Fault diagnosis method for electric energy metering box based on multi-source data fusion

Abstract

The application discloses a power metering box fault diagnosis method based on multi-source data fusion, relates to the technical field of power metering box fault diagnosis, and comprises the following steps: acquiring running signal sequences output by multi-source sensors in a power metering box in real time, performing time domain alignment and abnormal jump point cleaning to obtain a standardized input sequence; mapping the standardized input sequence to a graph structure with metering box running states as nodes, calculating a local fluctuation cumulative amount of each node as a node state base value; transferring and superimposing state influence coefficients of adjacent nodes in a step-by-step manner along the edge direction of the graph structure to generate a fusion state field reflecting a global coupling state; and determining a fault type and a fault severity level of the power metering box according to a response sequence and a state deviation amplitude between nodes in a fault propagation chain. The method can accurately capture a state coupling relationship between multiple nodes in the metering box and a fault propagation path, and realizes fault tracing and fine diagnosis.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology for electricity metering boxes, specifically a fault diagnosis method for electricity metering boxes based on multi-source data fusion. Background Technology

[0002] Electricity metering boxes integrate various electrical components, functional modules, and sensors. During long-term operation, they are susceptible to faults such as insulation degradation, poor contact, and component aging due to factors like electrical stress and environmental temperature and humidity. Once a fault occurs, it rarely exists in isolation at a single measuring point but spreads to surrounding components through electrical connections and spatial coupling, creating a chain reaction. Existing fault diagnosis methods for electricity metering boxes mostly rely on data from single-type sensors, using threshold comparisons or trend analysis, treating each measuring point signal as an independent information source. This fails to fully consider the electrical coupling relationships and interactions between measuring points, resulting in insensitivity to weak faults in their early stages of evolution or chain faults propagating across components, leading to high frequencies of missed and false alarms. Some solutions attempt to introduce multi-source data fusion, but these often remain at the level of simple time synchronization and feature stitching of different sensor signals, failing to construct a global state field at the fusion level that can characterize the dynamic influence relationships between components, and thus unable to depict the coupling and propagation mechanism of faults within the metering box. Furthermore, existing methods, when determining the type and severity of faults, typically rely solely on the abnormal signal amplitude of individual nodes at the current moment. This lack of precise reconstruction of the fault triggering point and tracing of the fault propagation path leads to coarse diagnostic results, making it difficult to distinguish between local primary faults and secondary anomalies caused by other components, thus significantly misleading maintenance decisions. Therefore, a method is needed that can fully express the state coupling relationships of multiple nodes within the metering box and accurately locate the fault trigger point and propagation path to address the problems of insufficient dynamic capture of fault propagation and inadequate evidence for fault tracing. Summary of the Invention

[0003] This paper presents a fault diagnosis method for power metering boxes based on multi-source data fusion. By constructing a graph structure to fuse the state field, the method expresses the global coupling relationship of multiple nodes in the metering box. Based on the signal trend deviation from the starting time and the propagation chain, the method realizes the location of the fault trigger point and the analysis of the propagation path, thereby improving the accuracy and traceability of fault diagnosis.

[0004] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a fault diagnosis method for power metering boxes based on multi-source data fusion, the method comprising: The system acquires real-time operating signal sequences from multiple sensors within the power metering box, performs time-domain alignment and abnormal transition point cleanup on each sequence to obtain a standardized input sequence. This eliminates sampling time deviations and unexpected pulse interference from multiple sources, ensuring a high-quality signal foundation with spatiotemporal consistency. The standardized input sequence is mapped onto a graph structure with the metering box's operating state as nodes. The cumulative local fluctuation of the signal sequence corresponding to each node is calculated as the node's state base value. The accumulated fluctuation energy within a sliding window is used to highlight the abnormal activity of a node compared to its historical levels. Based on the node state base value, the state influence coefficients of adjacent nodes are progressively propagated and superimposed along the edges of the graph structure to generate a fused state field reflecting the global coupling state. Through iterative diffusion and convergence control, the state value of each node incorporates electrical connections. The combined influence of topology and spatial proximity avoids information silos. High-gradient change regions are extracted from the fused state field as candidate fault influence domains. Within these candidate fault influence domains, the starting point where the signal evolution trend deviates from the nominal trajectory is located, serving as the fault trigger point. Gradient field analysis and trend backtracking are used to eliminate steady-state, slowly changing interference and accurately capture the fault initiation time. Based on the fault trigger point, the correlation strength of the signal sequences of each node within the candidate fault influence domain within a preset window before triggering is traced back to construct a fault propagation chain. Strong and weak correlations are used to filter and restore the fault's response order and diffusion path in the physical network. Based on the response order and state deviation magnitude between nodes in the fault propagation chain, the fault type and severity level of the power metering box are determined, and a diagnostic result integrating the fault origin location and diffusion trajectory is output.

[0005] As a preferred technical solution of the present invention, the step of performing time-domain alignment and abnormal jump point cleaning on each operating signal sequence to obtain a standardized input sequence specifically includes: uniformly interpolating each operating signal sequence output by multiple source sensors to the same sampling time reference to ensure that each sequence has a corresponding amplitude at each sampling time; for each signal amplitude at each sampling time, calculating the absolute value of the difference between it and the amplitude at the previous time, and marking sampling points whose absolute value of the difference exceeds a preset jump threshold as candidate abnormal points; for the marked candidate abnormal points, using the amplitude of the adjacent normal sampling points before and after it to perform linear interpolation replacement to complete the abnormal jump point cleaning, thereby effectively removing glitch while preserving the normal fluctuation profile.

[0006] As a preferred technical solution of the present invention, mapping the standardized input sequence to a graph structure with the metering box operating status as nodes specifically includes: defining each electrical node inside the power metering box, the installation position of each sensor, and the interface of each functional module as an independent node in the graph structure, and recording the spatial coordinates of each independent node; establishing an undirected edge between any two independent nodes based on the electrical connection relationship between them, or the relationship that the spatial distance between the two nodes is less than a preset adjacency distance threshold, and assigning an edge weight reflecting the tightness of the connection to each undirected edge; constructing the graph structure with all independent nodes and their undirected edges and edge weights, and using the standardized input sequence corresponding to each independent node as the temporal feature vector of that node, so that the graph structure simultaneously embeds electrical membership and spatial distribution information.

[0007] As a preferred embodiment of the present invention, the step of calculating the cumulative local fluctuation of the signal sequence corresponding to each node as the node state base value specifically includes: for each independent node in the graph structure, calculating the sum of the absolute values ​​of the amplitude differences between adjacent nodes in the amplitude sequence of its temporal feature vector within a preset sliding window to obtain the total local fluctuation of the independent node; obtaining the fluctuation deviation factor based on the ratio of the total local fluctuation to the average fluctuation in a preset historical period; and using the product of the total local fluctuation and the fluctuation deviation factor as the node state base value of the independent node. The node state base value is used to sensitively reflect the activity level of the independent node at the current moment and to amplify and prompt abnormal fluctuations.

[0008] As a preferred embodiment of the present invention, the step of transmitting and superimposing the state influence coefficients of adjacent nodes along the edge direction of the graph structure according to the node state base value to generate a fused state field reflecting the global coupling state specifically includes: taking each independent node in the graph structure as the starting source node, multiplying the node state base value of the independent node by the edge weight between the independent node and each adjacent node to obtain the initial influence amount transmitted to each adjacent node; after receiving the initial influence amounts transmitted by all adjacent source nodes, each adjacent node accumulates the received initial influence amounts to obtain the cumulative influence amount of the adjacent node, and superimposes the cumulative influence amount onto the node state base value of the adjacent node itself to form the updated state value after one transmission; using the obtained updated state value as the new transmission source value, the transmission and superposition operation to the next layer of adjacent nodes is repeated until the sum of the changes in the state values ​​of each independent node in the entire graph structure in two consecutive iterations is lower than a preset convergence threshold, and the converged state values ​​of each independent node form the fused state field, thereby expressing the direct and indirect coupling effects in a balanced manner in the global scope.

[0009] As a preferred embodiment of the present invention, the step of extracting high gradient change regions from the fused state field as candidate fault influence regions specifically includes: calculating the absolute value of the difference between the state value of each independent node and the state values ​​of all its neighboring nodes in the fused state field, and taking the maximum value among all the absolute differences as the local gradient value of the independent node; marking independent nodes whose local gradient values ​​exceed a preset gradient threshold as high gradient nodes, and using each high gradient node as the center, including all independent nodes whose graph distance from the high gradient node is less than a preset propagation radius into the same initial influence region; merging all initial influence regions that overlap or are adjacent into the same high gradient change region, and using all the high gradient change regions obtained after merging as candidate fault influence regions, thereby automatically aggregating related regions caused by the same fault source and eliminating blind spots caused by manual segmentation.

[0010] As a preferred embodiment of the present invention, the step of locating the starting moment when the signal evolution trend deviates from the nominal trajectory in the candidate fault influence domain as the fault trigger point specifically includes: for each candidate fault influence domain, extracting the standardized input sequence of all independent nodes within the influence domain in a preset historical time interval, and calculating the mean amplitude of all independent nodes within the influence domain at each moment to obtain the global mean curve of the influence domain; using the fitted trend line of the global mean curve during normal operation as the nominal trajectory of the influence domain; comparing the difference between the global mean curve and the nominal trajectory point by point backward from the end of the preset historical time interval, and positioning the moment when the difference first exceeds the preset allowable deviation range as the starting moment when the signal evolution trend deviates from the nominal trajectory, as the fault trigger point of the candidate fault influence domain. This method is universally applicable to trend change detection under various gradually changing backgrounds.

[0011] As a preferred technical solution of the present invention, the step of using the fault trigger point as a benchmark to trace back the correlation strength of the signal sequences of each node in the candidate fault influence domain within a preset window before triggering, and constructing a fault propagation chain, specifically includes: taking the fault trigger point as the end time, forward by a preset time window length to obtain the pre-trigger signal segment of each independent node in the candidate fault influence domain; calculating the Pearson correlation coefficient between the pre-trigger signal segments of any two independent nodes in the candidate fault influence domain, and marking the two nodes with a correlation coefficient greater than a preset correlation threshold as a strong correlation pair; taking the independent node where the fault trigger point is located as the starting node, gradually expanding outward along each strong correlation pair, arranging all independent nodes traversed in the continuous expansion process according to the order of expansion, forming the fault propagation chain, thereby restoring the temporal chain of orderly transmission of the fault from the starting point to the surrounding area.

[0012] As a preferred embodiment of the present invention, the step of determining the fault type and severity level of the power metering box based on the response sequence and state deviation magnitude among nodes in the fault propagation chain specifically includes: for each independent node in the fault propagation chain, calculating the difference between the updated state value of the independent node and the node state base value of the independent node as the state deviation magnitude of the independent node; sequentially accumulating the state deviation magnitudes of all independent nodes in the fault propagation chain in order of arrangement to obtain the total deviation accumulation, and determining the fault severity level based on the preset interval in which the total deviation accumulation falls; comparing the response sequence of each independent node in the fault propagation chain with the standard node response sequence template corresponding to each fault type in the preset fault database, and taking the preset fault type with the highest matching degree with the standard node response sequence template as the determined fault type, thereby improving the fault type differentiation and level quantification accuracy by means of the dual constraints of propagation sequence and deviation magnitude.

[0013] As a preferred technical solution of the present invention, after determining the fault type and severity level of the power metering box based on the response sequence and state deviation magnitude between nodes in the fault propagation chain, the spatial coordinates corresponding to the first independent node in the fault propagation chain are marked as the fault starting position, and the spatial coordinates of all independent nodes in the fault propagation chain are connected in order to generate a fault propagation path trajectory. The fault type, the fault severity level, the fault starting position, and the fault propagation path trajectory are combined to form a diagnostic result data package and output, providing visualized fault source pointing and affected range information directly for operation and maintenance troubleshooting.

[0014] The technical effects and advantages provided by the present invention in the above technical solution are as follows: Multi-source sensor signal sequences are mapped onto a graph structure with the metering box's operating status as nodes. The cumulative local fluctuation of the signal sequence corresponding to each node is calculated as the node's state base value. Based on this, the state influence coefficients of adjacent nodes are progressively propagated and superimposed along the edges of the graph structure to form a fused state field. Graph nodes are defined according to internal electrical nodes, sensor installation locations, and functional module interfaces. Edge weights reflect the tightness of electrical connections or spatial proximity. The state influence coefficients are obtained by multiplying the node's state base value by the edge weights. The propagation process iterates until the global state value converges. This fused state field not only preserves the real-time fluctuation characteristics of each measuring point but also incorporates single-point fluctuations into the mutual interference expression of neighboring and even all network nodes through graph iterative diffusion. This allows weak anomalies that are not easily identifiable on a single sequence to form identifiable high-gradient change regions in tightly coupled areas after propagation and superposition. Thus, even if the fault signal does not significantly exceed the conventional threshold at individual measuring points, it will cause gradient anomaly aggregation in the relevant node group due to coupling propagation, achieving sensitive capture of early faults and cross-component cascading faults, overcoming the problem that isolated threshold judgments easily overlook fault precursors. After extracting the high-gradient change region from the fused state field as the candidate fault influence domain, a global mean curve is calculated for the standardized input sequence of all independent nodes within the influence domain. Using the trend line fitted during normal operation as the nominal trajectory, point-by-point comparisons are made from the end of the historical time interval backwards to locate the starting moment when the global mean curve deviates from the nominal trajectory as the fault trigger point. Based on this trigger point, signal segments within a preset window are extracted, and the Pearson correlation coefficient between any two node signal segments within the domain is calculated. Strongly correlated pairings are extended into sequentially arranged fault propagation chains. The fault type and severity level are then determined based on the order of node responses and the magnitude of state deviation in the propagation chain. Through precise calibration of the deviation from the starting moment, the original fault moment and subsequent propagation evolution stage are distinguished, avoiding misjudging secondary anomalies as the fault source. The propagation chain clearly shows the time sequence path of the fault gradually spreading to surrounding nodes from the starting position along the coupling relationship. Combined with the matching with the standard node response sequence template in the preset fault database, the nature of the fault can be accurately identified. At the same time, the severity level can be determined according to the cumulative amount of state deviation, so that the diagnosis results have both source tracing path and quantitative classification information, which facilitates fixed-point maintenance and risk assessment. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0016] Figure 1 This is a flowchart of a fault diagnosis method for electricity metering boxes based on multi-source data fusion; Figure 2 This is a flowchart of the standardized processing of multi-source sensor operating signals for an electricity metering box; Figure 3 These are the signal magnitude sequences of independent nodes and the curves showing the variation of the node state base values; Figure 4 This is a schematic diagram of the high gradient change region of the fused state field and the influence domain of candidate faults. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] See Figure 1 This invention provides a fault diagnosis method for electricity metering boxes based on multi-source data fusion. The method acquires real-time operating signal sequences from multiple sensors within the electricity metering box, performs time-domain alignment and abnormal transition point cleaning on each operating signal sequence to obtain a standardized input sequence. The standardized input sequence is mapped to a graph structure with the metering box's operating state as nodes. The cumulative local fluctuation of the signal sequence corresponding to each node is calculated as the node's state base value. Based on the node's state base value, the state influence coefficients of adjacent nodes are progressively propagated and superimposed along the edges of the graph structure to generate a fused state field reflecting the global coupling state. High-gradient change regions are extracted from the fused state field as candidate fault influence domains, and the starting moment when the signal evolution trend deviates from the nominal trajectory within the candidate fault influence domain is located as the fault trigger point. Using the fault trigger point as a reference, the inter-correlation strength of the signal sequences of each node within the candidate fault influence domain within a preset window before triggering is traced back to construct a fault propagation chain. Based on the response sequence and state deviation magnitude between nodes in the fault propagation chain, the fault type and severity level of the electricity metering box are determined, and the fault diagnosis result is output.

[0019] In specific implementation, please refer to Figure 2The system acquires real-time operating signal sequences from multiple sensors within the energy metering box. These sensors include at least two types: current sensors, voltage sensors, temperature sensors, and humidity sensors. Since the operating signal sequences output by different sensors differ at their original sampling times, time-domain alignment is required. The time-domain alignment is achieved by: establishing a unified sampling time reference, using an equally spaced time point sequence, with the time interval between adjacent time points set based on the highest original sampling frequency among all the multiple sensors; and for each operating signal sequence output by a multiple sensor, using the original amplitude-time pairs in the operating signal sequence as interpolation nodes, employing a linear interpolation method to calculate the corresponding amplitude of the operating signal sequence at each sampling time of the sampling time reference, thus obtaining the aligned operating signal sequence. This ensures that all operating signal sequences have a corresponding amplitude at each sampling time.

[0020] After time-domain alignment, abnormal transition point cleaning is performed. For each aligned operating signal sequence, every sampling moment except the first sampling moment is traversed, and the absolute value of the difference between the signal amplitude at that sampling moment and the signal amplitude at the previous sampling moment is calculated as the amplitude transition variable. The amplitude transition variable is compared with a preset transition threshold. The preset transition threshold is determined based on the statistical upper limit of the historical signal amplitude transition variables collected by each multi-source sensor under stable and fault-free operation of the power metering box. The maximum value of the historical signal amplitude transition variable can be used as the preset transition threshold. If the amplitude transition variable at a certain sampling moment exceeds the preset transition threshold, that sampling moment is marked as a candidate abnormal point in the operating signal sequence.

[0021] In each running signal sequence, linear interpolation replacement is performed on each marked candidate outlier. The method of linear interpolation replacement is as follows: in the running signal sequence containing the candidate outlier, find the nearest previous unmarked sampling point as the previous normal point; find the nearest next unmarked sampling point as the next normal point. If there are no previous normal points before or after the candidate outlier, the original amplitude of the candidate outlier is retained without replacement. Obtain the sampling time and signal amplitude of the previous normal point and the sampling time and signal amplitude of the next normal point, substitute them into the linear interpolation formula, and calculate the replaced signal amplitude at the candidate outlier. The linear interpolation formula is: in, This represents the signal amplitude after the candidate outlier is replaced. This indicates the signal amplitude before the normal point. This indicates the signal amplitude at the normal point. Indicates the sampling time before the normal point. Indicates the sampling time after the normal point. This represents the sampling time of the candidate outlier. The calculated... The signal amplitude at the candidate outlier is assigned to complete the cleaning of that candidate outlier. After traversing all candidate outliers, all abnormal transition points in each running signal sequence are cleaned, resulting in a standardized input sequence.

[0022] In practical implementation, each electrical node, sensor installation location, and functional module interface within the energy metering box is defined as an independent node in the diagram structure. Each electrical node includes the main incoming line terminal, branch outgoing line terminal, and grounding terminal within the energy metering box; each sensor installation location includes the current transformer installation location, voltage acquisition terminal location, temperature sensor contact location, and humidity sensor placement location; each functional module interface includes the metering module data interface, communication module interface, and protection device interface. Each independent node is defined with its spatial coordinates recorded. The spatial coordinates are established using a preset vertex on the outer shell of the energy metering box as the origin, along the length, width, and height of the box, forming a three-dimensional coordinate system. The three-dimensional coordinate values ​​of the geometric center of each independent node relative to the origin are measured and used as the spatial coordinates of the independent node.

[0023] An undirected edge is established between any two independent nodes based on their electrical connection or spatial distance being less than a preset adjacency distance threshold. Electrical connection refers to a direct wire connection between two independent nodes in the internal electrical circuit of the energy metering box, automatically determined by reading the electrical wiring diagram. Spatial distance is determined between independent nodes without electrical connection by calculating the Euclidean distance between their spatial coordinates and comparing it to the preset adjacency distance threshold. The preset adjacency distance threshold is set based on the internal space dimensions of the energy metering box and the sensor's sensitivity range, and is set to one-tenth of the largest of the box's length, width, and height. If the Euclidean distance between two independent nodes is less than the preset adjacency distance threshold, an undirected edge is established between them.

[0024] While establishing undirected edges, each undirected edge is assigned a weight, reflecting the tightness of the connection between the two independent nodes. For undirected edges established based on electrical connections, the edge weight is determined according to the type and impedance parameters of the connecting wires, with a value ranging from 0.5 to 1. The larger the cross-sectional area and the shorter the length of the wire, the closer the edge weight is to 1. For undirected edges established based on spatial distances, the edge weight is determined by the ratio of the Euclidean distance between the two independent nodes to a preset adjacency distance threshold. The edge weight is set to 1 minus the ratio of the Euclidean distance to the preset adjacency distance threshold; the smaller the ratio, the closer the edge weight is to 1. Using all independent nodes as the vertex set, all undirected edges between independent nodes as the edge set, and the edge weight corresponding to each undirected edge as the edge attribute, a graph structure is formed.

[0025] The normalized input sequence corresponding to each independent node is used as the temporal feature vector of that independent node. The normalized input sequence is the running signal sequence after time-domain alignment and cleaning of abnormal transition points. If an independent node corresponds to multiple multi-source sensors, the multiple normalized input sequences at the independent node are concatenated into a multi-dimensional temporal feature vector according to a fixed arrangement order of sensor types; if an independent node corresponds to only one multi-source sensor, the normalized input sequence at the independent node is directly used as a one-dimensional temporal feature vector.

[0026] When calculating the baseline values ​​of node states, for each independent node in the graph structure, a preset sliding window is used to slide along the temporal feature vector of the independent node to obtain the amplitude sequence within the preset sliding window. If the temporal feature vector is multidimensional, the vector magnitude is calculated time-by-time for the multidimensional data, transforming the multidimensional temporal feature vector into a single-dimensional magnitude sequence before applying the preset sliding window. The length of the preset sliding window is set according to the signal fluctuation period inside the energy metering box, taking an integer multiple of the power frequency period. The number of sampling points corresponding to the sliding window length is denoted as... The sliding window moves forward one sampling point at a time.

[0027] The sum of the absolute values ​​of the differences between adjacent amplitudes of an independent node within a preset sliding window is calculated as the total local fluctuation of that independent node. The calculation formula is as follows: in, This represents the total local fluctuation of an independent node. This indicates the number of sampling points included in the preset sliding window. This indicates the index of the sampling point within the preset sliding window. The value range is 1 to , This indicates that the index of the preset sliding window is... The amplitude corresponding to the sampling point, This indicates that the index of the preset sliding window is... The amplitude corresponding to the sampling point.

[0028] The average fluctuation of an independent node within a preset historical period is obtained. The preset historical period length is 50 to 200 times the preset sliding window length, and is a continuous historical sampling period during which the power metering box is confirmed to be in normal operating condition. Within the preset historical period, the preset sliding window is slid point by point, and the total local fluctuation corresponding to each window position is calculated. The arithmetic mean of all local fluctuation totals is taken as the average fluctuation of the independent node within the preset historical period. The ratio of the current local fluctuation total to the average fluctuation is used as the fluctuation deviation factor. The larger the value of the fluctuation deviation factor deviates from 1, the more significant the difference between the current signal fluctuation level and the historical normal fluctuation level. The product of the local fluctuation total and the fluctuation deviation factor is used as the node state base value of the independent node, which is used to characterize the activity level of the independent node at the current moment.

[0029] See Figure 3 In the graph, the horizontal axis represents time in seconds. The left side of the vertical axis, marked in blue, indicates the range of signal amplitude (magnitude), while the right side, marked in orange, indicates the range of total local fluctuations and the node state baseline value. The solid blue line represents the signal amplitude sequence corresponding to a specific node, fluctuating between approximately 213 and 223. The curve exhibits a clear periodic fluctuation trend, reflecting the dynamic changes in the sensor's timing characteristics at that node. The solid red line represents the node state baseline value calculated for that independent node, ranging from approximately 18 to 34. This baseline value fluctuates over time, maintaining a certain degree of synchronization with the overall trend of the signal amplitude sequence, but with a narrow fluctuation amplitude, indicating that the node state baseline value effectively reflects the activity level of local signal fluctuations. The dashed orange line represents the total local fluctuation, ranging from 21 to 29. Its fluctuation trend is highly consistent with the node state baseline value curve, verifying the crucial role of the total local fluctuation in calculating the node state baseline value. The dashed green line represents the average fluctuation value, a constant of approximately 24, serving as a baseline for fluctuations during historical normal operation periods, used to compare the deviation of the total local fluctuation value at the current moment.

[0030] In practice, each independent node in the graph structure is taken as the starting source node. The base state value of the starting source node is obtained, along with the edge weights of the undirected edges between the starting source node and each adjacent node. For any adjacent node, the base state value of the starting source node is multiplied by the edge weight from the starting source node to the adjacent node; the product is used as the initial influence amount transmitted from the starting source node to the adjacent node. Each starting source node transmits its initial influence amount to all its adjacent nodes in the same manner.

[0031] At each adjacent node, collect the initial influence values ​​transmitted from all neighboring source nodes. Accumulate all received initial influence values ​​to obtain the cumulative influence value of the adjacent node. Add the cumulative influence value to the adjacent node's own node state base value to obtain the updated state value of the adjacent node after one transmission. The calculation formula for this process is as follows: in, Indicates that the index is The updated state value of an independent node after one transmission; Indicates that the index is The node state base value of the independent node; Indicates and index is Each independent node establishes an index set of all adjacent source nodes with undirected edges; express The index of an adjacent source node in the set; Indicates adjacent source node The node state base value; Indicates that the index is The independent nodes and indexes are The edge weights of undirected edges between independent nodes.

[0032] After each transmission, the updated state value of each independent node is used as the new source value for transmission, and the transmission and superposition operation is repeated for the next level of adjacent nodes. At the beginning of each iteration, all independent nodes use the updated state value obtained in the previous round as the source value for transmission in this round, synchronously transmit the influence to their respective adjacent nodes, and recalculate the cumulative influence and the updated state value for the new round. The iteration terminates when the sum of the changes in the state values ​​of all independent nodes in the entire graph structure between two adjacent iterations is calculated. This is done by summing the absolute values ​​of the difference between the updated state value of each independent node in this round and the updated state value in the previous round. If this sum is lower than a preset convergence threshold, the iteration stops. The preset convergence threshold is set according to the magnitude of the node state base value, and is set to 0.001. After iteration convergence, the final state values ​​of all independent nodes together constitute the fused state field.

[0033] Extract candidate fault influence domains from the fused state field. Calculate the local gradient value of each independent node in the fused state field, as follows: for the node with index 1... Independent nodes, obtain the index as The state value of the independent node, and obtain the index as Each independent node has state values ​​for all its adjacent nodes with direct undirected edges, and the indices of these indices are calculated. The absolute difference between the state value of an independent node and the state value of each of its neighboring nodes is taken, and the maximum value among the multiple absolute differences is used as the index. The local gradient values ​​of independent nodes.

[0034] The local gradient values ​​are compared with a preset gradient threshold. The preset gradient threshold is set as the arithmetic mean of the local gradient values ​​of all independent nodes in the fused state field multiplied by a coefficient greater than 1, which is 2. Independent nodes whose local gradient values ​​exceed the preset gradient threshold are marked as high-gradient nodes. Using each high-gradient node as the center, the graph distance from other independent nodes in the graph structure to that high-gradient node is calculated. The graph distance is defined as the number of undirected edges traversed by the shortest path between two nodes. All independent nodes whose graph distance is less than the preset propagation radius, along with the high-gradient node itself, are included in the same initial influence region. The preset propagation radius is set based on the average shortest path distance between all independent nodes in the graph structure; the result of multiplying the average shortest path distance by 0.5 and rounding down is used as the preset propagation radius.

[0035] One or more initial influence regions formed around all high-gradient nodes are merged. If two initial influence regions contain at least one identical independent node, they are considered to overlap; if at least one independent node in one initial influence region is connected to at least one independent node in another initial influence region by a direct undirected edge, the two initial influence regions are considered to be adjacent. All overlapping or adjacent initial influence regions are merged into a single high-gradient change region. The number of high-gradient change regions obtained after merging is one or more, and all high-gradient change regions constitute the candidate fault influence domain.

[0036] See Figure 4The horizontal axis of the graph represents the spatial position of the electricity metering box along its length (in meters), and the vertical axis represents its spatial position along its width (in meters). The graph uses colored contour lines to show the distribution of state values ​​corresponding to each independent node in the fused state field. The color bar on the right indicates the range of fused state values, gradually increasing from 0.08 to 0.80, with the color transitioning from dark purple to yellow, indicating a trend of increasing state values. Nodes are marked with solid black dots. Nodes marked with red pentagrams are high-gradient nodes whose local gradient values ​​exceed a preset gradient threshold. These high-gradient nodes are concentrated in the central area of ​​the box, indicating that the activity level of independent nodes in this area is significantly higher than in other areas. A black dashed line connects some of the red high-gradient nodes to form a closed boundary, which is the boundary of the candidate fault influence domain. This boundary covers the area with higher state values, indicating that the highly active area of ​​the fused state field has been effectively aggregated into the fault influence domain. Observing the spatial variation trend of the fusion state value, it can be seen that the state value gradually decreases from the vicinity of the maximum value in the center of the candidate fault influence domain outwards, reflecting the global coupling effect formed by the cumulative amount of local fluctuations of the state value through multiple rounds of neighborhood transmission and superposition.

[0037] In practice, candidate fault influence domains are obtained, with one or more domains. For each candidate fault influence domain, a fault trigger point location operation is performed. Standardized input sequences of all independent nodes within the candidate fault influence domain are extracted within a preset historical time interval. The preset historical time interval is a fixed time period traced back from the current diagnostic time. This fixed time interval is set based on the fault development characteristics of the electricity metering box and is set to 600 seconds. For the multidimensional standardized input sequence, the magnitude of the multidimensional amplitude is calculated time-by-time, converting the multidimensional standardized input sequence into a single-dimensional magnitude sequence. The mean amplitude of all independent nodes at each sampling time within the candidate fault influence domain is calculated. The mean amplitude is the arithmetic mean of the single-dimensional magnitudes of all independent nodes at the corresponding sampling time. The mean amplitudes of all sampling times are arranged in chronological order to form the global mean curve of the candidate fault influence domain.

[0038] The system divides the historical time interval into normal operation periods, with the first third of the interval designated as the normal operation period. Data points of the global mean curve within these normal operation periods are extracted, and a least-squares method is used to perform linear fitting on the data points, resulting in a fitted trend line. This fitted trend line is then used as the nominal trajectory of the candidate fault influence domain. At each subsequent sampling time, the nominal trajectory outputs a corresponding nominal amplitude, which is determined by extrapolation from the fitted trend line equation.

[0039] Starting from the end of a preset historical time interval, the difference between the mean amplitude of the global mean curve at each sampling moment is calculated along the reverse time direction, and the nominal amplitude at the same sampling moment is taken as the absolute value of the difference. The absolute value of the difference is compared with a preset allowable deviation range, which is set to ±5% of the nominal amplitude. If the absolute value of the difference at a certain sampling moment exceeds the preset allowable deviation range, and the absolute value of the difference at all subsequent sampling moments along the reverse time direction continues to exceed the preset allowable deviation range, then that sampling moment is positioned as the starting moment when the signal evolution trend deviates from the nominal trajectory, and is used as the fault trigger point of the candidate fault influence domain. If no situation exceeding the preset allowable deviation range is found after traversing back to the starting moment of the preset historical time interval, then it is determined that there is no fault trigger point within the candidate fault influence domain, and the processing of the candidate fault influence domain ends.

[0040] The independent node containing the fault trigger point is obtained as follows: Among all independent nodes within the candidate fault influence domain, calculate the absolute value of the difference between the single-dimensional modulus value of each independent node at the sampling time of the fault trigger point and the average single-dimensional modulus value of the independent node during normal operation. The independent node with the largest absolute value is taken as the independent node containing the fault trigger point. The independent node containing the fault trigger point is used as the starting node.

[0041] Using the fault trigger point as the end time, a preset time window length is extracted to obtain the pre-trigger signal segment for each independent node within the candidate fault influence domain. The preset time window length is set to 100 sampling points to allow sufficient backtracking analysis length based on the fault propagation speed. The pre-trigger signal segment contains the standardized input sequence amplitude of each independent node at each sampling time within the time period of backtracking from the fault trigger point to the specified window length; if it is multidimensional, a single-dimensional modulus value is used.

[0042] Calculate the Pearson correlation coefficient between the pre-trigger signal segments of any two independent nodes within the candidate fault influence domain. The formula for calculating the Pearson correlation coefficient is as follows: in, Represents an independent node With independent nodes The Pearson correlation coefficient between the signal segments prior to triggering. and These represent the index identifiers of two different independent nodes within the candidate fault's influence domain; This indicates the total number of sampling points contained in the signal segment before triggering. Equal to the preset time window length; This indicates the index of the sampling point within the signal segment before triggering. The value range is 1 to ; Represents an independent node The sequence index of the pre-trigger signal segment is The amplitude of the sampling points, Represents an independent node The sequence index of the pre-trigger signal segment is The amplitude of the sampling points; Represents an independent node All of the pre-trigger signal segments The arithmetic mean of the amplitudes at each sampling point Represents an independent node All of the pre-trigger signal segments The arithmetic mean of the amplitudes at each sampling point.

[0043] A preset correlation threshold is set, with a value of 0.85. For each pair of independent nodes within the candidate fault influence domain, the calculated Pearson correlation coefficient... If the correlation value is greater than the preset correlation threshold of 0.85, then the pair of independent nodes is marked as a strongly correlated pair.

[0044] Starting from the independent node where the fault trigger point is located, the fault propagation chain is constructed by progressively expanding outward along strongly associated pairs. The expansion process uses a breadth-first traversal: an empty fault propagation chain list is created, the starting node is added to the list, and marked as visited; in each expansion, all strongly associated adjacent nodes of the current last independent node in the fault propagation chain list are retrieved, and the independent nodes marked as visited are removed. If any remaining adjacent nodes exist, they are added to the fault propagation chain list in descending order of their Pearson correlation coefficient with the current independent node and marked as visited; if the number of remaining adjacent nodes is zero, the process backtracks to the previous independent node to continue searching for unvisited strongly associated adjacent nodes; this expansion process is repeated until all strongly associated adjacent nodes connected to the starting node have been visited, or the traversal depth reaches a preset upper limit. The preset upper limit of the expansion depth is set to twice the total number of independent nodes within the candidate fault's influence domain. During the continuous expansion process, all independent nodes traversed are arranged in the order of their first visit, that is, in the order they are added to the fault propagation chain list, thus forming a fault propagation chain.

[0045] In the specific implementation, the constructed fault propagation chain is obtained, with independent nodes arranged in the order of expansion. For each independent node in the fault propagation chain, the updated state value of the independent node in the fused state field is obtained, along with its own node state base value. The difference between the updated state value and the node state base value is calculated, and this difference is taken as the state deviation magnitude of the independent node. The state deviation magnitude represents the state increment of the independent node caused by the superposition of neighboring state propagation after the fused state field iteratively converges. If the state deviation magnitude is positive, it indicates that the independent node is affected by the state injection from adjacent fault sources, and its own activity level is pushed up; if the state deviation magnitude is negative, it indicates that the original fluctuation level of the independent node is suppressed by the normal state of the neighboring region in the global coupling.

[0046] The state deviations of all independent nodes in the fault propagation chain are accumulated sequentially according to the order of the fault propagation chain. The accumulation formula is: in, This indicates the total accumulated deviation. This represents the total number of independent nodes in the fault propagation chain. This represents the index of the order of independent nodes in the fault propagation chain. The value ranges from 1 to , Corresponding to the first independent node in the fault propagation chain, This corresponds to the last independent node in the fault propagation chain. The index representing the order of events in the fault propagation chain is: The magnitude of the state deviation of the independent node.

[0047] Based on total deviation accumulation The severity level of the fault is determined by the preset interval. The preset interval is divided into three levels: the first preset interval is... Less than the preset first threshold corresponds to a minor fault level; the second preset range is... A fault level greater than or equal to a preset first threshold and less than a preset second threshold corresponds to a medium fault level; the third preset range is... A value greater than or equal to the preset second threshold corresponds to a severe fault level. The preset first and second thresholds are set based on the correspondence between the total accumulated deviation of the electricity metering box in the historical fault samples and the actual maintenance records. The boundary value between minor and moderate faults in the historical fault samples is used as the preset first threshold, and the boundary value between moderate and severe faults is used as the preset second threshold.

[0048] When determining the fault type, a preset fault library is invoked. This library stores standard node response sequence templates for each fault type, defining the standard response order of independent nodes in the fault propagation chain when a specific fault type occurs. The standard node response sequence templates are constructed as follows: from several historical fault cases of known fault types, the arrangement order of independent nodes in the fault propagation chain constructed in each historical fault case is extracted; for several arrangements under the same fault type, the longest common subsequence method is used to extract the common independent node order fragments, which are then stored in the preset fault library as standard node response sequence templates. The response order of each independent node in the current fault propagation chain is obtained; this response order is the arrangement order of independent nodes in the fault propagation chain. The arrangement order of the current fault propagation chain is matched with the standard node response sequence templates for each fault type in the preset fault library. The matching degree is calculated using sequence similarity based on edit distance; the smaller the edit distance, the higher the matching degree. The preset fault type with the highest matching degree to the standard node response sequence template is selected as the fault type to be determined.

[0049] The spatial coordinates corresponding to the first independent node in the fault propagation chain are marked as the fault initiation position. The first independent node is the node in the fault propagation chain sequence. The corresponding independent nodes. Obtain the spatial coordinates of all independent nodes in the fault propagation chain, and connect them sequentially according to the order of the fault propagation chain. The connection method is as follows: starting from the spatial coordinates of the first independent node in the order, connect the line segment to the spatial coordinates of the second independent node, then connect the line segment from the spatial coordinates of the second independent node to the spatial coordinates of the third independent node, and so on, until connecting to the spatial coordinates of the last independent node, forming a continuous spatial polyline. This spatial polyline is the fault propagation path trajectory.

[0050] The fault type, fault severity level, fault starting location, and fault propagation path trajectory are combined to form a diagnostic result data packet. The diagnostic result data packet is organized in a structured data format and includes fields for fault type, fault severity level, three-dimensional coordinates of the fault starting location, and coordinate sequence of the fault propagation path trajectory. The diagnostic result data packet is output to the power metering box operation and maintenance management system or a local display terminal.

[0051] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A fault diagnosis method for electricity metering boxes based on multi-source data fusion, characterized in that, The method includes: The real-time output sequence of the operating signal from the multi-source sensors in the power metering box is obtained, and the time-domain alignment and abnormal transition point cleaning of each operating signal sequence are performed to obtain the standardized input sequence. The standardized input sequence is mapped to a graph structure with the metering box operating status as nodes, and the local fluctuation accumulation of the signal sequence corresponding to each node is calculated as the node state base value. Based on the node state base value, the state influence coefficients of adjacent nodes are passed and superimposed along the edge direction of the graph structure to generate a fusion state field that reflects the global coupling state. High gradient change regions are extracted from the fused state field as candidate fault influence regions, and the starting moment when the signal evolution trend deviates from the nominal trajectory is located in the candidate fault influence region as the fault trigger point. Based on the fault trigger point, the correlation strength of the signal sequences of each node in the candidate fault influence domain within the preset window before triggering is traced back to construct the fault propagation chain. Based on the order of responses and the magnitude of state deviation between nodes in the fault propagation chain, the fault type and severity level of the power metering box are determined, and the fault diagnosis results are output.

2. The method for fault diagnosis of power metering boxes based on multi-source data fusion according to claim 1, characterized in that, The process of performing time-domain alignment and abnormal transition point cleaning on each running signal sequence to obtain a standardized input sequence includes: The various operating signal sequences output by the multi-source sensors are uniformly interpolated to the same sampling time reference, so that all signal sequences have a corresponding amplitude at each sampling time. For each signal amplitude at each sampling time, calculate the absolute value of the difference between it and the amplitude at the previous time, and mark the sampling points whose absolute value of the difference exceeds the preset jump threshold as candidate anomalies. For a sampling point marked as a candidate anomaly, linear interpolation is performed using the amplitude of the adjacent normal sampling points before and after the sampling point to replace it, thus completing the cleaning of abnormal jump points and obtaining a standardized input sequence.

3. The method for fault diagnosis of electricity metering boxes based on multi-source data fusion according to claim 1, characterized in that, The step of mapping the standardized input sequence to a graph structure with the metering box operating status as nodes includes: Each electrical node, sensor installation location, and functional module interface inside the power metering box is defined as an independent node in the graph structure, and the spatial coordinates of each independent node are recorded. Based on the electrical connection relationship or spatial distance between any two independent nodes that is less than a preset adjacency distance threshold, an undirected edge is established between the corresponding two nodes, and each undirected edge is assigned an edge weight that reflects the degree of connection tightness. The graph structure is constructed using all independent nodes, their undirected edges, and edge weights, and the standardized input sequence corresponding to each independent node is used as the temporal feature vector of that node.

4. The method for fault diagnosis of power metering boxes based on multi-source data fusion according to claim 3, characterized in that, The calculation of the cumulative local fluctuation of the signal sequence corresponding to each node, as the node state base value, includes: For each independent node in the graph structure, the sum of the absolute values ​​of the amplitude differences between adjacent nodes is calculated using the amplitude sequence of its temporal feature vector within a preset sliding window to obtain the total local fluctuation of that independent node. The volatility deviation factor is obtained by comparing the total local volatility of the independent node with its average volatility over a preset historical period. The product of the total local fluctuation of the independent node and the fluctuation deviation factor is used as the node state base value of the independent node, which is used to characterize the state activity level of the independent node at the current moment.

5. The method for fault diagnosis of power metering boxes based on multi-source data fusion according to claim 4, characterized in that, The process of generating a fused state field reflecting the global coupling state by progressively passing and superimposing the state influence coefficients of adjacent nodes along the edge direction of the graph structure based on the node state base value includes: Using each independent node in the graph structure as the starting source node, multiply the node state base value of the independent node by the edge weights between the independent node and each neighboring node to obtain the initial influence amount transmitted to each neighboring node. After receiving the initial influence values ​​transmitted by all neighboring source nodes, each neighboring node sums up the received initial influence values ​​to obtain the cumulative influence value of the neighboring node, and then adds the cumulative influence value to the node state base value of the neighboring node to form the updated state value after one transmission. Using the updated state value as the new source value, the transmission and superposition operation to the next layer of adjacent nodes is repeated until the sum of the changes in the state values ​​of each independent node in the entire graph structure in two consecutive iterations is lower than the preset convergence threshold. The converged state values ​​of each independent node are then used to form the fused state field.

6. The method for fault diagnosis of power metering boxes based on multi-source data fusion according to claim 5, characterized in that, The step of extracting high gradient change regions from the fused state field as candidate fault influence regions includes: Calculate the absolute value of the difference between the state value of each independent node and the state values ​​of all its neighboring nodes in the fused state field, and take the maximum value among all the absolute values ​​of the differences as the local gradient value of the independent node. Independent nodes whose local gradient values ​​exceed a preset gradient threshold are marked as high gradient nodes. Then, taking each high gradient node as the center, all independent nodes whose graph distance from the high gradient node is less than a preset propagation radius are assigned to the same initial influence region. All overlapping or adjacent initial influence regions are merged into the same high gradient change region, and all the high gradient change regions obtained after merging are used as candidate fault influence domains.

7. The method for fault diagnosis of power metering boxes based on multi-source data fusion according to claim 6, characterized in that, The starting moment when the evolution trend of the location signal deviates from the nominal trajectory in the candidate fault influence domain is taken as the fault trigger point, including: For each candidate fault influence domain, the standardized input sequence of all independent nodes in the influence domain within a preset historical time interval is extracted, and the mean amplitude of all independent nodes in the influence domain at each moment is calculated to obtain the global mean curve of the influence domain. The fitting trend line of the global mean curve during the normal operation period is used as the nominal trajectory of the influence domain; The difference between the global mean curve and the nominal trajectory is compared point by point backward from the end of the preset historical time interval. The moment when the difference first exceeds the preset allowable deviation range is positioned as the starting moment when the signal evolution trend deviates from the nominal trajectory, and is used as the fault trigger point of the candidate fault influence domain.

8. The method for fault diagnosis of power metering boxes based on multi-source data fusion according to claim 7, characterized in that, The step of constructing a fault propagation chain by backtracking the correlation strength of signal sequences of each node within the candidate fault influence domain within a preset window before triggering, based on the fault trigger point, includes: Using the fault trigger point as the end time, a preset time window length is extracted forward to obtain the pre-trigger signal segment of each independent node within the candidate fault influence domain; Calculate the Pearson correlation coefficient between the pre-trigger signal segments of any two independent nodes within the influence domain of the candidate fault, and mark the two nodes with a correlation coefficient greater than a preset correlation threshold as strongly correlated pairs; Starting from the independent node where the fault trigger point is located, the fault propagation chain is formed by gradually expanding outward along each strongly associated pair and arranging all independent nodes traversed during the continuous expansion process in the order of expansion.

9. The method for fault diagnosis of power metering boxes based on multi-source data fusion according to claim 8, characterized in that, The determination of the fault type and severity level of the power metering box based on the sequence of responses and the magnitude of state deviation between nodes in the fault propagation chain includes: For each independent node in the fault propagation chain, the difference between the updated state value of the independent node and the node state base value of the independent node is calculated as the state deviation magnitude of the independent node. The state deviation amplitudes of all independent nodes in the fault propagation chain are sequentially accumulated in order to obtain the total deviation accumulation, and the fault severity level is determined according to the preset interval in which the total deviation accumulation is located. The response order of each independent node in the fault propagation chain is compared with the standard node response order template corresponding to each fault type in the preset fault database. The preset fault type with the highest matching degree with the standard node response order template is taken as the fault type to be determined.

10. The method for fault diagnosis of power metering boxes based on multi-source data fusion according to claim 9, characterized in that, After determining the fault type and severity level of the power metering box based on the response sequence and state deviation magnitude between nodes in the fault propagation chain, the method further includes: The spatial coordinates of the first independent node in the fault propagation chain are marked as the fault starting position, and the spatial coordinates of all independent nodes in the fault propagation chain are connected in order to generate the fault propagation path trajectory. The fault type, the fault severity level, the fault starting location, and the fault propagation path trajectory are combined to form a diagnostic result data packet, and the diagnostic result data packet is output.

Images