Low-altitude unmanned aerial vehicle fault detection method and system for railway section

Abstract

This invention discloses a method and system for fault detection on railway sections using a low-altitude unmanned aerial vehicle (UAV). The invention relates to the technical field of fault detection methods. It identifies abnormal railway sections based on the location of multiple abnormal parameters and the environmental type of the detected section; determines the anomaly level of multiple abnormal sections based on a distribution map of the railway sections; and determines a fault detection path based on the anomaly level of multiple abnormal sections, the railway section's operating status map, and the location of the low-altitude UAV, thereby improving the accuracy of fault detection path detection. Therefore, it determines surface anomaly images and installation anomaly images based on the anomaly detection of abnormal sections; determines the corresponding fault content of the abnormal section based on the surface anomaly images, installation anomaly images, and past anomaly events of the abnormal section; and determines a real-time maintenance plan for each abnormal section based on its operating status and the fault content of multiple abnormal sections, thereby improving the accuracy of the real-time maintenance plan for each abnormal section.

Description

Technical Field

[0001] This invention relates to the technical field of fault detection methods, and more particularly to a method and system for detecting faults in railway sections using a low-altitude unmanned aerial vehicle (UAV). Background Technology

[0002] With the development of technology, railway sections, as part of the main railway line, are set up along the direction of the main railway line. Trains travel along the extension direction of the railway section and travel to different positions of the railway section at different times. Railway sections are generally inspected within a preset time period, such as once a month or once every six months. In the current technology, the inspection of railway sections is not real-time and online inspection of railway sections cannot be realized, which affects the accuracy of abnormal sections and makes it impossible to reasonably form a real-time maintenance plan for each abnormal section. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art. This invention provides a method and system for detecting faults in railway sections using low-altitude unmanned aerial vehicles (UAVs).

[0004] This invention provides a method for fault detection on railway sections using a low-altitude unmanned aerial vehicle (UAV), comprising: determining multiple sub-detection sections based on a railway section distribution map; determining a combination of section parameters for each detection section based on online detection of the multiple sub-detection sections; determining multiple abnormal parameters based on traversal of each section parameter combination; determining abnormal sections based on the location of the multiple abnormal parameters and the environmental type of the detection section; determining the abnormality level of the multiple abnormal sections based on the multiple abnormal sections and the railway section distribution map; determining a fault detection path based on the abnormality level of the multiple abnormal sections, the railway section's operating status map, and the location of the low-altitude UAV; the low-altitude UAV flying along the fault detection path to perform anomaly detection on each abnormal section; determining surface anomaly images and installation anomaly images based on the anomaly detection of the abnormal sections; determining the fault content corresponding to the abnormal section based on the surface anomaly images, installation anomaly images, and past anomaly events of the abnormal section; and determining a real-time maintenance plan for each abnormal section based on the operating status of each abnormal section and the fault content of the multiple abnormal sections.

[0005] This invention provides a fault detection system for railway sections using a low-altitude unmanned aerial vehicle (UAV). This system is applied to the aforementioned fault detection method for railway sections using a low-altitude UAV. The fault detection system for railway sections using a low-altitude UAV includes:

[0006] The segment parameter combination module is used to determine multiple sub-detection segments based on the distribution map of railway segments, and to determine the segment parameter combination of each detection segment based on the online detection of multiple sub-detection segments;

[0007] The abnormal road segment module is used to determine multiple abnormal parameters in each combination of road segment parameters based on the traversal of that combination of parameters, and to determine the abnormal road segment based on the location of the multiple abnormal parameters and the environmental type of the detected road segment.

[0008] The fault detection path module is used to determine the anomaly level of multiple abnormal road segments based on the distribution map of multiple abnormal road segments and railway segments, and to determine the fault detection path based on the anomaly level of multiple abnormal road segments, the driving status map of railway segments and the location of low-altitude UAVs.

[0009] Anomaly image module is used by low-altitude UAVs to fly along the fault detection path, perform anomaly detection on each abnormal section, and determine surface anomaly images and installation anomaly images based on the anomaly detection of the abnormal sections.

[0010] The real-time maintenance plan module is used to determine the fault content corresponding to the abnormal road segment based on surface abnormality images, installation abnormality images, and past abnormal events of the abnormal road segment, and to determine the real-time maintenance plan for each abnormal road segment based on the operating status of each abnormal road segment and the fault content of multiple abnormal road segments.

[0011] Compared with the prior art, the beneficial effects of the present invention are:

[0012] In this embodiment of the invention, the method is used to determine the anomaly level of multiple abnormal road segments based on the distribution map of multiple abnormal road segments and railway segments. The fault detection path is determined based on the anomaly level of multiple abnormal road segments, the driving status map of railway segments, and the position of low-altitude UAVs. Multiple anomaly parameters are introduced, and abnormal road segments are controlled. This method takes into account the anomaly level of multiple abnormal road segments, the driving status map of railway segments, and the position of low-altitude UAVs, thereby improving the detection accuracy of the fault detection path.

[0013] Therefore, the low-altitude UAV flies along the fault detection path to detect anomalies in various road sections. Based on the anomaly detection, it determines surface anomaly images and installation anomaly images. Based on the surface anomaly images, installation anomaly images, and past anomaly events of the anomaly road section, it determines the corresponding fault content of the anomaly road section. Based on the operational status of each anomaly road section and the fault content of multiple anomaly road sections, it determines the real-time maintenance plan for each anomaly road section. The introduction of surface anomaly images and installation anomaly images ensures multi-dimensional anomaly detection of anomaly road sections by the UAV, realizes the overall consideration of the operational status of each anomaly road section and the fault content of multiple anomaly road sections, and improves the accuracy of the real-time maintenance plan for each anomaly road section. Attached Figure Description

[0014] Figure 1This is a flowchart illustrating the fault detection method for railway sections using a low-altitude unmanned aerial vehicle (UAV) according to an embodiment of the present invention.

[0015] Figure 2 This is a schematic diagram of the structural composition of a low-altitude unmanned aerial vehicle (UAV) fault detection system for railway sections in an embodiment of the present invention. Detailed Implementation

[0016] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0017] Please see Figure 1 and Figure 2 A method for fault detection on railway sections using low-altitude unmanned aerial vehicles (UAVs), applied to scenarios involving fault detection on railway sections using low-altitude UAVs; the method includes:

[0018] Step S11: Determine multiple sub-detection sections based on the distribution map of railway sections, and determine the combination of section parameters for each detection section based on the online detection of multiple sub-detection sections;

[0019] Step S12: In each combination of road segment parameters, determine multiple abnormal parameters based on the traversal of the road segment parameter combination, and determine the abnormal road segment based on the location of the multiple abnormal parameters and the environmental type of the detected road segment.

[0020] Step S13: Determine the anomaly level of multiple abnormal road sections based on the distribution map of multiple abnormal road sections and railway sections, and determine the fault detection path according to the anomaly level of multiple abnormal road sections, the driving status map of railway sections and the position of low-altitude UAV.

[0021] Step S14: The low-altitude UAV flies along the fault detection path, performs anomaly detection on each abnormal section, and determines the surface anomaly image and installation anomaly image based on the anomaly detection of the abnormal section.

[0022] Step S15: Based on surface anomaly images, installation anomaly images, and past anomaly events of the anomaly road segment, determine the fault content corresponding to the anomaly road segment, and determine the real-time maintenance plan for each anomaly road segment according to the operating status of each anomaly road segment and the fault content of multiple anomaly road segments.

[0023] In step S11, the specific steps are as follows:

[0024] S111: Collect the names of railway sections, determine the distribution map of railway sections based on the names of railway sections and railway database, trigger the initial detection of low-altitude UAVs based on the distribution map of railway sections, and collect multiple section boundary points of railway sections based on the initial detection of UAVs.

[0025] S112: Based on the distribution map of multiple road segment boundaries and railway segments, determine multiple sub-detection segments, and perform online detection on each sub-detection segment;

[0026] S113: Collect multiple road segment parameters based on online detection of each sub-detection road segment; mark the corresponding road segment detection positions for the multiple road segment parameters; and determine the combination of road segment parameters for each detection road segment based on the multiple road segment parameters and the corresponding road segment detection positions.

[0027] In the embodiments of this application, the collected railway section name is used to initiate a query request to the railway database. The database returns all key geographic and structural information related to the name. The extracted geographic coordinates, line structure and other information are visualized in geographic information system (GIS) software or UAV mission planning software to generate a "distribution map" containing the line centerline, key structure locations, curve radii, slope changes, etc. This distribution map is digital and contains precise spatial coordinates. The output is a digital geospatial dataset, which usually exists in the form of vector data (such as line features, point features) or raster data containing coordinate information.

[0028] The mission planning software plans the initial detection flight path of the UAV. The initial detection flight path covers the entire target railway section and takes into account the UAV's flight performance (such as speed, endurance, turning radius), sensor field of view, overlap ratio requirements (for subsequent image stitching or 3D modeling), and obstacle avoidance (avoiding known tall obstacles). The planned flight path, flight altitude, flight speed, sensor working mode (such as camera photography, LiDAR scanning) and other parameters are converted into flight mission commands that the UAV can understand.

[0029] The mission instructions are uploaded to the UAV through the ground control station (GCS) software, and the mission is initiated; or, through a preset scheduling system, the UAV is automatically triggered to take off and execute the mission at a specified time. The main purpose of this "initial inspection" is to conduct a rapid and comprehensive data collection of the entire target railway section, with a focus on identifying and accurately locating the "section boundary points" on the line; "section boundary points" usually refer to locations with clear physical or logical boundaries.

[0030] The drone performs the initial inspection according to the initial inspection flight path; the onboard sensors (mainly cameras, and also LiDAR) continuously collect data along the railway line, generating a series of images, point clouds, or other sensor data. At the same time, using computer vision technology, the collected images are analyzed to automatically identify features such as mileage markers, boundary markers, curve start and end points, bridge / tunnel entrances / exits, etc., and record their coordinates; the precise geographic coordinates (latitude, longitude, and elevation) of each identified section boundary point and its corresponding boundary type (such as "mileage K1050 mark", "straight-to-curve transition point", "bridge start point") are recorded to form a boundary point list; the output is a dataset containing the coordinates and types of all identified section boundary points, usually in the form of GIS point elements or coordinate lists; the system realizes the complete process from specifying a target railway line, to obtaining its precise geographic information, to planning and executing the initial drone inspection, and finally accurately collecting and recording key physical or logical boundary points (section boundary points) along the line.

[0031] Furthermore, multiple segment boundary points and railway segment distribution maps are introduced. Segment boundary points define key locations on the railway line, while distribution maps provide the overall geographical background and other reference information (such as the locations of bridges, tunnels, and stations). The system divides railway lines based on these boundary points. The most direct method is to define the railway line segment between two adjacent segment boundary points as a "sub-detection segment," which is similar to dividing a line into several segments using marker points.

[0032] First, ensure that these boundary points are sorted according to the actual order on the railway line (e.g., from the starting point to the end point, increasing or decreasing by mileage). Then, the system pairs the sorted adjacent boundary points. For example, if the boundary point list is [P1,P2,P3,P4], the resulting sub-detection segments are P1-P2, P2-P3, and P3-P4.

[0033] For each newly generated sub-detection segment, the system assigns it attributes, which typically include: a unique identifier, such as "sub-segment_001", "sub-segment_002", etc.; a geographical range, recording the precise coordinates of the start and end points of the sub-segment; a length, calculating the length of the sub-segment (e.g., the curved distance along the railway line); associated boundary point information, recording which boundary point the sub-segment starts at and ends at (e.g., sub-segment_001 starts at "mileage K1050 mark" and ends at "bridge entrance (XX bridge)"); and environmental information, if the distribution map contains relevant information, initially marking the environmental type of the sub-segment, such as "bridge segment", "tunnel segment", "straight line segment", "curved segment", etc. The system outputs a list of "sub-detection segments", each containing its definition, range, and attribute information.

[0034] A list of "multiple sub-detection segments" is introduced; real-time, data-driven detection is performed on each sub-detection segment to collect its status information. Here, "online detection" emphasizes real-time data acquisition and analysis during the drone's flight. For each sub-detection segment, the drone plans a detailed flight path to ensure coverage of all key areas of the segment. This includes flying along the track centerline, hovering or circling around key equipment (such as signal lights and overhead contact line supports), and taking photos of bridges or tunnels from specific angles. The path planning takes into account the drone's sensor field of view and image overlap requirements. The drone flies along the planned path and activates its onboard sensors in real time to collect data.

[0035] The sensors include: a high-definition visible light camera: used to capture images of tracks, ballast, overhead contact lines, signaling equipment, etc., for visual inspection and surface defect identification; an infrared thermal imager: used to detect abnormal heating of components such as tracks, overhead contact lines, and insulators; a lidar (LiDAR) system: used to acquire high-precision three-dimensional point cloud data for measuring track geometry (gauge, elevation, orientation, level), detecting ballast settlement, and measuring overhead contact line geometric parameters; and a multispectral / hyperspectral camera (optional): used to detect material aging, corrosion, etc.

[0036] All collected data (images, point clouds, heat maps, etc.) are precisely labeled with their corresponding geographical location (usually through the GPS / INS system on the drone) and timestamp, and associated with the "sub-detection segment" currently being detected. This ensures that subsequent analysis can accurately determine which segment and location each data point belongs to. For each sub-detection segment, all raw data (images, point clouds, heat maps, etc.) collected during this online detection process are output, along with related metadata (such as GPS coordinates, timestamps, sensor parameters, etc.).

[0037] Therefore, various sub-detection segments were introduced. The raw data collected by online detection in each sub-detection segment (e.g., sub-segment 001, 002, 003, etc.) includes: high-resolution visible light images, infrared thermal imaging images, LiDAR point cloud data, laser ranging data, millimeter-wave radar data, IMU (inertial measurement unit) data, GPS / RTK positioning data, etc.

[0038] "Segment parameters" are not the raw data itself, but rather feature values ​​or indicators extracted from the raw data that reflect the current state of the sub-segment. These parameters should be representative and able to quantify the health status or potential risks of the segment. Optionally, this includes extracting track clarity, presence of foreign objects (such as gravel or debris), sleeper condition (damage or dirt), and ballast accumulation from visible light images; extracting temperature distribution of the track and sleepers from infrared images to indicate abnormal current or component overheating; extracting track geometric parameters such as gauge, track elevation, direction (horizontal displacement), and gauge change rate; detecting sleeper positional deviations, ballast geometry, and accumulation height; analyzing vibration frequency and amplitude if speed sensor data is available; and accurately calculating the distance of the detection point relative to the starting point of the sub-segment by combining GPS / RTK data.

[0039] Multiple road segment parameters were introduced, and precise location information recorded by low-altitude UAVs during the inspection was collected. Each road segment parameter was collected at one or more specific locations during the inspection process. In order to accurately locate the problem later, the parameter must be associated with its collection location. For example, the track gauge deviation at a specific point is +2mm, rather than the average deviation of the entire sub-segment.

[0040] In the data record or database, add a location marker field for each road segment parameter. This location is: precise GPS coordinates (longitude, latitude, elevation); distance relative to the starting point of the sub-segment (e.g., 75 meters from point P1); combined with mileage marker information; for continuously changing parameters (such as track geometry), record the parameter values ​​of a series of sampling points and their corresponding locations.

[0041] Multiple road segment parameters and corresponding road segment detection locations are introduced. "Road segment parameter combination" refers to the aggregation of all parameters (and their location information) belonging to the same sub-detection road segment to form a comprehensive description of the sub-road segment's status. This combination is regarded as a "digital profile" or "health record" of the sub-road segment. The system organizes all parameters and their location information belonging to the same sub-detection road segment (e.g., sub-road segment 001), which is usually achieved through database queries or data structures (such as dictionaries or lists). Finally, each sub-detection road segment corresponds to a "parameter combination" record containing all its parameters and location information. The massive amount of raw data obtained from UAV online detection is transformed into structured "road segment parameters" with precise location information, and these parameters are organized by sub-detection road segments to form "road segment parameter combinations." This combination not only contains the status information of the sub-road segment but also retains spatial positioning capabilities.

[0042] In step S12, the specific steps are as follows:

[0043] S121: Real-time monitoring of parameter combinations for each road segment, traversal of parameter combinations for each road segment, determination of multiple parameters for road segments to be detected based on the traversal of parameter combinations for each road segment, and determination of multiple associated parameters based on the mapping relationship between multiple parameters for road segments to be detected and associated parameters.

[0044] S122: If a road segment parameter exceeds a preset safety parameter threshold among multiple road segment parameters to be detected, the ratio calculation between the road segment parameter and multiple associated parameters is triggered to determine the corresponding matching coefficient. If the matching coefficient exceeds the preset matching coefficient threshold, the road segment parameter to be detected is treated as an abnormal parameter to collect multiple abnormal parameters.

[0045] S123: Mark the location corresponding to the abnormal parameter, determine the corresponding detection section based on the matching of the location corresponding to the abnormal parameter and the distribution map of the railway section, determine the environmental type of the detection section based on the environmental detection of the detection section, and determine the abnormal section based on the values, locations and environmental types of multiple abnormal parameters.

[0046] In the embodiments of this application, data is transmitted back to the ground station or cloud platform via wired network, wireless network (such as 4G / 5G), or satellite communication; monitoring is continuous and polled at set time intervals (e.g., every 5 minutes, every 10 minutes); it monitors not only whether the parameter combination itself arrives on time, but also its integrity, whether the data format is correct, and whether there are obvious anomalies in key parameters (e.g., whether the values ​​are within the physical range, such as whether the temperature suddenly changes to -100℃ or 5000℃); real-time monitoring is a prerequisite for subsequent traversal and analysis; once the data arrives and is confirmed to be valid, the traversal process is triggered.

[0047] For each "road segment parameter combination" that has just been monitored in real time and confirmed to be valid, examine all the "road segment parameters" contained within that combination. These parameters include quantitative data (such as numerical values ​​and percentages) and qualitative data (such as status descriptions and image features). This is not just a matter of viewing, but also a preparation for the next step of screening "road segment parameters to be detected". The traversal process identifies which parameters are of key concern and which parameter changes indicate potential problems. This is usually implemented through programming, using loop structures (such as for loops) to access the parameter list or database records one by one.

[0048] Not all parameters encountered will become "parameters to be detected"; they are selected according to preset rules or strategies. Common criteria include: Importance: Some parameters (such as conductor height, pull-out value, insulator status) are more important than other parameters (such as image background brightness) and are more likely to cause faults, so they are given priority as parameters to be detected; Variability: Parameters with large value changes or those in a dynamic process (such as instantaneous changes in parameters caused by vibration when a train passes) are given attention; Historical data comparison: Parameters with large deviations compared to the historical normal values ​​or recent average values ​​of the road segment; Parameter type: Only quantitative parameters are selected, or those parameters that are directly used for subsequent proportional calculations are selected; From a combination of road segment parameters, a portion of those considered to require further in-depth analysis and comparison are selected as "parameters to be detected", and their specific values ​​and location information are retained.

[0049] The associated parameter mapping relationship is a pre-established knowledge base or rule base that defines which "road segment parameters to be detected" are associated with or influence each other with which other "road segment parameters" (i.e., "associated parameters"). This association is physical (e.g., changes in conductor height affect the pull-out value) and also logical (e.g., poor insulator condition affects the clarity of that area in the image).

[0050] For each selected "parameter of the road segment to be detected", the system queries the mapping relationship to find other related parameters. These related parameters are not "to be detected" themselves, but they are crucial for assessing the degree of abnormality or authenticity of the parameter to be detected. They provide the calculation objects for the proportional calculation in the next step S122, so that the anomaly judgment is not based on the absolute value of a single parameter, but considers the relative relationship between parameters, thereby improving the accuracy and robustness of the judgment.

[0051] Furthermore, the system compares the current value of each parameter of the track segment to be tested with its corresponding preset safety parameter threshold. This threshold is set based on railway safety regulations, equipment design standards, and historical experience data. Threshold setting examples: Contact line pull-out value: Standard requirements are usually within ±150mm; assuming a preset safety threshold of ±200mm (relatively lenient, used for initial screening); Track gauge in track geometry: Standard requirements are usually 1435mm ±1mm; assuming a preset safety threshold of 1435mm ±2mm; Insulator surface temperature: Under ambient temperature, if the insulator temperature is significantly higher than the surrounding environment or adjacent insulators, it indicates an anomaly; assuming a preset safety threshold of 15℃ higher than the ambient temperature; Result judgment: For each parameter to be tested, the system will obtain a Boolean result (yes / no); if the parameter value exceeds the threshold range, it is marked "yes," indicating that the parameter has a problem and requires further analysis; otherwise, it is marked "no," indicating that the parameter is within the safe range.

[0052] The system will only perform a ratio calculation between the parameter to be detected and multiple associated parameters when a certain parameter to be detected is determined to "exceed the threshold". For the parameter to be detected that exceeds the threshold, the system will retrieve the multiple associated parameters found for it in step S121. The "ratio calculation" here needs to be specifically defined and is not necessarily a simple division. Depending on the application scenario, it can take the following forms: Simple ratio method: Parameter value to be detected / Standard value of associated parameters or Parameter value to be detected / Current value of associated parameters. For example, when the pull-out value exceeds the threshold, it is related to the tilt of the support post. The calculation is the amount of pull-out value exceeding the threshold / tilt of the support post.

[0053] By using proportional calculations, we attempt to quantify the "fit" or "amplification effect" between the parameter being tested exceeding the standard and its related factors. For example, if the pull-out value exceeds the standard by a large margin, but the pillar tilt is also very serious, it means that the problem is more serious or more complex. Conversely, if the pull-out value exceeds the standard, but the pillar is straight, it is just a localized minor problem.

[0054] The matching coefficient is a comprehensive index calculated based on the results of proportional calculations. It is used to measure the "degree of matching" or "strength of association" between the exceeding of a parameter and its associated parameters. This coefficient aims to determine whether the exceeding phenomenon is "reasonably" caused by the associated factors or whether it is abnormally severe. The calculation method of the matching coefficient needs to be predefined. It is: simple average or weighted average: averaging multiple proportional calculation results for a parameter to be detected; assigning different weights to different associated parameters; maximum value method: taking the maximum value among all proportional calculation results as the matching coefficient; composite function method: using more complex functions (such as exponential or logarithmic) to combine with the proportional calculation results to calculate the matching coefficient; threshold mapping method: mapping the proportional calculation results to a fixed matching coefficient value according to the different intervals they fall into. The higher the matching coefficient, the stronger the relationship between the exceeding of the parameter and its associated factors, or the more "prominent" the exceeding phenomenon, and the more likely it is to be suspected as a true anomaly.

[0055] The preset matching coefficient threshold is an empirical value or a value derived from historical data analysis. It represents the threshold at which the system considers a match to be "sufficiently suspicious." If the value is below this threshold, even if the parameter exceeds the limit, it is considered a false alarm or a slight deviation. If the value is above this threshold, it is more indicative of a real anomaly.

[0056] The system compares the calculated matching coefficient with a preset matching coefficient threshold. If the matching coefficient > the preset matching coefficient threshold, the parameter to be detected is determined to be an "abnormal parameter." The system records this and triggers a data acquisition or storage mechanism for subsequent analysis and processing. If the matching coefficient ≤ the preset matching coefficient threshold, the parameter to be detected is determined to be out of range, but its relationship with related parameters is insufficient to indicate a serious abnormality, and it is temporarily not marked as an abnormal parameter. In this case, for catenary parameters (such as pull-out values), a higher threshold is set because slight out-of-range values ​​have little impact, but serious out-of-range values ​​accompanied by related problems such as support tilting must be monitored. The threshold is assumed to be 15. For track geometry parameters (such as track gauge), a lower threshold is set because even small deviations affect the smoothness of train operation. The threshold is assumed to be 3.

[0057] Therefore, the location corresponding to the abnormal parameter is marked, and the corresponding detection section is determined by matching the location of the abnormal parameter with the distribution map of railway sections. The environmental type of the detection section is determined based on the environmental detection of the detection section. The abnormal section is determined based on the values, locations and environmental types of multiple abnormal parameters. This overall consideration of the values, locations and environmental types of multiple abnormal parameters ensures the accuracy of the abnormal section.

[0058] At this point, in step S122, we have identified some abnormal parameters; each abnormal parameter was collected online within a specific sub-detection section, and its original data (such as images taken by drones or point clouds scanned by lasers) has precise geographical coordinates or mileage markers relative to the railway starting point; these abnormal parameters are precisely associated with their original collection locations and clearly marked.

[0059] The data includes the UAV's GPS / RTK positioning data, inertial navigation system (INS) data, odometer data, and registration results with image / point cloud data; the precise latitude and longitude coordinates and / or mileage markers (e.g., K123+450, indicating a distance of 123 km 450 meters from the railway starting point) are recorded in the database for each anomalous parameter; and the data is also directly visualized on the digitized railway distribution map, such as placing a marker point or highlighting the anomalous area.

[0060] Although we have divided the railway into sub-detection sections in S112, the location of an abnormal parameter may be near the boundary between two sub-detection sections. Alternatively, we may need to associate multiple scattered abnormal parameters with a logically more relevant larger area (i.e., detection section) for unified analysis. This step involves using the precise location of the abnormal parameters and the railway distribution map (which includes the boundary information of the sub-detection sections) to determine which detection sections these abnormal parameters mainly belong to.

[0061] At this point, the location (mileage marker or coordinates) of the abnormal parameter is compared with the boundary of the sub-detection segment. Typically, an abnormal parameter is assigned to the sub-detection segment where its location is located. If multiple abnormal parameters are located close to each other, they are grouped into the same detection segment for analysis. Scattered abnormal points are aggregated into logically related segment units, which facilitates subsequent environmental analysis and comprehensive judgment.

[0062] The environment of a railway section has a significant impact on its condition and potential faults; for example, higher humidity in tunnels leads to equipment corrosion; vibration-induced loosening is more likely to occur on bridges; and rockfalls or landslides are more likely to occur in mountainous sections. This step uses environmental data collected previously (in the online inspection of S112) or combines it with geographic information system (GIS) data to determine the type of environment in which the current inspection section is located.

[0063] The drone is equipped with environmental sensors (temperature, humidity, wind speed, air pressure, etc.), image analysis (identifying vegetation cover, soil type, buildings, etc.), lidar data (identifying terrain and obstacles), and pre-loaded GIS data (such as map-marked tunnels, bridges, rivers, geological zoning, etc.); several typical railway environment types are predefined, such as: open plains, hilly areas, mountainous areas, inside tunnels, on bridges, urban areas, near rivers / lakes, etc.; by comprehensively analyzing the environmental monitoring data and GIS data, the monitored railway section is classified into one or more of the most suitable environmental types.

[0064] Optionally, assuming that during the online detection of S112, the drone conducted an environmental scan of the road section: image analysis shows that most of the road section is a bare soil roadbed with a small number of shrubs on both sides; lidar data indicates that the terrain of the road section is flat and there are no tall obstacles; GIS data marks the area as a plain; based on this information, the system determines that the environmental type of the detected road section 003 is: open plain.

[0065] If the road section being inspected is from K100+500 to K101+000, and this section includes a long tunnel: image and lidar data will show the tunnel entrance and internal features; environmental sensors will measure high humidity and low temperature inside the tunnel; GIS data will clearly mark the tunnel location; the system will determine the environmental type of the road section being inspected as: inside the tunnel (or a composite environment including the tunnel entrance).

[0066] The system now has comprehensive information about the anomalies: the specific values ​​(severity) of the anomaly parameters, their distribution on the road segment, and the environmental context of the road segment. Based on this information, the system needs to determine whether these anomalies are concentrated or severe enough to mark the entire detected road segment (or a part thereof) as an "anomaly segment" that requires special attention.

[0067] The study introduces the concepts of anomaly concentration, anomaly severity, and environmental impact. Regarding anomaly concentration, it considers whether multiple anomalous parameters are concentrated within a small area of ​​the detected road segment or dispersed throughout the entire segment. Regarding anomaly severity, it determines whether the deviation of anomalous parameter values ​​from safety thresholds is slight or severe. Regarding environmental impact, it considers whether the current environmental conditions exacerbate these anomalies. Pre-defined rules are needed to assist in the judgment. For example, "If more than three different types of parameter anomalies are found within the same detected road segment, or if any key parameter (such as track gauge or catenary height) is severely exceeded, then the detected road segment is marked as an anomalous road segment," or "If image features indicating a risk of rockfall are found in a mountainous road segment, even if only one is present, it should be marked as an anomalous road segment." One or more "abnormal road segments" are identified, and their mileage range, the main anomalous parameter information they contain, and the environmental type are recorded.

[0068] Optional, inspection section: sub-inspection section 003 (K100+000 to K100+500); abnormal parameters: A: catenary pull-out value 180mm (location K100+320); B: track gauge 1437mm (location K100+380); environment type: open plain.

[0069] The system makes the following judgments: Number of abnormal parameters: 2; Severity of abnormality: The pull-out values ​​exceed the standard by a large margin (180mm > 150mm standard, and also > 200mm threshold setting is too conservative), while the track gauge exceeds the standard by a small margin (1437mm > 1435mm standard, but did not trigger the S122 abnormality marker because the matching coefficient is low); Abnormality concentration: K100+320 and K100+380 are 60 meters apart, which is not very concentrated within the 500-meter sub-detection section, but not too dispersed either; Environmental impact: The open plain environment has little impact on this type of geometric parameter abnormality; Preset rule: The assumed rule is "If two or more parameters marked as abnormal by S122 are found in the same sub-detection section, then the sub-detection section is marked as an abnormal section".

[0070] In step S13, the specific steps are as follows:

[0071] S131: Collect multiple abnormal road segments, and determine the length of each abnormal road segment by detecting the length of each abnormal road segment. Determine the first abnormal parameter based on the location and length of the multiple abnormal road segments.

[0072] S132: Determine the second anomaly parameter based on the location of multiple abnormal road segments and the distribution map of railway segments; determine the anomaly level of multiple abnormal road segments based on the mapping relationship between the first anomaly parameter, the second anomaly parameter and the anomaly level;

[0073] S133: Determine the running status map of the railway segment based on the name of the railway segment, the railway database, and the current time, and mark the running status of each abnormal segment. At the same time, collect the position of the low-altitude UAV. Based on the abnormality level, running status, and position of the low-altitude UAV of multiple abnormal segments, determine the fault detection path. At this time, the fault detection path presents the sub-low-altitude flight paths of multiple abnormal segments, and determine the flight sequence of each sub-low-altitude flight path based on the sub-low-altitude flight paths and the running status of multiple abnormal segments.

[0074] In the embodiments of this application, information on all abnormal road segments that have been identified and marked in step S12 (especially S123) is obtained; the system needs to extract the identification information of each abnormal road segment from the previously stored or generated abnormal list, which typically includes the name, unique ID, and precise geographical coordinate range (starting and ending coordinates) of the abnormal road segment on the railway segment distribution map. This step ensures that we have a clear list of all objects that need further analysis.

[0075] For each identified abnormal road segment, its physical length needs to be accurately calculated. This length is determined in the following ways: Method 1 (based on coordinates): Using the geographic coordinates of the starting and ending points obtained in step S131, geospatial calculation methods (such as the Haversine formula or distance calculation in a more precise projected coordinate system) are used to calculate the straight-line distance between the two points or the curved distance along the railway line (if railway line data is available); for complex curved road segments, segmented calculations are required and then summed.

[0076] By combining the location and length information of the abnormal road segment, one or more "first anomaly parameters" are generated for subsequent analysis. The purpose of this parameter is to quantify a certain characteristic of the anomaly, such as its scope of influence, severity, or its positional importance in the railway network. The specific method for determining the "first anomaly parameter" depends on the design objectives. The following are some definition methods:

[0077] Definition Method 1: Abnormal Impact Range Index; directly use the road segment length itself as the first abnormal parameter; the longer the length, the greater the potential impact; first abnormal parameter = abnormal road segment length;

[0078] Definition Method 2: Location-weighted length; If the anomalies in certain areas (such as near stations, tunnel entrances, and bridge sections) are more critical than those in other areas, different weight coefficients are assigned to the length of road segments at different locations; for example, the length of road segments near stations is *1.5, and the length of ordinary road segments is *1.0; First anomaly parameter = anomaly road segment length * location weight coefficient.

[0079] The system first identifies the objects to be processed (collecting abnormal road sections); then, it accurately measures the physical impact range of each abnormality (calculating the road section length); finally, it combines the basic information of length with other factors such as location to generate the "first abnormality parameter," which provides an important quantitative basis for subsequent assessment of the severity and priority of abnormalities and for planning detection resources.

[0080] Furthermore, a second anomaly parameter is determined based on the location of multiple abnormal road segments and the distribution map of railway segments; the anomaly level of multiple abnormal road segments is determined based on the mapping relationship between the first anomaly parameter, the second anomaly parameter, and the anomaly level, which takes into account the overall consideration of the mapping relationship between the first anomaly parameter, the second anomaly parameter, and the anomaly level, and ensures the accuracy of the anomaly level of multiple abnormal road segments.

[0081] At this point, by utilizing the spatial location information of the anomalous section within the railway network, combined with an understanding of the entire railway distribution map (which typically includes information on important geographical features such as line grade, station locations, hubs, bridges, and tunnels), the strategic importance or potential impact range of the anomalous section can be assessed. The second anomaly parameter aims to quantify this importance or impact. It no longer focuses solely on the physical attributes of the anomaly itself (such as length, as reflected by the first anomaly parameter), but rather on the additional risks or value brought about by its location within the entire railway system.

[0082] The system analyzes the geographical location of each abnormal road segment to see if it is close to or crosses important nodes (such as major stations, marshalling yards, and hubs), critical engineering structures (such as long bridges, tunnels, and high embankment roadbeds), or is located in a busy traffic area. It also considers the line's classification (trunk line, branch line) and whether it is close to residential areas or important facilities. Based on these analysis results, a second abnormal parameter value is defined for each abnormal road segment. Definition method:

[0083] Based on proximity to important facilities: Define a score based on the distance of the abnormal section from the nearest station, hub, bridge, etc. (the closer the distance, the higher the score); Based on line class and location: Define a score based on the line class of the abnormal section (main lines score higher) and whether it is located in the middle (greater impact) or at both ends of the section; Based on potential impact range: Define a score considering the extent of the impact on traffic if the section fails (e.g., whether it will cause the entire section to be shut down).

[0084] By combining the two quantitative indicators (the first anomaly parameter and the second anomaly parameter) obtained from the previous calculation, and through a preset anomaly level mapping relationship (usually a rule base, scoring model, or decision tree), a clear anomaly level (e.g., level 1, level 2, level 3, or high, medium, low) is finally determined for each abnormal road segment. This level represents the urgency, severity, and priority of the anomaly.

[0085] The anomaly level mapping relationship is predefined, specifying which anomaly level should correspond to different combinations of the first and second anomaly parameters. For example, the rule states: "If the first anomaly parameter is >1000 meters and the second anomaly parameter is >8 minutes, it is judged as a Level 1 anomaly; if the first anomaly parameter is >800 meters and the second anomaly parameter is >5 minutes, it is judged as a Level 2 anomaly; otherwise, it is a Level 3 anomaly." This mapping relationship is adjusted according to actual needs and security strategies. By comprehensively assessing the physical scale and strategic location of the anomaly, it is classified into levels, providing a clear basis for subsequent resource allocation (such as dispatching drones and arranging maintenance personnel) and decision-making (such as whether to immediately block the line).

[0086] Therefore, based on the railway segment name, railway database, and current time, the operating status map of the railway segment is determined, and the operating status of each abnormal segment is marked. Simultaneously, the location of low-altitude UAVs is collected. Based on the anomaly level, operating status, and UAV location of multiple abnormal segments, a fault detection path is determined. This fault detection path presents multiple sub-low-altitude flight paths for multiple abnormal segments. The flight sequence of each sub-low-altitude flight path is determined based on the sub-low-altitude flight paths and operating status of multiple abnormal segments. This comprehensive consideration of the sub-low-altitude flight paths and operating status of multiple abnormal segments ensures the accuracy of the flight sequence. Furthermore, multiple anomaly parameters are introduced, and abnormal segments are controlled. This comprehensive consideration of the anomaly level, railway segment operating status map, and UAV location improves the detection accuracy of the fault detection path.

[0087] At this point, the system needs to obtain the current actual operating status of the railway. It needs to know which sections of the railway line are idle (to be inspected) and which sections have trains passing through or about to pass through (not to be inspected) at the planned time of drone inspection. This requires querying the railway database, which contains information such as train timetables, current train locations, and line occupancy. Combined with the current precise time, a dynamic "travel status diagram" is generated. This diagram is a data structure that marks the status of each section at the time of the query. Simultaneously, the system inputs: a list of railway section names and the current time (accurate to the second); the system queries: the railway operation database (containing train timetables, real-time tracking data, dispatch instructions, etc.); the system outputs: a "travel status diagram," such as a dictionary or database table, in the format {section ID: status, ...}, where the status is "idle," "occupied (train number XX, estimated passing time YY:ZZ)," "under maintenance (not allowed to pass)," etc.

[0088] The system matches the "Driving Status Map" generated in the previous step with the previously identified list of "Abnormal Road Segments," labeling each abnormal road segment as either currently detectable or undetectable. This directly affects which abnormal road segments are visited within the current time window when planning the flight path. The system inputs are: a list of abnormal road segments (including segment ID / location information) and the "Driving Status Map" generated in the previous step. The system processes: iterates through each abnormal road segment, searching for its corresponding status in the "Driving Status Map" based on its segment ID / location information. The system outputs: updates the abnormal road segment information, adding a "Current Driving Status" field.

[0089] The drone's current location is the starting point for planning its flight path. The system needs to know the drone's specific coordinates when it started planning its path in real time or near real time in order to calculate the flight distance and time from the current location to each target abnormal section. The system obtains the drone's real-time latitude and longitude coordinates through GPS or other positioning systems integrated on the drone. This data is usually transmitted back to the ground control station or cloud platform via a wireless link.

[0090] The fault detection path is determined based on the anomaly level, driving status, and position of the low-altitude UAV in multiple abnormal road segments. At this time, the higher-level abnormal road segments are visited first; only abnormal road segments with the current status of "idle" can be visited (road segments with the status of "occupied" or "under maintenance" are temporarily excluded from the current plan); the path planning needs to start from the current position of the UAV and consider the feasibility of returning to the base or the next take-off and landing point; the goal of the system is to generate an overall "fault detection path" that includes a series of "sub-low-altitude flight paths"; each sub-path corresponds to an abnormal road segment to be detected.

[0091] For path planning of "fault detection paths", graph theory (such as Dijkstra's algorithm) or heuristic methods (such as genetic algorithm or ant colony algorithm) are used to solve the problem. A graph is constructed, where nodes include the current position of the UAV, the start / end point of all accessible (idle) abnormal road segments, and the base / landing point. Edges represent flyable paths, with weights based on flight distance, estimated flight time, or the combined cost considering fuel / electricity. The goal of this approach is to find a path that covers all high-priority and accessible abnormal road segments while minimizing total cost (time, distance, etc.) or maximizing priority.

[0092] Optionally, assuming the current time is 10:10 AM, we only consider the abnormal road segments A and C, which are in the "idle" state; Abnormal level: A is level two, C is level three; UAV location: 116.3912, 39.9087; Path 1: UAV location > Abnormal road segment A > Abnormal road segment C > Return to base; *Path 2: UAV location > Abnormal road segment C > Abnormal road segment A > Return to base; Since the level of abnormal road segment A (level two) is higher than that of abnormal road segment C (level three), the algorithm will tend to choose path 1, that is, fly to the higher-level abnormal road segment A first; The fault detection path is manifested as: UAV location > Sub-low-altitude flight path 1 (covering A) > Sub-low-altitude flight path 2 (covering C) > Return to base; At this time, sub-low-altitude flight paths 1 and 2 have been determined.

[0093] The system rechecks the latest driving status (which may change briefly) and strictly determines the execution order of each sub-low-altitude flight path in the overall path according to the anomaly level (higher level takes priority), distance / time cost (when the levels are the same, the closer / less time-consuming path is selected first), and driving status (ensuring that the target abnormal road segment is indeed in an "idle" state when executing the sub-path).

[0094] Optionally, path 1 is UAV location > A > C > return to base; while the UAV is flying to A, the state of A remains "idle"; while the UAV completes the detection of A and is flying to C, the state of C remains "idle" (for example, checking whether train G1234 has passed B); the final determined flight sequence is: first execute sub-low-altitude flight path 1 (detect A), then execute sub-low-altitude flight path 2 (detect C).

[0095] In step S14, the specific steps are as follows:

[0096] S141: The low-altitude UAV acquires the fault detection path and flies along the fault detection path at low altitude. The low-altitude UAV detects each abnormal road segment in real time during the low-altitude flight. At this time, the low-altitude UAV performs anomaly detection on each abnormal road segment.

[0097] S142: In the anomaly detection of each abnormal road segment, the first camera of the drone collects surface images of the abnormal road segment, determines the surface defects of the abnormal road segment based on the detection of the surface images of the abnormal road segment, and determines the surface anomaly image based on the synthesis of the defect area and defect location of each surface defect.

[0098] S143: The second camera of the drone is set to one side of the first camera and dynamically captures the installation position of abnormal road sections. Based on the images of the installation positions of abnormal road sections, the installation defects of abnormal road sections are determined. Based on the synthesis of the defect areas and defect positions of each installation defect, an abnormal installation image is determined.

[0099] In the embodiments of this application, before or shortly after takeoff, the UAV needs to receive a "fault detection path" planned in step S13 from a ground control station, a preset database, or via wireless communication. This path usually exists in the form of a series of precise geographic coordinates (waypoints), each waypoint containing instructions such as location (latitude and longitude, altitude), flight altitude, flight speed, and hovering time. The flight control system on the UAV receives and parses this path data, stores it in memory, and uses it as the basis for subsequent autonomous flight. The process of obtaining the path needs to ensure the integrity and accuracy of the data.

[0100] After acquiring the path, the UAV activates its flight control system and begins autonomous flight according to the stored path instructions. The flight control system automatically controls the UAV's attitude (pitch, roll, yaw) and power (throttle) based on the current GPS location, target waypoint location, set flight altitude, and speed, guiding the UAV precisely from one waypoint to the next. Throughout the flight, the UAV maintains a predetermined low altitude (e.g., 5-15 meters above the orbital surface) to ensure that the camera can obtain clear and detailed ground images, while avoiding obstacles and complying with airspace management regulations. The flight control system continuously monitors the flight status and makes fine adjustments as necessary to cope with environmental interference such as wind.

[0101] While performing low-altitude flight, the drone's sensor system (mainly cameras, but also including lidar, infrared sensors, etc., but based on the context, this mainly refers to cameras) is activated, beginning continuous or on-demand detection of the railway section passing below. When the flight path passes through a pre-marked "abnormal section," the system automatically triggers a detailed detection of the current section below based on instructions in the path planning (such as hovering after reaching a specific waypoint, adjusting the camera angle, etc.) or through real-time image recognition technology. This process involves using cameras to capture high-resolution images or videos, which are then stored in real time for subsequent analysis (e.g., surface and installation defect identification in S142 and S143). The main task of the drone itself in this step is to perform physical approach and observation, collecting raw data.

[0102] Furthermore, when the drone flies to a marked abnormal road segment (e.g., the K101153 segment identified by GPS coordinates or mileage markers) according to the low-altitude flight path planned by S13, the system will automatically trigger the first camera (usually a high-resolution visible light camera with a certain zoom capability) to start working.

[0103] The drone needs to maintain a stable, level flight attitude, flying at a preset constant altitude (e.g., 5-10 meters above the orbital plane) and speed (e.g., 5-8 km / h); the first camera will shoot at a vertically downward (Nadir) or slightly tilted angle to ensure that the image covers the entire anomalous area and a certain range around it; the shooting frequency and resolution will be dynamically adjusted according to the length and importance of the anomalous segment to ensure that no details are missed; for example, for shorter severe anomalies, the flight speed will be reduced and the shooting resolution will be increased.

[0104] An overlapping shooting strategy is employed to ensure 30%-50% overlap between adjacent images, which is crucial for subsequent image stitching and 3D reconstruction (if needed). Simultaneously, metadata such as the precise shooting time, GPS coordinates, and camera attitude (pitch, yaw, roll angle) of each image is recorded to accurately locate defects. The system automatically detects ambient lighting conditions; if insufficient light or strong shadows exist, it automatically switches to an infrared camera (if equipped) for assistance, or adjusts flight time to avoid unfavorable lighting. Furthermore, it avoids low-altitude hovering shooting in strong winds to ensure image clarity.

[0105] The acquired raw images are first preprocessed, including noise reduction (such as using Gaussian filtering), contrast enhancement (such as histogram equalization), and color correction (if necessary) to improve the accuracy of subsequent detection. This usually relies on advanced computer vision and machine learning algorithms. The system loads pre-trained deep learning models (such as CNN-based image segmentation or object detection models) for railway surface defects (such as rail corrugation, peeling, side wear, fish scale pattern, ballast defects, sleeper damage, etc.).

[0106] The preprocessed image is input into the model; the model automatically scans the image and identifies regions that match preset feature patterns. For example, for rail corrugation, the model will look for wavy textures with specific wavelengths and amplitudes; for peeling, it will look for blocky regions with clear edges and colors / textures different from the surroundings. Each detected region is classified into a specific defect type (such as "rail corrugation" or "sleeper crack") and marked on the image with bounding boxes or segmentation masks of different colors or shapes. At the same time, a confidence score is output (indicating how likely the model is to identify this type of defect). For defects with low confidence or complex defects, the system marks them for human engineers to review and confirm, improving the accuracy of detection.

[0107] Therefore, the second camera of the drone is set to one side of the first camera and dynamically captures the installation position of abnormal road sections. Based on the images of the installation positions of abnormal road sections, the installation defects of abnormal road sections are determined. Based on the synthesis of the defect areas and defect positions of each installation defect, an installation anomaly image is determined. This overall consideration of the synthesis of the defect areas and defect positions of each installation defect ensures the accuracy of the installation anomaly image.

[0108] At this point, in order to achieve multi-angle and multi-faceted observation, drones are usually equipped with multiple cameras; the second camera is physically mounted next to the first camera, but usually has a different perspective or focal length. This layout design allows the second camera to capture a different view than the first camera (which is usually mainly aimed at the track surface), which is particularly suitable for observing the track support and fixing system, which is what we often call the "installation position"; for example, the second camera is set to be slightly tilted, or at a fixed angle to the first camera, so as to better observe the connection between the track and the sleeper, the status of the fasteners, the side of the sleeper, and the contact between the sleeper and the ballast.

[0109] When the drone flies along the preset fault detection path and passes through the abnormal section previously marked by the system, the second camera will automatically start or adjust its angle to continuously capture the installation location of this section. Here, "dynamic capture" means that the camera will adjust its focus, angle or shooting frequency as the drone flies to ensure that the details of the installation structure can be clearly recorded. It focuses on the "skeleton" and "joints" of the track, that is, the parts that support and fix the track.

[0110] The UAV control system analyzes the installation location images captured by the second camera, which typically involves image processing and pattern recognition technologies. The system searches for predefined defect patterns, such as analyzing texture changes, color anomalies, and geometric deviations (e.g., deformation of fastener holes, direction of sleeper cracks) in the image to determine whether there are installation defects. For example, the system sets rules: if the image brightness or shape of a fastener in the image differs from the surrounding normal fasteners by more than a threshold, or if linear or mesh-like structural changes are detected on the sleeper surface, it is determined that there is an installation defect.

[0111] After identifying installation defects, the system needs to visually overlay this defect information onto the original installation location image to generate a new "installation anomaly image." This includes precisely marking the location of each defect on the image (e.g., using rectangles, circles, or arrows) and labeling the type of defect with text. At the same time, the system will also record or display the size of the area affected by the defect (e.g., the length and width of the crack, or the extent of the loose area). This process is to "synthesize" the "area" and "location" information of the defect onto the original image.

[0112] Optionally, assuming the first camera (camera A) is a wide-angle camera, mainly aimed at the center line of the track and shooting downwards at the rail surface to observe surface defects such as cracks and peeling; then, the second camera (camera B) is installed to the left of camera A, slightly tilted to the lower left, or at a small angle (e.g., 15 degrees) to camera A. In this way, when the drone flies over the track, camera A mainly shoots the installation position of camera B; when the drone flies to the previously marked abnormal section (e.g., K101150 to K101155), the second camera (B) will continuously shoot the fastening system of this section of the track. It will capture details such as the side of the sleeper, whether the fastener is loose, whether the pad is displaced, and whether the spike is missing or broken; the images include close-ups of the connection between the sleeper and the rail, as well as the condition of the ballast around the sleeper.

[0113] The system analyzes images of track section K101150 to K101155 captured by camera B. At K101152, the image analysis algorithm detects that the image of a certain rail spike's nut is darker or has blurred edges than other nuts. Combining this with image clarity information, the system determines that the rail spike is loose. At K101153, the algorithm detects a thin, slightly darker linear structure in the side image of the sleeper, which differs from normal texture. This is determined to be a sleeper crack. Thus, the system identifies the specific installation defects: the loose rail spike at K101152 and the sleeper crack at K101153. Regarding the two installation defects identified above (the loose rail spike at K101152...), For example, for the sleeper crack at K101153, the system will generate an installation anomaly image on the original image captured by camera B. At K101152, a rectangle will be superimposed on the image to enclose the loose nut, and labeled "Loose track spike K101152". At K101153, a mark (such as an arrow or curve) will be superimposed on the image to point to the location of the sleeper crack, and labeled "Sleeper crack K101153". The final image with the mark is the "installation anomaly image". It not only includes the original appearance of the installation structure, but also clearly points out the specific installation defects detected and their precise location, providing clear visual reference and positioning information for subsequent maintenance personnel.

[0114] In step S15, the specific steps are as follows:

[0115] S151: Based on the anomaly tracing of each abnormal road segment, the past abnormal events of the abnormal road segment are determined. At the same time, the surface abnormality features are determined based on the recognition of surface abnormal images, and the installation abnormality features are determined based on the recognition of installation abnormal images. The fault content corresponding to the abnormal road segment is determined by matching the surface abnormality features, installation abnormality features and the past abnormal events of the abnormal road segment.

[0116] S152: Collect the railway section operation plan table, and determine the operation status of each abnormal section based on the railway section operation plan table, current time and current position of the train;

[0117] S153: Based on the operating status of each abnormal road segment and the fault content of multiple abnormal road segments, determine the maintenance sequence and maintenance time of each abnormal road segment, and determine the real-time maintenance plan of each abnormal road segment according to the maintenance sequence, maintenance time and corresponding maintenance measures of each abnormal road segment.

[0118] In the embodiments of this application, past abnormal events of abnormal road segments are determined based on the anomaly tracing of each abnormal road segment. At the same time, surface abnormal features are determined based on the recognition of surface abnormal images, and installation abnormal features are determined based on the recognition of installation abnormal images. The fault content corresponding to the abnormal road segment is determined by matching the surface abnormal features, installation abnormal features, and past abnormal events of the abnormal road segment. This approach takes into account the overall consideration of matching surface abnormal features, installation abnormal features, and past abnormal events of the abnormal road segment, ensuring the accuracy of the fault content corresponding to the abnormal road segment.

[0119] At this point, the system will access a historical database or maintenance record system that stores all past inspection reports, maintenance records, accident records, etc. for the railway section. For each abnormal section detected now (for example, the section marked with track mileage K101153), the system will query all events recorded for that section in the past period of time (whether it is several months, a year or longer, depending on the system design).

[0120] These "past anomalies" include: Defect type: such as rail cracks, sleeper damage, ballast settlement, loose fasteners, signal equipment failure, etc.; Occurrence time: the date the event was first recorded or discovered; Handling measures: the repair or handling methods taken at the time, such as replacing sleepers, grinding rails, filling ballast, tightening bolts, etc.; Handling results: whether the repair was effective and whether it recurred; Frequency: the frequency of the same or similar types of events occurring in this section of track; understanding the historical "health status" of this section of track helps determine whether the current anomaly is "new" or "recurring," and whether there is some kind of "inertial problem."

[0121] The system introduces surface anomaly images and uses image recognition algorithms (based on deep learning models) to analyze them. The algorithm attempts to identify specific patterns, shapes, colors, or textures in the image and classify them as known defect types. "Surface anomaly features" typically include: Defect type: as mentioned earlier, rail corrugation, peeling, abrasion, corrosion, rail head crushing, etc.; Defect location: coordinates in the image, mapped to the precise location on the actual track (e.g., distance from the starting point, and on which side of the rail); Defect size: length, width, and depth (if estimated from multiple images or a specific algorithm); Defect severity: a preliminary assessment of its severity (e.g., minor, moderate, severe) based on size, shape, or comparison with standard specifications.

[0122] Anomaly images were introduced, and image recognition or sensor data analysis algorithms were used. For installation anomalies, it is necessary to detect whether components are missing, loose, misaligned, corroded, or damaged. "Installation anomaly characteristics" include: anomaly type: loose or missing fasteners, cracked or displaced sleepers, uneven or lost ballast, increased or decreased track gauge, damaged insulation joints, etc.; anomaly location: the specific location on the track structure; anomaly state: such as "severely loose", "slightly corroded", "completely missing", etc.

[0123] Obtain information on anomalies in the installation of the track structure and its auxiliary equipment, which are related to surface defects (e.g., loose fasteners causing rail misalignment, leading to uneven wear).

[0124] The system uses a pre-set rule base or machine learning model for "matching"; for example, whether the currently detected "rail corrugation" and "fastener loosening" are related to the historical record of "fastener loosening (February 2025)," and whether fastener loosening is the cause of this increased corrugation; whether similar problems (such as corrugation and fastener issues) have occurred repeatedly in the history of this section, which is related to the design, materials, train load, or environmental factors of this section; and the causal relationship chain inferred by the system based on the railway engineering knowledge base; for example, "fastener loosening > unstable rail position > uneven stress generated when the train passes > accelerated rail corrugation"; "fault content" is a more in-depth and specific diagnostic result, which describes the root cause or main feature combination of the anomaly. It is a specific fault name or a descriptive phrase, extracting the most valuable fault diagnosis conclusions for maintenance decisions from information from multiple dimensions.

[0125] Optionally, suppose the drone detects an anomaly in track section K101153; the system traces back and finds: November 2024: recorded "minor ballast settlement", which has been "partially filled"; February 2025: recorded "loose fasteners (left rail)", which has been "tightened"; April 2025: recorded "rail surface scratches", which has been "grinding repair". These historical records indicate that this section has recently experienced several minor problems, but all of them have been dealt with.

[0126] For section K101153, the system analyzed the surface anomaly image generated by S142 and identified the following defects: Defect type: rail corrugation (a periodic rail surface irregularity); Defect location: 4.5 meters from the starting point of the section, located on the downline (hypothetical) rail; Defect size: wavelength approximately 30 cm, wave depth approximately 0.3 mm; Defect severity: medium (judged according to a preset severity threshold, for example, a wave depth exceeding 0.2 mm is considered medium).

[0127] The drone also captured images of the installation status on section K101153; after system analysis, it was identified as follows: Anomaly type: loose fasteners; Anomaly location: also at a distance of 4.5 meters from the starting point, corresponding to the above-mentioned rail corrugation location, the fasteners below (e.g., a certain type of spring clip) are loose; Anomaly state: severely loose (the spring clip has lost its preload and is obviously loose).

[0128] The system matched and analyzed the information of section K101153: Surface anomaly: rail corrugation (moderate); Installation anomaly: fasteners are seriously loose; Historical events: there are recent records of loose fasteners and rail abrasions, and the ballast has also settled.

[0129] The current fastener loosening is directly related to historical fastener problems; the loose fasteners cause instability in the rail position, which in turn causes or exacerbates corrugation; although there has been ballast settlement in the past, no significant settlement was observed in this inspection, so the main cause of corrugation is more likely to be an installation problem; the confirmed fault is "uneven rail wear (corrugation) caused by loose fasteners". This conclusion is more instructive than simply saying "there is corrugation" or "there is a fastener problem". It points out the priority direction of maintenance - first, the loose fastener problem needs to be addressed, and then the rail needs to be ground.

[0130] Furthermore, the system collects the railway section's operation plan, and determines the operation status of each abnormal section based on the railway section's operation plan, current time, and the current position of the train. This comprehensive approach takes into account the railway section's operation plan, current time, and the current position of the train, ensuring the accuracy of the operation status of each abnormal section.

[0131] At this point, the system needs to connect to the railway dispatch center or operation management system to obtain the latest train timetable. This timetable details which trains will run on which lines and at what times within a specific time period. The timetable typically includes: train number: such as G1234 (high-speed train), K567 (ordinary passenger train), freight 123 (freight train); route: the name or number of the line the train will pass through; stations along the way: the stations the train is scheduled to stop at; arrival and departure times: the planned times at each station or specific location; speed limits: speed regulations for certain sections. The timetable needs to be real-time or near real-time because train times can change due to weather, dispatch adjustments, etc.

[0132] The system needs to obtain the current precise time, which is usually obtained from a Network Time Protocol (NTP) server or the system's built-in real-time clock and synchronized to ensure accuracy; the current time is the basis for judging the train schedule status; by comparing with the time points in the operation schedule, it is determined whether the planned events (such as train arrival, passing) are about to happen, are in progress, or have already passed relative to the current time.

[0133] The system needs to connect to a train tracking system (such as a GPS-based, beacon-based, or communication-based positioning system), which can provide real-time updated train location information. For each train in operation, the system will receive its current precise geographical location (e.g., distance from the origin, such as K101200) and operating status (e.g., in motion, stopped). The train's current location information is key to determining the real-time occupancy of the track, revealing whether the train is currently on an abnormal section or is about to enter that section.

[0134] The system now has three pieces of information: the train schedule, the current time (14:25), and the current position of each train. For each abnormal section (e.g., K101153), the system will make the following judgments: check if any train is currently located on the K101153 section; if so, the current status of the section is "occupied" or "train passing through"; check the train schedule to see which trains are scheduled to pass through K101153 shortly after the current time (e.g., within the next 5-10 minutes); if the schedule shows that a train is about to enter, the status of the section is "about to be occupied" or "approaching passage".

[0135] If no train is currently occupying the track, no train is about to enter, and the schedule does not show any trains passing through in the near future, the track segment's status is "idle" or "detectable / maintainable." If the abnormal track segment is located within a station, it is also necessary to consider whether any trains are currently stopping at that station. For example, if K567 is stopping at station K101000, and K101153 includes part of that station, the status is "occupied."

[0136] By integrating static operation plans and dynamic train location information, a real-time operational status label (such as "idle," "occupied," or "about to be occupied") is provided for each abnormal section. This status information is crucial, as it directly determines whether further inspection (such as detailed image acquisition in S14) or maintenance work should be carried out on the section at the current moment. For example, if the system determines that K101153 is currently "idle," then drones or maintenance teams can safely enter the area to work. If it is determined to be "occupied" or "about to be occupied," then it is necessary to wait for the train to pass before operation can be carried out, or the work plan should be adjusted. This ensures that inspection and maintenance activities do not interfere with normal railway operations, while also ensuring the safety of the operators.

[0137] Therefore, based on the operational status of each abnormal road segment and the fault content of multiple abnormal road segments, the maintenance sequence and maintenance time of each abnormal road segment are determined. Real-time maintenance plans for each abnormal road segment are then determined based on the maintenance sequence, maintenance time, and corresponding maintenance measures. This comprehensive approach considers the maintenance sequence, maintenance time, and corresponding maintenance measures of each abnormal road segment, ensuring the accuracy of the real-time maintenance plans. Furthermore, the introduction of surface anomaly images and installation anomaly images ensures multi-dimensional anomaly detection of abnormal road segments by UAVs, achieving a comprehensive consideration of the operational status and fault content of each abnormal road segment, and improving the accuracy of the real-time maintenance plans for each abnormal road segment.

[0138] At this point, the operational status of each abnormal section and the fault details of multiple abnormal sections are introduced. The operational status of each abnormal section: This is the result of S152, telling us which sections are currently free, which are occupied, and which are about to be occupied; for example, abnormal section A is currently free, abnormal section B is occupied, and abnormal section C is about to be occupied. The fault details of multiple abnormal sections: This is the result of S151, telling us what the specific problem is for each abnormal section; for example, abnormal section A is rail corrugation, abnormal section B is turnout switch rail wear, and abnormal section C is loose fasteners.

[0139] Determining the maintenance sequence is usually a priority ranking process; the system will decide which to repair first and which to repair later based on the following factors: At this time, the severity / urgency of the fault: Which faults, if not dealt with in time, will lead to more serious consequences (such as derailment risk, train interruption)? For example, switch point rail wear (B) is usually more urgent than fastener loosening (C) because it directly affects train turning safety; rail corrugation (A), although affecting comfort and track life, is somewhere in between in terms of urgency.

[0140] Operational status: Which road sections are currently or about to be idle, utilizing available "maintenance windows" (i.e., the time period during which line operation is interrupted for maintenance)? For example, if A is currently idle, while B and C both need to wait, then A's priority will be increased;

[0141] Resource availability: Are there specific repair equipment or professionals only suitable for certain types of faults? For example, a grinding machine is only suitable for handling rail corrugation (A);

[0142] Scope of impact: The type and number of trains affected by the malfunction; for example, malfunctions affecting high-speed trains have a higher priority.

[0143] Once the maintenance sequence is determined, the specific maintenance time needs to be arranged in conjunction with the operational status. For high-priority sections, if they are currently free, maintenance should be arranged immediately. If they are not currently free, the next available "maintenance window" time should be predicted based on the train operation plan. For example, if section B is occupied, it is necessary to check when the train will pass through and when the line will be free again. The maintenance time also needs to take into account the time required for the maintenance work itself. For example, grinding the rails (A) takes 1 hour, and replacing fasteners (C) takes 30 minutes. A preliminary maintenance sequence (e.g., A>C>B) and a preliminary maintenance time arrangement (e.g., A: start immediately; C: after the train passes through at 3 pm; B: maintenance window time from 2 am to 4 am tomorrow) should be generated for each abnormal section.

[0144] The document introduces the maintenance sequence, maintenance time, and corresponding maintenance measures for each abnormal road section. The maintenance sequence and maintenance time are the results of S153, telling us which section to repair first, which section to repair later, and approximately when to repair it.

[0145] Maintenance measures for each abnormal section: Based on the fault content determined by S151 and the railway maintenance specifications, clarify the specific maintenance methods required for each fault; for example, rail corrugation requires a rail grinding machine; worn switch rails require replacement of the switch rails or professional grinding; loose fasteners require tightening with a torque wrench or replacement with new fasteners.

[0146] Transform the preceding decisions into a detailed, actionable plan. At this stage, further refine the initial maintenance schedule, specifying the precise start and end times, taking into account preparation and cleanup time. Based on the maintenance measures, identify the required personnel, equipment, materials, and tools. For example, handling section C requires a fastener maintenance team, wrenches, and spare fasteners; handling section A requires a grinding machine and its operators; handling section B requires turnout maintenance experts and equipment replacement, etc. Develop specific operating procedures and safety protocols for the maintenance work of each section. For example, fastener maintenance: set up protection > inspect > tighten / replace > check torque > remove protection. If multiple maintenance tasks need to be performed simultaneously or sequentially, consider the transfer time of personnel and equipment between different locations. For example, after completing section C, how can the fastener team quickly move to section A? Integrate all this information into a clear maintenance plan, distribute it to relevant maintenance teams and the dispatch center, and output a detailed real-time maintenance plan, including for each abnormal section: priority, specific maintenance time (start and end), required personnel, required equipment, required materials, specific operating steps, safety precautions, etc.

[0147] By combining fault diagnosis results, real-time track status, and maintenance resource requirements, the priority and schedule of maintenance work were scientifically determined, and a detailed execution plan was developed. This not only ensured that maintenance work could be carried out in a timely and effective manner, minimizing the impact of faults on railway transportation, but also improved the utilization efficiency of maintenance resources and ensured the safety of operations. For example, through this step, the railway department accurately knows "who" will go "where" and "what method" to repair "what problem," thus carrying out maintenance work in an orderly manner.

[0148] Optional abnormal sections: K101153 (A: rail corrugation), K101200 (B: turnout switch rail wear), K101300 (C: loose fasteners); Operating status (S152 result): A: Idle; B: Occupied (freight train is passing); C: About to be occupied (K567 train is about to enter); Fault details (S151 result): A: Rail corrugation, affecting comfort, medium urgency; B: Turnout switch rail wear, affecting steering safety, high urgency; C: Loose fasteners, causing track deformation, high urgency.

[0149] Assessing urgency: B (turnout) and C (fastener) are more urgent than A (corrugating); Considering operational status: Section A is currently empty; although the urgency is moderate, it should be addressed immediately; the corrugating is expected to take 1 hour; Section B is occupied; extremely urgent, but must wait for freight trains to pass; the line is expected to be empty for about 30 minutes after the freight trains pass; 30 minutes is insufficient to complete turnout repair (it would take several hours), but preliminary inspection or temporary reinforcement can be performed; the main repair needs to be scheduled for the next long track maintenance window; Section C is about to be occupied; extremely urgent, requires waiting for K567 to pass; assuming the line will be empty after K567 passes. 1 hour of downtime; 1 hour is sufficient to handle loose fasteners (replace or tighten); Inspection sequence: C>A>B (prioritize urgent issues that can be handled immediately, C, then A, and finally B, which must wait and is complex to repair); Inspection time: C: Start immediately after K567 passes, estimated time 30-60 minutes; A: Start immediately after C is completed, utilizing the remaining downtime or scheduled later, estimated time 1 hour; B: Schedule for a thorough repair during the next long window (e.g., 1:00-5:00 AM) that night or early the next morning.

[0150] Inspection sequence and time (S153-1 result): C (immediately after K567 passes, about 30-60 minutes), A (immediately after C is completed, about 1 hour), B (1:00-5:00 AM the next day); Inspection measures: C (loose fasteners): tighten or replace fasteners, requiring wrenches, torque testers, and spare fasteners; A (rail corrugation): rail grinding, requiring a rail grinding vehicle and its operators and maintenance personnel; B (turnout switch rail wear): replace the switch rail, requiring a turnout maintenance team, lifting equipment, new switch rails, and turnout adjustment tools.

[0151] Determine the real-time maintenance plan:

[0152] Section C (K101300): Estimated train passage time is 14:50; Maintenance time: 14:55-15:25 (assuming 30 minutes); Required resources: Fastener maintenance team (2 people), wrenches, torque tester, several spare fasteners; Work steps: 14:50 Confirm train passage > 14:55 Set up work protection > Check loose fasteners > Tighten or replace > Test torque > 15:25 Remove protection > Restore track to normal;

[0153] Section A (K101153): Maintenance time: 15:25-16:25 (immediately following Section C, utilizing remaining idle time); Required resources: 1 track grinding machine, several grinding machine operators and maintenance personnel; Work steps: 15:20 Grinding machine arrives and prepares > 15:25 Set up work protection > Grinding machine begins operation > 16:15 Grinding completed > 16:25 Remove protection > Restore normal track conditions;

[0154] Section B (K101200): Maintenance time: 1:00-5:00 AM the next day (long maintenance window); Required resources: turnout maintenance team (5 people), lifting equipment, new switch rails, turnout adjustment tools, safety personnel; Work steps: Personnel and equipment arrive nearby and stand by before 11:00 PM > Enter the work area to prepare at 1:50 AM > Set up work protection at 1:00 AM > Dismantle the old switch rail > Install the new switch rail > Adjust the turnout geometry > Check and confirm multiple times > Remove protection at 4:50 AM > Restore the track to normal at 5:00 AM.

[0155] Generated real-time maintenance plan (summary):

[0156] Task 1: K101300 Fastener Repair: Time: Today 14:55-15:25; Personnel: AA, BB; Equipment: Wrench, torque meter; Materials: 8 sets of M24 high-strength bolts; Precautions: Must evacuate before the train passes, and pay attention to oncoming trains on adjacent tracks during the operation;

[0157] Task 2: K101153 rail grinding: Time: Today 15:25-16:25; Personnel: Grinding machine team; Equipment: Rail grinding machine No. 01; Precautions: Grinding parameters should be set according to Class A corrugated grinding, and the rail temperature should be checked before and after the operation;

[0158] Task 3: Replacement of switch rail for K101200 turnout: Time: Tomorrow 1:00-5:00; Personnel: All turnout maintenance team; Equipment: Lifting equipment, turnout adjustment tools; Materials: 1 set of new switch rail for P60-1 / 12 movable frog; Precautions: All preparations must be completed before the maintenance window. The replacement process must strictly follow the technical specifications. After replacement, the geometric dimensions must be measured multiple times to confirm that they meet the standards.

[0159] Please see Figure 2 , Figure 2 This is a schematic diagram of the structural composition of a low-altitude UAV fault detection system for railway sections according to an embodiment of the present invention; the low-altitude UAV fault detection system for railway sections includes:

[0160] The segment parameter combination module 21 is used to determine multiple sub-detection segments based on the distribution map of railway segments, and to determine the segment parameter combination of each detection segment based on the online detection of multiple sub-detection segments;

[0161] The abnormal road segment module 22 is used to determine multiple abnormal parameters based on the traversal of the road segment parameter combination in each road segment parameter combination, and to determine the abnormal road segment according to the location of the multiple abnormal parameters and the environmental type of the detected road segment.

[0162] The fault detection path module 23 is used to determine the abnormality level of multiple abnormal road sections based on the distribution map of multiple abnormal road sections and railway sections, and to determine the fault detection path based on the abnormality level of multiple abnormal road sections, the driving status map of railway sections and the position of low-altitude UAV.

[0163] The abnormal image module 24 is used for the low-altitude UAV to fly along the fault detection path, perform abnormal detection on each abnormal section, and determine the surface abnormal image and installation abnormal image based on the abnormal detection of the abnormal section.

[0164] The real-time maintenance plan module 25 is used to determine the fault content corresponding to the abnormal road segment based on surface abnormality images, installation abnormality images and past abnormal events of the abnormal road segment, and to determine the real-time maintenance plan for each abnormal road segment based on the operating status of each abnormal road segment and the fault content of multiple abnormal road segments.

[0165] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for fault detection on railway sections using a low-altitude unmanned aerial vehicle (UAV), characterized in that, include: Based on the distribution map of railway sections, multiple sub-detection sections are determined. The combination of section parameters for each detection section is then determined based on the online detection of these sub-detection sections. This process includes: collecting the names of railway sections; determining the distribution map of railway sections based on the names and a railway database; triggering initial detection by a low-altitude UAV based on the distribution map; collecting multiple section boundary points based on the initial UAV detection; determining multiple sub-detection sections based on the multiple section boundary points and the railway section distribution map; performing online detection on each sub-detection section; collecting multiple section parameters based on the online detection of each sub-detection section; marking the corresponding section detection locations with the multiple section parameters; and determining the combination of section parameters for each detection section based on the multiple section parameters and the corresponding section detection locations. In each combination of road segment parameters, multiple abnormal parameters are determined based on the traversal of the road segment parameter combination, and abnormal road segments are determined based on the location of multiple abnormal parameters and the environmental type of the detected road segment. Based on the distribution maps of multiple abnormal road sections and railway sections, the abnormality levels of multiple abnormal road sections are determined, and the fault detection path is determined based on the abnormality levels of multiple abnormal road sections, the driving status map of railway sections, and the location of low-altitude UAVs. The low-altitude drone flies along the fault detection path, performs anomaly detection on each abnormal section, and determines surface anomaly images and installation anomaly images based on the anomaly detection of the abnormal sections. Based on surface anomaly images, installation anomaly images, and past anomaly events of the abnormal road segment, the fault content corresponding to the abnormal road segment is determined. Based on the operating status of each abnormal road segment and the fault content of multiple abnormal road segments, a real-time maintenance plan for each abnormal road segment is determined.

2. The method for fault detection of railway sections by low-altitude unmanned aerial vehicles according to claim 1, characterized in that, In each road segment parameter combination, multiple abnormal parameters are determined based on the traversal of that road segment parameter combination. Abnormal road segments are then determined based on the location of these multiple abnormal parameters and the environmental type of the detected road segment, including: Real-time monitoring of parameter combinations for each road segment; traversal of parameter combinations for each road segment; determination of multiple road segment parameters to be detected based on the traversal of parameter combinations for each road segment; determination of multiple associated parameters based on the mapping relationship between multiple road segment parameters to be detected and associated parameters. If a road segment parameter exceeds a preset safety parameter threshold among multiple road segment parameters to be detected, a proportional calculation between the road segment parameter and multiple associated parameters is triggered to determine the corresponding matching coefficient. If the matching coefficient exceeds a preset matching coefficient threshold, the road segment parameter to be detected is treated as an abnormal parameter, and multiple abnormal parameters are collected. The location corresponding to the abnormal parameter is marked. The corresponding detection section is determined by matching the location of the abnormal parameter with the distribution map of railway sections. The environmental type of the detection section is determined based on the environmental detection of the detection section. The abnormal section is determined based on the values ​​and locations of multiple abnormal parameters and the environmental type of the detection section.

3. The method for fault detection of railway sections by low-altitude unmanned aerial vehicles according to claim 1, characterized in that, The method involves determining the anomaly levels of multiple abnormal road sections based on the distribution map of multiple abnormal road sections and railway sections, and determining the fault detection path based on the anomaly levels of multiple abnormal road sections, the driving status map of the railway sections, and the location of the low-altitude UAV, including: Multiple abnormal road segments are collected, and the length of each abnormal road segment is determined by detecting the length of each abnormal road segment. The first abnormal parameter is determined based on the location and length of the multiple abnormal road segments. The second anomaly parameter is determined based on the location of multiple abnormal road segments and the distribution map of railway segments; the anomaly level of multiple abnormal road segments is determined based on the mapping relationship between the first anomaly parameter, the second anomaly parameter and the anomaly level.

4. The method for fault detection of railway sections by low-altitude unmanned aerial vehicles according to claim 3, characterized in that, The method of determining the anomaly level of multiple abnormal road sections based on the distribution map of multiple abnormal road sections and railway sections, and determining the fault detection path based on the anomaly level of multiple abnormal road sections, the driving status map of railway sections, and the position of low-altitude UAVs, also includes: The system determines the railway segment's operating status map based on the railway segment name, railway database, and current time, and marks the operating status of each abnormal segment. Simultaneously, it collects the location of low-altitude UAVs. Based on the anomaly level, operating status, and UAV location of multiple abnormal segments, a fault detection path is determined. At this point, the fault detection path presents multiple sub-low-altitude flight paths for multiple abnormal segments, and the flight sequence of each sub-low-altitude flight path is determined based on the multiple sub-low-altitude flight paths and the operating status of multiple abnormal segments.

5. The method for fault detection of railway sections by low-altitude unmanned aerial vehicles according to claim 1, characterized in that, The low-altitude UAV flies along the fault detection path, performs anomaly detection on each abnormal section, and determines surface anomaly images and installation anomaly images based on the anomaly detection of the abnormal sections, including: The low-altitude drone acquires the fault detection path and flies along the fault detection path at low altitude. During the low-altitude flight, the low-altitude drone detects each abnormal road segment in real time. At this time, the low-altitude drone performs anomaly detection on each abnormal road segment.

6. The method for fault detection of railway sections by low-altitude unmanned aerial vehicles according to claim 5, characterized in that, The low-altitude UAV flies along the fault detection path, performs anomaly detection on each abnormal section, and determines surface anomaly images and installation anomaly images based on the anomaly detection of the abnormal sections. It also includes: In the anomaly detection of each abnormal road segment, the first camera of the drone collects surface images of the abnormal road segment, determines the surface defects of the abnormal road segment based on the detection of the surface images of the abnormal road segment, and determines the surface anomaly image based on the synthesis of the defect area and defect location of each surface defect. The drone's second camera is positioned to one side of the first camera and dynamically captures the installation location of abnormal road sections. Based on the images captured at the installation locations of abnormal road sections, the installation defects of the abnormal road sections are determined. An abnormal installation image is then determined by synthesizing the defect areas and locations of each installation defect.

7. The method for fault detection of railway sections by low-altitude unmanned aerial vehicles according to claim 1, characterized in that, The method involves determining the fault content corresponding to the abnormal road segment based on surface anomaly images, installation anomaly images, and past anomaly events of the abnormal road segment. Based on the operational status of each abnormal road segment and the fault content of multiple abnormal road segments, a real-time maintenance plan for each abnormal road segment is determined, including: Based on the anomaly tracing of each abnormal road segment, the past abnormal events of the abnormal road segment are determined. At the same time, surface abnormal features are determined based on the identification of surface abnormal images, and installation abnormal features are determined based on the identification of installation abnormal images. The fault content corresponding to the abnormal road segment is determined by matching the surface abnormal features, installation abnormal features and past abnormal events of the abnormal road segment.

8. The method for fault detection of railway sections by low-altitude unmanned aerial vehicles according to claim 7, characterized in that, The method of determining the fault content corresponding to the abnormal road segment based on surface anomaly images, installation anomaly images, and past anomaly events of the abnormal road segment, and determining the real-time maintenance plan for each abnormal road segment based on the operating status of each abnormal road segment and the fault content of multiple abnormal road segments, also includes: Collect the railway section's operation plan, and determine the operation status of each abnormal section based on the railway section's operation plan, current time, and the current position of the train; Based on the operating status of each abnormal road segment and the fault content of multiple abnormal road segments, the maintenance sequence and maintenance time of each abnormal road segment are determined. Based on the maintenance sequence, maintenance time and corresponding maintenance measures of each abnormal road segment, a real-time maintenance plan for each abnormal road segment is determined.

9. A fault detection system for railway sections using a low-altitude unmanned aerial vehicle (UAV), characterized in that, The low-altitude UAV fault detection system for railway sections is applied to the low-altitude UAV fault detection method for railway sections as described in any one of claims 1-8, wherein the low-altitude UAV fault detection system for railway sections includes: The segment parameter combination module is used to determine multiple sub-detection segments based on the distribution map of railway segments, and to determine the segment parameter combination of each detection segment based on the online detection of multiple sub-detection segments; The abnormal road segment module is used to determine multiple abnormal parameters in each combination of road segment parameters based on the traversal of that combination of parameters, and to determine the abnormal road segment based on the location of the multiple abnormal parameters and the environmental type of the detected road segment. The fault detection path module is used to determine the anomaly level of multiple abnormal road segments based on the distribution map of multiple abnormal road segments and railway segments, and to determine the fault detection path based on the anomaly level of multiple abnormal road segments, the driving status map of railway segments and the location of low-altitude UAVs. Anomaly image module is used by low-altitude UAVs to fly along the fault detection path, perform anomaly detection on each abnormal section, and determine surface anomaly images and installation anomaly images based on the anomaly detection of the abnormal sections. The real-time maintenance plan module is used to determine the fault content corresponding to the abnormal road segment based on surface abnormality images, installation abnormality images, and past abnormal events of the abnormal road segment, and to determine the real-time maintenance plan for each abnormal road segment based on the operating status of each abnormal road segment and the fault content of multiple abnormal road segments.

Images