding to the branch topology based on historical normal temporal reflection data; The waveform decoupling component is used to decompose the time-domain reflection waveform under test by the branch topology constraint to obtain the normal topology reflection component and the abnormal residual reflection component. A reflection event graph construction component is used to generate a reflection event graph based on the normal topology reflection component, the abnormal residual reflection component, and the branch topology. A fault classification component is used to intelligently classify the fault categories of the cable branch box output by the reflection event graph.

[0007] Preferably, the topology modeling component is configured to: map the trunk segment, the branch segment, the connection node and the end position as topology nodes respectively, map the electrical connection relationship between adjacent cable segments as topology edges, and associate each topology edge with a propagation path identifier, equivalent cable length and reference propagation speed; Based on the path combination from the topology node to the time-domain reflection test access terminal, a candidate reflection arrival sequence is generated, and candidate reflection arrival sequences with the same or similar reflection arrival times are merged into a topology overlap event set to form branch topology constraint data for limiting subsequent waveform decoupling.

[0008] Preferably, the normal reflection dictionary construction component is configured to: perform baseline drift correction, amplitude scale normalization and sampling clock alignment on historical normal time domain reflection data, and extract normal reflection segments corresponding to each connection node and end position based on the branch topology; A dictionary entry is established for the normal reflection segment, including reflection delay interval, reflection polarity, amplitude attenuation range, waveform width range, and adjacent reflection shielding relationship. The dictionary entries are indexed according to the branch path hierarchy so that the set of normal reflection events can characterize the echo distribution of normal impedance discontinuities in the cable branch box.

[0009] Preferably, the waveform decoupling component is configured to: set a candidate reflection time window corresponding to the branch topology in the time-domain reflection waveform to be tested, limit the normal topology reflection component to be formed by a combination of dictionary entries in the normal reflection event set, and limit the abnormal residual reflection component to a range related to the candidate reflection time window and its adjacent unexplained waveform interval; The normal topological reflection component and the abnormal residual reflection component are solved by simultaneously constraining the normal reflection reconstruction error, the abnormal residual sparsity, and the temporal order consistency between adjacent branch reflection events.

[0010] Preferably, the reflection event graph construction component is configured to: extract candidate echo events from the normal topology reflection component and the abnormal residual reflection component, use each candidate echo event as a graph node, and record the echo arrival time, local peak shape, reflection polarity, residual energy ratio and corresponding topology node identifier for each graph node; Graph edges are established for adjacent candidate echo events on the same path, candidate echo events located in the set of topologically overlapping events, and candidate echo events with anomalous residual continuity. Path adjacency, time interval, and energy transfer relationship are written into the graph edges.

[0011] Preferably, the topology modeling component is further configured to: after generating the candidate reflection arrival sequence, mark the mirror arrival events formed by multiple reflections in the branch path, and establish exclusion constraints for the mirror arrival events according to the path combination relationship of the topology edges; When different candidate reflection arrival sequences fall into the same sampling resolution window, an overlap priority label is generated according to the path level, propagation distance and node type, and the overlap priority label is written into the branch topology constraint data to limit the candidate range of the graph nodes in the subsequent reflection event graph.

[0012] Preferably, the normal reflection dictionary construction component and the waveform decoupling component are jointly configured to generate adaptive confidence boundaries based on the reflection delay interval and amplitude attenuation range of the dictionary entries; Within the candidate reflection time window corresponding to the time-domain reflection waveform to be tested, echo segments falling into the adaptive confidence boundary are preferentially included in the normal topology reflection component, while echo segments deviating from the adaptive confidence boundary and not interpretable by the combination of adjacent dictionary entries are included in the abnormal residual reflection component, and the corresponding deviation dimension identifier is retained as the fault classification input feature.

[0013] Preferably, the fault classification component is configured to: process the reflection event graph using a graph attention classification network, and calculate topological adjacency attention, residual abnormal attention, and polarity change attention respectively during the graph node update process; The topological adjacency attention is limited to the path adjacency relationship defined by the graph edges, the residual abnormal attention is limited to the graph nodes with abnormal residual continuity, and the polarity change attention is limited to the graph nodes where the reflection polarity changes. Based on the graph representation after the fusion of the three types of attention, the fault categories of poor contact, local insulation degradation, open circuit, low-resistance short circuit and gradual change of impedance due to moisture are output.

[0014] Preferably, the waveform decoupling component is further configured to: establish an abnormal segment chain for the echo segment corresponding to the deviation dimension identifier, and filter the abnormal segment chain according to time continuity, topological path continuity and residual energy decrease consistency; When the same abnormal fragment chain spans multiple candidate reflection time windows, the primary topological node of the abnormal fragment chain is determined based on the set of topological overlap events and the overlap priority label, and the overlapping echoes not belonging to the primary topological node are input as auxiliary abnormal evidence into the reflection event graph construction component.

[0015] Preferably, the fault classification component is further configured to: when training the graph attention classification network, the reconstruction error of the normal topology reflection component, the sparsity constraint of the abnormal residual reflection component, the consistency of the primary topology node of the abnormal fragment chain, and the fault category label are jointly configured as a training objective. During the inference phase, the class probabilities, primary topology nodes, and auxiliary anomaly evidence output by the graph attention classification network are checked for consistency, and an intelligent classification result containing the fault category and the corresponding branch location interval is output.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By mapping the main section, branch sections, connection nodes, and end positions of the cable branch box to a branch topology, and constructing a set of normal reflection events based on historical normal time-domain reflection data, the measured time-domain reflection waveform can be decomposed into normal topology reflection components and abnormal residual reflection components under branch topology constraints. Echoes generated at normal connection nodes and end positions are no longer treated as ordinary background noise, but rather as interpretable reflection events with reflection delay, reflection polarity, amplitude attenuation range, and waveform width range for reconstruction. Waveform segments that cannot be interpreted by the set of normal reflection events are retained as abnormal residual reflection components. Therefore, when branch paths are close, reflection peaks overlap, or weak fault echoes are attached to the vicinity of normal echoes, the system can separate normal topology echoes from faulty abnormal echoes and construct a reflection event graph containing path adjacency, time interval, and energy transfer relationships for candidate echo events. The fault classification process is based on the reflection event map, no longer relying solely on peak thresholds or template differentials. The classification of categories such as poor contact, local insulation degradation, open circuit, low-resistance short circuit, and damp impedance gradual change can be constrained by both echo morphology and branch propagation path, reducing classification bias caused by misidentifying normal branch echoes as fault echoes.

[0017] 2. In further data processing, candidate reflection arrival sequences, topological overlap event sets, mirror arrival event exclusion constraints, and overlap priority labels are written into the branch topological constraint data. This limits the attribution range of echo segments within similar time windows, reducing the possibility of the same anomalous waveform being repeatedly interpreted by multiple branch paths. The normal reflection dictionary and waveform decoupling process jointly generate adaptive confidence boundaries, including echo segments falling within the normal delay and attenuation range into the normal topological reflection component, and including segments deviating from the corresponding boundaries and not interpretable by a combination of adjacent dictionary entries into the anomalous residual reflection component, while retaining the deviation dimension identifier. After filtering the anomalous segment chain according to temporal continuity, topological path continuity, and residual energy decrease consistency, its primary attribution topological node and auxiliary anomalous evidence can participate in the training and inference verification of the graph attention classification network. Through the above processing, mirror reflections, overlapping reflections, and continuous anomalous residuals can maintain their respective source identifiers in the same reflection event graph, reducing mutual contamination between classification input features. The classification result can simultaneously include the fault category and the corresponding branch position interval. Waveform segments with low similarity or strong overlapping conflicts can also be retained as auxiliary anomalous evidence, facilitating subsequent verification by waveform segment and topological node. Attached Figure Description

[0018] Figure 1 This is the overall operation flowchart of the intelligent fault classification system for cable branch boxes of the present invention; Figure 2 This is a flowchart of the topology modeling component processing of the present invention; Figure 3 This is a flowchart of the normal reflection dictionary construction component processing of the present invention; Figure 4 This is a flowchart of the waveform decoupling and abnormal segment chain processing of the present invention; Figure 5 This is a flowchart of the reflection event graph construction and fault classification reasoning process of the present invention. Detailed Implementation

[0019] 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, not all, of the embodiments of the present invention. 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.

[0020] Please refer to Figure 1This embodiment provides an intelligent fault classification system for cable branch boxes based on time-domain reflection data. This system classifies the time-domain reflection waveforms formed by the main cable, branch cables, connection nodes, and end positions within the cable branch box. The system receives data including the time-domain reflection waveform to be tested, historical normal time-domain reflection data, the branch connection relationships of the cable branch box, length records of the main and branch sections, path start information corresponding to the test access point, and category labels corresponding to historical fault samples. The time-domain reflection waveform to be tested is a discrete amplitude sequence collected along the time axis, and the historical normal time-domain reflection data consists of similar discrete waveforms formed by the cable branch box in a fault-free state or a state confirmed to be normal after maintenance. The system includes a topology modeling component, a normal reflection dictionary construction component, a waveform decoupling component, a reflection event graph construction component, and a fault classification component. The topology modeling component converts the main trunk, branch segments, connection nodes, and end positions into a graph-structured branch topology; the normal reflection dictionary construction component extracts the echo events corresponding to each normal impedance discontinuity point based on historical normal time-domain reflection data; the waveform decoupling component decomposes the measured time-domain reflection waveform into normal topology reflection components and abnormal residual reflection components; the reflection event graph construction component maps the decomposition results to the branch topology to form a reflection event graph; and the fault classification component outputs the fault category and corresponding branch location interval based on the reflection event graph. These components transmit intermediate data through a data interface without altering the structure of the cable branch box, and the processing objects are limited to time-domain reflection data and its derived data.

[0021] In this embodiment, the branch topology is represented using a directed graph data structure. The test access point serves as the starting topology node, and the trunk segment, branch segment, connecting node, and end position are mapped as topology nodes or topology edges, respectively. Connecting nodes and end positions are topology nodes capable of generating normal reflection events, while the trunk segment and branch segment are topology edges capable of generating propagation delay and amplitude attenuation. Each topology edge records the path identifier, starting topology node, ending topology node, equivalent cable length, and reference propagation speed. The topology modeling component generates propagation paths from the test access point to each candidate reflection point based on electrical connections and calculates the candidate reflection arrival time. For path p, the candidate reflection arrival time is determined according to the following formula:

[0022] in, This indicates the time when the candidate reflection event corresponding to path p arrives at the test access terminal, and e represents a topological edge in path p. This represents the equivalent cable length corresponding to topological edge e. Let represent the reference propagation speed corresponding to topological edge e. The summation symbol indicates the accumulation of propagation times for all topological edges in path p. The coefficient 2 represents the outbound and return paths of the time-domain reflected signal. For example, the path corresponding to a candidate reflection point may contain a main segment of 120 meters in length and a branch segment of 80 meters in length, both with a reference propagation speed of ____ per second. If the propagation path length is 200 meters, then substituting this value yields a candidate reflection arrival time of 2.5 microseconds. This calculation result serves as the center position of the candidate reflection time window, used to limit the candidate range of normal reflection events and abnormal residual reflection events during subsequent waveform decomposition.

[0023] Table 1 shows the field organization method of the branch topology data in this embodiment. The fields in Table 1 are not limited to a specific storage format and can be stored in a relational data table, graph database, or in-memory graph structure, as long as they can be read by subsequent components and used for reflection event constraints.

[0024] Table 1. Field organization of branch topology data

[0025] In one embodiment, the normal reflection dictionary construction component performs unified preprocessing on historical normal time-domain reflection data. Specifically, the component reads the time-domain reflection waveforms generated in multiple normal state acquisition batches from the same cable branch box, and performs baseline drift correction, amplitude scale normalization, and sampling clock alignment on each waveform. Baseline drift correction can use the median or low-order smooth curve of the non-fault quiescent interval as the baseline estimate, and subtract the baseline estimate from the original waveform to obtain the zero-center waveform; amplitude scale normalization can use the stable amplitude near the reference injection pulse or the normal end echo amplitude as the scale factor to make the waveforms of different batches have comparable amplitude ranges; sampling clock alignment can select the direct response near the test access point, the echo of the fixed connection node, or the main reference peak in the historical normal waveform as the alignment benchmark, obtain the time offset through cross-correlation search, and uniformly move the historical normal waveforms to the same time coordinate system. After the above preprocessing is completed, the normal reflection dictionary construction component extracts normal reflection segments within the time window corresponding to each connection node and end position based on the candidate reflection arrival time calculated in the branch topology, and calculates the average shape, reflection polarity, amplitude attenuation range, waveform width range and adjacent reflection occlusion relationship of the reflection segments.

[0026] In this embodiment, each dictionary entry in the normal reflection dictionary corresponds to a normal impedance discontinuity point or a set of topological overlap events. A dictionary entry includes at least a dictionary entry identifier, a corresponding topological node identifier, a candidate reflection time window, a reflection polarity code, a reference waveform segment, an amplitude attenuation range, a waveform width range, a path level index, and a masking relationship marker. The reflection polarity code can be represented using positive, negative, and bipolar transition markers without altering the cable structure. The waveform width range can be expressed using half-peak width, energy concentration intervals, or the duration of a local main lobe. The masking relationship marker indicates multiple normal reflection events that may occur simultaneously within the same sampling resolution window. During subsequent waveform decoupling, this marker allows multiple dictionary entries to be combined to interpret the same local waveform, rather than directly identifying mixed peaks as abnormal residuals. The normal reflection dictionary is not constructed using a simple template; instead, normal echoes are decomposed into multiple reflection events corresponding to branch topologies, enabling normal echoes at connection nodes and end positions to participate in reconstruction in a callable dictionary format.

[0027] Table 2 shows one optional field organization method for the normal reflection dictionary. The reflection delay range, amplitude attenuation range, and waveform width range in Table 2 are all obtained from historical normal time-domain reflection data, and their values ​​can be updated as the number of normal samples in the same cable branch box increases.

[0028] Table 2. One optional field organization method for the normal reflection dictionary.

[0029] In one embodiment, the normal reflection dictionary construction component generates an adaptive confidence boundary for each dictionary entry. The adaptive confidence boundary expresses the allowable fluctuation range of a normal reflection event in reflection delay, amplitude, and waveform width, avoiding the direct identification of slight batch differences in normal connection node echoes as anomalous residuals. For dictionary entry j, the adaptive confidence boundary is determined according to the following formula:

[0030] in, This represents the adaptive confidence boundary for dictionary entry j. This represents the average reflection delay of dictionary entry j. Indicates the degree of dispersion of reflection delay. Indicates the average reflection amplitude. Indicates the degree of dispersion of the reflection amplitude. Indicates the average waveform width. Indicates the degree of dispersion in waveform width. This represents the boundary coefficient related to the stability of the corresponding topological node. For example, if the average reflection delay of a connected node is 4.00 microseconds, the reflection delay dispersion is 0.05 microseconds, and the boundary coefficient is 2, then the reflection delay boundary is 3.90 microseconds to 4.10 microseconds; if the average reflection amplitude is 0.60 and the amplitude dispersion is 0.04, then the amplitude boundary is 0.52 to 0.68. When the delay, amplitude, and width of the echo segment to be measured all fall within the corresponding boundaries, it can be used as a candidate segment for the normal topological reflection component; when there are one-dimensional or multi-dimensional deviations, the deviation dimension is retained as a subsequent classification feature.

[0031] In one embodiment, after receiving the time-domain reflected waveform to be measured, the waveform decoupling component performs baseline correction, noise smoothing, amplitude normalization, and candidate window matching on the waveform to be measured according to a unified time coordinate. The waveform decoupling component embeds reference waveform segments from the normal reflection dictionary into the sampling sequence corresponding to the waveform to be measured according to the candidate reflection time window, forming a dictionary matrix; each column in the dictionary matrix corresponds to a time-domain segment of a normal reflection event, and the position of the column vector on the time axis is constrained by the candidate reflection arrival time calculated by the branch topology. The waveform to be measured is represented as a combination of normal topological reflection components, anomalous residual reflection components, and unexplained disturbances, wherein the normal topological reflection components are obtained by combining the dictionary matrix and dictionary coefficients, and the anomalous residual reflection components are confined within the candidate reflection time window and its adjacent unexplained waveform interval. For anomalous waveforms falling near the echo of a normal connection node, the waveform decoupling component does not directly delete the corresponding normal dictionary entry, but retains the local differences that cannot be explained by the combination of normal dictionary entries as anomalous residuals while maintaining the interpretability of the normal topological reflection components.

[0032] In this embodiment, the waveform decoupling component uses a constrained optimization method to solve for the normal topological reflection component and the anomalous residual reflection component. The time-domain reflection waveform to be measured is denoted as... The normal reflection dictionary matrix is ​​denoted as The dictionary coefficient vector is denoted as The abnormal residual vector is denoted as The candidate region mask that allows abnormal residuals to occur is denoted as The mask for non-candidate regions that suppress the occurrence of anomalous residuals is denoted as The set of reflection event pairs in the topological path that need to maintain the arrival time order is denoted as . The estimated arrival time of the reflected event r is denoted as . The waveform decoupling target is determined according to the following formula:

[0033] in, This represents the dictionary coefficient vector obtained from the solution. This represents the abnormal residual vector obtained from the solution. , and represents the weights of the residual penalty for non-candidate regions, the residual sparsity penalty for candidate regions, and the temporal order consistency penalty, respectively. This represents the order penalty incurred when an event *r*, which should arrive first on the same path, is later than event *s*. For example, if the sampling vector of a local waveform under test is [0.4, 1.1, 0.9], the normal dictionary reconstruction vector is [0.3, 1.0, 0.7], and the abnormal residual vector is [0.1, 0.0, 0.2], then the reconstruction error vector is [0.0, 0.1, 0.0], and the squared error term is 0.01. If this local region belongs to a candidate region and the time order penalty is 0, the optimization process will trade off between residual sparsity and normal dictionary reconstruction, retaining the local abnormal components. Instead of being forcibly incorporated into the normal reflection component, it is reflected in the middle.

[0034] In a preferred embodiment, the waveform decoupling component employs an alternating solution method to handle the aforementioned optimization objectives. When the anomalous residual vector is fixed, the system solves the dictionary coefficient vector using non-negative or signed-constrained least squares, ensuring that the normal topological reflection components are interpreted by the normal reflection dictionary as much as possible. When the dictionary coefficient vector is fixed, the system performs masked weighted sparse estimation on the reconstruction difference, preserving anomalous segments within candidate regions and suppressing isolated disturbances within non-candidate regions. For dictionary entries with occlusion relationship markers, the waveform decoupling component allows multiple normal dictionary entries to participate in reconstruction within the same sampling window and limits the combination range based on the path hierarchy index. For reflection events without occlusion relationships but with close time windows, the system requires that the dictionary coefficient combination conforms to the topological path order, disallowing far-end reflection events from forming normal topological reflection components independently without near-end propagation interpretation. Through this processing method, normal connection node echoes, terminal echoes, and anomalous residual echoes are separated into different data objects in the same waveform under test, allowing the subsequent classification process to access both normal topological reflection components and anomalous residual reflection components simultaneously.

[0035] In one embodiment, the reflection event graph construction component extracts candidate echo events from normal topological reflection components and anomalous residual reflection components. Candidate echo events can originate from echo segments activated by normal dictionary entries or from local energy concentration segments within anomalous residual reflection components. For each candidate echo event, the reflection event graph construction component records the echo arrival time, local peak shape, reflection polarity, waveform width, residual energy proportion, and corresponding topological node identifier. The local peak shape can be characterized by the main peak amplitude, the difference between the left and right slopes, the peak width, and the number of zero crossings; the reflection polarity can be encoded by the local main lobe direction; and the residual energy proportion can be obtained by the ratio of the energy of the anomalous residual segment to the total energy of the corresponding window of the waveform under test. Graph nodes in the reflection event graph represent candidate echo events, and graph edges represent topological path adjacency, overlap, or anomalous residual continuity between candidate echo events. The rules for establishing graph edges do not rely on manual empirical labels but are jointly determined by branch topology, the normal reflection dictionary, and decoupling results.

[0036] In this embodiment, the graph node feature vectors are constructed according to the following formula:

[0037] in, This represents the feature vector of the graph node corresponding to the m-th candidate echo event. This indicates the arrival time of the normalized echo. Represents the local peak shape coefficient. Indicates the reflection polarity code. Indicates the normalized waveform width. Indicates the percentage of residual energy. This represents the topology node encoding. For example, if a candidate echo event has a normalized arrival time of 0.42, a local peak shape coefficient of 0.73, a reflection polarity encoding of 1, a normalized waveform width of 0.18, and a residual energy percentage of 0.36, corresponding to a topology node encoding of 5, then the graph node feature vector is denoted as a column vector composed of the above six values. Each dimension of this vector comes from the time-domain reflection waveform or branch topology and does not contain data unrelated to the cable branch box structure.

[0038] Furthermore, the reflection event graph construction component establishes path adjacency edges between adjacent candidate echo events on the same path, overlap interpretation edges between candidate echo events in the same set of topologically overlapping events, and residual continuity edges between candidate echo events with temporal continuity in anomalous residual components. Path adjacency edges record path adjacency relationships, candidate arrival time intervals, and topological hierarchy differences; overlap interpretation edges record the degree of overlap of sampling windows, occlusion relationship markers, and overlap priority labels; residual continuity edges record the time intervals, residual energy changes, and polarity changes between anomalous residual segments. For the same candidate echo event, multiple graph edges can exist simultaneously; for example, an anomalous residual segment near a connection node may have path adjacency edges with a nearby normal connection node, overlap interpretation edges with the branch end echo within the same sampling window, and residual continuity edges with anomalous residual segments in adjacent windows. The reflection event graph is jointly represented by graph nodes and graph edges, enabling the fault classification component to read input data from two dimensions: local waveform morphology and branch topological relationships.

[0039] In one embodiment, the fault classification component employs a graph attention classification network to process the reflection event graph. The graph attention classification network calculates the information transfer weights between graph nodes according to topological adjacency attention, residual abnormality attention, and polarity change attention, respectively. Topological adjacency attention is calculated only between graph nodes connected by adjacent edges along a path; residual abnormality attention is calculated only between graph nodes connected by continuous residual edges; and polarity change attention is calculated only between graph nodes where the reflection polarity changes or the local zero-crossing pattern changes. For attention type c, the attention weight of the m-th graph node to the n-th adjacent graph node is determined by the following formula:

[0040] in, Let c represent the attention weight of node m to node n in the graph below. The projection vector representing attention type c. This represents the feature transformation matrix for attention type c. This represents the feature vector of the graph edge between graph node m and graph node n. This represents vector concatenation. This represents the set of adjacent graph nodes connected to graph node m under attention type c. For example, under residual abnormal attention, a graph node has two residual consecutive adjacent nodes with unnormalized attention scores of 0.7 and 0.2, respectively. After exponential normalization, the corresponding attention weights are approximately 0.62 and 0.38. The graph attention classification network fuses the node representations obtained from different attention types and outputs fault categories such as poor contact, local insulation degradation, open circuit, low-resistance short circuit, and moisture-induced impedance gradient.

[0041] In this embodiment, the fault category and branch location interval output by the fault classification component are jointly determined by the graph representation, the primary topology node, and auxiliary anomaly evidence. If the primary topology node of an abnormal residual segment is located near a connecting node, and the residual anomaly attention is concentrated between the connecting node and adjacent nodes, while the local peak shape exhibits amplitude fluctuations and width changes, the classification component can map the corresponding graph representation to the poor contact category. If the abnormal residual segment shows continuous low-amplitude energy accumulation along the same branch path, and the reflection polarity does not form a strong open circuit characteristic, the classification component can map the corresponding graph representation to the local insulation degradation or moisture-induced impedance gradient category. If the abnormal residual segment corresponds to a strong polarity reflection and forms a new main peak before the end, the classification component can map the corresponding graph representation to the open circuit category. If the abnormal reflection polarity is opposite to that of the open circuit category and the residual energy is concentrated in the low-resistance anomaly candidate interval, the classification component can map the corresponding graph representation to the low-resistance short circuit category. The above classification process does not use a single threshold as the judgment basis, but rather uses the node features, graph edge relationships, and attention constraints in the reflection event graph as input, and outputs an intelligent classification result containing the fault category and the corresponding branch location interval.

[0042] In this embodiment, the system operation process can be summarized as follows: the branch topology provides propagation path constraints for the time-domain reflection data; the normal reflection dictionary provides normal echo interpretations for connection nodes and end positions; the waveform decoupling component splits the waveform under test into normal topology reflection components and abnormal residual reflection components; the reflection event graph construction component organizes the two types of reflection components into graph-structured data; and the fault classification component completes the fault category output on the graph-structured data. This embodiment covers the overall processing logic of the intelligent fault classification system for cable branch boxes, and can distinguish normal branch echoes from faulty abnormal echoes from mixed time-domain reflection waveforms, providing a unified data foundation for subsequent hierarchical processing.

[0043] In one embodiment, the topology modeling component further extends the candidate reflection arrival sequences with topological constraints. Specifically, after generating an initial branch topology based on the main trunk, branch segments, connection nodes, and end positions of the cable branch box, the topology modeling component traverses all reachable paths from the test access point, forming multiple candidate reflection paths. Each candidate reflection path includes a path identifier, a set of topological edges traversed, the corresponding candidate reflection point, the path level, and the candidate reflection arrival time. For candidate reflection paths with the same or similar arrival times, the topology modeling component does not forcibly merge them into a single echo, but instead generates a set of topological overlap events. The set of topological overlap events stores multiple candidate reflection arrival sequences and their corresponding topological node identifiers, allowing these candidate reflection events to participate in the interpretation within the same local window during subsequent waveform decoupling. Through this data structure, echo overlap caused by close branch lengths can be saved in a topological constraint manner, preventing accidental deletion or simplification to a single peak during the preprocessing stage.

[0044] Furthermore, when generating candidate reflection arrival sequences, the topology modeling component marks mirror arrival events formed by multiple reflections. Mirror arrival events typically originate from signals returning to the test access point after multiple reflections between the connection node and the end position, and their arrival time may be close to the actual fault reflection arrival time. The topology modeling component identifies mirror arrival events based on path combination relationships, including whether there are repeated topological edges in the round-trip path, whether the reflection path passes through the same end position multiple times, and whether the candidate arrival time can be obtained by combining multiple normal nodes. Candidate sequences marked as mirror arrival events are written with exclusion constraints, which limit the mirror event from being used as primary evidence of a fault without a corresponding normal reflection event being activated. If a local abnormal peak in the waveform under test falls into both the actual branch candidate window and the mirror arrival window, the waveform decoupling component, based on the exclusion constraints, requires the normal topology reflection component to first interpret the necessary normal echoes in the mirror path, and then determines whether the remaining residual constitutes an abnormal residual reflection component.

[0045] Preferably, when different candidate reflection arrival sequences fall within the same sampling resolution window, the topology modeling component generates overlap priority labels for the corresponding candidate events. The overlap priority labels are jointly determined by path level, propagation distance, node type, and historical normal echo stability. Near-end connection nodes with lower path levels typically have stronger normal echo interpretation capabilities. End-position and far-end branch nodes require interpretation in conjunction with occlusion relationship markers when echoes overlap. Dictionary entries with higher historical normal echo stability have priority matching qualifications within the same window. The overlap priority labels do not directly determine the fault category but are used to limit the candidate range of graph nodes and the order of dictionary entry combinations. By writing mirror arrival events, exclusion constraints, and overlap priority labels into the branch topology constraint data, the topology modeling component provides finer constraints for subsequent waveform decoupling, enabling nearby branch echoes, mirror echoes, and anomalous echoes to have distinguishable data sources within the same time window. This embodiment extends the branch topology constraint data, making it suitable for time-domain reflection classification scenarios of cable branch boxes with many branch paths and frequent candidate echo overlap.

[0046] Table 3 illustrates one way of organizing the fields in the topology constraint extension data. Table 3 explains how mirror arrival events, exclusion constraints, and overlap priority labels are incorporated into the subsequent data processing flow.

[0047] Table 3. One method of organizing fields in topology constraint extended data.

[0048] In one embodiment, the normal reflection dictionary construction component further employs a hierarchical clustering approach to generate dictionary entries. For multiple historical normal reflection segments corresponding to the same topology node, the normal reflection dictionary construction component first calculates the similarity between segments within a time-aligned window. This similarity can be represented by waveform correlation, peak shape difference, and polarity consistency. For connected nodes with stable reflection patterns, multiple historical normal reflection segments can be grouped into the same dictionary entry; for connected nodes significantly affected by occlusion from adjacent branches, multiple sub-entries can be formed based on the occlusion status. Each sub-entry remains bound to the same topology node but has different reference waveform segments and occlusion relationship markers. This processing allows the normal reflection dictionary to express the echo patterns that may occur in the same connected node under different normal operating conditions, without misassigning normal occlusion differences to anomalous residuals.

[0049] In this embodiment, the normal reflection dictionary construction component also records the occlusion relationships between adjacent reflections. These occlusion relationships can be determined jointly by the arrival time intervals of candidate reflections and the local peak superposition relationships in the historical normal waveform. When the candidate arrival time intervals of two or more normal reflection events are less than the width of the local main lobe, the system marks these events as occlusion-capable combinations. When the historical normal waveform shows that the amplitude of one event changes with the activation level of another event, the system establishes occlusion dependencies between dictionary entries. These dependencies inform the waveform decoupling component that the same local mixed peak may be formed by multiple normal dictionary entries and cannot be directly extracted as an abnormal residual without attempting combination reconstruction. For dictionary entries with strong occlusion relationships, the system retains combination columns or linkage coefficient constraints when constructing the dictionary matrix, ensuring that they have group activation conditions during waveform reconstruction.

[0050] In a preferred embodiment, the waveform decoupling component determines the attribution of the echo segment under test based on adaptive confidence boundaries. Segments falling within the boundaries of the corresponding dictionary entries for reflection delay, amplitude, and waveform width are preferentially considered as candidates for normal topology reflection components. Segments deviating from the reflection delay boundary are marked as arrival time deviations; segments deviating from the amplitude attenuation range are marked as amplitude deviations; segments deviating from the waveform width range are marked as morphological deviations; segments exhibiting multiple deviations retain multiple deviation dimension identifiers. Deviation dimension identifiers are not directly equivalent to fault categories but serve as graph node features or auxiliary anomaly evidence in subsequent reflection event graphs. For example, if the arrival time of the echo under test for a certain connection node is still within the normal reflection delay boundary, but the waveform width exceeds the corresponding boundary and the residual energy proportion increases, then this segment will be split into a normal topology reflection component and a morphological deviation-type anomalous residual reflection component. If a new echo appears in a remote branch window that cannot be explained by the normal dictionary, the new echo is entered into the subsequent graph construction process as an anomalous residual reflection component.

[0051] Furthermore, the waveform decoupling component establishes anomaly fragment chains when processing deviation fragments. An anomaly fragment chain consists of multiple anomalous residual fragments that are temporally continuous, topologically adjacent, or exhibit continuity in residual energy variation. When filtering anomaly fragment chains, the system reads three conditions: whether the time interval between anomaly fragments conforms to the relationship between adjacent candidate reflection windows, whether the topological nodes corresponding to the anomaly fragments are located on the same or adjacent branch paths, and whether the residual energy of the anomaly fragments shows a trend consistent with propagation attenuation. Anomalous residual fragments that meet the above conditions are retained as the same anomaly fragment chain; isolated disturbances that do not meet the conditions and only appear near a single sampling point are not included in the classification network as the main anomaly fragment chain. For anomaly fragment chains spanning multiple candidate reflection time windows, the system determines the primary topological node based on the set of overlapping topological events and the overlap priority label, retaining other relevant candidate nodes as auxiliary anomaly evidence. The primary topological node is used to locate the fault branch location interval, and the auxiliary anomaly evidence is used for consistency verification during classification inference. This embodiment, through adaptive confidence boundaries, deviation dimension identifiers, and anomaly fragment chains, enables the waveform decoupling results to form a traceable data link.

[0052] In one embodiment, the reflection event graph construction component performs type-based processing on graph edges. Path adjacency edges are established based on the relationship between adjacent topological nodes in the branch topology; if two candidate echo events are located on the same propagation path and their corresponding topological nodes are adjacent in the graph structure, a path adjacency edge is established. Overlap interpretation edges are established based on the set of topologically overlapping events; if two candidate echo events fall into the same sampling resolution window, or if the normal dictionary entries of two candidate events have an occlusion relationship, an overlap interpretation edge is established. Residual continuity edges are established based on anomalous fragment chains; if two candidate echo events belong to the same anomalous fragment chain or have an adjacent chain relationship, a residual continuity edge is established. Polarity change edges are established based on reflection polarity encoding changes; if there is a positive / negative polarity conversion, bipolar transition, or local zero-crossing anomaly between two candidate echo events, a polarity change edge is established. Different types of graph edges correspond to different attention calculation ranges in the subsequent graph attention classification network, avoiding unconstrained fully connected propagation of all graph nodes.

[0053] In this embodiment, the graph edge feature vector may include time interval, topological hierarchy difference, path adjacency marker, overlap window ratio, residual energy change, and polarity conversion marker. The time interval expresses the distance between two candidate echo events in the time domain coordinates; the topological hierarchy difference expresses the propagation hierarchy relationship of two candidate echo events in the branch topology; the path adjacency marker distinguishes between adjacent nodes on the same path and overlapping nodes on different paths; the overlap window ratio describes the degree of overlap between the corresponding time windows of two candidate echo events; the residual energy change describes the continuous change of anomalous residuals along the time axis; and the polarity conversion marker describes the change in reflection polarity between adjacent events. After the reflection event graph construction component generates the graph structure, it sends the graph node features, graph edge features, primary topological node, auxiliary anomaly evidence, and candidate fault location intervals to the fault classification component. All of the above data objects can be derived from the time-domain reflection waveform, branch topology, and normal reflection dictionary.

[0054] In one embodiment, the fault classification component performs supervised training and physical consistency constraint training on the graph attention classification network. Training samples include historical fault temporal reflection waveforms, corresponding branch topologies, normal reflection dictionaries, normal topology reflection components and abnormal residual reflection components obtained through waveform decoupling, fault category labels confirmed manually or through maintenance, and fault location interval labels. During training, the graph attention classification network takes the reflection event graph as input and obtains graph-level and node-level representations through multi-layer node updates. The graph-level representation is used to output the fault category, and the node-level representation is used to output the primary topology node or candidate location interval. The training objective simultaneously includes fault category loss, normal topology reflection reconstruction error, abnormal residual sparsity constraint, and primary topology node consistency constraint. The joint training objective is determined according to the following formula:

[0055] in, Indicates the joint training objective. Indicates the loss categorized by fault type. , , and These represent the weights of each training item. , , and The meaning is consistent with the aforementioned waveform decoupling objective. This represents the set of candidate graph nodes in the chain of abnormal fragments. This indicates the primary topology node to which the abnormal fragment chain corresponding to graph node m belongs. In the graph below, node m represents residual abnormal attention and points to the primary home topology node. Attention weights are assigned. For example, if the fault category classification loss for a training sample is 0.28, the normal topology reflection reconstruction error is 0.04, the anomalous residual sparsity term is 0.15, and the primary-home topology node consistency penalty is 0.20, and the corresponding weights are 1, 0.5, 0.2, and 0.3 respectively, then the joint training objective is 0.39. Through this training objective, the classification network learns the category boundaries while preserving the constraints of normal reflection reconstruction, anomalous residual sparsity, and topology home consistency.

[0056] In a preferred embodiment, the fault classification component performs consistency verification on the output during the inference phase. The fault classification component reads the category probability, primary home topology node, auxiliary anomaly evidence, and candidate branch location interval from the graph attention classification network output, and checks whether the fault category with the highest category probability is consistent with the anomaly fragment chain corresponding to the primary home topology node. If the category probability points to the open circuit category, but only low-amplitude continuous residuals exist near the primary home topology node without strong polarity reflections, the system lowers the confidence level of the output for that category and retains the conflicting waveform fragments. If the category probability points to the poor contact category, and the primary home topology node is located near a connection node, while auxiliary anomaly evidence shows morphological deviations in adjacent echoes, the system outputs the poor contact category and the corresponding branch location interval. If the category probability is close between local insulation degradation and gradual change in damp impedance, the system distinguishes them based on the residual energy distribution and polarity change attention of the anomaly fragment chain. The verification process during the inference phase does not introduce new sensing objects or external service data; it only performs consistency checks on the time-domain reflection waveform and internal evidence of the reflection event graph.

[0057] Furthermore, the fault classification component can retain low-confidence level samples as samples to be reviewed, and use them to update the normal reflection dictionary or training sample library after maintenance confirmation. If the review result shows that a certain echo segment belongs to a stable morphological change under normal operating conditions, the normal reflection dictionary construction component will classify the segment into the normal sub-entry of the corresponding topology node and update the corresponding adaptive confidence boundary; if the review result confirms that the segment belongs to a faulty abnormal reflection, the training sample library will record the corresponding reflection event graph, abnormal segment chain, fault category, and fault location interval. The update process retains the original branch topology constraints and does not change the basic relationship between topology nodes and topology edges due to a single sample. In this way, the system can continuously accumulate normal reflection events and fault reflection events matching the same cable branch box during long-term use, so that subsequent waveform decoupling and classification inference have a more complete data foundation. This embodiment limits the training and inference process to ensure that the reflection event graph classification result is consistent with the waveform decoupling result, the primary topology node, and auxiliary abnormal evidence.

[0058] In one embodiment, when the system processes contact failures, the measured time-domain reflection waveform near the connection node typically exhibits local amplitude shifts, local peak broadening, or repetitive micro-reflections near the normal echo. The waveform decoupling component reads the normal reflection dictionary entry corresponding to the connection node, reconstructs the portion that falls within the adaptive confidence boundary as a normal topological reflection component, and extracts the amplitude or shape deviation portion as anomaly residual reflection components. The reflection event graph construction component establishes path adjacency edges and residual continuum edges between the corresponding connection node graph nodes, adjacent path graph nodes, and anomaly residual graph nodes. The fault classification component computes the graph representation under the joint constraints of residual abnormal attention and topological adjacency attention, and inputs the chain of anomalous fragments near the connection node as the main evidence into the classification network. If the anomalous residuals are mainly concentrated near the connection node and there is no new primary peak at a distant end, the output position interval is limited to the branch range where the connection node is located. This embodiment illustrates the processing method when the normal echo of the connection node and the abnormal echo of the contact failure occur superimposed, and the relevant data are all from the time-domain reflection waveform and branch topology.

[0059] In one embodiment, when the system processes local insulation degradation faults, the measured time-domain reflection waveform may not form an independent strong reflection peak, but rather exhibit continuous low-amplitude residuals, local baseline morphology shifts, or slow waveform width changes along a certain branch path. The waveform decoupling component extracts residual segments within multiple consecutive candidate reflection time windows along the same branch path and determines whether these residual segments satisfy time continuity, topological path continuity, and residual energy reduction consistency. Residual segments that meet the conditions are organized into anomaly segment chains. The reflection event graph construction component establishes continuous residual edges for the graph nodes in the anomaly segment chain and connects the normal reflection graph nodes on the corresponding branch path through path adjacency edges. The fault classification component reads the residual anomaly attention and polarity change attention. If the anomaly segment chain exhibits continuous residuals rather than single-point strong reflections, and the reflection polarity does not show strong polarity characteristics such as open circuits or low-resistance short circuits, then the graph representation is mapped to the local insulation degradation category. This embodiment discloses the processing logic when weak fault echoes are obscured by normal echoes, avoiding the neglect of anomaly segments due to reliance solely on peak search.

[0060] In one embodiment, when the system handles open-circuit faults, the measured time-domain reflection waveform may exhibit a strong new positive polarity peak within the time window corresponding to the fault location, causing a change in the original end-point normal echo morphology. The waveform decoupling component first reconstructs the normal end-point dictionary entries that may appear after the fault location. If the end-point normal echo cannot appear according to the existing delay and amplitude range, and the new peak cannot be explained by the combination of normal dictionary entries for the connecting nodes, the new peak is classified as an anomalous residual reflection component. The reflection event graph construction component establishes path adjacency edges between the new peak graph node, the upstream connecting node graph node, and the original end-point location graph node, and generates auxiliary anomalous evidence based on the absence or weakening of the original end-point normal echo. The fault classification component reads the polarity change attention and topological adjacency attention, using the strong polarity new peak, the change in the original end-point echo, and the primary topological node as classification inputs. When outputting the fault category, the location interval is determined based on the candidate reflection arrival time of the new peak and the corresponding branch path. This embodiment incorporates the topological relationship between the newly added reflections of open-circuit faults and the original normal reflections at the end into the classification input, reducing path assignment bias caused by judging solely based on the location of strong peaks.

[0061] In one embodiment, when the system handles low-impedance short-circuit faults, the measured time-domain reflection waveform may exhibit an abnormal reflection with the opposite polarity to that of the open-circuit type at a certain branch position. The waveform decoupling component reads the normal reflection dictionary entries within the corresponding candidate window. If the normal dictionary entries cannot explain the polarity change of the reflection, and the abnormal residual energy is concentrated in a local window, a low-impedance abnormal candidate echo event is generated. The reflection event graph construction component establishes path adjacency edges between the low-impedance abnormal candidate echo event and the upstream path node, and simultaneously establishes polarity change edges between candidate echo events where polarity has changed. The fault classification component processes these graph edges within the calculation range of polarity change attention, and outputs the low-impedance short-circuit category and branch position interval by combining the residual energy ratio, candidate reflection arrival time, and the primary topology node. If there are also mirror arrival events within the same window, the exclusion constraint provided by the topology modeling component requires that the mirror path corresponds to a normal reflection event with a matching basis; otherwise, the mirror event only participates in the verification as auxiliary abnormal evidence. This embodiment discloses a classification processing method when low-impedance short-circuit abnormal polarity echoes overlap with mirror echoes.

[0062] In one embodiment, when the system processes moisture-induced impedance gradient faults, the measured time-domain reflection waveform may exhibit a wide, slowly varying echo, residual energy spreading along branch paths, or waveform width deviations in multiple adjacent windows. The waveform decoupling component separates the normal topological reflection components from the morphological deviation-type anomalous residual reflection components based on the waveform width range in the normal reflection dictionary and adaptive confidence boundaries. The reflection event graph construction component organizes multiple adjacent morphological deviation segments into anomaly segment chains and establishes continuous residual edges between their graph nodes. If the anomaly segment chain spans multiple candidate reflection windows, the primary topological node is determined based on the overlap priority label, and the morphological deviations of other windows are used as auxiliary anomalous evidence. The fault classification component reads the residual anomalous attention and polarity change attention. If the residual energy distribution exhibits continuous, slowly varying characteristics and the polarity transition is not concentrated at a single point, the moisture-induced impedance gradient category is output. This embodiment illustrates the classification method when the gradient impedance anomaly does not form a single spike, enabling continuous residual features to enter the graph structure classification process.

[0063] In a preferred embodiment, the system can also perform source alignment between multiple detection data from the same cable branch box. All multiple detection data are processed using the same branch topology constraints and normal reflection dictionary, forming multiple reflection event graphs. The system aligns the graph nodes corresponding to the same topology node in terms of time coordinates and amplitude scales, comparing the occurrence position and deviation dimension of anomalous residual reflection components in multiple detections. If the same primary topology node repeatedly exhibits similar anomalous fragment chains in multiple detection batches, the fault classification component treats this anomalous fragment chain as stable anomalous evidence; if the anomalous residual appears only in a single detection and lacks topological continuity, the system treats it as low-confidence auxiliary anomalous evidence. This processing does not change the single-detection fault classification process; it only performs graph-level comparison of source-based detection results at the input data level. Through source alignment, the normal reflection dictionary can distinguish between long-term stable normal morphological changes and short-term anomalous residuals, allowing the fault classification component to obtain a clearer primary topology node input. This embodiment is applicable to scenarios involving periodic time-domain reflection detection of the same cable branch box, and the processing objects are still limited to time-domain reflection data and topology constraint data.

[0064] In one embodiment, the system's data storage can adopt a hierarchical data object organization method. The raw layer stores the time-domain reflection waveform to be tested, historical normal time-domain reflection data, and historical fault time-domain reflection data; the topology layer stores branch topology, candidate reflection arrival sequences, topological overlap event sets, mirror arrival event markers, exclusion constraints, and overlap priority labels; the dictionary layer stores the normal reflection dictionary, adaptive confidence boundaries, occlusion relationship markers, and path hierarchy indexes; the decoupling layer stores normal topological reflection components, abnormal residual reflection components, deviation dimension identifiers, and abnormal fragment chains; the graph structure layer stores graph node features, graph edge features, primary topology nodes, and auxiliary anomaly evidence; and the classification layer stores fault categories, category probabilities, branch position intervals, and consistency verification results. These hierarchical data objects have explicit referencing relationships, and any classification result can be traced back to the corresponding waveform fragment to be tested, normal dictionary entry, and branch topology node. Table 4 shows a preferred data object mapping relationship.

[0065] Table 4. A preferred data object mapping relationship

[0066] In this embodiment, the verifiability of the classification results is achieved through the reference relationships between data objects. When the fault classification component outputs a certain fault category, the classification layer records the corresponding reflection event graph identifier; the graph structure layer traces back to the graph nodes and edges involved in the classification based on the reflection event graph identifier; the decoupling layer traces back to the abnormal residual reflection components and normal topology reflection components based on the graph nodes; the dictionary layer traces back to the corresponding normal reflection dictionary entries based on the normal topology reflection components; the topology layer traces back to the branch paths and candidate reflection arrival sequences based on the primary topology node; and the original layer traces back to the time interval in the time-domain reflection waveform to be measured based on the abnormal residual fragments. This reference relationship allows the classification results to be jointly explained by the time-domain reflection waveform fragments, normal dictionary entries, and branch topology, preventing the classification output from deviating from the physical propagation path. Through the above multiple embodiments, the system can complete fault category classification and branch location interval output based on time-domain reflection data even when there is overlap of multiple branch echoes in cable branch boxes, mirror reflection interference, weak abnormal residuals attached to the echoes of normal connection nodes, and gradual impedance anomalies do not form independent spikes.

Claims

1. A cable branch box fault intelligent classification system based on time-domain reflectometry data, characterized in that, include: Topology modeling component, used to generate branch topology based on the trunk section, branch section, connection node and end position of the cable branch box; A normal reflection dictionary construction component is used to form a set of normal reflection events corresponding to the branch topology based on historical normal temporal reflection data; The waveform decoupling component is used to decompose the time-domain reflection waveform under test by the branch topology constraint to obtain the normal topology reflection component and the abnormal residual reflection component. A reflection event graph construction component is used to generate a reflection event graph based on the normal topology reflection component, the abnormal residual reflection component, and the branch topology. A fault classification component is used to intelligently classify the fault categories of the cable branch box output by the reflection event graph.

2. The intelligent fault classification system for cable branch boxes based on time-domain reflection data according to claim 1, characterized in that, The topology modeling component is configured to: map the trunk segment, the branch segment, the connection node and the end position as topology nodes respectively, map the electrical connection relationship between adjacent cable segments as topology edges, and associate each topology edge with a propagation path identifier, equivalent cable length and reference propagation speed; Based on the path combination from the topology node to the time-domain reflection test access terminal, a candidate reflection arrival sequence is generated, and candidate reflection arrival sequences with the same or similar reflection arrival times are merged into a topology overlap event set to form branch topology constraint data for limiting subsequent waveform decoupling.

3. The intelligent fault classification system for cable branch boxes based on time-domain reflection data according to claim 2, characterized in that, The normal reflection dictionary construction component is configured to: perform baseline drift correction, amplitude scale normalization, and sampling clock alignment on historical normal time domain reflection data, and extract normal reflection segments corresponding to each connection node and end position based on the branch topology; A dictionary entry is established for the normal reflection segment, including reflection delay interval, reflection polarity, amplitude attenuation range, waveform width range, and adjacent reflection shielding relationship. The dictionary entries are indexed according to the branch path hierarchy so that the set of normal reflection events can characterize the echo distribution of normal impedance discontinuities in the cable branch box.

4. The intelligent fault classification system for cable branch boxes based on time-domain reflection data according to claim 3, characterized in that, The waveform decoupling component is configured to: set a candidate reflection time window corresponding to the branch topology in the time-domain reflection waveform to be tested; limit the normal topology reflection component to be formed by a combination of dictionary entries in the normal reflection event set; and limit the abnormal residual reflection component to a range related to the candidate reflection time window and its adjacent unexplained waveform interval. The normal topological reflection component and the abnormal residual reflection component are solved by simultaneously constraining the normal reflection reconstruction error, the abnormal residual sparsity, and the temporal order consistency between adjacent branch reflection events.

5. The intelligent fault classification system for cable branch boxes based on time-domain reflection data according to claim 4, characterized in that, The reflection event graph construction component is configured to: extract candidate echo events from the normal topology reflection component and the abnormal residual reflection component, use each candidate echo event as a graph node, and record the echo arrival time, local peak shape, reflection polarity, residual energy ratio and corresponding topology node identifier for each graph node. Graph edges are established for adjacent candidate echo events on the same path, candidate echo events located in the set of topologically overlapping events, and candidate echo events with anomalous residual continuity. Path adjacency, time interval, and energy transfer relationship are written into the graph edges.

6. The intelligent fault classification system for cable branch boxes based on time-domain reflectometry data according to claim 5, characterized in that, The topology modeling component is further configured to: after generating the candidate reflection arrival sequence, mark the mirror arrival events formed by multiple reflections in the branch path, and establish exclusion constraints for the mirror arrival events according to the path combination relationship of the topology edges; When different candidate reflection arrival sequences fall into the same sampling resolution window, an overlap priority label is generated according to the path level, propagation distance and node type, and the overlap priority label is written into the branch topology constraint data to limit the candidate range of the graph nodes in the subsequent reflection event graph.

7. The intelligent fault classification system for cable branch boxes based on time-domain reflection data according to claim 6, characterized in that, The normal reflection dictionary construction component and the waveform decoupling component are jointly configured to generate adaptive confidence boundaries based on the reflection delay interval and amplitude attenuation range of the dictionary entries; Within the candidate reflection time window corresponding to the time-domain reflection waveform to be tested, echo segments falling into the adaptive confidence boundary are preferentially included in the normal topology reflection component, while echo segments deviating from the adaptive confidence boundary and not interpretable by the combination of adjacent dictionary entries are included in the abnormal residual reflection component, and the corresponding deviation dimension identifier is retained as the fault classification input feature.

8. The intelligent fault classification system for cable branch boxes based on time-domain reflection data according to claim 7, characterized in that, The fault classification component is configured to: use a graph attention classification network to process the reflection event graph, and calculate topological adjacency attention, residual abnormal attention and polarity change attention respectively during the graph node update process; The topological adjacency attention is limited to the path adjacency relationship defined by the graph edges, the residual abnormal attention is limited to the graph nodes with abnormal residual continuity, and the polarity change attention is limited to the graph nodes where the reflection polarity changes. Based on the graph representation after the fusion of the three types of attention, the fault categories of poor contact, local insulation degradation, open circuit, low-resistance short circuit and gradual change of impedance due to moisture are output.

9. The intelligent fault classification system for cable branch boxes based on time-domain reflectometry data according to claim 8, characterized in that, The waveform decoupling component is further configured to: establish an abnormal segment chain for the echo segment corresponding to the deviation dimension identifier, and filter the abnormal segment chain according to time continuity, topological path continuity and residual energy decrease consistency; When the same abnormal fragment chain spans multiple candidate reflection time windows, the primary topological node of the abnormal fragment chain is determined based on the set of topological overlap events and the overlap priority label, and the overlapping echoes not belonging to the primary topological node are input as auxiliary abnormal evidence into the reflection event graph construction component.

10. The intelligent fault classification system for cable branch boxes based on time-domain reflection data according to claim 9, characterized in that, The fault classification component is further configured to: when training the graph attention classification network, the reconstruction error of the normal topology reflection component, the sparsity constraint of the abnormal residual reflection component, the consistency of the primary topology node of the abnormal fragment chain, and the fault category label are jointly constituted as the joint training objective. During the inference phase, the class probabilities, primary topology nodes, and auxiliary anomaly evidence output by the graph attention classification network are checked for consistency, and an intelligent classification result containing the fault category and the corresponding branch location interval is output.

Images