A method and system for collaborative inspection of track-contact network UAVs

Abstract

This application relates to the field of intelligent operation and maintenance technology for rail transit, and discloses a method and system for collaborative inspection of track and overhead contact line using unmanned aerial vehicles (UAVs). It aims to solve problems in existing inspection modes, such as isolated multi-UAV task planning, fragmented perception modes, insufficient system autonomy, and low efficiency, incomplete coverage, and data silos caused by disconnection from train operation sequences. The method includes: receiving traffic control commands and selecting available UAVs, dividing inspection subgroups according to task type; and generating a globally optimal task sequence and flight path through multi-objective optimization. This application achieves synchronous perception of all elements of the track and overhead contact line, multi-UAV collaborative scheduling, and closed-loop control through the above scheme, improving data consistency to over 98%, increasing operational efficiency by 40%, supporting efficient inspection of 50 km of track, and achieving a defect detection rate of 96.7%. It is suitable for intelligent operation and maintenance of large-scale railway networks.

Description

Technical Field

[0001] This invention belongs to the field of intelligent operation and maintenance technology for rail transit, specifically relating to a method and system for collaborative inspection of rail-contact network drones. Background Technology

[0002] With the continuous advancement of intelligent rail transit construction, railway infrastructure inspection is gradually developing towards automation, refinement, and efficiency. As a core link in ensuring the safe operation of trains, the track geometry and the health status of the overhead contact system need to be monitored jointly with high frequency and high precision. Traditional inspection modes mainly rely on manual boarding or track inspection vehicle operations, which have problems such as high workload, limited coverage, and long data acquisition cycle, making it difficult to meet the requirements of modern high-speed railways for real-time and comprehensive operation and maintenance.

[0003] Among them, intelligent inspection technology based on drone platforms has been initially applied in the railway field in recent years. By carrying multiple types of sensors, it can realize non-contact inspection of track lines and overhead contact lines, and has advantages such as high flexibility and fast response speed. The basic goal of this technology is to break through the spatial limitations of ground inspection by using an aerial perspective, improve the observability of complex sections, and obtain key parameters through image recognition and point cloud analysis to provide data support for facility status assessment.

[0004] While existing technologies have developed single-drone inspection solutions for tracks or overhead contact lines, significant shortcomings remain: First, task planning is isolated, lacking a collaborative scheduling mechanism among multiple drones, resulting in low resource utilization and severe duplication of flights; second, perception modes are fragmented, with track geometry deformation and overhead contact line suspension status handled separately, failing to achieve spatiotemporally aligned data fusion analysis; third, system autonomy is insufficient, lacking a unified closed loop for task allocation, path optimization, and result feedback, making it difficult to adapt to dynamically changing inspection needs; finally, the overall architecture is loose, failing to incorporate the operational sequence of the traffic control system into the inspection scheduling decision-making process, making it impossible to avoid train operation windows or emergency situations, affecting inspection safety and effectiveness. These problems are particularly prominent in high-density operational lines, becoming key technical bottlenecks restricting drone inspection from "single-point demonstration" to "system integration." Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies by providing a collaborative inspection method and system for track-overhead catenary using unmanned aerial vehicles (UAVs), which can effectively solve the problems mentioned in the background. Currently, railway infrastructure inspections generally rely on manual labor or single-platform equipment for independent testing, resulting in low operational efficiency, limited coverage, and severe data silos. Track geometry and overhead catenary suspension parameters are collected and analyzed by different systems, lacking a unified scheduling and multi-source information fusion mechanism, making it difficult to achieve real-time and accurate assessment of the overall service status of the line. Furthermore, traditional inspection modes cannot dynamically respond to sudden faults or environmental disturbances, have rigid task allocation strategies, and low resource utilization. This invention constructs an integrated track-overhead catenary inspection system that integrates intelligent scheduling, multi-UAV collaboration, heterogeneous sensing, and data fusion, achieving comprehensive, high-efficiency, and closed-loop intelligent inspection of key railway facilities.

[0006] To achieve the above objectives, the present invention provides the following technical solution: In one aspect, a track-contact network UAV collaborative inspection system, comprising the following components: The orbital state perception module is mounted on the first type of UAV platform, integrating a high-precision inertial measurement unit and a line structured light sensor. This module collects geometric parameters such as track gauge, levelness, elevation irregularities, and track orientation deviation in real time, and generates three-dimensional track profile data with millimeter-level accuracy. The overhead contact line monitoring module is mounted on a Type II unmanned aerial vehicle platform and is equipped with a visible light camera, an infrared thermal imager, and a laser rangefinder. This module simultaneously acquires images of the contact wire height, pull-out value, deformation of suspension components, and temperature distribution at electrical connection points to identify physical anomalies and thermal defects in the contact wire. The task dispatch center receives line operation status instructions, weather warnings, and historical inspection reports from the railway traffic control system. It combines preset inspection priority rules with the current availability of UAVs to generate a globally optimal multi-UAV task allocation scheme, ensuring high-frequency coverage and emergency response capabilities in key sections. The communication relay network consists of wireless relay nodes deployed at base stations or mobile vehicle platforms along the route. This network supports a hybrid transmission mode combining 5G private networks and self-organizing networks, ensuring low-latency, highly reliable two-way communication links between multiple drone swarms and the dispatch center. The data fusion engine receives raw observation data uploaded from different types of UAVs. Based on a spatiotemporal alignment algorithm, it maps orbital geometry data and catenary status information to a unified geographic coordinate system, performs cross-modal feature correlation analysis, outputs a comprehensive health index, and triggers anomaly alarms. The path planner generates a safe, energy-efficient flight path for each UAV that meets time window constraints, based on the sequence of inspection target points issued by the mission scheduling center, combined with digital orbit maps, airspace restricted areas, and real-time meteorological data. The autonomous obstacle avoidance unit integrates binocular stereo vision and millimeter-wave radar to detect dynamic obstacles (such as birds and construction machinery) and static obstructions (such as utility poles and trees) on the flight path in real time, and achieves millisecond-level collision avoidance response through local replanning; Preferably, the task scheduling center adopts a distributed negotiation mechanism based on the improved contract network protocol, which supports multiple UAV agents to bid and negotiate for the execution rights of complex task packages, and achieves load balancing while ensuring the minimum total energy consumption, thus avoiding system paralysis caused by the failure of a single node. Furthermore, the task scheduling center has a built-in predictive maintenance model, which uses a long short-term memory neural network to model the deterioration trend in historical inspection data, predict potential fault areas 72 hours in advance, and automatically increase the inspection frequency level of the corresponding section. In addition, the data fusion engine introduces the weighted evidence synthesis theory to evaluate the confidence and resolve conflicts of repeated observations of the same location by multiple UAVs. The evidence weight threshold is set at 0.85. The defect is confirmed only when the overall support after synthesis exceeds the threshold, which significantly reduces the false alarm rate. Preferably, the path planner adopts a hierarchical A* search algorithm, where the upper layer quickly generates a coarse track in the grid map, and the lower layer uses B-spline curves to smooth the track and eliminate sharp turns, so that the UAV attitude changes continuously and reduces energy loss and vibration interference. Furthermore, the autonomous obstacle avoidance unit is equipped with a dynamic region of interest focusing mechanism, which adjusts the sensor scanning priority according to the current inspection task type—when performing contact wire inspection, it increases the sampling frequency of point cloud density in the vertical space within 15 meters in front to 20Hz, thereby improving the ability to identify small parts. In addition, the communication relay network has an adaptive switching function for link quality. When a relay node detects that the signal-to-noise ratio of the main channel is lower than 12dB, it automatically switches to the backup frequency band and starts forward error correction coding to ensure that the packet loss rate of critical control commands does not exceed 0.1%. Preferably, the orbital state perception module implements a motion compensation algorithm during data acquisition, and uses the attitude angle information provided by the IMU to perform real-time shaking processing on the structured light point cloud to compensate for pitch and roll disturbances during the flight of the UAV, so that the reconstruction error is controlled within ±0.3mm. Furthermore, the contact wire monitoring module is equipped with a multispectral imaging component, which can stably extract the contact wire edge contour under day / night / rain and fog conditions. It combines Hough transform and least squares fitting algorithm to invert the contact wire spatial equation and calculate its offset relative to the design position with an accuracy of ±1.5mm. In addition, the data fusion engine also integrates a causal reasoning module to analyze whether track irregularities cause abnormal dynamic tension of the overhead contact system. The causal strength threshold is set to 0.7. When a strong causal relationship is determined between the two, a joint handling suggestion is pushed to the upstream maintenance management system. On the other hand, a method for collaborative inspection of track-contact network UAVs includes the following specific steps: Step S110: Receive the daily inspection plan instruction issued by the railway traffic control system. The instruction includes mandatory inspection sections, priority levels, time limits, and no-fly zone information. Step S120: Based on the status data of all currently available drones, including remaining battery power, sensor working status, current location and the time of the most recent calibration, filter out a set of candidate drones that meet the mission conditions; Step S130: Based on the spatial distribution characteristics and task type of the section to be inspected, divide the candidate UAVs into an orbit inspection subgroup and an overhead contact line inspection subgroup, and assign a corresponding sensing payload working mode to each subgroup. Step S140: Based on the multi-objective optimization scheduling method for emergency mapping UAV resources, the task scheduling center performs optimization solutions, comprehensively considering flight distance, energy consumption, task deadline, and risk coefficient, to generate the task sequence and execution time of each UAV.

[0007] Step S150: After receiving the mission instructions, each UAV calls the path planner to generate the optimal flight path from its current position to the first target point. The path must avoid known obstacles and airspace control areas. Step S160: During the flight of the UAV along the planned path, the UAV collects real-time status data of the track or catenary and transmits it back to the data fusion engine through the communication relay network. Step S170: The data fusion engine performs timestamp alignment and coordinate normalization on the received multi-source heterogeneous data to establish a state database under a unified spatiotemporal reference. Step S180: Perform joint diagnostic analysis on the normalized data to determine whether there are geometric deviations, temperature anomalies or structural loosening phenomena that exceed the threshold. Step S190: If an abnormal event is detected, an alarm message is immediately generated and fed back to the task scheduling center, triggering a nearby drone to go for verification or start an emergency inspection task. Step S200: After completing all inspection tasks, summarize all data to form a structured inspection report and upload it to the railway operation and maintenance management platform for subsequent decision-making. Preferably, in step S140, a mixed-integer linear programming model is used for task allocation. The objective function is defined as minimizing the weighted sum of the maximum completion time and the total energy consumption, with weight coefficients of 0.6 and 0.4, respectively. The constraints include the UAV's endurance, sensor tolerance limit, and task mutual exclusion rules. Furthermore, in step S160, the UAV records the GPS timestamp and inertial navigation position while collecting data, ensuring that the time synchronization error between each frame of image and point cloud data is less than 50ms. In addition, in step S170, a registration method based on the ICP (Iterative Closest Point) algorithm is used to stitch together the point clouds of consecutive frames to construct a three-dimensional track model of the complete segment, with a point cloud density of not less than 1,000 effective points per square meter. Preferably, in step S180, a hierarchical criterion is set: a level 1 alarm corresponds to a parameter deviation exceeding 1.2 times but less than 1.5 times the allowable limit; a level 2 alarm corresponds to exceeding 1.5 times but less than 2.0 times; and a level 3 alarm corresponds to exceeding 2.0 times. Different levels trigger different response procedures. Furthermore, in step S190, for a level 3 alarm event, the system automatically freezes the relevant section's operating permit and notifies the train dispatching system to implement temporary speed limits or service suspension measures until on-site confirmation and repair. In addition, the inspection report generated in step S200 includes a spatial location map, trend change curve, defect classification statistics table and maintenance priority sorting list, and supports multi-dimensional query by line, section and equipment type; Compared with the prior art, the present invention has the following beneficial effects: By constructing a dual-dimensional collaborative inspection architecture for both track and overhead contact system, synchronous perception of the status of all elements of railway infrastructure has been achieved, solving the problems of spatiotemporal misalignment and information fragmentation caused by traditional separate inspection, and improving data consistency to over 98%. By introducing a task scheduling mechanism based on traffic control system linkage, the inspection activities are deeply coupled with the train operation plan. This enables high-density inspections to be automatically activated during non-operational periods and switched to a low-interference mode during operation, resulting in an overall increase in operational efficiency of 40% and a decrease in resource idle rate of 35%. By employing a multi-UAV swarm collaboration and data fusion strategy, the system's robustness and coverage capabilities are significantly enhanced. A single mission can complete comprehensive inspection of a route up to 50 kilometers long, with an average defect detection rate of 96.7% and a missed detection rate of less than 0.8%. It has achieved a shift from passive response to proactive prevention. By integrating historical data with real-time observation to establish a health evolution model, the proportion of potential hidden dangers detected in advance has increased to 82%, significantly reducing the probability of sudden shutdown accidents. The system has the ability to operate completely autonomously, and can complete task decomposition, path planning, obstacle avoidance flight and result reporting without human intervention. It can support up to 120 drones working together per day and is suitable for routine intelligent operation and maintenance of large-scale railway networks. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the overall technical solution architecture proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of multi-source heterogeneous data fusion and task collaborative scheduling in this invention. Detailed Implementation

[0009] Please refer to Figure 1 and Figure 2 To further illustrate the technical means and effects of the present invention in order to achieve the intended purpose, the following detailed description is provided in conjunction with the accompanying drawings and preferred embodiments, based on the specific implementation methods, structures, features and effects of the present invention.

[0010] Example 1 In a busy section of the high-speed railway network, from K120+300 to K170+800, more than 180 trains run daily in this section, which includes multiple tunnel groups and viaducts. The track geometry is significantly affected by temperature deformation and wheel-rail impact, and the catenary suspension system is in a dynamic tension fluctuation environment for a long time. Although traditional manual inspections are conducted every two weeks, the long cycle makes it difficult to detect sudden defects in a timely manner. In contrast, the application of unmanned aerial vehicle (UAV) railway inspection technology has significantly improved the efficiency and accuracy of inspections. UAVs can quickly cover large areas, are not limited by terrain, and can carry a variety of sensors and equipment to achieve multiple inspection functions. However, existing single-platform UAV systems still have problems with asynchronous data acquisition and spatial matching errors when performing track or catenary inspection tasks, which affects the accurate judgment of the relationship between track irregularities and catenary offset. To address this, the track-overhead contact network UAV collaborative inspection system described in this invention was deployed, constructing an intelligent inspection cluster of 56 heterogeneous UAVs. Of these, 28 are Class I UAVs (model UAV-TG1) equipped with track status sensing modules, and the other 28 are Class II UAVs (model UAV-CM1) equipped with overhead contact network monitoring modules. During smart grid inspection, the UAVs access a hybrid communication architecture of 5G private network and self-organizing network through 42 communication relay nodes deployed along the line. This achieved an end-to-end average latency of less than 80ms and a control command packet loss rate consistently below 0.09%, performance indicators consistent with current technological applications.

[0011] The task dispatch center receives the daily inspection plan instruction from the railway traffic control system. This instruction automatically designates the non-operational period (00:00 to 04:30 daily) as the main inspection window based on the train timetable. The section from K145+200 to K148+600 is marked as a first-priority mandatory inspection section due to recent pantograph-catenary sparking issues. The remaining sections are covered according to the second-priority standard. Simultaneously, a yellow thunderstorm warning is issued, indicating a risk of instantaneous strong winds in the area from K130+100 to K135+400, and this area is designated as a temporary no-fly zone. The system initiates step S110, parsing the instruction content and extracting the coordinate set of the mandatory inspection section, the priority level matrix, the time constraint vector, and the three-dimensional airspace no-fly boundary parameters. Proceed to step S120, access the UAV status database to obtain real-time information for all 56 UAVs, including current battery percentage, IMU calibration validity period (not exceeding 7 days), operating temperature range of the line structured light sensor (-20℃ to +60℃), infrared thermal imager self-test results, GPS positioning accuracy (better than 1.5 meters), current location latitude and longitude, and the timestamp of the last mission completion. The screening criteria are set as follows: remaining battery power not less than 75%, most recent calibration within 72 hours, no sensor fault alarms, and a straight-line distance of less than 15 kilometers from the target section. After screening, a candidate UAV set is formed, with a total of 49 UAVs meeting the criteria, including 23 track inspection UAVs and 26 overhead contact line inspection UAVs.

[0012] Step S130 is executed, and the area to be inspected is divided into subgroups based on its spatial topology. The 50.5-kilometer-long line is divided into 10 logical inspection units, each approximately 5 kilometers long. The complexity coefficient of each unit is calculated based on characteristics such as the radius of the track curve, gradient, bridge ratio, and tunnel density. For units with a curve radius less than 800 meters or a tunnel ratio exceeding 60%, a higher proportion of catenary inspection resources are allocated. Ultimately, the track inspection subgroup consists of 23 UAV-TG1 aircraft, and the catenary inspection subgroup consists of 26 UAV-CM1 aircraft. The sensing payload operating modes are configured for each subgroup: the track subgroup uses a high frame rate line structured light scanning mode (sampling frequency 120Hz) and simultaneously activates the IMU motion compensation algorithm; the catenary subgroup switches to a multispectral imaging mode, superimposing near-infrared band (850nm±10nm) on top of the visible light channel (resolution 4096×3072) to enhance nighttime imaging capabilities, and the laser ranging device is set to continuously emit pulses at a frequency of 10Hz.

[0013] In step S140, the task scheduling center uses a mixed-integer linear programming model to solve for the optimal task allocation scheme. The decision variables are defined as follows: , indicating the first Was the drone assigned to perform the mission? The task of each inspection unit; the objective function is: ,in The maximum completion time among all tasks. This is an estimate of total energy consumption. The energy consumption model takes into account flight distance. Climbing height drag coefficient Battery discharge efficiency The calculation formula is as follows, taking into account factors such as: in, For the first The drone carried out the first Total energy consumption of the task For the first The drone carried out the first The duration of the task, For hovering power, Forward flight power (related to speed). This refers to the power of climb (related to vertical speed). The constraints for sensor operating power include: the total flight time of each UAV is limited by its design and battery performance; for example, the UAV-TG1 model has a flight time of no more than 38 minutes, while the UAV-CM1 model has a flight time of 42 minutes. Only one UAV is allowed to enter adjacent airspace within a 200-meter radius at a time to avoid electromagnetic interference; contact wire inspection tasks must be performed under light intensity above 50 lux (excluding late night hours) to ensure accuracy and safety; orbit inspection tasks should avoid weather conditions with rainfall exceeding 2 mm / h to prevent adverse weather conditions from affecting the reliability of the inspection results. Through iterative calculations by the solver, a task sequence containing 107 specific action instructions is generated, with execution timing accurate to the second, ensuring that the two subgroups advance in a spatially staggered manner while maintaining synchronous observation in time.

[0014] After receiving the mission packet, each UAV proceeds to step S150, invoking its built-in path planner to generate a flight path from its current location to the first target point. The path planner employs a hierarchical A-search algorithm: the upper layer performs global pathfinding on a pre-loaded 1m×1m resolution raster map, integrating a digital orbital model (including bridge, tunnel, and utility pole coordinates), airspace control zones (circular buffer zone r=150m), and real-time weather cloud maps (dynamically labeled strong wind areas); in UAV trajectory planning, the lower layer's generated polyline trajectory is fitted using a non-uniform cubic B-spline curve to ensure curvature continuity constraint κ≤0.08rad / m, while controlling the UAV's yaw angle change rate to not exceed 15° / s. This processing method has been proven to effectively reduce fuselage vibration, thereby improving sensor accuracy. For example, the UAV numbered TG1-17 departs from the K120+000 takeoff point. The original path A contains three sharp turns (112°, 105°, and 118° respectively). After B spline smoothing, the maximum turn angle is reduced to 83°, and the flight energy consumption is predicted to be reduced by 14.6%.

[0015] The UAV begins flying along the optimized path, proceeding to step S160, where data acquisition and transmission are performed simultaneously. Taking UAV-CM1-09 as an example, when flying over the K145+500 suspension bridge, a dynamic region of interest focusing mechanism is activated, designating a 15-meter vertical space in front of it as a high-interest area. The binocular vision system's point cloud acquisition frequency is increased from the default 10Hz to 20Hz, and the millimeter-wave radar beamwidth is reduced from 60° to 30° to enhance resolution. When a flock of migratory birds is detected entering the flight corridor, the autonomous obstacle avoidance unit completes obstacle trajectory prediction within 120 milliseconds, determining a collision probability of 0.73. Local replanning is immediately triggered, generating a new trajectory with a detour offset of 18 meters. The detour process takes 9.3 seconds, with a maximum deviation of 21.4 meters from the original path. After completion, it automatically returns to the main flight path. During this process, the overhead contact line monitoring module continuously operates, capturing a set of multimodal data packets every 0.8 seconds. These packets include: a 4096×3072 pixel visible light image, a 640×512 resolution infrared thermal image (temperature sensitivity 0.05℃), a set of laser ranging points (12 points in total, sampling interval 0.6 seconds), and the corresponding GPS timestamp (UTC+8 standard time, accuracy ±10ms) and inertial navigation position (fused RTK differential signal, horizontal positioning error ≤0.3 meters). All data is uploaded through the communication relay network. The link quality monitoring module evaluates the channel SNR in real time. When the SNR of the main frequency band 1.8GHz drops to 11.8dB, it automatically switches to the backup 2.6GHz frequency band and enables LDPC forward error correction coding to ensure the complete delivery of critical data frames.

[0016] After receiving the multi-source heterogeneous data stream, the data fusion engine executes step S170, first performing timestamp alignment and coordinate normalization. For the track point cloud data, a registration method based on the ICP algorithm is used to stitch consecutive frames together, with an initial matching error set to 5 cm and an iterative convergence threshold set to 0.5 mm, allowing a maximum of 200 iterations. During the point cloud stitching process, IMU attitude angles are introduced as prior estimates to compensate for high-frequency noise caused by flight jitter, resulting in a point cloud density of 1137 effective points per square meter for the reconstructed 3D track model, far exceeding industry benchmark requirements. For the catenary image sequence, georegistration is achieved using initial GPS location values ​​combined with feature point matching (SIFT algorithm extracts no less than 50 corresponding points), establishing a spatial index table in a unified WGS-84 coordinate system. All data are aggregated in 500ms time windows, and linear interpolation is used to align the timestamps of asynchronous acquisitions, ensuring that the time synchronization error between track elevation irregularity measurements and catenary temperature readings is controlled within 42ms.

[0017] Proceed to step S180 to perform joint diagnostic analysis. Based on the technical specifications for alarm information and fault classification in urban rail transit signaling systems, the system establishes a classification criterion system: Level 1 alarm trigger condition is a parameter deviation exceeding the national standard limit by 1.2 times but less than 1.5 times; Level 2 is 1.5 to 2.0 times; and Level 3 is exceeding 2.0 times. Taking the detection data at K146+320 as an example, the track condition sensing module measured a gauge deviation of +14.3mm (standard limit ±12mm), which is a Level 1 exceedance; simultaneously, the contact wire monitoring module recorded a contact wire pull-out value of 487mm at this location (design value 450mm, allowable deviation ±30mm), also exceeding the upper limit by 1.23 times. The data fusion engine calls the causal reasoning module to analyze whether there is a correlation between the two. Input variables include: track lateral irregularity amplitude, simulated car body sway acceleration, contact wire tension fluctuation spectrum, and simulation results of dropper force distribution. Using Bayesian network modeling, the causal strength of the track anomaly on the catenary deviation was calculated, yielding a posterior probability P(catenline deviation | track deviation) of 0.78, exceeding the preset threshold of 0.7, indicating a strong causal relationship. Although both conditions were triggering Level 1 alarms, the system, due to its clear linkage mechanism, still pushed joint handling suggestions to the upstream maintenance management system, prompting priority to addressing track geometric defects to eliminate cascading risks.

[0018] If an abnormal event is detected, step S190 is executed. When UAV-TG1-03 identifies rail surface peeling damage at K147+150 with a depth of 3.8mm, exceeding the line depth limit (3mm) for speeds greater than 120km / h by 2.0 times, a level three alarm is triggered. The system immediately freezes all subsequent non-emergency flight missions in this section and sends a verification instruction to UAV-CM1-12 and UAV-CM1-13, which are currently performing missions in the adjacent airspace, requiring them to change their original flight paths and conduct cross-verification imaging of the location from upstream and downstream directions, respectively. Within 27 seconds of receiving the instruction, both aircraft complete path replanning and arrive at the designated observation point, uploading high-resolution images to confirm the damage. The mission dispatch center simultaneously notifies the train dispatch system, and in accordance with Article 3.4.7 of the "Railway Traffic Organization Rules," a temporary speed limit of 45km / h is implemented in this section, and heavy-haul freight trains are prohibited from passing until the engineering section confirms the repair on-site. The entire response process, from detection to the issuance of the handling instruction, took less than 3 minutes, which meets the timeliness requirements for emergency response.

[0019] After completing all inspection tasks, proceed to step S200 to summarize all data and generate a structured inspection report. The report uses an XML Schema-defined data structure and includes four core components: a spatial positioning layer (GeoJSON format, marking the latitude and longitude and impact range of all anomalies), a trend change curve set (SVG vector image, displaying the evolution trajectory of key parameters over the past 30 days), a defect classification statistics table (CSV format, summarized by type, level, and section), and a maintenance priority sorting list (JSON array, weighted and scored considering three indicators: safety impact, diffusion rate, and maintenance difficulty). The report supports multi-dimensional queries by line number, station section, equipment category (such as turnouts, anchor joints), and defect nature (mechanical damage, electrical overheating). For example, in EMU maintenance, by inputting specific conditions via a web terminal, such as "K145-K148" and "contact wire thermal defect," the system can identify and return temperature anomalies. For instance, in a certain section, the system detected 6 temperature anomalies, 3 of which were located at disconnector switch connections, with an average temperature rise of 42.5℃ at these points. The discovery of such anomalies corresponds to the fault causes mentioned in the high-speed train air conditioning system failure cases, such as problems with the control program of the built-in inverter of the air conditioning unit or poor cooling function of the air conditioning unit. All of these should be marked as level two alarms and included in the maintenance plan.

[0020] Throughout the inspection process, the task scheduling center continuously runs a predictive maintenance model. This model employs a Long Short-Term Memory (LSTM) neural network architecture, processing inspection data from the past 90 days through its input layer, including key temporal features such as track wear rate, fastener loosening frequency, contact wire wear, and dropper fatigue index. The model's hidden layers consist of three stacked LSTM units (128, 64, and 32 neurons respectively) to predict the probability of functional degradation in each kilometer section within the next 72 hours. Model training uses a sliding window method, fine-tuning the weights with the latest 7 days of data each time to ensure adaptation to seasonal changes. When the predicted fault probability of a certain section exceeds 0.65 for three consecutive days, the system automatically increases its inspection frequency from twice a week to once a day and prioritizes high-performance UAVs for inspection in the scheduling plan, achieving a fundamental shift from passive response to proactive prevention.

[0021] In addition, the data fusion engine performs cross-device consistency checks. For repeated observations at the same location, an improved DS evidence theory is used, employing a weighted evidence synthesis method to resolve conflicts. A basic confidence assignment function m(A) is defined to represent the degree of support a UAV provides for proposition A, and the evidence weight ω_i is jointly determined by the sensor calibration age, flight stability index, and image sharpness score. When three or more UAVs provide judgments on the same defect, a synthesis operation is initiated. The defect is confirmed only when the final synthesis confidence level Bel(A) > 0.85 and the conflict factor K < 0.15. For example, at K138+760, two UAVs, using their onboard high-definition cameras and deep learning algorithms, reported insulator damage (support levels of 0.7 and 0.65, respectively). However, due to blurred images from a third UAV in rainy or foggy weather, the insulator condition was not identified. Through the weighted synthesis algorithm, the system synthesizes the data from the three UAVs, resulting in an overall support level of 0.82, which does not reach the preset alarm threshold. Therefore, the system decided not to issue an alarm for the time being, but instead scheduled a focused review of the suspected area in the next cycle, effectively avoiding false alarm disturbances caused by single-point misidentification.

[0022] The communication relay network maintained a stable connection throughout the operation. Each relay node was equipped with dual radio frequency modules, primarily using the 5G NR Sub-6GHz band for large data backhaul, supplemented by an IEEE 802.11ad millimeter-wave link as a short-range high-speed backup. A mesh topology was formed between the nodes, and the routing protocol adopted was OLSR (Optimized Link State Routing), periodically broadcasting link status messages with an update interval of 2 seconds. At 02:18 on the same day, the K152+400 base station switched to UPS power due to a mains power outage, and its signal transmission power was automatically reduced by 20%, causing the SNR at the coverage edge to drop to 11.5dB. At this time, the relay node immediately triggered an adaptive handover mechanism, transferring the control channel to the 2.4GHz ISM band and increasing the FEC coding redundancy from 1 / 2 to 2 / 3, successfully maintaining the command link without causing any mission interruption.

[0023] This embodiment achieves closed-loop operation without human intervention throughout the entire process. A comprehensive inspection of the 50.5-kilometer line took 3 hours and 52 minutes, collecting a total of 1.87TB of track point cloud data, 96,000 frames of catenary image sequences, and infrared thermal imaging data. Figure 2 30,000 images and 1.47 million sets of laser ranging records were generated. Post-event sampling verification showed that the accuracy of track geometry measurement (gauge ±0.4mm, levelness ±0.35mm, unevenness ±0.6mm) and catenary parameters (contact wire height ±1.2mm, pull-out value ±1.8mm, temperature measurement ±1.5℃) all exceeded design specifications. The defect detection rate, as assessed by a third party, reached 96.9%, with a missed detection rate of 0.7% and a false alarm rate reduced from 12% using traditional methods to 3.1%. The system supports an average of 118 UAV sorties per day, increasing resource utilization to 89.4%, fully validating the technical advantages and engineering feasibility of this invention in the routine operation and maintenance of large-scale railway networks.

[0024] Example 2 To address the unique service environment of railway lines in high-altitude permafrost regions, this embodiment constructs a track-overhead contact network UAV collaborative inspection system suitable for extreme climatic conditions. This scenario is located on the Ando-Nagqu section of the Qinghai-Tibet Railway, at an altitude above 4500 meters, with an average annual temperature of -3.2℃. Frequent frost heave and thawing settlement occur, resulting in strong seasonal deformation of the track foundation. Furthermore, uneven settlement of the overhead contact network supports due to frost heave leads to instability in the suspension height. Conventional inspection methods are limited by frequent severe weather and difficulties in manual access, making it difficult to guarantee the frequency of inspections. This embodiment introduces three key technological changes based on Embodiment 1: First, the first type of UAV is upgraded to a low-temperature resistant composite platform (model UAV-TG2), whose orbital state perception module integrates differential GPS (dGPS) and fiber optic gyroscope IMU, which can maintain heading stability in areas with magnetic declination as high as 18°; Second, the second type of UAV (model UAV-CM2) is equipped with a terahertz imager to replace some infrared equipment, which can penetrate the ice and snow cover layer to detect broken strands inside the contact wire; Third, a permafrost active layer monitoring submodule is added to the mission scheduling center, which integrates ground temperature sensor network data to optimize the inspection strategy.

[0025] After system startup, the entire process from S110 to S200 is followed. However, in step S120, when screening candidate UAVs, an environmental adaptability criterion is added: the UAVs are required to have an IP67 protection rating, a lithium battery with low-temperature heating function (operating lower limit -40℃), and an anti-icing coating on the propellers. Only 21 UAVs meet the criteria after screening, including 10 UAV-TG2s and 11 UAV-CM2s. In step S130, when dividing the UAVs into subgroups, given the high coupling between track deformation and catenary settlement in permafrost areas, a joint inspection mode is adopted. Each UAV-TG2 is paired with a UAV-CM2 to form a "dual-aircraft formation," maintaining a spatial configuration with a longitudinal spacing of 80 meters and a lateral offset of 15 meters to ensure synchronous observation of the same geological unit. A dedicated microwave communication link (frequency 5.8GHz, bandwidth 100Mbps) is established within the formation to achieve real-time data exchange and collaborative positioning correction.

[0026] The path planning phase faces new challenges: Permafrost regions contain numerous seasonal thaw settlement pits, ranging from 5 to 15 meters in diameter and reaching depths of 0.9 meters, yet these are not marked on standard digital maps. The path planner introduces a terrain accessibility assessment model in S150. This model constructs a surface stiffness index layer based on a historical synthetic aperture radar (SAR) image database, classifying areas with a stiffness threshold I_s < 0.65 as potential soft soil zones, prohibiting low-altitude drone crossings. The minimum safe flight path altitude is increased from 35 meters in plains areas to 60 meters, with an additional 15 meters of redundancy to handle sudden downdrafts. Furthermore, due to the thin air at high altitudes, motor thrust decreases by approximately 30%. The path planner incorporates an atmospheric density correction coefficient ρ / ρ0 = 0.58 (measured value) in its energy consumption calculations, re-optimizing task allocation weights to prevent drones from returning due to insufficient power.

[0027] During step S160 of the data acquisition process, the orbital state sensing module activates differential motion compensation mode. Since the IMU is susceptible to interference in strong magnetic field environments, the system uses dGPS carrier phase observations to assist attitude calculation, increasing the sampling frequency to 200Hz. The compensation algorithm expression is as follows: in, This is the corrected attitude data vector. This is the original attitude data vector. The corrected attitude angles. This is the gyroscope output angle. The attitude angle deviation retrieved by dGPS The rotation matrix is ​​used. This method further reduces the point cloud reconstruction error to ±0.25mm, overcoming the cumulative drift problem caused by strong magnetic interference at high altitudes. The contact wire monitoring module activates terahertz transmission imaging mode, transmitting at a frequency of 0.3THz, penetrating the snow layer (thickness ≤30cm) to obtain images of the internal structure of the contact wire. Combined with pulse echo delay analysis, the location of the broken strand is determined, achieving a detection sensitivity up to single-strand fracture (cross-sectional area loss of 3.125%). When UAV-CM2-07 detected signs of fracture in the 5th strand of the contact wire in a certain anchor section at K982+610 (a sudden drop of 42% in the reflected signal amplitude), it was immediately marked as a potential hazard point.

[0028] In step S170, the data fusion engine added a permafrost geological correlation analysis function. In addition to conventional spatiotemporal alignment, a geothermal field coupling model was introduced to perform cross-correlation analysis between the track settlement rate and the measured temperature sequence at a depth of -2 meters. When a positive correlation (correlation coefficient r=0.83) was found between accelerated settlement and rising geothermal temperature with a lag of approximately 14 hours, it was determined to be in an active thawing and settlement period, and the inspection frequency for that section was automatically increased to three times per day. In step S180, joint diagnosis, a new alarm logic was set: if the settlement increase of the same support foundation exceeds 2mm / day for two consecutive days, and the contact line height decreases by ≥3mm simultaneously, a level two alarm is triggered; even if no single item exceeds the limit, it is considered a systemic risk. Three such cases were identified in the section from K976+340 to K978+120, providing timely warnings and preventing suspension failure accidents.

[0029] This embodiment particularly strengthens the energy management mechanism. Given the weak charging infrastructure in high-altitude areas, all drones use hydrogen fuel cells as their main power source, boasting an energy density of up to 1200Wh / kg, extending their flight time to 78 minutes. The ground support vehicle is equipped with a mobile hydrogen production unit, producing high-purity hydrogen on-site through water electrolysis to ensure continuous operation. The communication relay network employs a satellite relay backup scheme; when the ground link is interrupted, it automatically switches to the BeiDou short message channel to transmit critical alarm information, ensuring basic command functions are maintained even under extreme conditions. This inspection completed 42.3 kilometers of line testing, overcoming adverse conditions such as continuous snowfall and instantaneous wind speeds reaching 28m / s. A total of 5 structural hazards were discovered, 2 of which were verified as early signs of melting and subsidence. Early intervention and remediation saved approximately 1.7 million yuan in maintenance costs, demonstrating the adaptability and reliability of this invention in special geographical environments.

[0030] Example 3 This embodiment proposes a fully autonomous indoor inspection system architecture that does not rely on GNSS, targeting densely packed underground tunnel sections of urban rail transit. The application scenario is the Z1-Z8 section of a subway line, 28.6 kilometers long, entirely within tunnels where GPS signals are completely blocked, making traditional UAVs unable to locate the equipment. This embodiment abandons external satellite navigation and constructs a novel navigation system based on the fusion of UWB (Ultra-Wideband) indoor positioning network and SLAM (Simultaneous Localization and Mapping). The system deployment method undergoes a fundamental change: UWB anchor points are installed every 30 meters on the tunnel sidewalls to construct a three-dimensional positioning framework, achieving high-precision positioning services. This technology can provide a ranging accuracy of ±10cm, ensuring accurate personnel and equipment positioning even in complex tunnel environments. The UAV eliminates the GPS module, replacing it with UWB tags and a solid-state LiDAR, with a field of view of 270°×30° and a scanning frequency of 10Hz.

[0031] The system process still strictly follows S110 to S200, but the S120 state screening emphasizes the performance indicators of inertial devices: it requires that, according to the technical specifications of the inertial measurement unit (IMU), its zero-bias stability be better than 0.5° / hr and its random walk noise be lower than 0.05° / √hr. This ensures that, without external correction, the accumulation rate of dead reckoning error can be controlled within 0.8% of the flight distance. Step S130 no longer simply divides the process by track / catenment, but instead implements zoned cooperation based on the tunnel cross-section structure: the circular tunnel section adopts a "ring scan" mode, with one UAV hovering in the center and the LiDAR rotating horizontally to scan the entire circumference; the rectangular section adopts a "dual-sided wall-hugging flight" strategy, with two UAVs flying parallel along the left and right side walls respectively to ensure full coverage of the catenary support structure.

[0032] The path planner refactors its algorithm logic in S150. Due to the lack of a global coordinate reference, a relative path planning method is adopted: a local coordinate system is established with the UWB anchor point at the entrance as the origin, and the path is defined by the relative displacement vectors between a series of waypoints. The A* algorithm runs on a discretized 3D voxel map (voxel size = 0.5 cubic meters), avoiding fixed obstacles such as ventilation ducts and fire hoses. During B-spline smoothing, an additional jerk constraint |j| ≤ 2 m / s³ is applied to prevent vibrations caused by violent maneuvers from interfering with the sensors. The flight speed is limited to below 12 km / h to ensure that the LiDAR point cloud density is not less than 800 points per cubic meter.

[0033] In the S160 system, due to the lack of a unified time standard, each UAV uses the IEEE 1588 Precision Time Protocol (PTP) for nanosecond-level clock synchronization. The master clock is located in the section control room and distributes time signals to each UWB anchor point via a fiber optic network, which are then transmitted wirelessly to the UAVs. To ensure the accuracy of time synchronization, each frame of point cloud data is embedded with an absolute timestamp (UTC), with the error controlled within ±200 nanoseconds. The track status perception module uses a laser triangulation ranging array, consisting of a linear array of 16 transmit-receive units with a projection angle of 75° and a sampling interval of 2cm. With IMU tilt compensation, the track gauge measurement accuracy is ±0.5mm. The catenary monitoring module uses a narrow-angle high-definition camera (135mm focal length) and an array of supplementary lights (color temperature 5600K, illuminance ≥500lux) to photograph the insulator strings. The images are enhanced by a dehazing algorithm and then fed into a convolutional neural network for palmprint recognition.

[0034] The data fusion engine performs registration in S170 without a reference coordinate system. The NDT (Normal Distributions Transform) algorithm is used instead of ICP because it is less sensitive to initial pose and suitable for weakly textured tunnel environments. Point cloud frames within a continuous 5-second timeframe are aggregated into a "super keyframe," and the optimal transformation matrix is ​​solved through maximum likelihood estimation to construct a unified 3D model of the entire tunnel. After spatial alignment, S180 joint diagnostics are performed: focusing on analyzing the correlation between track centerline offset and lateral displacement of the contact wire suspension points. When the two deviations are in the same direction and the distance is less than 1.5 meters, it is determined to be an overall structural offset, triggering a special assessment process. In section Z5, a track offset of 18mm to the left was found, and the left-side contact wire support device also shifted 21mm to the left during the same period. The system inferred that this might be caused by unbalanced earth pressure on the tunnel sidewalls and recommended conducting ground-penetrating radar detection for verification.

[0035] This embodiment establishes a tiered evacuation mechanism in the S190 anomaly response. Upon detecting a Level 3 alarm, an emergency procedure is immediately initiated: first, nearby drones are ordered to switch to low-power standby mode to conserve power; second, the shortest safe return route is planned, prioritizing communication channels with emergency helipads; finally, if communication is interrupted for more than 30 seconds, an audible and visual beacon transmitter is activated to guide rescue efforts. The inspection report generated in step S200 includes a dedicated "structural convergence deformation" analysis module, using multi-period scanning data to calculate the tunnel cross-sectional contraction rate and provide early warning of potential encroachment risks. This embodiment successfully achieves fully automated inspection of underground tunnels, accumulating a flight mileage of 1200 kilometers, with positioning drift consistently controlled within ±0.8 meters, filling a technological gap in intelligent detection of underground spaces in urban rail transit.

[0036] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention are within the scope of the present invention.

Claims

1. A method for collaborative inspection of track-contact network UAVs, characterized in that, include: Receive inspection plan instructions issued by the railway traffic control system, which include mandatory inspection sections, priority levels, time limits, and no-fly zone information; Based on the status data of all currently available drones, a set of candidate drones that meet the mission requirements is selected; Based on the spatial distribution characteristics and task type of the sections to be inspected, the candidate UAV set is divided into an orbit inspection subgroup and an overhead contact line inspection subgroup, and a corresponding sensing payload working mode is assigned to each subgroup. The task scheduling center is used to perform multi-objective optimization solutions to generate task sequences and execution timing for each UAV. After receiving the mission instructions, each UAV calls the path planner to generate the optimal flight path from its current position to the first target point; During the flight of the UAV along the optimal flight path, the UAV collects real-time status data of the orbit or overhead contact line and transmits it back to the data fusion engine through the communication relay network; The data fusion engine performs timestamp alignment and coordinate normalization on the received multi-source heterogeneous data to establish a state database under a unified spatiotemporal reference. Perform joint diagnostic analysis on the normalized data to determine if any abnormal events exist; If an abnormal event is detected, an alarm message is immediately generated and fed back to the task scheduling center, triggering a nearby drone to go for verification or start an emergency inspection task; After completing all inspection tasks, all data are compiled into a structured inspection report and uploaded to the railway operation and maintenance management platform.

2. The method for collaborative inspection of track-contact network UAVs according to claim 1, characterized in that, The process of using a task scheduling center to perform multi-objective optimization and generate task sequences and execution times for each UAV includes: Receive line operation status instructions, weather warning information and historical inspection reports from the railway traffic control system; By combining preset inspection priority rules with the current availability of drones, a globally optimal multi-drone task allocation scheme is generated to ensure high-frequency coverage and emergency response capabilities in key areas.

3. The method for collaborative inspection of track-contact network UAVs according to claim 2, characterized in that, The task scheduling center adopts a distributed negotiation mechanism based on the improved contract network protocol, which supports bidding and negotiation among multiple UAV agents for the execution rights of complex task packages, and achieves load balancing while ensuring the minimum total energy consumption.

4. The method for collaborative inspection of track-contact network UAVs according to claim 1, characterized in that, The invocation of the path planner generates the optimal flight path from the current position to the first target point, including: Based on the inspection target point sequence issued by the mission scheduling center, and combined with digital orbit maps, airspace restricted areas, and real-time meteorological data, a safe, energy-efficient flight path that meets time window constraints is generated for each UAV.

5. The track-contact network UAV collaborative inspection method according to claim 4, characterized in that, The path planner employs a hierarchical A* search algorithm. The upper layer quickly generates a coarse trajectory in a grid map, while the lower layer uses B-spline curves for smoothing to eliminate sharp turns and ensure continuous changes in UAV attitude.

6. The method for collaborative inspection of track-contact network UAVs according to claim 1, characterized in that, During the flight of the UAV along the optimal flight path, the UAV collects real-time status data of the orbit or overhead contact line, including: For the track inspection subgroup, high-precision inertial measurement units and line structured light sensors are integrated to collect track geometric parameters in real time and generate three-dimensional track contour data. For the overhead contact line inspection subgroup, visible light cameras, infrared thermal imagers, and laser rangefinders are configured to simultaneously acquire images of the physical state and temperature distribution of the overhead contact line.

7. The method for collaborative inspection of track-contact network UAVs according to claim 6, characterized in that, During the data acquisition process, a motion compensation algorithm is implemented to perform real-time de-jitter processing on the structured light point cloud using attitude angle information provided by the inertial measurement unit, thereby compensating for attitude disturbances during the UAV's flight.

8. The method for collaborative inspection of track-contact network UAVs according to claim 1, characterized in that, The data fusion engine performs timestamp alignment and coordinate normalization processing on the received multi-source heterogeneous data, including: Based on the spatiotemporal alignment algorithm, track geometry data and catenary status information are mapped to a unified geographic coordinate system, and cross-modal feature correlation analysis is performed.

9. A track-contact network UAV collaborative inspection system, characterized in that, include: The task scheduling center is used to receive inspection plan instructions from the railway traffic control system and generate multi-drone task allocation schemes by combining the drone status with preset rules. A communication relay network is used to ensure low-latency, highly reliable two-way communication links between multiple UAV clusters and the dispatch center; The data fusion engine is used to receive raw observation data from different types of UAVs and perform spatiotemporal alignment and cross-modal feature correlation analysis. Multiple drones, including a first-type drone equipped with an orbital state perception module and a second-type drone equipped with a catenary monitoring module; The first type of UAV is used to collect orbital geometric parameters, and the second type of UAV is used to collect images of the physical state and temperature distribution of the overhead contact line. Each UAV is also used to generate flight paths, perform inspections, transmit data back, and trigger verification or emergency tasks when an anomaly is detected.

10. The track-contact network UAV collaborative inspection system according to claim 9, characterized in that, The multiple UAVs are also integrated with an autonomous obstacle avoidance unit, which is used to detect dynamic obstacles and static obstructions on the flight path in real time and achieve millisecond-level collision avoidance response through local replanning; the task scheduling center is also equipped with a predictive maintenance model, which is used to model the deterioration trend in historical inspection data, predict potential fault areas and automatically increase the inspection frequency level of the corresponding sections.

Images