A fault early warning method for urban rail power supply system based on multi-source data fusion

Abstract

This invention relates to the field of power monitoring technology, specifically to a fault early warning method for urban rail transit power supply systems based on multi-source data fusion. The method involves acquiring thermal images, extracting temperature drift trajectories, and marking thermal anomalies. Combined with voltage sampling, it identifies abnormal data, analyzes the overlap of anomaly time and direction, locates overlapping fault points, and identifies linked early warning areas through spatial statistics. In this invention, by identifying the continuous drift trajectory of the temperature centroid in the traction transformer thermal image and extracting thermal anomalies, and combining this with voltage sampling sequences within the same time period to analyze voltage mutation characteristics, a composite identification method combining time synchronization and trend direction comparison is achieved. This effectively improves the accuracy of fault symptom location. By aggregating overlapping fault points from multiple power supply sections and performing spatiotemporal clustering, the ability to perceive common anomalies within the region is significantly improved. Based on the joint calculation of spatial proximity structure and statistical clustering, potential linked fault areas can be clearly identified.

Description

Technical Field

[0001] This invention relates to the field of power monitoring technology, and in particular to a fault early warning method for urban rail power supply systems based on multi-source data fusion. Background Technology

[0002] The field of power monitoring technology involves the real-time acquisition, analysis, and status assessment of various electrical parameters in power systems, such as current, voltage, frequency, and power, to achieve effective control over the operating status of power equipment and risk warning.

[0003] Among them, the fault early warning method for urban rail power supply system refers to identifying potential abnormal situations by analyzing monitoring data of the contact network and traction power supply circuit in response to possible fault symptoms such as voltage anomalies, current surges, and contact network disconnections in the rail transit power supply network.

[0004] Existing technologies rely solely on monitoring and analyzing single parameters such as voltage and current in the overhead contact system and traction power supply circuits. This makes it difficult to capture the multi-dimensional characteristics of different types of faults in a timely manner before they occur. In the early stages when traction transformers experience abnormal surface heating before significant voltage fluctuations occur, it is often difficult to issue early warnings. As a result, the faults are only discovered when they reach a critical state, which can easily lead to power outages or system cascading failures. In the case of synchronous anomalies at multiple points in a region, existing technologies lack the ability to identify the spatiotemporal distribution patterns between anomalies and cannot form a basis for multi-segment linkage judgments. This results in delayed early warnings and a limited response range for the overall system. A typical scenario is that multiple transformers exhibit similar abnormal trends in a short period of time but are not identified by the system, ultimately forming a chain of hidden faults across segments. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a fault early warning method for urban rail power supply systems based on multi-source data fusion.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a fault early warning method for urban rail transit power supply systems based on multi-source data fusion, comprising the following steps:

[0007] S1: Collect continuous thermal images of the surface of the track traction transformer within a specified period, identify the continuous drift trajectory of the temperature centroid of the traction transformer surface, and mark the set of thermal anomaly points;

[0008] S2: Obtain the time period of the concentrated recording of the thermal anomaly points, extract the voltage sampling sequence of the high-voltage side of the traction transformer within the same time period, and identify the voltage anomaly data group;

[0009] S3: Analyze whether the time points when the thermal anomaly point set and the voltage anomaly data group produce anomalies within the same time period overlap, and determine whether the temperature centroid drift direction is consistent with the voltage change direction, and mark the traction transformer fault overlap point.

[0010] S4: Summarize the overlapping fault points of the traction transformers in multiple power supply sections, analyze whether multiple transformers have the same sudden behavior within a specified time, and construct a regional anomaly aggregation set;

[0011] S5: Using the coordinates and fault occurrence time of each abnormal traction transformer in the regional anomaly aggregation set as input, construct a spatial proximity structure, analyze the distribution of fault points in the structure using Ripley's K function, and mark the linkage early warning area.

[0012] As a further aspect of the present invention, the step of obtaining the thermal anomaly point set specifically includes:

[0013] S111: Collect continuous thermal images of the surface of the track traction transformer within a specified period, select the pixel group in the central region of each frame where the temperature exceeds the ambient temperature, determine the temperature distribution centroid of each frame based on the relationship between the temperature of the pixel and the image position, and generate a temperature centroid trajectory sequence.

[0014] S112: Based on the temperature centroid trajectory sequence, track the change path of the centroid position between consecutive frames, identify the position where the direction changes in the path, locate the temporal position of the change according to the image frame number, extract the coordinate points and time information representing the change behavior, and generate a set of drift change positions.

[0015] S113: Based on the coordinates and time data of each change point in the set of drift change locations, and combined with the boundary of the traction transformer surface monitoring area, invalid points outside the area are screened out to generate a set of thermal anomaly points.

[0016] As a further aspect of the present invention, the step of acquiring the voltage anomaly data group specifically includes:

[0017] S211: Call the time data corresponding to each anomaly point in the thermal anomaly point set, determine the start and end time range, and extract the voltage side sampling data sequence of the traction transformer voltage side within the same time period;

[0018] S212: Based on the time distribution and voltage change of continuous data points in the voltage-side sampling data sequence, identify the starting position where the voltage exceeds the voltage change threshold, extract the change indicators of the fluctuation direction trend, duration and numerical amplitude corresponding to each change segment, and generate a voltage fluctuation feature set.

[0019] S213: Filter the change segments of each change index in the voltage fluctuation feature set, classify them in chronological order, and generate voltage anomaly data groups.

[0020] As a further aspect of the present invention, the step of obtaining the fault overlap point of the traction transformer specifically includes:

[0021] S311: Call up all the time information recorded in the thermal anomaly point set and the voltage anomaly data group, establish corresponding time axis sequences respectively, and filter out the anomaly records that occur at the same time in the same time period by comparing the interval relationship between each time point in the two sequences, and generate an anomaly time overlap data table.

[0022] S312: Based on the thermal anomaly point and abnormal voltage corresponding to each overlapping time point in the abnormal time overlap data table, extract the direction of continuous coordinate change and the direction of corresponding voltage data change, determine whether the change directions are the same, and filter the direction consistency marker list.

[0023] S313: Based on the matching point information with consistent direction and overlapping time in the directional consistency mark list, integrate them according to the traction transformer number and time sequence to generate traction transformer fault overlap points.

[0024] As a further aspect of the present invention, the step of obtaining the regional anomaly aggregation set specifically includes:

[0025] S411: Call the time information and physical coordinates recorded in the fault overlap point of the traction transformer, extract the power supply section identifier of each fault point and divide it into groups, count the time distribution range and spatial location set of the fault occurrence in each group, and generate a section spatiotemporal distribution data table.

[0026] S412: Cluster the coordinates and time of each fault point in the segment's spatiotemporal distribution data table using the DBSCAN clustering algorithm to generate a spatiotemporal clustering boundary index group;

[0027] S413: Based on the cluster information divided in the spatiotemporal clustering boundary index group, extract the corresponding power supply section number and transformer number in the cluster, merge the cluster sets that simultaneously satisfy section concentration and fault type consistency, and generate a regional anomaly aggregation set.

[0028] As a further aspect of the present invention, the step of obtaining the linkage early warning area specifically includes:

[0029] S511: Call the coordinate information and corresponding fault occurrence time of each abnormal traction transformer in the regional anomaly aggregation set, record the straight-line distance and time tag sequence between all transformers, and generate a spatiotemporal relationship structure data group.

[0030] S512: Based on the spatiotemporal relationship structure data set, set the clustering radius interval and time weight factor based on spatial distance, statistically analyze the distribution characteristics between fault points, input Ripley's K function to calculate the clustering degree of traction transformer fault overlap points in geographic space, and generate a set of spatial clustering statistical curves.

[0031] S513: Determine the degree of deviation between the spatial aggregation statistical curve group and the theoretical uniform distribution curve, filter spatial areas whose deviation exceeds the set aggregation intensity judgment benchmark value, and output them as linkage warning areas.

[0032] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0033] In this invention, by identifying the continuous drift trajectory of the temperature centroid in the thermal image of the traction transformer and extracting thermal anomaly points, and combining the voltage sampling sequence within the same time period to analyze voltage mutation characteristics, a composite identification method of time synchronization and trend direction comparison is realized, which effectively improves the accuracy of fault symptom location. By aggregating fault overlap points in multiple power supply sections and performing spatiotemporal clustering operations, the ability to perceive common anomalies in the region can be significantly improved. Based on the joint calculation of spatial proximity structure and statistical clustering degree, potential linkage fault areas can be clearly identified, strengthening the linkage early warning capability of spatial correlation and temporal evolution trend among multiple transformers, avoiding the misjudgment and omission problems caused by single-point anomaly judgment, and improving the efficiency and accuracy of response to potential fault spread in urban rail power supply systems under large-scale, multi-section backgrounds. Attached Figure Description

[0034] Figure 1 This is a flowchart illustrating the fault early warning method for urban rail power supply systems based on multi-source data fusion according to the present invention.

[0035] Figure 2 This is a schematic diagram of the process for marking thermal anomaly point sets according to the present invention;

[0036] Figure 3 This is a schematic diagram of the process for identifying voltage anomaly data groups according to the present invention;

[0037] Figure 4 This is a schematic diagram of the process for marking the fault overlap point of the traction transformer according to the present invention;

[0038] Figure 5 This is a schematic diagram illustrating the process of constructing a regional anomaly aggregation set according to the present invention;

[0039] Figure 6 This is a schematic diagram of the process for marking the linkage early warning area in this invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0041] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0042] Please see Figure 1 This invention provides a technical solution: a fault early warning method for urban rail power supply systems based on multi-source data fusion, comprising the following steps:

[0043] S1: Collect continuous thermal images of the surface of the track traction transformer within a specified period, identify the continuous drift trajectory of the temperature centroid of the traction transformer surface, and mark the set of thermal anomaly points;

[0044] S2: Obtain the concentrated recording time period of thermal anomaly points, extract the voltage sampling sequence of the high-voltage side of the traction transformer within the same time period, and identify the voltage anomaly data group;

[0045] S3: Analyze whether the time points when the thermal anomaly point set and the voltage anomaly data group produce anomalies within the same time period overlap, and determine whether the direction of temperature centroid drift is consistent with the direction of voltage change, and mark the traction transformer fault overlap point.

[0046] S4: Summarize the overlapping fault points of traction transformers in multiple power supply sections, analyze whether multiple transformers experience the same sudden change behavior within a specified time, and construct a regional anomaly aggregation set;

[0047] S5: Using the coordinates and fault occurrence time of each abnormal traction transformer in the regional anomaly aggregation set as input, a spatial proximity structure is constructed. The distribution of fault points in the structure is analyzed using Ripley's K-function, and the linkage early warning area is marked.

[0048] Please see Figure 2 The specific steps for obtaining the hot anomaly point set are as follows:

[0049] S111: Collect continuous thermal images of the surface of the track traction transformer within a specified period, select the pixel group in the central region of each frame where the temperature exceeds the ambient temperature, determine the temperature distribution centroid of each frame based on the relationship between the temperature of the pixel and the image position, and generate a temperature centroid trajectory sequence.

[0050] When collecting continuous thermal images of the surface of a rail traction transformer, a thermal imager is installed at the monitoring site and set to continuously capture images of the transformer at a frequency of 5 frames per second for 10 minutes, resulting in a total of 3000 thermal images. For each frame, the central area is selected as the analysis area. For example, if the overall image size is 800×600 pixels, the central area is set as a square area with a width and height of 200 pixels each. The infrared temperature value of each pixel is obtained, and then the temperature value of each pixel is compared with the currently measured ambient temperature. For example, if the ambient temperature is measured to be 32 degrees Celsius, then pixels with a temperature greater than 32 degrees Celsius are selected as high-temperature pixels. For each high-temperature pixel, the coordinates and temperature values ​​of all high-temperature pixels are statistically analyzed. By using a temperature-weighted method, the concentration of these pixels in the image is calculated to determine the centroid of temperature. For example, if one pixel has a temperature of 36 degrees Celsius and another has a temperature of 38 degrees Celsius, their corresponding positions are located slightly to the right and slightly below the center of the image, respectively. By multiplying the temperature value by the pixel position for all high-temperature pixels, summing all the pixels and dividing by the total temperature value, the coordinates of the centroid of temperature in this frame are obtained. This process is repeated for all 3000 frames of images to form a continuous centroid trajectory sequence, which represents the path of the center of the transformer heat source as it changes over time.

[0051] S112: Based on the temperature centroid trajectory sequence, track the process path of the change in centroid position between consecutive frames, identify the position where the direction changes in the path, locate the temporal position of the change according to the image frame number, extract the coordinate points and time information representing the change behavior, and generate a set of drift change positions.

[0052] Based on the above centroid trajectory sequence, the centroid coordinates are read frame by frame at each instant, and the corresponding time point or frame number is recorded. For example, the centroid in frame 1 is on the left side of the image, in frame 2 it is slightly to the right of the center, and in frame 3 it changes to the upper right. The center trajectory path is recorded. During the continuous change of the path, the directional trend between adjacent centroid positions is observed segment by segment. For example, from frame 1 to frame 2 the centroid shifts to the right, and from frame 2 to frame 3 the centroid shifts to the upper right, so the overall direction is upper right. If the position suddenly changes to upper left starting from frame 4, it is determined that a turning point has occurred. When identifying the turning point... Based on the significance of changes in the horizontal or vertical direction, a significant turning point is considered if the angle of change exceeds 30 degrees from the original direction. According to this criterion, all image frames where turning points occur are gradually searched and the frame number and the centroid coordinate information at that time are recorded. For example, if a significant change occurs from the lower left to the upper right direction in frame 50, and its centroid coordinate position is slightly above the center of the image, then this coordinate and the frame number together form a drift change point. Finally, after integrating all turning points, a set of drift change positions is constructed.

[0053] S113: Based on the coordinates and time data of each change point in the drift change location set, combined with the boundary of the traction transformer surface monitoring area, invalid points outside the area range are screened out to generate a thermal anomaly point set;

[0054] After obtaining the set of drift change locations, each data point in the set needs to be compared with the boundary of the monitoring area. The boundary is usually mapped based on the actual physical size of the traction transformer. For example, the monitoring area is defined as the area on the image from horizontal coordinate 100 to 300 and vertical coordinate 50 to 250. Each drift point's coordinate value is checked to see if it falls within this area. For example, if the coordinates of a drift point are horizontal 120 and vertical 130, it is obviously within the specified area and is retained. If a point is horizontal 320, it exceeds the boundary and needs to be deleted. In addition, time information is also needed to screen whether the abnormal points have temporal continuity. For example, if a drift point occurs at frame 1500, which is exactly 5 minutes, and another drift point occurs at frame 1510, which is 2 seconds after 5 minutes, the time interval between the two drift points is 2 seconds. If the change time interval is less than 2 seconds, it may be a normal heat flow fluctuation rather than an abnormal drift, and the point can be removed. Through the dual verification method of area boundary and time, the effective drift points that are both within the area and meet the change time sequence requirements are finally retained, and these points are integrated to generate a set of thermal anomaly points.

[0055] Please see Figure 3 The specific steps for obtaining the voltage anomaly data set are as follows:

[0056] S211: Call the time data corresponding to each anomaly point in the thermal anomaly point set, determine the start and end time range, and extract the voltage side sampling data sequence of the traction transformer voltage side within the same time period;

[0057] The time data from the hot anomaly point set is retrieved first. Time is extracted from all anomaly points, and the timestamp corresponding to each anomaly point is recorded in a time series. For example, if anomalies occur at 300 seconds, 304 seconds, and 309 seconds, the minimum time point is determined to be 300 seconds, and the maximum to be 309 seconds. The analysis start and end time range is set to 300 seconds to 309 seconds, serving as the extraction interval for subsequent voltage sampling data. Then, voltage-side data within the corresponding time period is read from the voltage measurement system. The sampling frequency is set to 100 times per second, resulting in 900 voltage sampling points extracted within this 9-second time period. These points are arranged in chronological order, forming a voltage-side sampling data sequence. This sequence is the basic information source for subsequent voltage anomaly judgment. During this process, it is necessary to ensure that the sampling data is completely aligned with the time of the hot anomaly points; that is, the time corresponding to the first point in the sampling data is exactly 300 seconds, and the last point is exactly 309 seconds. If microsecond-level time information exists in the hot anomaly point set...

[0058] S212: Based on the time distribution and voltage change of continuous data points in the voltage side sampling data sequence, identify the starting position where the voltage exceeds the voltage change threshold, extract the change index of fluctuation direction trend, duration and numerical amplitude corresponding to each change segment, and generate a voltage fluctuation feature set.

[0059] In the voltage-side sampling data sequence, the continuous change of voltage value at each moment is analyzed point by point. By calculating the voltage difference between two adjacent sampling points, it is identified when the voltage change exceeds the mutation threshold. For example, if the mutation threshold is set to 2 volts, if the voltage at a certain sampling point is 220 volts and the next point is 217.8 volts, the change is 2.2 volts, which exceeds the set threshold. This sampling point is recorded as the mutation start point, and the next sampling point is tracked. It is recorded whether the voltage is continuously decreasing, continuously increasing, or fluctuating. The duration is calculated in combination with the time interval. If the mutation lasts for 0.6 seconds, the direction is decreasing, and the total voltage drop is 5 volts, then this segment is recorded as a complete voltage fluctuation event. Subsequently, all mutation start points in the subsequent data are analyzed in turn, and their direction, duration, and amplitude are extracted. Finally, all fluctuation events are organized to generate a voltage fluctuation feature set. Each set of feature data includes a direction label (increasing or decreasing), start time, end time, and numerical change.

[0060] S213: Filter the change segments of each change index in the voltage fluctuation feature set, classify them in chronological order, and generate voltage anomaly data groups;

[0061] For each record in the voltage fluctuation feature set, each item is filtered according to the change index. The filtering rules are based on dual criteria of change amplitude and duration. For example, a segment with a voltage change amplitude greater than 3 volts and a duration of more than 0.3 seconds is identified as a significant voltage fluctuation event. If a voltage drop amplitude of 4.1 volts and a duration of 0.5 seconds meets the criteria, it is retained, while another segment with a voltage rise of only 1.5 volts is removed. After filtering, all retained items are arranged and classified in order of start time. If multiple fluctuation events occur with a time interval of less than 0.5 seconds, they are merged into a continuous event for processing. Finally, the voltage anomaly data group is generated according to the classification criteria and time.

[0062] Please see Figure 4 The specific steps for obtaining the fault overlap point of the traction transformer are as follows:

[0063] S311: Call all the time information recorded in the thermal anomaly point set and voltage anomaly data group, establish corresponding time axis sequences respectively, and filter the anomaly records that occur at the same time in the same time period by comparing the interval relationship between each time point in the two sequences, and generate an anomaly time overlap data table.

[0064] The system retrieves the time information contained in the thermal anomaly point set and the voltage anomaly data set. First, it extracts the timestamps of all records in both datasets into two independent timeline sequences. Each point in the thermal anomaly point set contains timestamp information; for example, the first time point in the thermal anomaly point set is 302 seconds, and the second is 305 seconds. If a voltage anomaly interval in the data set starts at 303 seconds and continues to 306 seconds, then the thermal anomaly point at 305 seconds falls within this anomaly interval, constituting an overlap. However, 302 seconds is earlier than the start of this interval and does not meet the overlap condition. The judgment criteria are: if a thermal anomaly time point is located between the start and end times of any voltage anomaly interval (including the start and end times themselves), it is considered a time overlap; if it is earlier than the start time or later than the end time, it does not constitute an overlap. Using this logic, the time relationship between all thermal anomaly points and voltage anomaly segments is compared, and records constituting overlap are filtered out and integrated into an anomaly time overlap data table. Each record indicates the overlapping time point, the corresponding thermal anomaly information, and the voltage anomaly content for that time period.

[0065] S312: Based on the thermal anomaly point and abnormal voltage corresponding to each overlapping time point in the abnormal time overlap data table, extract the direction of continuous coordinate change and the direction of corresponding voltage data change, determine whether the change direction is the same, and filter the list of direction consistency markers.

[0066] Based on the overlapping records in the abnormal time overlap data table, the coordinate information of the thermal anomaly point and the voltage value change information of each record are extracted. For the thermal anomaly part, the two-dimensional direction vector of the heat source movement is calculated by the coordinate difference between the current frame and the previous frame, and the polar angle corresponding to this vector direction (calculated using the arctangent function, in degrees) is used as the reference for the direction of heat source movement. To standardize the direction determination, the 360° circle is divided into 8 directional regions, specifically as follows: Right direction (→): angle range is... The upper right direction (↗): the angle range is Upward (↑): Angle range is The upper left direction (↖): the angle range is Left direction (←): Angle range is and The lower left direction (↙): the angle range is Downward (↓): Angle range is The lower right direction (↘): the angle range is To ensure that each major direction (except for the directly left) covers a 45° range, based on the periodicity of angles from -180° to +180°, the left direction (←) is defined as the union of two intervals near +180° and -180°. This area extends 22.5° "up" (counterclockwise) from the center line, reaching +157.5°. Therefore, the range from +157.5° to +180° is considered the left direction, corresponding to [+157.5°, +180°]. Similarly, this area extends 22.5° "down" (clockwise) from the center line, reaching -157.5°. Therefore, the range from -180° to -157.5° is also considered the left direction, corresponding to (-180°, -157.5°).

[0067] For voltage data, if the voltage value in the current frame decreases relative to the previous frame, the voltage direction is marked as "decreasing"; if the voltage value increases, the direction is marked as "increasing". The direction of the heat source is matched with the direction of voltage change: when the heat source moves in the direction of "down", "down-right", or "down-left" and the voltage direction is "decreasing", or when the heat source moves in the direction of "up", "up-right", or "up-left" and the voltage direction is "increasing", the directions are considered consistent; otherwise, they are considered inconsistent. By checking the consistency of direction for each overlapping record, consistent items are selected, and a list of consistent direction markers is constructed. This list includes the overlapping time point, the direction of thermal anomaly movement, the direction of voltage change, and the consistency judgment result.

[0068] S313: Based on the matching point information with consistent direction and overlapping time in the direction consistency mark list, integrate them according to the traction transformer number and time sequence to generate traction transformer fault overlap points;

[0069] Points with consistent direction and time overlap are extracted from the list of directional consistency markers. They are grouped according to the traction transformer number in the record, and each group of data is arranged in chronological order. All points that meet the conditions are integrated to form a set of fault overlap points of the traction transformer. Each fault overlap point has the traction transformer identifier, thermal anomaly coordinates, voltage change amplitude and occurrence time. After integration, it can form overlap points such as a certain transformer number having a heat source moving to the upper right and a voltage drop at 305 seconds, and another transformer number having a leftward movement and a voltage rise at 312 seconds. Finally, all overlap points are systematically generated and summarized in this structured recording method.

[0070] Please see Figure 5 The specific steps for obtaining the regional anomaly aggregation set are as follows:

[0071] S411: Call the time information and physical coordinates recorded in the fault overlap point of the traction transformer, extract the power supply section identifier of each fault point and divide it into groups, count the time distribution range and spatial location set of the fault in each group, and generate a section spatiotemporal distribution data table.

[0072] The system retrieves the time information and physical coordinates recorded in the traction transformer fault overlap points. First, based on the spatial coordinates of each fault point, its physical location in the track power supply system is mapped to the corresponding power supply section identifier. For example, if the coordinates of a fault point fall within the coordinate range of the 3rd power supply section, its power supply section identifier is set to 3. All fault points are categorized and grouped according to power supply sections, with each group representing the fault set of a power supply section. Then, time statistics are performed on the fault points within each group, extracting the earliest and latest fault occurrence times within the group to form the time distribution range of that power supply section. For example, if the fault start time of a group is 600 seconds and the end time is 660 seconds, then the time range of that group is 60 seconds. Further, the spatial coordinates of all fault points in the group are extracted, and these coordinates are unified as the spatial distribution set of the group, forming the spatiotemporal record of faults within each power supply section. Finally, all section numbers, corresponding time ranges, and spatial coordinate sets are integrated and output to construct a section spatiotemporal distribution data table.

[0073] S412: Cluster the coordinates and time of each fault point in the segment spatiotemporal distribution data table using the DBSCAN clustering algorithm to generate a spatiotemporal clustering boundary index group;

[0074] To perform spatiotemporal clustering analysis on traction transformer fault overlap points, the density-based spatial clustering algorithm DBSCAN was employed. The fault points contain three data dimensions: horizontal coordinates... (Unit: meters), Vertical coordinate (Unit: meters) and time information (Unit: seconds). Since the three dimensions have different physical dimensions, the three-dimensional data must first be normalized to unify them into a dimensionless standardized space before distance calculation and density determination can be performed.

[0075] For each fault point Its three-dimensional properties are the original vector. The vector is converted to a dimensionless standard vector using the following normalization formula. :

[0076] , , ;

[0077] in, : indicates the first The original horizontal and vertical coordinates and timestamps of each fault point; The minimum and maximum values ​​of the x-axis in the dataset; : Minimum and maximum values ​​of the ordinate; : Minimum and maximum time values; : These are the normalized horizontal coordinate, vertical coordinate, and time, respectively.

[0078] After normalization, any two fault points and The joint spatiotemporal distance between them is defined as follows:

[0079] ;

[0080] in, : No. The normalized three-dimensional vector of each fault point; : No. The normalized three-dimensional vector of each fault point; : Represents a point With point The joint distance in the standardized space; all difference units are unified to "dimensionless", therefore the distance value is also a unitless scalar, with a numerical range of... between.

[0081] There are three traction transformer fault points, and their original data are as follows: Point A: , , Point B: , , Point C: , , .

[0082] Calculate the global minimum and maximum values: , → Horizontal range: 90 meters , → Longitudinal range 100 meters, , →Time range: 40 seconds.

[0083] Normalize the three points separately:

[0084] Point A: , , ;

[0085] Point B: , , ;

[0086] Point C: , , ;

[0087] Calculate the joint distance between points:

[0088] Points A and B: ;

[0089] Points A and C: ;

[0090] Points B and C: ;

[0091] Set DBSCAN parameters:

[0092] Neighborhood radius (Unitless, applicable to normalized spaces), minimum number of neighbors ;

[0093] Clustering result: The distance between A and B is The two are neighbors; the distances between A and C, and between B and C, are both greater than 1. A and B do not constitute a neighborhood; both neighborhoods of A and B contain at least two points (including themselves), satisfying the condition that... Therefore, all of them are core points; C has no neighbors, does not meet the core point condition, and is not in any neighborhood, so it is determined to be an outlier.

[0094] Based on the above analysis, the DBSCAN algorithm outputs the following clustering boundary information:

[0095] Cluster 1: Contains two points A and B, indicating that their densities are similar in the normalized spatiotemporal space;

[0096] Noise point: Point C is considered an abnormal isolated point because its distance from other points is too large.

[0097] First, the lateral, longitudinal, and temporal data of each fault point are standardized to a unified dimensionless interval using a normalization formula. This resolves the inconsistency of physical dimensions in the original data and ensures comparability in subsequent calculations across spatial and temporal dimensions. Next, based on the normalized values, a standardized three-dimensional Euclidean distance formula is used to calculate the relative proximity between any two points in space and time. This calculation result indicates whether two fault points are sufficiently close in both spatial and temporal dimensions to potentially belong to the same density cluster. Each distance value is used to determine whether a point falls within a specified radius, thereby determining the number of neighbors for that point. This is then combined with a minimum neighbor threshold to identify core points. Finally, a clustering algorithm is used to determine which points can be connected to form clusters and which points are identified as isolated outliers due to insufficient neighborhood. The entire process, with progressive standardization, distance measurement, neighborhood determination, core point identification, and cluster assignment as its main computational threads, achieves efficient clustering and identification of traction transformer fault points in both time and space dimensions.

[0098] S413: Based on the cluster information divided in the spatiotemporal clustering boundary index group, extract the corresponding power supply section number and transformer quantity in the cluster, merge the cluster sets that simultaneously satisfy section concentration and fault type consistency, and generate regional anomaly aggregation set;

[0099] First, read the list of all fault points contained in each cluster, and extract the power supply section number corresponding to each fault point. This can be matched with a preset section mapping table by the coordinate range. For example, if the coordinates of all fault points in a cluster fall between 100 and 160 meters horizontally, and this coordinate range is mapped to the 3rd power supply section in the system, then the section number corresponding to this cluster is 3. Next, count the number of transformer numbers associated with the fault points in this cluster. If a cluster contains multiple fault point records, but some transformer numbers are duplicated, the duplicate numbers need to be removed, and only the total number of unique numbers is counted. For example, if a cluster has six fault points, corresponding to A01, A02, A01, A03, A03, and A04 respectively, then the count is... There are four transformers. After extracting these data, the clusters are judged for segment concentration, that is, whether all fault points in the cluster belong to the same power supply segment. This can be achieved by judging whether there are multiple different segment numbers. If only one number appears, it is considered concentrated; if multiple numbers appear, it is considered non-concentrated. Then, the clusters that meet the segment concentration criteria are judged for fault type consistency. The method is to extract the fault type field marked by each fault point in the cluster and perform classification statistics. If the proportion of a certain type reaches a preset consistency threshold, such as 90%, it is judged as consistent. If it does not reach the threshold, the cluster is excluded. Finally, the clusters that meet both conditions after the above screening are merged and uniformly grouped into the regional anomaly aggregation set.

[0100] Please see Figure 6The specific steps for obtaining the linked early warning area are as follows:

[0101] S511: Call the coordinate information and corresponding fault occurrence time of each abnormal traction transformer in the regional anomaly aggregation set, record the straight-line distance and time label sequence between all transformers, and generate a spatiotemporal relationship structure data group.

[0102] The algorithm retrieves the coordinates and fault occurrence time of each anomalous traction transformer from the regional anomaly aggregation set. First, it extracts the two-dimensional spatial coordinates and corresponding fault occurrence time of each traction transformer within the aggregation set. For example, if a transformer is numbered T001 with coordinates (128, 92) and a fault time of 620 seconds, this information is used as the input data point. After iterating through all anomalous transformers, a raw dataset containing the coordinates and times of all transformers is constructed. Then, the linear distance between any two transformers is calculated sequentially. The two-dimensional Euclidean distance formula is used to combine and calculate all point pairs, generating a complete distance matrix. Each row represents the spatial distance between a transformer and all other transformers. For example, if there are five abnormal transformers, a 5x5 symmetrical distance matrix is ​​formed. The main diagonal is zero, indicating that the distance to itself is zero. At the same time, the fault occurrence time of each transformer is recorded in chronological order, and the corresponding time values ​​are used to form a time tag sequence to ensure that the spatial distance and time point match one by one. Finally, the distance matrix and the time sequence are combined to construct a structured data group to represent the spatial relative positional relationship between abnormal transformers and the order of events in time, forming a spatiotemporal relationship structured data group.

[0103] S512: Based on the spatiotemporal relationship structure data set, set the clustering radius interval and time weight factor based on spatial distance, statistically analyze the distribution characteristics between fault points, input Ripley's K function to calculate the clustering degree of traction transformer fault overlap points in geographic space, and generate a set of spatial clustering statistical curves.

[0104] Based on the spatiotemporal relational structure data set, a quantitative analysis of the clustering characteristics of traction transformer fault points in geographic space needs to be conducted using spatial statistical methods, employing Ripley's K-function as the analytical tool. First, the analysis radius interval and time weighting factor need to be defined. The analysis radius represents the spatial threshold at which points are considered "nearest neighbors," while the time weighting factor is used to adjust for the impact of fault time sequence in spatial distance statistics, thus accurately reflecting spatiotemporal density. The closer the faulty transformers are to each other and the closer their fault times, the greater their weight and the higher their contribution to clustering in spatial statistics; conversely, if the time interval is large, the effect of the point pair needs to be attenuated.

[0105] The expression for Ripley's K function with time weighting is as follows:

[0106] ;

[0107] in, At the aggregation scale Under the condition of meters, the statistical value of spatial clustering between abnormal traction transformers, the larger the value, the stronger the clustering; The total area of ​​the observation area, for example, 40,000 square meters for a 200m × 200m track segment; The number of traction transformers included in the current area's anomaly aggregation set, for example, a total of 3 anomaly transformers were collected; : indicates the first The transformer and the first The straight-line Euclidean distance between the transformers, in meters, is calculated based on their horizontal and vertical coordinates. Distance threshold judgment function: when the distance between two transformers is less than or equal to a set radius. When the time condition is met, the value is 1; otherwise, it is 0. : No. With the The time weighting factor among the transformers is expressed as follows: , : respectively the first With the The time point at which the transformer failed is indicated in seconds. Time decay coefficient: controls the intensity of the effect of time difference on clustering, measured in reciprocals of a second; weighting factor values ​​range from 0 to 1, indicating that the closer the time between fault points, the greater the impact. A value closer to 1 indicates a higher contribution; a larger time difference results in a weight value closer to 0. Set the analysis radius. 10 meters, 20 meters, and 30 meters are used as multi-scale clustering observation ranges. Time decay coefficient. It can be set to 0.1, which means that when the time difference is 10 seconds, the weight drops to about 0.367. This value can be determined according to the characteristics of historical accidents. For example, if the linkage between traction transformers within 10 seconds is more common, it can be set to 0.1. If the response is faster, the value can be increased.

[0108] By using multiple different radii The values ​​are calculated to form a spatial clustering statistical curve. Each point on this curve represents the spatial clustering intensity of transformer fault points within the current clustering radius. Comparing this statistical curve with the theoretical uniform distribution curve will be used in subsequent steps to determine the degree of clustering deviation, thereby identifying spatial areas with abnormal clustering risk.

[0109] First, a complete spatiotemporal data structure is constructed based on the spatial coordinates and fault time of each abnormal traction transformer. Then, within multiple defined clustering radii, the straight-line distance and fault time difference between all transformers are analyzed pairwise. For each pair of transformers whose distance does not exceed the current clustering radius, the weight value corresponding to their time difference is calculated. The closer the times are, the closer the weight value is to 1; the larger the time difference, the exponentially the weight decays, approaching 0. All point pairs that meet the spatial proximity condition are weighted and counted according to their time difference to form a clustering intensity evaluation with spatiotemporal density as the weight. Next, all weighted count values ​​are normalized and combined with the area of ​​the entire observation area and the total number of transformers to calculate the clustering statistics under the current clustering radius. This process is repeated at multiple radius scales to form a complete spatial clustering statistics curve, which is used to characterize the overall spatial distribution characteristics of fault points.

[0110] The theoretical uniform distribution curve, used for comparison, is obtained by assuming that the same number of transformers are completely randomly and uniformly distributed within the same region. It is calculated by multiplying the transformer density by the neighborhood area corresponding to the current cluster radius. Under this uniformity assumption, the probability of adjacency between transformers exhibits a continuous and smooth distribution, with the statistical result increasing exponentially with the square of the radius. In other words, at each radius scale, the theoretically counted neighboring points increase with the area (i.e., the square of the radius), constructing a theoretical distribution curve that is regular and non-clustering. This theoretical curve serves as a baseline standard for actual clustering behavior, providing a reference for measuring the degree of deviation from actual clustering.

[0111] Ultimately, by comparing the calculated cluster statistics for each cluster radius with the theoretical values ​​for the corresponding radius, it is possible to determine whether the spatial clustering intensity at the current scale exceeds the normal random fluctuation range, thus identifying whether there are significant anomalous clustering regions in space.

[0112] S513: Determine the degree of deviation between the spatial clustering statistical curve group and the theoretical uniform distribution curve, filter spatial areas whose deviation exceeds the set clustering intensity judgment benchmark value, and output as linkage warning areas;

[0113] First, the actual clustering statistical curve and the theoretical uniform distribution curve are read to ensure that they are constructed based on the same clustering radius sequence. Then, the difference between corresponding values ​​is calculated for each clustering radius point. The difference is calculated by subtracting the theoretical curve value from the actual curve value to obtain the deviation at each clustering scale. For example, when the clustering radius is 20 meters, the actual statistical value is 8500, and the theoretical value is 7000, so the deviation is 1500. Next, a benchmark value for judging clustering intensity is set to define whether the deviation is significant. This benchmark value can be set as a percentage, such as 20%, indicating that when the actual value is significantly different from the theoretical value, the deviation is considered significant. A deviation of 20% or more above the theoretical value is considered significant. The deviation magnitude at each aggregation scale is converted into a percentage of the theoretical value. For example, the deviation of 1500 corresponds to a theoretical value of 7000, so the percentage is 21.4%, which is higher than the set threshold of 20%. Therefore, this scale is considered a significant deviation. Then, the spatial radius corresponding to all significant deviation aggregation scales is mapped back to the spatial area range, and the spatial coordinate coverage area is marked in combination with the fault point location. Finally, the spatial segments formed under all significant deviation aggregation scales are merged and these spatial areas are summarized and output as the linkage early warning area.

[0114] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A fault early warning method for urban rail transit power supply systems based on multi-source data fusion, characterized in that, Includes the following steps: S1: Collect continuous thermal images of the surface of the track traction transformer within a specified period, identify the continuous drift trajectory of the temperature centroid of the traction transformer surface, and mark the set of thermal anomaly points; S2: Obtain the time period of the concentrated recording of the thermal anomaly points, extract the voltage sampling sequence of the high-voltage side of the traction transformer within the same time period, and identify the voltage anomaly data group; S3: Analyze whether the time points when the thermal anomaly point set and the voltage anomaly data group produce anomalies within the same time period overlap, and determine whether the temperature centroid drift direction is consistent with the voltage change direction, and mark the traction transformer fault overlap point. S4: Summarize the overlapping fault points of the traction transformers in multiple power supply sections, analyze whether multiple transformers have the same sudden behavior within a specified time, and construct a regional anomaly aggregation set; S5: Using the coordinates and fault occurrence time of each abnormal traction transformer in the regional anomaly aggregation set as input, construct a spatial proximity structure, analyze the distribution of fault points in the structure using Ripley's K function, and mark the linkage early warning area.

2. The fault early warning method for urban rail transit power supply system based on multi-source data fusion according to claim 1, characterized in that, The specific steps for obtaining the thermal anomaly point set are as follows: S111: Collect continuous thermal images of the surface of the track traction transformer within a specified period, select the pixel group in the central region of each frame where the temperature exceeds the ambient temperature, determine the temperature distribution centroid of each frame based on the relationship between the temperature of the pixel and the image position, and generate a temperature centroid trajectory sequence. S112: Based on the temperature centroid trajectory sequence, track the change path of the centroid position between consecutive frames, identify the position where the direction changes in the path, locate the temporal position of the change according to the image frame number, extract the coordinate points and time information representing the change behavior, and generate a set of drift change positions. S113: Based on the coordinates and time data of each change point in the set of drift change locations, and combined with the boundary of the traction transformer surface monitoring area, invalid points outside the area are screened out to generate a set of thermal anomaly points.

3. The fault early warning method for urban rail transit power supply system based on multi-source data fusion according to claim 2, characterized in that, The specific steps for acquiring the voltage anomaly data group are as follows: S211: Call the time data corresponding to each anomaly point in the thermal anomaly point set, determine the start and end time range, and extract the voltage side sampling data sequence of the traction transformer voltage side within the same time period; S212: Based on the time distribution and voltage change of continuous data points in the voltage-side sampling data sequence, identify the starting position where the voltage exceeds the voltage change threshold, extract the change indicators of the fluctuation direction trend, duration and numerical amplitude corresponding to each change segment, and generate a voltage fluctuation feature set. S213: Filter the change segments of each change index in the voltage fluctuation feature set, classify them in chronological order, and generate voltage anomaly data groups.

4. The fault early warning method for urban rail transit power supply system based on multi-source data fusion according to claim 3, characterized in that, The specific steps for obtaining the fault overlap point of the traction transformer are as follows: S311: Call up all the time information recorded in the thermal anomaly point set and the voltage anomaly data group, establish corresponding time axis sequences respectively, and filter out the anomaly records that occur at the same time in the same time period by comparing the interval relationship between each time point in the two sequences, and generate an anomaly time overlap data table. S312: Based on the thermal anomaly point and abnormal voltage corresponding to each overlapping time point in the abnormal time overlap data table, extract the direction of continuous coordinate change and the direction of corresponding voltage data change, determine whether the change directions are the same, and filter the direction consistency marker list. S313: Based on the matching point information with consistent direction and overlapping time in the directional consistency mark list, integrate them according to the traction transformer number and time sequence to generate traction transformer fault overlap points.

5. The fault early warning method for urban rail transit power supply system based on multi-source data fusion according to claim 4, characterized in that, The specific steps for obtaining the regional anomaly aggregation set are as follows: S411: Call the time information and physical coordinates recorded in the fault overlap point of the traction transformer, extract the power supply section identifier of each fault point and divide it into groups, count the time distribution range and spatial location set of the fault occurrence in each group, and generate a section spatiotemporal distribution data table. S412: Cluster the coordinates and time of each fault point in the segment's spatiotemporal distribution data table using the DBSCAN clustering algorithm to generate a spatiotemporal clustering boundary index group; S413: Based on the cluster information divided in the spatiotemporal clustering boundary index group, extract the corresponding power supply section number and transformer number in the cluster, merge the cluster sets that simultaneously satisfy section concentration and fault type consistency, and generate a regional anomaly aggregation set.

6. The fault early warning method for urban rail transit power supply system based on multi-source data fusion according to claim 5, characterized in that, The specific steps for obtaining the coordinated early warning area are as follows: S511: Call the coordinate information and corresponding fault occurrence time of each abnormal traction transformer in the regional anomaly aggregation set, record the straight-line distance and time tag sequence between all transformers, and generate a spatiotemporal relationship structure data group. S512: Based on the spatiotemporal relationship structure data set, set the clustering radius interval and time weight factor based on spatial distance, statistically analyze the distribution characteristics between fault points, input Ripley's K function to calculate the clustering degree of traction transformer fault overlap points in geographic space, and generate a set of spatial clustering statistical curves. S513: Determine the degree of deviation between the spatial aggregation statistical curve group and the theoretical uniform distribution curve, filter spatial areas whose deviation exceeds the set aggregation intensity judgment benchmark value, and output them as linkage warning areas.

Images