Railway corridor low-altitude intelligent inspection system and device

Abstract

The present application provides a railway corridor low-altitude intelligent inspection system and device, relating to the technical field of railway corridor low-altitude intelligent inspection. The system comprises: a path planning module, which generates a reference fixed flight path template through a multi-period adaptive fixed flight path optimization algorithm, and introduces a space-time consistency constraint to ensure that multi-period images meet the pixel-level registration requirements; a flight control and data acquisition module controls the unmanned aerial vehicle to fly according to the path and collects images; an edge analysis module identifies suspected disaster areas in real time and extracts key images; a cloud intelligent analysis module performs multi-period change detection based on key images and historical images to determine the type of hidden danger; a risk assessment module fuses multi-dimensional features to calculate a risk index and determine a risk level; and a warning release module releases warning information, achieving long-term consistency of the inspection path, real-time identification of hidden dangers, and accurate risk assessment, and improving the automation and intelligence level of railway corridor inspection.

Description

Technical Field

[0001] This invention relates to the field of low-altitude intelligent inspection technology for railway corridors, and in particular to a low-altitude intelligent inspection system and device for railway corridors. Background Technology

[0002] Railway corridors, as crucial transportation infrastructure, traverse complex environments and are susceptible to geological disasters and human activities, leading to safety hazards such as slope collapses and foreign object intrusion, threatening train safety. Traditional manual inspection methods suffer from low efficiency, limited coverage, strong subjectivity, and high risk, making them unsuitable for large-scale, high-frequency routine inspections. While drone technology has led to the gradual application of drone-based inspections, existing solutions still have significant shortcomings: Firstly, drone inspections often rely on single manual operations or simple pre-set routes, resulting in significant differences in flight paths and acquisition parameters across different periods. This leads to substantial spatial and radiometric variations in the acquired image data, hindering high-precision, pixel-level change detection and analysis. Secondly, data processing typically depends on post-event centralized transmission and manual interpretation, resulting in poor real-time performance. It fails to identify potential hazards and focus on key areas during flight, wasting data transmission and computing resources and causing delayed early warning responses. Furthermore, existing systems lack a closed-loop intelligent analysis process from image acquisition to risk assessment, hindering automatic hazard identification, quantitative evaluation, and tiered early warning.

[0003] Therefore, there is an urgent need for a low-altitude intelligent inspection technology solution for railway corridors that can ensure the comparability of data from multiple periods, achieve real-time edge processing and cloud-based intelligent analysis collaboration, and ultimately complete automated risk assessment and early warning. Summary of the Invention

[0004] To address the aforementioned problems in the existing technology, the first aspect of this invention proposes a low-altitude intelligent inspection system for railway corridors, comprising:

[0005] The path planning module automatically generates flight paths using a multi-phase adaptive fixed route optimization algorithm. This includes generating a baseline fixed route template for multi-phase inspections based on flight path data from historical inspection phases and geographic information system data of the railway corridor. In subsequent inspection tasks, flight paths are generated using this baseline fixed route template as a constraint. The baseline fixed route template is generated by minimizing the average deviation between candidate paths and historical inspection paths. Furthermore, a spatiotemporal consistency constraint is introduced when generating flight paths to ensure that the spatial position deviation of corresponding path points and the deviation of image acquisition parameters in different inspection phases do not exceed a preset threshold, thus guaranteeing that the multi-phase image data meets pixel-level or sub-pixel-level registration requirements. When acquiring real-time meteorological data or obstacle information, the path planning module adjusts local route segments in the baseline fixed route template only under the premise of satisfying the path offset penalty constraint, and limits the maximum offset of the adjusted flight path relative to the baseline fixed route template to avoid global reconstruction of the overall flight path.

[0006] The flight control and data acquisition module is configured to control the UAV to fly autonomously according to the flight path and to acquire multiple phases of image data of the railway corridor during the flight.

[0007] The edge analysis module, deployed on the drone, is configured to perform real-time target detection on the image data, identify suspected disaster areas, and extract key image data containing suspected disaster areas based on the identification results.

[0008] The cloud-based intelligent analysis module is configured to perform multi-period change detection on the key image data and historical image data from the same period, extract surface change features, and determine the type of hidden danger based on the surface change features.

[0009] The risk assessment module is configured to, after determining the type of hazard, calculate a risk index based on the multidimensional features obtained from remote sensing image inversion using a feature fusion model, and determine the risk level of the hazard based on the risk index;

[0010] The early warning release module is configured to generate and release early warning information based on the risk level.

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

[0012] Through the collaborative operation of multiple modules, the problems of incomparability of multi-period data, poor real-time performance, and lack of intelligent closed-loop analysis pointed out in the background technology are effectively solved. First, the path planning module, through a multi-period adaptive fixed route optimization algorithm, can generate a baseline fixed route template based on historical data and geographic information, and use this template as a constraint to generate flight paths in subsequent inspections. This process introduces spatiotemporal consistency constraints to ensure that the spatial location of inspection path points in different periods is highly consistent with the image acquisition parameters, ensuring from the data source that multi-period image data meets pixel-level or sub-pixel-level registration requirements, laying a solid foundation for subsequent accurate change detection. At the same time, this module supports local adaptive fine-tuning when encountering real-time weather or obstacles, rather than global reconstruction, balancing flight safety and flexibility while maintaining the overall stability of the route.

[0013] Secondly, the flight control and data acquisition module precisely executes the planned path and autonomously completes image data acquisition. Next, the edge analysis module deployed on the UAV performs real-time target detection on the acquired image data, quickly identifying suspected disaster areas and extracting and uploading only key image data containing these areas. This step significantly reduces the amount of data that needs to be transmitted to the cloud, lowers communication bandwidth pressure and cloud processing load, and significantly improves the system's immediate response speed to potential hazards.

[0014] Then, the cloud-based intelligent analysis module receives key image data, performs multi-period change detection on it and compares it with historical data from the same period, extracts surface change characteristics, and determines the type of hazard based on these characteristics. Next, the risk assessment module calculates a risk index using a feature fusion model based on the multi-dimensional characteristics retrieved from the remote sensing image, and determines the risk level accordingly. Finally, the early warning issuance module generates and issues early warning information based on the risk level.

[0015] The entire process, from data collection, real-time screening, in-depth analysis to risk assessment and early warning issuance, forms a complete automated intelligent closed loop. The modules are closely integrated and work synergistically to achieve intelligent management and control of railway corridor hazards from early detection and accurate identification to quantitative assessment and timely early warning, significantly improving inspection efficiency, analysis accuracy, and emergency response capabilities. Attached Figure Description

[0016] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1The diagram shown is a flowchart illustrating the execution steps of a low-altitude intelligent inspection system for railway corridors according to an embodiment of the present invention.

[0018] Figure 2 The diagram shown is a structural schematic of a low-altitude intelligent inspection system for railway corridors provided in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0020] The specific embodiments of the present invention will be described below.

[0021] Example 1

[0022] like Figure 1 and Figure 2 As shown, the first aspect of this invention proposes a low-altitude intelligent inspection system for railway corridors, comprising:

[0023] The path planning module automatically generates flight paths using a multi-phase adaptive fixed route optimization algorithm. This includes generating a baseline fixed route template for multi-phase inspections based on flight path data from historical inspection phases and geographic information system data of the railway corridor. In subsequent inspection tasks, flight paths are generated using this baseline fixed route template as a constraint. The baseline fixed route template is generated by minimizing the average deviation between candidate paths and historical inspection paths. Furthermore, a spatiotemporal consistency constraint is introduced when generating flight paths to ensure that the spatial position deviation of corresponding path points and the deviation of image acquisition parameters in different inspection phases do not exceed a preset threshold, thus guaranteeing that the multi-phase image data meets pixel-level or sub-pixel-level registration requirements. When acquiring real-time meteorological data or obstacle information, the path planning module adjusts local route segments in the baseline fixed route template only under the premise of satisfying the path offset penalty constraint, and limits the maximum offset of the adjusted flight path relative to the baseline fixed route template to avoid global reconstruction of the overall flight path.

[0024] The flight control and data acquisition module is configured to control the UAV to fly autonomously according to the flight path and to acquire multiple phases of image data of the railway corridor during the flight.

[0025] The edge analysis module, deployed on the drone, is configured to perform real-time target detection on the image data, identify suspected disaster areas, and extract key image data containing suspected disaster areas based on the identification results.

[0026] The cloud-based intelligent analysis module is configured to perform multi-period change detection on the key image data and historical image data from the same period, extract surface change features, and determine the type of hidden danger based on the surface change features.

[0027] The risk assessment module is configured to, after determining the type of hazard, calculate a risk index based on the multidimensional features obtained from remote sensing image inversion using a feature fusion model, and determine the risk level of the hazard based on the risk index;

[0028] The early warning release module is configured to generate and release early warning information based on the risk level.

[0029] The railway corridor low-altitude intelligent inspection system constructs a complete technology chain from automated data acquisition to intelligent early warning issuance. The first link in this system is the path planning module. The core algorithm of this module is the multi-period adaptive fixed flight path optimization algorithm. Here, "multi-period" emphasizes that the algorithm processes multiple inspection tasks accumulated over time; "adaptive" means that the algorithm can adjust according to limited new information; and "fixed flight path optimization" means that its goal is to generate and maintain a reusable and optimized benchmark flight path. The algorithm first uses the actual flight path data of UAVs in historical inspection periods, combined with the geographic information system data of the railway corridor (such as digital maps containing terrain elevation, slope, coordinates of fixed structures, and boundaries), to generate a benchmark fixed flight path template for future long-term inspections through specific optimization calculations. This template is not a simple copy of a single flight, but a standard path derived after comprehensively balancing historical flight experience and geographic spatial constraints. In each subsequent inspection task, the process of generating flight paths is based on this benchmark fixed flight path template as the fundamental constraint, that is, the newly planned path must be carried out within the framework of this template. To ensure accurate comparison and analysis of images acquired at different times, the algorithm proactively applies spatiotemporal consistency constraints during path planning. Specifically, this constraint requires that the actual position in three-dimensional space of path points in the same sequence along the UAV's flight path in different inspection missions, and the parameters used by the UAV's image acquisition equipment when passing those points (such as focal length, aperture, and shutter speed), must be controlled within a pre-set, extremely small threshold. This mechanism ensures high stability in spatial geometry and imaging radiometric characteristics of multi-phase image data from the source of data acquisition, an indispensable prerequisite for subsequent pixel-level or sub-pixel-level image registration (i.e., precise alignment of images from different periods). In actual flight, the environment is dynamically changing. When the path planning module obtains real-time meteorological data (such as sudden changes in wind speed and direction) or newly discovered obstacle information via the data link, it has the ability to adapt along the flight path. This adaptation is designed as local adaptive fine-tuning. Its core principle is that fine-tuning must be performed under a rule called "path deviation penalty constraint," which aims to quantify the potential costs of deviating from the baseline route (such as decreased data comparability). Under this premise, the algorithm only makes necessary, minor adjustments to the affected route segments, and never globally replans the flight routes of the entire inspection area. This design, while responding to emergencies and ensuring flight safety, maximizes the historical consistency and repeatability of the overall route.

[0030] In this module, the phrase "adjusting the path of local route segments in the baseline fixed route template and limiting the maximum offset of the adjusted flight path relative to the baseline fixed route template to avoid global reconstruction of the overall flight path" further defines the specific adjustment strategy adopted by the path planning module in response to real-time environmental changes. The core of this statement is to emphasize that the system possesses limited and controlled local adaptability while maintaining consistency across multiple flight paths, rather than undertaking large-scale route replanning.

[0031] Specifically, the "local route segment" refers to the portion of the flight path that is directly affected by real-time weather conditions (such as sudden changes in wind speed) or newly discovered obstacles (such as temporary construction facilities) during the inspection flight. This adjustment is performed within the calculation framework of "path deviation penalty constraint," meaning that the system will evaluate the loss of data comparability and the flight safety benefits caused by deviating from the baseline route, ensuring that adjustments only occur when necessary.

[0032] Furthermore, the "maximum offset of the adjusted flight path relative to the baseline fixed flight path template" sets a clear physical or logical upper limit for this local adjustment. This limit ensures that even when dealing with environmental interference, the actual flight trajectory of the UAV will not deviate excessively from the baseline flight path generated by optimizing historical data, thereby fundamentally maintaining the spatiotemporal consistency of the image acquisition points and attitudes across multiple periods, and preserving a data foundation for subsequent high-precision change detection.

[0033] Ultimately, the principle of "avoiding global reconstruction of the entire flight path" clarifies the design principle and effect of this adjustment strategy. This indicates that the system does not recalculate the entire flight path every time environmental changes occur, but rather adopts a "fine-tuning" strategy. This not only significantly reduces the real-time burden of planning calculations and improves system response speed, but more importantly, it avoids the interruption of historical data sequences caused by large-scale changes in flight paths, ensuring the long-term comparability and analytical continuity of inspection data, which meets the fundamental requirement of long-term data stability in the routine and periodic inspections of railway corridors.

[0034] Therefore, this technical feature embodies a balanced design: while adhering to the core objectives of long-term route consistency and data comparability, it introduces a restricted local adaptive mechanism to give the system the necessary operational flexibility and security, thereby achieving a balance between reliability, efficiency and adaptability in practice.

[0035] The flight control and data acquisition module receives the final flight path command from the path planning module. Its core function is to precisely control the UAV's flight control system, driving it to autonomously fly along a predetermined path. During this process, it controls the onboard optical or remote sensing payloads to continuously or selectively acquire image data of the railway corridor according to preset acquisition parameters. This acquired raw image data forms the basis for all subsequent analyses. To improve system response time and optimize data flow, an edge analysis module is deployed on the UAV itself or its tightly coupled onboard computing unit. "Edge" here refers to the computing location close to the data source. This module embeds an optimized, lightweight target detection algorithm model, enabling it to process the real-time image stream during UAV flight using onboard computing resources. Its task is to quickly analyze each frame or set of images, identifying potential areas that may indicate disasters or hazards, such as slope cracks, deformation of slope protection structures, or illegal accumulation of materials on the tracks. After identification is complete, the edge analysis module performs a crucial filtering operation: instead of transmitting all the original image data back to the backend, it extracts only a limited number of key image data fragments containing suspected disaster areas from the massive dataset based on the identification results. This process completes the first condensation from "all data" to "key data" at the edge of data generation.

[0036] The extracted key image data is transmitted via wireless communication network to a remote cloud-based intelligent analysis module. The cloud server possesses powerful storage capacity and high-performance computing resources. Upon receiving the current key image data, this module retrieves historical image data from the same area and similar periods from its database for professional Earth observation analysis, i.e., multi-period change detection. Through a series of image processing algorithms (such as image registration, difference analysis, and feature extraction), the module can quantitatively identify and extract surface change features from time-series image sequences. These features provide quantitative evidence of surface object movement, deformation, or changes in cover. Based on the morphology, scale, spatial distribution patterns, and temporal evolution of these extracted surface change features, the cloud-based intelligent analysis module uses pre-set or machine learning-based discrimination logic to further determine the types of hazards causing these changes, such as classifying them as "subsidence," "slippage," "water accumulation," or "foreign object intrusion." After the hazard type is determined, the risk assessment module is activated. This module operates based on remote sensing image inversion technology, extracting or calculating multi-dimensional, risk-related feature parameters from multiple images of the affected area. It then calls a pre-built feature fusion model (e.g., a weighted comprehensive evaluation model), using these multi-dimensional features as input for comprehensive calculation, ultimately outputting a quantified risk index. This index is a comprehensive numerical expression of the current hazard level. Finally, the early warning release module, based on the risk index output by the risk assessment module and referring to preset risk level classification standards, determines the specific risk level (e.g., "low risk," "medium risk," "high risk," "extremely high risk") corresponding to the hazard. Subsequently, the module automatically generates structured early warning information, typically including key elements such as hazard location coordinates, type, risk level, and discovery time. Through an integrated communication interface, the early warning information is published to designated monitoring platforms or the terminal devices of relevant responsible personnel, thus completing the entire automated closed loop from hazard discovery, analysis, assessment to information delivery.

[0037] In summary, this system ensures long-term data source consistency through the path planning module, laying the foundation for high-precision time-series analysis; automates operations through the flight control and data acquisition module; enables real-time front-end preprocessing of data through the edge analysis module, significantly improving data value density and system response speed; achieves in-depth identification and qualitative analysis of potential hazards through the cloud-based intelligent analysis module; enables quantitative risk assessment through the risk assessment module; and finally, achieves effective information output through the early warning release module.

[0038] The modules are sequentially connected and their functions complement each other, transforming the traditional periodic manual inspection mode into a continuous, intelligent, data-driven automated monitoring and early warning system. This effectively addresses the urgent need for long-term data comparability, real-time early warning, and intelligent analysis in railway corridor safety management.

[0039] In some implementations, the objective function of the baseline fixed route template satisfies the following relationship: ,in, Indicates a fixed baseline route template. Indicates candidate flight paths, Let N represent the historical inspection path for period i, and let N represent the number of historical inspection periods. This represents the geographical constraint cost function based on terrain slope and structure distribution, where α and β are weighting coefficients.

[0040] This embodiment further elaborates on the specific mathematical optimization model used to generate the baseline fixed flight path template, namely, the composition of its objective function. The objective function is designed to find an optimal flight path that can serve as a long-term benchmark from numerous historical flight paths through quantitative calculations. The objective function mainly consists of two parts added together, with weight coefficients adjusted to balance their importance. The first part is the "historical fit" term, which mathematically calculates the deviation between any candidate baseline fixed flight path template and the actual inspection paths executed in each historical period. Then, the average deviation over all historical periods is taken, and the algorithm seeks to minimize this average. Here, a candidate flight path refers to a path that the algorithm is evaluating during the optimization process that could potentially become a benchmark; the historical inspection path of period i is the actual trajectory of the UAV's flight in the past task i; N represents the total number of historical inspection periods. Minimizing the average deviation means that the algorithm tends to select a path that is spatially most "similar" or "closest" to all past flight trajectories as the benchmark. This allows the generated benchmark template to inherit and summarize reasonable experience and habitual flight paths from historical flights.

[0041] However, merely pursuing similarity to historical routes may be insufficient, as historical routes themselves may be influenced by specific conditions at the time and may not be entirely ideal or safe. Therefore, the objective function introduces a second part: the geographical constraint cost function. This function quantifies the "suitability" or "cost" of a given flight path in the actual geographical environment. Based on high-precision Geographic Information System (GIS) data, it calculates factors such as the terrain slope (excessively steep slopes may affect flight stability or image quality), the distance between the path and along-line structures (such as power towers, bridges, tunnel entrances, and stations), and converts these factors into a cost. A path that traverses complex terrain or is too close to obstacles will have a significantly higher geographical constraint cost. By incorporating this into the objective function and seeking to minimize it, the algorithm automatically avoids high-risk or unflyable areas during optimization, making the generated baseline fixed flight path template more spatially safe and reasonable.

[0042] α and β, as weighting coefficients, are key parameters that adjust the proportion of the two factors mentioned above in the final decision. α determines the algorithm's emphasis on the goal of "faithfulness to history," while β determines its emphasis on the goal of "compliance with geographical safety constraints." By reasonably setting the values ​​of α and β (e.g., based on domain expert experience or historical flight safety data training), the optimization direction can be precisely guided. Finally, by solving for the minimum value of this weighted objective function, the solution obtained is the optimal baseline fixed flight path template. This template mathematically achieves the best balance between generalizing historical flight patterns and adhering to geographical constraints. It is not a simple geometric mean, but a standardized flight corridor with theoretical basis generated after multi-objective optimization decision-making, providing a stable, reliable, and optimized spatial benchmark framework for all subsequent inspection tasks.

[0043] In some implementations, the spatiotemporal consistency constraint is used to limit the spatial position deviation of corresponding path points during different inspection periods to satisfy the following relationship:

[0044] ,in, This represents the inspection path for period t. This represents the average path across multiple periods. The preset path variance threshold;

[0045] The local adaptive fine-tuning is achieved in the following way:

[0046] ,in, The adjusted flight path, As a baseline fixed route template, This is the path adjustment vector calculated based on real-time weather indices and obstacle indices. For sensitivity parameters, , Used to limit the magnitude of path adjustment.

[0047] This embodiment provides specific mathematical models and implementation methods for the two abstract concepts of spatiotemporal consistency constraints and local adaptive fine-tuning, giving them clear operability and quantifiable boundaries. The core of spatiotemporal consistency constraints lies in controlling the dispersion between flight paths of different periods, implemented through a mathematical relationship based on the concept of variance statistics. Specifically, the inspection path of period t represents the actual flight trajectory of the UAV when performing the inspection task at time point t, which consists of a series of ordered spatial coordinate points. The average path of multiple periods is not a real flight path, but a virtual statistical center line, obtained by calculating the average of the spatial coordinates of path points of the same order in all historical and current inspection paths. The constraint requires that for any inspection path, the sum of the squares of the spatial position deviations of all path points on it from the corresponding points on the average path must be less than a pre-set path variance threshold. This threshold is a strict technical standard, and its value directly determines the range of spatial fluctuations allowed for each period's flight path. By enforcing this inequality constraint, it is ensured that the flight trajectory of each period is strongly "anchored" within a narrow statistical tolerance band around the average path determined by the paths of all periods. This mathematically enforces the long-term spatial stability of the flight trajectory, which is a fundamental technical measure to ensure the high geometrical consistency of multi-period image data and meet the requirements of high-precision registration.

[0048] When flight conditions change and route adjustments are necessary, a local adaptive fine-tuning mechanism defines how controlled and limited path modifications are made. This mechanism clearly describes how the adjusted flight path is constructed through a mathematical expression. The adjusted flight path is the sum of a baseline fixed route template and an adjustment amount. This adjustment amount is determined by two parts: a path adjustment vector and a sensitivity parameter. The path adjustment vector is a directional value calculated based on real-time input meteorological indices (such as wind speed and visibility combined into a risk index) and obstacle indices (such as the risk vector formed by the position and size of obstacles). It indicates the theoretical direction and magnitude in which points on the baseline path should be moved to cope with the current environment. However, to avoid excessive and destructive changes to the path due to strong environmental interference signals, the system introduces a sensitivity parameter to dynamically modulate the adjustment magnitude. This sensitivity parameter is designed as a function of the magnitude of the path adjustment vector. Its key characteristic is that it is a function that decreases as the magnitude of the adjustment vector increases. This means that when the environmental interference detected by the sensor is small (small path adjustment vector magnitude), the sensitivity parameter is relatively large, allowing for relatively "sensitive" fine-tuning close to the theoretical value. When the environmental interference signal is very strong (large path adjustment vector magnitude), the sensitivity parameter automatically decreases, thereby suppressing the theoretical adjustment range and limiting the final applied adjustment to a reasonable range. The entire fine-tuning process is also constrained by a maximum adjustment range, setting an insurmountable absolute upper limit for path deviation. Therefore, local adaptive fine-tuning is an intelligent adjustment strategy with a built-in negative feedback mechanism. It enables the system to respond to dynamic environments as necessary, while ensuring that any adjustment is local, small-amplitude, and controlled through the adjustment of the sensitivity parameter and the upper limit of the amplitude. This steadfastly maintains the core value of multi-period data comparability guaranteed by spatiotemporal consistency constraints during long-term operation and maintenance.

[0049] In some implementations, when identifying suspected disaster areas, the edge analysis module generates label information containing spatial location and confidence level for each suspected disaster area, and extracts and uploads the corresponding image data only when the confidence level is higher than a preset threshold.

[0050] This embodiment significantly refines the working mechanism of the edge analysis module, introducing a confidence-based decision-making logic, which is crucial for achieving intelligent data filtering and improving the overall efficiency and reliability of the system. When the edge analysis module runs its target detection algorithm to analyze real-time acquired images, its output not only marks the bounding boxes of suspected disaster areas but also calculates and associates a numerical attribute called "confidence" for each identified suspected area. Confidence is a probability value or score generated internally by the target detection algorithm model, typically ranging from 0 to 1 or expressed as a percentage. It comprehensively reflects the algorithm's certainty, based on learned features, that a target object (i.e., a specific type of disaster or hazard) exists in the current image block. For example, if the features of an area in the image highly match the features of landslide signs and the image is clear, the confidence may be close to 1; if the features are blurry or ambiguous, the confidence may be lower. Generating label information containing spatial location and confidence means that each circled suspected object is accompanied by two types of core metadata: where it is (location) and how similar it is (confidence).

[0051] The edge analysis module does not indiscriminately process all identification results, but rather performs a filtering decision based on a preset threshold. System administrators or algorithm developers pre-set a confidence threshold based on actual application needs and tolerance for false alarms. Before deciding whether to upload data for a suspected area to the cloud, the module compares the confidence value in the area's tagging information with this preset threshold in real time. The decision rule is: only when the confidence of a suspected disaster area is higher than the preset threshold is the identification result deemed sufficiently reliable to warrant further analysis using subsequent network transmission and cloud computing resources. At this point, the edge analysis module triggers two actions: first, it precisely extracts the image data block containing the high-confidence suspected area (i.e., key image data) from the original image data stream; second, it packages this image data block along with its corresponding tagging information (spatial location, confidence, etc.) and uploads it to the cloud-based intelligent analysis module via the communication link. Conversely, for suspected areas with a confidence level lower than or equal to the preset threshold, the edge analysis module filters them out directly, neither extracting the corresponding image data nor uploading it. This mechanism has significant technical advantages. First, it implements "pre-load filtering" of the data stream, eliminating a large number of low-confidence, potentially false alarm, and redundant data at the source of data generation, greatly saving wireless communication bandwidth. This is particularly suitable for scenarios along railway lines where network bandwidth may be limited or transmission costs may be high. Second, it effectively reduces the data processing pressure on cloud servers, allowing the cloud to concentrate its powerful computing capabilities on accurately performing in-depth change detection and detailed analysis of high-confidence, high-potential-risk targets, improving the efficiency and targeting of the entire system's analysis work. Finally, this is also a crucial step in improving the overall accuracy of the system's output warning information. By suppressing low-confidence suspected targets from entering subsequent processes at the edge, it can significantly reduce the downstream transmission of false alarms caused by fuzzy algorithm recognition, making the warning information generated by the warning release module more credible and practical, and reducing the ineffective verification work of maintenance personnel. Therefore, confidence-based intelligent screening is the core manifestation of the evolution of edge computing from "simple recognition" to "possessing preliminary judgment and decision-making capabilities," and is a crucial technical link in building an efficient, economical, and reliable edge-cloud collaborative processing architecture.

[0052] In some implementations, the surface change characteristics include at least one of the following or a combination thereof: surface displacement, surface displacement rate, area growth rate of change zone, spatial clustering, and temporal evolution acceleration;

[0053] The types of hazards are determined by a machine learning classification model, which identifies the types of hazards based on the morphological features, spatial distribution features, and temporal evolution features of surface change characteristics.

[0054] This embodiment further specifies the methods for determining surface change features and hazard types extracted by the cloud-based intelligent analysis module. Surface change features refer to a series of physical quantities or indicators that can characterize changes in surface morphology, location, or attributes, quantitatively derived from multi-period remote sensing image data through professional Earth observation analysis methods. These features do not exist in isolation but constitute a multi-dimensional feature set to comprehensively describe an area of ​​abnormal change. Among them, surface displacement refers to the linear distance or vector change in the spatial location of a specific point or area in images at two different time points. It can be calculated using high-precision image matching technology or synthetic aperture radar interferometry technology, and it directly reflects the absolute magnitude of surface movement. Surface displacement rate is the rate of change of surface displacement relative to the time span, i.e., the magnitude of displacement per unit time. It describes the speed of surface deformation and is a key dynamic indicator for judging the intensity of disaster activity. The rate of change in the area focuses on the extent of the identified anomalous change areas themselves. It measures the speed at which the boundaries of these areas expand outward or contract inward over time, effectively indicating the evolution trend of the disaster's impact range, such as the extension of cracks at the rear edge of a landslide or the expansion of a subsidence basin. Spatial clustering is a statistical measure used to describe the degree of concentration of multiple independent change points or small change areas in spatial distribution. High clustering means that the change points are closely clustered in space, possibly indicating a contiguous, integrated unstable area; while low clustering may indicate that the changes are scattered and isolated, with potentially different causes and risks. Temporal evolution acceleration, based on displacement rate, further analyzes the trend of the rate value itself over time, i.e., whether the rate is increasing, decreasing, or remaining stable. This helps to capture the nonlinear characteristics of the deformation process and precursory signals of accelerated destruction. These features, from multiple independent yet interrelated dimensions such as spatial amplitude, temporal dynamics, extent evolution, and distribution patterns, collectively construct a refined and digital characterization of surface anomalies.

[0055] Based on the extracted multidimensional surface change features, the cloud-based intelligent analysis module employs a machine learning classification model to automatically determine the type of hazard. A machine learning classification model is a data-driven artificial intelligence algorithm that can automatically establish a complex mapping relationship between input features and output categories by learning and summarizing the inherent patterns in a large number of known samples. In the implementation of this system, a training dataset is first required. This dataset contains various disaster cases (such as landslides, collapses, debris flow accumulation, roadbed settlement, and slope crack development) that have occurred historically in railway corridor areas and have been verified on-site. For each historical case, a corresponding multi-period image sequence needs to be prepared, and a set of surface change feature values ​​(such as displacement, rate, area growth rate, aggregation degree, and acceleration) needs to be extracted from these images. These feature values ​​constitute the model's input vector; simultaneously, each case has a verified and clearly defined disaster type label, which serves as the target output for model training. Using this training dataset, an appropriate classification algorithm model can be selected and trained, such as a support vector machine, random forest, or neural network. During training, the model automatically learns typical patterns and differences in surface change characteristics corresponding to different disaster types, including morphological features (such as the consistency of displacement vector direction patterns), spatial distribution features (such as whether change points are distributed along a geological structural line or slope trend line), and temporal evolution features (such as whether the displacement rate shows a phased acceleration). The trained model then possesses the ability to identify disaster types based on newly input feature combinations. In actual inspection and analysis, when the cloud-based intelligent analysis module extracts a set of surface change features for a suspected hazard area from the current imagery, it inputs this feature set into the pre-trained machine learning classification model. The model calculates based on its internally learned complex discrimination rules and ultimately outputs a prediction result indicating the most likely type of hazard, such as "potential landslide," "gradual subsidence," or "local collapse." This method avoids the limitations of relying on manually set single, fixed thresholds for judgment, and can handle complex, non-linear interactions between multiple features, thus significantly improving the accuracy, robustness, and generalization ability to new and unknown patterns in automatic hazard type identification.

[0056] Therefore, clearly defining a series of multi-dimensional features, from surface displacement to temporal evolution acceleration, provides a comprehensive and in-depth data foundation for the quantitative description of potential hazards. This allows the analysis to move beyond the vague level of "change" and delve into the precise level of "how change occurs." Applying machine learning classification models to determine hazard types represents an evolution from rule-based expert systems to data-driven intelligent decision-making. By training with historical disaster data, the system can mimic and even surpass the ability of human experts to synthesize multiple clues for judgment, achieving the goal of automated and intelligent classification of complex and diverse railway corridor hazards. This step is crucial in transforming raw image difference data into conclusions with clear engineering geological significance and risk orientation, laying an essential foundation for subsequent targeted and categorized risk assessments, and greatly enhancing the depth and practicality of the entire system's analysis.

[0057] In some implementations, the risk assessment module employs a multi-feature remote sensing risk assessment model. Based on the type of potential hazard, it extracts multi-dimensional features from multiple remote sensing images and calculates the risk index through weighted fusion. The calculation method satisfies the following relationship:

[0058] ,in, Here, V is the risk index, A is the standardized surface displacement rate, S is the area growth rate of the changed area, D is the spatial clustering degree, and w1, w2, w3, and w4 are weighting coefficients. The weighting coefficients are determined based on training with historical disaster sample data, and different hazard types correspond to different weighting combinations. V, A, S, and D are all dimensionless parameters after standardization.

[0059] This embodiment details the specific calculation logic of the multi-feature remote sensing risk assessment model used in the risk assessment module, specifically how the risk index is generated by weighted fusion of multiple features. The multi-feature remote sensing risk assessment model is a mathematical model designed to integrate multiple independent indicators representing different aspects of risk into a single quantitative value that comprehensively reflects the degree of danger of a potential hazard. The model first guides the selection of features and weighting based on the hazard type identified by the cloud-based intelligent analysis module. For example, for hazards like "landslides," the model might focus on features such as surface displacement rate, spatial clustering, and temporal evolution acceleration; while for "ground subsidence," it might emphasize features such as surface displacement and the growth rate of the changed area. The core calculation formula of the model is a weighted summation: the risk index R equals the sum of the products of each selected, standardized feature value and its corresponding weight coefficient. Specifically, the formula is expressed as follows: In this formula, V, A, S, and D represent the standardized surface displacement rate, the rate of increase in the area of ​​change, the spatial clustering, and the temporal evolution acceleration, respectively. Standardization is a crucial data preprocessing step, aiming to eliminate the incomparability between different features caused by differences in their original dimensions (e.g., mm / year, km² / year) and numerical ranges. Common standardization methods include min-max normalization or Z-score standardization. After processing, the value of each feature is transformed to a uniform, dimensionless scale (e.g., between 0 and 1 or conforming to a standard normal distribution), thus allowing feature values ​​from different physical meanings to be weighted fairly.

[0060] Weighting coefficient , , , These are the core parameters in the model, and their magnitudes directly determine the relative importance or contribution of each feature in the final risk index. These weight coefficients are not subjectively determined but objectively determined through a training process based on historical disaster sample data. The training process requires collecting a batch of historical disaster cases. Each case not only needs its corresponding, unstandardized, original multidimensional feature values ​​but also a "label" or "target value" that represents its true risk level. This target value can be the actual loss level caused by the disaster, the emergency response level, or a risk score comprehensively assessed by experts based on the case details. Using this sample data, mathematical optimization algorithms (such as multiple linear regression, gradient descent, etc.) are used to fit and find a set of optimal weight coefficients, minimizing the overall error (such as mean square error) between the model's predicted risk index calculated from these weights and the sample's true risk label. More importantly, different hazard types correspond to different weight combinations. This means that during model training or system configuration, a separate set of independent weight coefficient combinations needs to be trained or set for each predefined hazard type (such as landslide, collapse, settlement, etc.). For example, in the landslide risk model, the weight of the surface displacement rate V... It may be overweighted because rapid displacement is a key indicator of impending landslides; while in the risk model of slow settlement, the area growth rate A has a higher weight. This may be even more pronounced, as the continuous expansion of the scope often means that the impact is intensifying. This "one policy for one type" weighting mechanism enables the risk assessment model to accurately reflect the differences in the inherent physical mechanisms of risk composition among different types of disasters, thereby greatly improving the scientific rigor and accuracy of the assessment results.

[0061] Ultimately, the risk index R calculated by this model is a comprehensive scalar value that integrates specific types of hazards, multiple key feature intensity information, and their relative importance judgments. It condenses complex information originally scattered across a multi-dimensional feature space into a single, intuitive value that can be mathematically compared and ranked. This risk index provides an objective and unified quantitative benchmark for subsequent risk level classification. It enables the system to conduct cross-object horizontal risk comparison and ranking of hazards of different geographical locations, types, and stages of development, achieving a key transformation from multi-dimensional feature analysis to single-dimensional risk decision-making. This quantitative risk assessment method, which uses historical data training to determine weights and performs multi-feature weighted fusion, greatly reduces subjective arbitrariness and improves the transparency, repeatability, and precision of the risk assessment process, providing direct and powerful scientific decision-making support for railway maintenance departments to conduct priority ranking, differentiated resource allocation, and precise early warning response.

[0062] In some implementations, the risk level is determined by comparing the risk index with a preset threshold range.

[0063] This embodiment clarifies how the risk level is ultimately determined based on the risk index calculated by the risk assessment module. This is a crucial step in mapping continuous numerical risk assessment results to discrete, easily understood, and operable warning levels. This determination process is essentially a classification decision based on threshold interval comparison. Predefined threshold intervals are a series of numerical ranges predefined during system deployment, calibration, or strategy formulation, with each interval corresponding to a specific risk level label. The setting of these threshold intervals requires comprehensive consideration of multiple factors, including statistical analysis of the relationship between risk indices and disaster consequences in historical disaster cases, relevant safety operation norms and standards in the railway industry, management's acceptability or tolerance for different risk levels, and the numerical distribution range output by the risk assessment model itself. For example, after expert evaluation and model validation, the risk index R can be defined as follows: when the value of R is in the range of [0, 0.3], it is defined as "low risk" (or "blue alert" level); when R is in the range of (0.3, 0.6], it is defined as "medium risk" (or "yellow alert" level); when R is in the range of (0.6, 0.8], it is defined as "high risk" (or "orange alert" level); and when R is in the range of (0.8, 1], it is defined as "extremely high risk" (or "red alert" level). The division of the range can be linear or non-linear; the number of levels can be set to three, four, or more according to the needs of refined management.

[0064] During system operation, whenever the risk assessment module calculates a specific risk index R value for a potential hazard, the early warning release module (or a software unit integrating this logic) immediately initiates this comparison process. This process involves a simple numerical judgment logic: the program sequentially compares the value of R with the upper and lower boundaries of various preset threshold intervals stored in the database to determine which consecutive numerical interval R falls into. Once a matching interval is found, the program assigns a pre-bound risk level label (such as "high risk") to the currently being addressed hazard. This determined risk level is one of the core pieces of information produced by the entire intelligent inspection system after a series of complex calculations, including data collection, edge detection, cloud-based in-depth analysis, and risk assessment. It is the final, most concise, and most directly relevant information for user decision-making.

[0065] Based on this defined risk level, the subsequent generation and dissemination strategies for early warning information by the early warning release module have a clear basis. The risk level directly determines the urgency, detail, format, scope and recipients of the early warning information, as well as the recommended initial response measures. For example, for a hazard deemed "high-risk," the system may automatically trigger the highest level alarm, generating a detailed early warning report including precise location, on-site snapshot, and recommended handling measures. This report will be simultaneously pushed to relevant personnel such as line maintenance managers and safety inspectors through multiple channels, including SMS, application push notifications, and monitoring screens, and may even initiate emergency response procedures. For "low-risk" hazards, the system may only list them in the daily or weekly inspection summary report, suggesting routine observation, without requiring immediate emergency action. This level determination method based on preset threshold ranges establishes a standardized and automated decision-making conversion mechanism. It efficiently and accurately converts the numerical results generated by complex, quantitative analysis models in the backend into the existing, qualitative risk management language and action guidelines in the front-end management system. This not only enables operations and management personnel unfamiliar with technical details to quickly understand and utilize the results of intelligent analysis, but also ensures that the output of the entire system can seamlessly integrate with existing safety production management systems and emergency plans that respond based on risk levels. Simultaneously, the preset threshold ranges provide an interface for system adaptability and strategy adjustments. Managers can prudently adjust the thresholds based on the importance of different routes, seasonal climate characteristics, or changes in policy requirements, thereby achieving flexible and macro-level control over the overall risk prevention and control scale without altering the core algorithm.

[0066] Example 2

[0067] Secondly, the present invention proposes a low-altitude intelligent inspection device for railway corridors, comprising a drone, a ground relay station and a cloud server, wherein the device is configured to run a low-altitude intelligent inspection system for railway corridors proposed in any of the above embodiments.

[0068] This device is a system-level entity composed of specific hardware components working together according to their functional divisions. It includes three core parts: a drone, a ground relay station, and a cloud server. The device is configured to run the railway corridor low-altitude intelligent inspection system proposed in any of the above embodiments. This means that all software functional modules and algorithm processes in the above embodiments are deployed and run on this physical hardware device, and are implemented through the cooperation between the various components.

[0069] Unmanned aerial vehicles (UAVs) serve as aerial mobile platforms and front-end sensing and computing nodes for a system. They typically consist of an aircraft platform, a flight control system, a high-precision global navigation satellite system (GNSS) receiver, a mission payload, and onboard edge computing devices. The functions of the flight control and data acquisition module are primarily executed by the UAV's flight control system and mission payload. The flight control system receives and interprets flight path commands generated by the path planning module, controlling the UAV to achieve autonomous takeoff and landing, route tracking, and hovering. The mission payload mainly refers to devices such as optical cameras, multispectral sensors, or lidar used to acquire image data; these devices operate automatically according to commands during flight. Meanwhile, the software for the edge analysis module is deployed on embedded edge computing devices (such as high-performance onboard computers) carried by the UAV. This equipment processes the raw image data transmitted from the cameras in real time during flight, performing target detection and key data extraction tasks. Ground relay stations are auxiliary facilities deployed at fixed points along railway lines, primarily undertaking functions such as communication enhancement, data relay, energy supply, and local monitoring. In long-distance or complex terrain (such as tunnel complexes or mountainous areas) inspection missions, direct wireless communication (such as 4G / 5G) between drones and the remote cloud may be unstable or have blind spots. Ground relay stations, equipped with long-range radio communication equipment or connected to wired networks, can act as communication relay nodes, receiving data (including key images and status information) sent by drones and reliably forwarding it to the cloud server; simultaneously, they can reliably transmit control commands and path information from the cloud to the drones. Some ground relay stations can also be designed as automatic take-off and landing pads and charging docks for drones, enabling automatic energy replenishment and supporting unmanned operation. The cloud server is the data processing and command center of the device, usually located in a remote data center, and consists of high-performance computing clusters, large-capacity storage systems, and network equipment. The core algorithms and applications of the path planning module, cloud intelligent analysis module, risk assessment module, and early warning release module are all deployed and run on the cloud server. The cloud server is responsible for receiving and storing key image data and metadata uploaded from various inspection terminals (via ground relay stations), scheduling powerful computing resources to perform intensive computing tasks such as multi-period change detection of massive data, machine learning model inference, and risk assessment calculation, and managing all user interfaces, databases, and early warning information release interfaces.

[0070] Drones, ground relay stations, and cloud servers are organically connected through wireless communication networks (which may include private networks, cellular networks, satellite communications, etc.), forming a complete "air-ground-cloud" collaborative operation system. Drones act as flexible "airborne sentinels," responsible for performing specific reconnaissance and preliminary intelligence processing tasks; ground relay stations act as robust "frontline communication fortresses," ensuring the smooth flow and extension of information transmission links; and cloud servers act as powerful "rear command and intelligent brains," conducting overall scheduling, in-depth analysis, and decision-making.

[0071] This system deeply integrates specific hardware infrastructure with the software logic defined in Example 1, transforming low-altitude intelligent inspection of the entire railway corridor from a conceptual approach into an engineering solution that can be practically deployed, operated, and maintained. It clarifies the physical infrastructure and resource configuration required for implementing the system, providing complete and feasible equipment support for achieving routine, automated, and intelligent low-altitude inspection and risk warning in a real railway operating environment.

[0072] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A low-altitude intelligent inspection system for railway corridors, characterized in that, include: The path planning module automatically generates flight paths using a multi-phase adaptive fixed route optimization algorithm. This includes generating a baseline fixed route template for multi-phase inspections based on flight path data from historical inspection phases and geographic information system data of the railway corridor. In subsequent inspection tasks, flight paths are generated using this baseline fixed route template as a constraint. The baseline fixed route template is generated by minimizing the average deviation between candidate paths and historical inspection paths. Furthermore, a spatiotemporal consistency constraint is introduced when generating flight paths to ensure that the spatial position deviation of corresponding path points and the deviation of image acquisition parameters in different inspection phases do not exceed a preset threshold, thus guaranteeing that the multi-phase image data meets pixel-level or sub-pixel-level registration requirements. When acquiring real-time meteorological data or obstacle information, the path planning module adjusts local route segments in the baseline fixed route template only under the premise of satisfying the path offset penalty constraint, and limits the maximum offset of the adjusted flight path relative to the baseline fixed route template to avoid global reconstruction of the overall flight path. The flight control and data acquisition module is configured to control the UAV to fly autonomously according to the flight path and to acquire multiple phases of image data of the railway corridor during the flight. The edge analysis module, deployed on the drone, is configured to perform real-time target detection on the image data, identify suspected disaster areas, and extract key image data containing suspected disaster areas based on the identification results. The cloud-based intelligent analysis module is configured to perform multi-period change detection on the key image data and historical image data from the same period, extract surface change features, and determine the type of hidden danger based on the surface change features. The risk assessment module is configured to, after determining the type of hazard, calculate a risk index based on the multidimensional features obtained from remote sensing image inversion using a feature fusion model, and determine the risk level of the hazard based on the risk index; The early warning release module is configured to generate and release early warning information based on the risk level.

2. The intelligent low-altitude inspection system for railway corridors according to claim 1, characterized in that, The objective function of the baseline fixed route template satisfies the following relationship: ,in, Indicates a fixed baseline route template. Indicates candidate flight paths, Let N represent the historical inspection path for period i, and let N represent the number of historical inspection periods. This represents the geographical constraint cost function based on terrain slope and structure distribution, where α and β are weighting coefficients.

3. The intelligent low-altitude inspection system for railway corridors according to claim 1, characterized in that, The spatiotemporal consistency constraint is used to limit the spatial position deviation of corresponding path points in different inspection periods to satisfy the following relationship: ,in, This represents the inspection path for period t. This represents the average path across multiple periods. The preset path variance threshold; Local adaptive fine-tuning is achieved in the following way: ,in, The adjusted flight path, As a baseline fixed route template, This is the path adjustment vector calculated based on real-time weather indices and obstacle indices. For sensitivity parameters, , Used to limit the magnitude of path adjustment.

4. The intelligent low-altitude inspection system for railway corridors according to claim 1, characterized in that, When identifying suspected disaster areas, the edge analysis module generates label information containing spatial location and confidence level for each suspected disaster area, and extracts and uploads the corresponding image data only when the confidence level is higher than a preset threshold.

5. The intelligent low-altitude inspection system for railway corridors according to claim 1, characterized in that, The surface change characteristics include at least one of the following or a combination thereof: surface displacement, surface displacement rate, area growth rate of change zone, spatial clustering, and temporal evolution acceleration. The types of hazards are determined by a machine learning classification model, which identifies the types of hazards based on the morphological features, spatial distribution features, and temporal evolution features of surface change characteristics.

6. The intelligent low-altitude inspection system for railway corridors according to claim 1, characterized in that, The risk assessment module employs a multi-feature remote sensing risk assessment model, extracting multi-dimensional features from multiple remote sensing images based on the type of potential hazard, and calculating the risk index through a weighted fusion method. The calculation method satisfies the following relationship: ,in, Here, V is the risk index, A is the standardized surface displacement rate, S is the area growth rate of the changed area, D is the spatial clustering degree, and w1, w2, w3, and w4 are weighting coefficients. The weighting coefficients are determined based on training with historical disaster sample data, and different hazard types correspond to different weighting combinations. V, A, S, and D are all dimensionless parameters after standardization.

7. The intelligent low-altitude inspection system for railway corridors according to claim 1, characterized in that, The risk level is determined by comparing the risk index with a preset threshold range.

8. A low-altitude intelligent inspection device for railway corridors, characterized in that, The device includes a drone, a ground relay station, and a cloud server, and is configured to operate a low-altitude intelligent inspection system for railway corridors as described in any one of claims 1 to 7.

Images