Unmanned aerial vehicle-based tower line safety hazard monitoring method and system

Abstract

The application provides a tower line safety hidden danger monitoring method and system based on a UAV, relates to the technical field of safety monitoring, and comprises the following steps: acquiring tower three-dimensional model data, determining a cooperative inspection path; establishing a hierarchical collection strategy, dynamically distributing collection density and frequency according to a deformation sensitive area; collecting multi-view depth images to construct a real-time point cloud model and extract deformation feature points; performing Kalman filtering dynamic tracking on key monitoring feature points, and generating a warning when a threshold is exceeded. The application improves monitoring precision and efficiency, and realizes early warning of tower safety hidden dangers.

Description

Technical Field

[0001] This invention relates to safety monitoring technology, and more particularly to a method and system for monitoring safety hazards along railway towers based on unmanned aerial vehicles (UAVs). Background Technology

[0002] As a crucial component of transmission networks, the structural safety and stability of transmission towers directly impact the reliable operation of the system. Traditional tower inspections rely primarily on manual methods, requiring inspectors to climb to great heights to examine the tower structure. This approach is not only inefficient but also poses significant safety hazards. With the rapid development of drone technology, drone-based tower inspections are gradually becoming a focus of industry attention. Drones, equipped with high-definition cameras or other sensing devices, can observe and collect data from towers from multiple angles, providing more comprehensive information for assessing the safety status of transmission towers.

[0003] Current drone inspection technologies typically employ a single-drone operation mode, lacking a mechanism for multi-drone collaborative work. Single-drone inspections struggle to acquire comprehensive information about the towers in a short time, resulting in low inspection efficiency and the potential to overlook safety hazards in critical areas, thus affecting the comprehensiveness and accuracy of the inspection.

[0004] Existing drone inspection methods often employ a uniform data collection strategy, neglecting the structural characteristics and stress variations of different parts of the tower. This method fails to focus on monitoring deformation-sensitive areas of the tower, resulting in blind spots or redundancy in data collection and making it difficult to accurately capture critical deformation information of the tower.

[0005] Existing technologies primarily focus on visible anomalies on the tower surface, such as static defects like loosening, detachment, or corrosion, lacking dynamic monitoring and early warning mechanisms for tower structural deformation. This makes it impossible to effectively identify and track the development trend of structural deformation, hindering early warning of potential safety hazards and thus reducing the preventative and proactive nature of inspections. Summary of the Invention

[0006] This invention provides a method and system for monitoring safety hazards along railway towers based on unmanned aerial vehicles (UAVs), which can solve the problems in the prior art.

[0007] A first aspect of the present invention provides a method for monitoring safety hazards along power transmission towers based on unmanned aerial vehicles (UAVs), comprising:

[0008] Acquire 3D model data of the target tower, and determine the collaborative inspection path and data collection location of multiple drones based on the 3D model data; control multiple drones to fly to the corresponding data collection location according to the collaborative inspection path, wherein each drone is equipped with a depth camera;

[0009] Based on the structural stress analysis results of the target tower, a layered acquisition strategy is established, dividing the target tower into multiple deformation-sensitive regions, and dynamically allocating the acquisition density and acquisition frequency of the UAV according to the stress characteristics of each deformation-sensitive region; and controlling multiple UAVs to simultaneously acquire multi-view depth images of the target tower according to the layered acquisition strategy.

[0010] Based on the multi-view depth images, a real-time 3D point cloud model of the target tower is constructed. The real-time 3D point cloud model is registered with the 3D model data of the target tower, deformation feature points are extracted, and the deformation feature points are graded and screened using a dual threshold iteration method. Based on the screening results, key monitoring feature points are determined.

[0011] The key monitoring feature points are dynamically tracked using a Kalman filter algorithm to obtain the motion trajectory of the deformation feature points. When the displacement of the motion trajectory exceeds a preset evaluation threshold, an early warning message is generated.

[0012] Based on the 3D model data, the collaborative inspection path and data collection location of multiple UAVs are determined; controlling the multiple UAVs to fly to the corresponding data collection location according to the collaborative inspection path includes:

[0013] Based on the three-dimensional model data, a target spatial feature model is constructed, and the structural features of the tower are analyzed according to the target spatial feature model. Based on the structural features, inspection points are determined, and the inspection points are mapped to three-dimensional space to form a set of collection locations. Based on the set of collection locations, a collaborative inspection path for multiple drones is generated.

[0014] A distributed communication network is constructed for multiple UAVs, and a communication topology matrix is ​​generated based on the spatial distribution of the multiple UAVs; the position vector, velocity vector, and obstacle perception information of each UAV are obtained through the distributed communication network, and the position vector, velocity vector, and obstacle perception information are shared in real time among the multiple UAVs;

[0015] Based on the obstacle perception information, a target point gravitational potential field term is constructed, and an obstacle repulsive potential field term and a cooperative obstacle avoidance potential field term are constructed based on the position vector. The target point gravitational potential field term, the obstacle repulsive potential field term, and the cooperative obstacle avoidance potential field term are weighted and combined to form a comprehensive potential field function.

[0016] The gradient of the integrated potential field function is calculated, and the gradient is multiplied by the step size factor to obtain the path adjustment amount. The collaborative inspection path is dynamically optimized according to the path adjustment amount to obtain an optimized inspection path that meets the safety distance requirements between UAVs and the connectivity requirements of the communication network. Multiple UAVs are controlled to fly to the corresponding collection positions according to the optimized inspection path.

[0017] A real-time 3D point cloud model of the target tower is constructed based on the multi-view depth images. The real-time 3D point cloud model is registered with the 3D model data of the target tower, and deformation feature points are extracted, including:

[0018] The stress distribution field of the tower at different heights and wind speeds is calculated to generate a stress distribution map. A three-dimensional stress variation curve is constructed based on the stress distribution map. The stress gradient is calculated in the three-dimensional stress variation curve, and the area with the largest stress gradient is identified as the key monitoring area.

[0019] Multiple depth cameras are deployed in the key monitoring area, and structural feature points are extracted from the multi-view depth images using the depth cameras.

[0020] The spatial location information of the structural feature points is fused with the stress distribution map to establish a correspondence between the feature point locations and the stress distribution, and a weighting coefficient is assigned to the structural feature points according to the correspondence.

[0021] Based on the weighting coefficients, the multi-view depth image is registered, and the registration accuracy requirement is improved in areas where the stress is greater than the preset registration threshold to generate a real-time three-dimensional point cloud model. The structural feature points in the real-time three-dimensional point cloud model are compared with the corresponding positions in the standard three-dimensional model, and the displacement direction and displacement amplitude of each structural feature point are calculated.

[0022] Based on the displacement direction and displacement amplitude, a connectivity analysis is performed on the structural feature points to identify deformation-related point groups. Feature points in the deformation-related point groups that do not conform to the stress distribution law are removed to obtain the final deformation feature points.

[0023] Based on the displacement direction and displacement amplitude, connectivity analysis is performed on the structural feature points to identify deformation-related point groups. Feature points in the deformation-related point groups that do not conform to the stress distribution law are removed, resulting in the final deformation feature points, including:

[0024] Based on the displacement direction and displacement amplitude, a displacement feature vector is constructed. The Euclidean distance and spatial distance between the displacement feature vectors are calculated. A feature correlation matrix is ​​constructed by combining the Euclidean distance and the spatial distance. The connectivity strength of the structural feature points is calculated based on the feature correlation matrix. The structural feature points are clustered based on the connectivity strength to obtain multiple deformation correlation point groups.

[0025] Calculate the mean and deviation of the displacement feature vectors of the structural feature points in the deformation-related point group, and construct the deformation distribution features by combining the mean and deviation of the displacement feature vectors.

[0026] Obtain stress distribution data of the tower, calculate stress values ​​and stress gradients at structural feature points based on the stress distribution data, construct stress distribution features using the stress values ​​and stress gradients, input the stress distribution features and deformation distribution features into a scoring function, and filter the structural feature points based on the output value of the scoring function to obtain the final deformation feature points.

[0027] The deformation feature points are hierarchically screened using a dual-threshold iteration method. Based on the screening results, the key monitoring feature points are determined to include:

[0028] An initial threshold and a target threshold are constructed. The initial threshold is used to perform a first screening of deformation feature points to obtain candidate feature points. The deformation amplitude distribution of the candidate feature points is calculated. The target threshold is dynamically adjusted according to the deformation amplitude distribution.

[0029] Using the target threshold as the iterative target, the candidate feature points are progressively screened. When the screening results are stable, the candidate feature points that are greater than the target threshold are identified as key monitoring feature points.

[0030] The key monitoring feature points are dynamically tracked using a Kalman filter algorithm to obtain the motion trajectory of the deformation feature points. When the displacement of the motion trajectory exceeds a preset evaluation threshold, an early warning message is generated, including:

[0031] The three-dimensional coordinates and displacement velocity of key monitoring feature points are obtained. The three-dimensional coordinates and displacement velocity are used to construct a state vector of the key monitoring feature points. Based on the state vector, historical state sampling is performed on the key monitoring feature points to obtain a historical state sampling sequence.

[0032] A spatiotemporal sequence matrix is ​​constructed based on the historical state sampling sequence. The spatiotemporal sequence matrix is ​​analyzed using a spatiotemporal feature extraction operator to obtain the evolution pattern features of the key monitoring feature points. The evolution pattern features are then used as correction parameters for the Kalman filter algorithm.

[0033] The state vector is input into the Kalman filter algorithm for state prediction. The state prediction result is corrected based on the evolutionary pattern characteristics to obtain the predicted state vector. Real-time observation data of the key monitoring feature points are obtained. The predicted state vector and the real-time observation data are weighted and fused to obtain the motion trajectory of the key monitoring feature points.

[0034] The displacement between adjacent sampling points in the motion trajectory is calculated, and the standard deviation and mean of the displacement are obtained by statistical analysis. A preset evaluation threshold is constructed based on the standard deviation and the mean. When the displacement is greater than the preset evaluation threshold, an early warning message is generated.

[0035] A spatiotemporal sequence matrix is ​​constructed based on the historical state sampling sequence. The spatiotemporal sequence matrix is ​​then analyzed using a spatiotemporal feature extraction operator to obtain the evolutionary pattern features of the key monitoring feature points, including:

[0036] The historical state sampling sequence is slide-divided according to the preset time window length, and the state vectors in each time window are arranged in time sequence to construct a spatiotemporal sequence matrix.

[0037] The temporal correlation of the state vector is calculated based on the spatiotemporal sequence matrix. The temporal correlation is characterized by the covariance between the state vector and its mean to obtain the temporal correlation feature. The position difference and velocity difference of adjacent state vectors in the spatiotemporal sequence matrix are calculated, and the position difference and velocity difference are combined to obtain the spatial difference feature.

[0038] The temporal correlation features and the spatial difference features are weighted and fused according to preset weights to obtain spatiotemporal fusion features. Singular value decomposition is performed on the spatiotemporal fusion features to obtain feature vector values ​​and feature values. The proportion of the feature vector values ​​and feature values ​​to the total feature values ​​is calculated to obtain the feature contribution. The feature contribution is sorted from largest to smallest. Feature vectors whose cumulative feature contribution reaches a preset contribution threshold are selected to construct evolutionary pattern features. The evolutionary pattern features are used to characterize the evolutionary pattern features of the key monitoring feature points.

[0039] The method further includes:

[0040] Deploy drone airports on iron towers, where drones are parked, charged, and communicated.

[0041] The inspection path is determined based on the preset inspection strategy. The inspection strategy includes a leapfrog inspection path, which is a flight path in which the drone takes off from the drone airport of the current tower, flies to the drone airport of the adjacent tower to park and charge, and then flies to the drone airport of the next tower.

[0042] The drone is controlled to inspect the area surrounding the tower along the inspection path, collecting image data of the area. The image data is analyzed to identify illegal land reclamation, unauthorized grazing, fire hazards, and abnormal animal intrusion in the area surrounding the tower. When a safety hazard is detected, a hazard alarm is generated and sent to the monitoring terminal.

[0043] A second aspect of the present invention provides a monitoring system for safety hazards along power transmission towers based on unmanned aerial vehicles (UAVs), comprising:

[0044] The first unit is used to acquire three-dimensional model data of the target tower, and determine the collaborative inspection path and data collection position of multiple drones based on the three-dimensional model data; control multiple drones to fly to the corresponding data collection position according to the collaborative inspection path, wherein each drone is equipped with a depth camera;

[0045] The second unit is used to establish a hierarchical acquisition strategy based on the structural stress analysis results of the target tower, divide the target tower into multiple deformation-sensitive areas, and dynamically allocate the acquisition density and acquisition frequency of the UAV according to the stress characteristics of each deformation-sensitive area; and control multiple UAVs to simultaneously acquire multi-view depth images of the target tower according to the hierarchical acquisition strategy.

[0046] The third unit is used to construct a real-time three-dimensional point cloud model of the target tower based on the multi-view depth image, register the real-time three-dimensional point cloud model with the three-dimensional model data of the target tower, extract deformation feature points, and use a dual threshold iteration method to classify and filter the deformation feature points, and determine the key monitoring feature points based on the filtering results.

[0047] The fourth unit is used to dynamically track the key monitoring feature points using the Kalman filter algorithm to obtain the motion trajectory of the deformation feature points. When the displacement of the motion trajectory exceeds a preset evaluation threshold, an early warning message is generated.

[0048] A third aspect of the present invention provides an electronic device, comprising:

[0049] processor;

[0050] Memory used to store processor-executable instructions;

[0051] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0052] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0053] The beneficial effects of this application are as follows:

[0054] This invention develops collaborative inspection paths based on three-dimensional tower model data and historical deformation data, and establishes a layered data acquisition strategy by combining structural stress analysis. This enables multiple drones to dynamically adjust the acquisition density and frequency according to the stress characteristics of the tower's deformation-sensitive areas, achieving accurate monitoring of tower safety hazards and significantly improving monitoring efficiency and resource utilization.

[0055] This invention employs a multi-drone collaborative operation method to acquire multi-view depth images, and combines a dual-threshold iterative hierarchical screening algorithm to filter deformation feature points, thereby achieving comprehensive monitoring of key structural parts of the tower. It can promptly detect minor deformations and potential safety hazards, significantly improving the accuracy and reliability of monitoring.

[0056] This invention uses the Kalman filter algorithm to dynamically track key monitoring feature points, obtain the motion trajectory of deformation feature points and compare it with a preset evaluation threshold, thus constructing a complete safety early warning mechanism. This mechanism can provide early warning of abnormal deformation of the tower, effectively prevent tower safety accidents, and ensure the safe and stable operation of the system. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating the method for monitoring safety hazards along power transmission towers based on unmanned aerial vehicles (UAVs) according to an embodiment of the present invention.

[0058] Figure 2 This is a schematic diagram illustrating the influence of stress gradient analysis on the identification of deformation feature points in an embodiment of the present invention;

[0059] Figure 3 This is a flowchart illustrating the dynamic tracking and early warning process for key monitoring feature points in this embodiment of the invention. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0062] Figure 1 This is a flowchart illustrating the method for monitoring safety hazards along power transmission towers based on unmanned aerial vehicles (UAVs) according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0063] Acquire 3D model data of the target tower, and determine the collaborative inspection path and data collection location of multiple drones based on the 3D model data; control multiple drones to fly to the corresponding data collection location according to the collaborative inspection path, wherein each drone is equipped with a depth camera;

[0064] Based on the structural stress analysis results of the target tower, a layered acquisition strategy is established, dividing the target tower into multiple deformation-sensitive regions, and dynamically allocating the acquisition density and acquisition frequency of the UAV according to the stress characteristics of each deformation-sensitive region; and controlling multiple UAVs to simultaneously acquire multi-view depth images of the target tower according to the layered acquisition strategy.

[0065] Based on the multi-view depth images, a real-time 3D point cloud model of the target tower is constructed. The real-time 3D point cloud model is registered with the 3D model data of the target tower, deformation feature points are extracted, and the deformation feature points are graded and screened using a dual threshold iteration method. Based on the screening results, key monitoring feature points are determined.

[0066] The key monitoring feature points are dynamically tracked using a Kalman filter algorithm to obtain the motion trajectory of the deformation feature points. When the displacement of the motion trajectory exceeds a preset evaluation threshold, an early warning message is generated.

[0067] In one optional implementation, the collaborative inspection path and data collection location of multiple UAVs are determined based on the 3D model data; controlling the multiple UAVs to fly to the corresponding data collection location according to the collaborative inspection path includes:

[0068] Based on the three-dimensional model data, a target spatial feature model is constructed, and the structural features of the tower are analyzed according to the target spatial feature model. Based on the structural features, inspection points are determined, and the inspection points are mapped to three-dimensional space to form a set of collection locations. Based on the set of collection locations, a collaborative inspection path for multiple drones is generated.

[0069] A distributed communication network is constructed for multiple UAVs, and a communication topology matrix is ​​generated based on the spatial distribution of the multiple UAVs; the position vector, velocity vector, and obstacle perception information of each UAV are obtained through the distributed communication network, and the position vector, velocity vector, and obstacle perception information are shared in real time among the multiple UAVs;

[0070] Based on the obstacle perception information, a target point gravitational potential field term is constructed, and an obstacle repulsive potential field term and a cooperative obstacle avoidance potential field term are constructed based on the position vector. The target point gravitational potential field term, the obstacle repulsive potential field term, and the cooperative obstacle avoidance potential field term are weighted and combined to form a comprehensive potential field function.

[0071] The gradient of the integrated potential field function is calculated, and the gradient is multiplied by the step size factor to obtain the path adjustment amount. The collaborative inspection path is dynamically optimized according to the path adjustment amount to obtain an optimized inspection path that meets the safety distance requirements between UAVs and the connectivity requirements of the communication network. Multiple UAVs are controlled to fly to the corresponding collection positions according to the optimized inspection path.

[0072] For the processing and analysis of 3D model data, the 3D model data of the iron towers is imported into the processing system. These iron towers include various types, such as power transmission towers and communication towers. An octree partitioning algorithm is used to spatially divide the model, with each leaf node measuring 0.5m × 0.5m × 0.5m. Through traversal and merging of the leaf nodes, the system extracts the spatial outline of the iron tower, including the geometric shape and position information of components such as the main structure, crossarms, and insulator strings. The system organizes this information into a structured target spatial feature model, which includes the main structure point set P. main Crossarm point set P cross and auxiliary equipment point set P accessory .

[0073] In the structural feature analysis of the iron tower, a graph-based structural analysis algorithm is used to identify key nodes. Specifically, the iron tower model is converted into an undirected graph G = (V, E), where vertices V represent structural nodes and edges E represent structural components. The system calculates the degree centrality C of each node. d (v) and betweenness centrality C b (v) Degree centrality reflects the number of connections a node has, while betweenness centrality reflects the importance of a node in the structure. For angle steel towers, the degree centrality of main column nodes is typically greater than 4, and the betweenness centrality is typically above 0.15; for important connection points, the degree centrality is typically greater than 3, and the betweenness centrality is typically above 0.08. The system determines the set of key nodes (Key) of the structure based on these index values. nodes .

[0074] When determining inspection points, the system uses the key node set Key nodes Applying a density-weighted algorithm, the system sets a criticality threshold T. key For nodes with a criticality level exceeding 0.7, six inspection points with different observation angles are generated around them; for nodes with a criticality level between 0.4 and 0.7, four inspection points with different observation angles are generated; and for nodes with a criticality level below 0.4, two inspection points with different observation angles are generated. The distance from the inspection points to the structural surface is dynamically adjusted according to the component size, with a minimum distance of 3 meters and a maximum distance of 10 meters. The system maps these inspection points into three-dimensional space, forming a set of acquisition locations P. collection Each acquisition location contains coordinates (x, y, z) and camera orientation vector (dx, dy, dz).

[0075] When generating a collaborative inspection path, the system first uses the K-means clustering algorithm to cluster the collection location set P. collectionThe system is divided into k subsets, where k equals the number of drones. For a common scenario of 4 drones coordinating inspections, the initial cluster centers are set at the upper, middle, lower parts of the tower, and the auxiliary equipment area, respectively. The number of clustering iterations is set to 20. After clustering, the system obtains 4 subsets P. sub 1,P sub 2,P su b 3,P sub 4. The data is allocated to four drones.

[0076] Within each subset, the system uses the TSP algorithm to generate an initial inspection path. Specifically, the system employs the 2-opt optimization algorithm, starting with a random initial path and attempting to obtain a shorter path by swapping two edges in the path. This iterative optimization continues until the path cannot be shortened further or the maximum number of iterations (100) is reached. For subset P... sub In scenario 1, which includes 20 data collection locations, the optimized total path length can typically be reduced by approximately 25%. The system generates an initial inspection path for each drone. initiali It contains a complete waypoint sequence from the starting position to each collection position and back to the starting position.

[0077] When constructing a distributed communication network, the system equips each UAV with a 2.4GHz wireless communication module, with a maximum transmission rate of 20Mbps and an effective communication distance of 200 meters. The system calculates the communication topology matrix T based on the UAV's initial position. comm The element T in the matrix comm [i][j] represents the communication status between UAV i and UAV j, with a value of 1 indicating direct communication and a value of 0 indicating no direct communication. The system uses TDMA (Time Division Multiple Access) to allocate a 20ms time slice to each UAV, during which it broadcasts its position and sensing information.

[0078] The UAV's position vector is represented by three-dimensional coordinates, such as UAV1's position vector P1 = (x1, y1, z1); the velocity vector is represented by velocity components in three directions, such as UAV1's velocity vector V1 = (vx1, vy1, vz1). Obstacle perception information is represented using a three-dimensional cylindrical model, with each obstacle represented by its center point coordinates (x1, y1, z1). obs ,y obs ,z obs ) and radius r obs Description. Each drone packages this information into a predefined data packet format and broadcasts it across the network. Other drones receive the information, parse it, and store it in their local environmental awareness database.

[0079] When constructing the artificial potential field, the system first calculates the gravitational potential field term of the target point based on the acquisition location. For the distance d from the current position Pi of UAV i to the target point Pgoal... goal Gravity magnitude U attr According to d goal 2 The inverse calculation of the gravitational coefficient k attr Set it to 0.5. When d goal When the distance is less than 1 meter, the gravitational force is set to a constant value to prevent the system from becoming unstable due to excessive gravity when approaching the target point.

[0080] The obstacle repulsive potential field term is calculated based on obstacle perception information, and is the distance d from the current position Pi of UAV i to the center point of the obstacle. obs When d obs Less than the safety threshold d safe When the distance is set to 15 meters, the system calculates the magnitude of the repulsive force U. rep Repulsive force and (1 / d) obs -1 / d safe ) 2 Proportional, repulsion coefficient k rep Set to 0.3. When multiple obstacles exist simultaneously, the system calculates the repulsive force generated by each obstacle, and then performs vector superposition to obtain the composite repulsive force.

[0081] The cooperative obstacle avoidance potential field term is calculated based on the relative positional relationship between the drones, for the distance d between drone i and drone j. uav When d uav Less than the safe distance d min When the distance is set to 10 meters, a repulsive force U is generated. cooprep The magnitude of the force is related to (1 / d) uav -1 / d min ) 2 Proportional, mutual repulsion coefficient k cooprep Set it to 0.2; when d uav Greater than the communication distance d comm When the distance is set to 200 meters, an attractive force U is generated. coopattr The magnitude of the force is related to (d) uav -d comm ) 2 Proportional, attraction coefficient k coopattr Set it to 0.1.

[0082] The system combines these three potential fields in a weighted manner to form a comprehensive potential field function U. total =w1×U attr +w2×U rep +w3×U coop The weighting coefficients are w1 = 0.5, w2 = 0.3, and w3 = 0.2. The system calculates U. totalgradient with respect to position (grad) U This gives the direction and magnitude of the potential force at the current location. (The grad...) U Multiplying this by the step size factor α (set to 0.8), we obtain the path adjustment amount ΔP = α × grad. U .

[0083] The application path adjustment is used to optimize the initial inspection path, and the initial path is adjusted accordingly. initiali Each waypoint P in waypoint Calculate the new position P after adjustment. new =P waypoint -△P. To ensure path smoothness, the system uses cubic spline interpolation to smooth the adjusted waypoint sequence, setting the maximum curvature between control points to 0.2, thus generating a continuous and smooth optimized inspection path. optimizedi .

[0084] In actual tower inspections, the system recalculates the environmental potential field and path adjustment parameters at a frequency of 10Hz to achieve dynamic path optimization. When environmental changes are detected, such as the appearance of new obstacles or changes in the communication topology, the system immediately adjusts the corresponding potential field parameters and recalculates the path. In this way, the system can respond to various changes in the flight environment in real time, ensuring the safe and collaborative operation of multiple UAV systems.

[0085] For the inspection of angle steel towers of a 220kV transmission line (45 meters high), the system deployed four drones, each with a maximum flight speed of 5 m / s and a hovering accuracy of ±0.3 meters. The system planned an inspection path for each drone including approximately 20 data collection points, with an average flight distance of about 300 meters. A complete inspection took approximately 8 minutes. Through this method, the system achieved collaborative inspection by multiple drones, effectively improving the efficiency and comprehensiveness of tower inspection and providing a reliable data foundation for subsequent safety hazard monitoring.

[0086] In one optional implementation, a real-time 3D point cloud model of the target tower is constructed based on the multi-view depth images. The real-time 3D point cloud model is then registered with the 3D model data of the target tower, and deformation feature points are extracted, including:

[0087] The stress distribution field of the tower at different heights and wind speeds is calculated to generate a stress distribution map. A three-dimensional stress variation curve is constructed based on the stress distribution map. The stress gradient is calculated in the three-dimensional stress variation curve, and the area with the largest stress gradient is identified as the key monitoring area.

[0088] Multiple depth cameras are deployed in the key monitoring area, and structural feature points are extracted from the multi-view depth images using the depth cameras.

[0089] The spatial location information of the structural feature points is fused with the stress distribution map to establish a correspondence between the feature point locations and the stress distribution, and a weighting coefficient is assigned to the structural feature points according to the correspondence.

[0090] Based on the weighting coefficients, the multi-view depth image is registered, and the registration accuracy requirement is improved in areas where the stress is greater than the preset registration threshold to generate a real-time three-dimensional point cloud model. The structural feature points in the real-time three-dimensional point cloud model are compared with the corresponding positions in the standard three-dimensional model, and the displacement direction and displacement amplitude of each structural feature point are calculated.

[0091] Based on the displacement direction and displacement amplitude, a connectivity analysis is performed on the structural feature points to identify deformation-related point groups. Feature points in the deformation-related point groups that do not conform to the stress distribution law are removed to obtain the final deformation feature points.

[0092] In the method for constructing a real-time 3D point cloud model of a target iron tower based on multi-view depth images, stress distribution calculation is a crucial step. This method employs finite element analysis (FEM) technology to establish a refined model of the tower. The towers include various types, such as power transmission towers and communication towers. The model includes major structural components such as the main tower body, crossarms, and diagonal braces. Taking a 220kV angle steel tower as an example, the tower height is 45 meters, the base width is 12 meters, and the top width is 2 meters. The model considers the material properties of the tower, such as the elastic modulus of Q345 steel being 210 GPa, Poisson's ratio being 0.3, and density being 7850 kg / m³. 3 In the finite element model, the tower is divided into 12,000 mesh elements, with each element having an average size of 0.3 meters, ensuring a balance between computational accuracy and efficiency.

[0093] The wind load on the tower varies significantly at different heights, and the relationship between wind speed and height is described using an exponential distribution model. A baseline wind speed of 10 m / s is set at 10 meters above ground level, and wind speeds at different heights are calculated based on this value. For example, at a height of 20 meters, the wind speed is approximately 11.5 m / s; at 30 meters, it is approximately 12.8 m / s; and at 45 meters, it is approximately 14.3 m / s. Based on these wind speed values, the wind load acting on various parts of the tower is calculated. The wind load calculation considers the structure's windward area, wind pressure coefficient, and height variation. At the top crossarm at a height of 45 meters, the wind pressure per unit area reaches 1.02 kPa; at 30 meters, it is approximately 0.82 kPa; and at 15 meters, it is approximately 0.62 kPa.

[0094] Wind load data is input into finite element analysis software to calculate the stress distribution of the tower under different wind speeds. The calculation results are output as a stress distribution map, with color coding indicating the stress magnitude at different locations. For example, under a baseline wind speed of 10 m / s, the maximum stress at the bottom connection point of the tower can reach 120 MPa, the stress at the middle diagonal brace connection point is approximately 80 MPa, and the stress at the top crossarm connection point is approximately 65 MPa. When the wind speed increases to 15 m / s, the maximum stress at the bottom connection point increases to 180 MPa, the stress at the middle diagonal brace connection point increases to 120 MPa, and the stress at the top crossarm connection point increases to 95 MPa.

[0095] Based on stress distribution maps under multiple wind speed conditions, a three-dimensional stress variation curve was constructed. This curve uses tower height, horizontal position, and wind speed as three dimensions, with stress value as the dependent variable. For example, within the 0-10 meter height range of the main column on the northeast side of the tower, the stress value increases from 50 MPa to 200 MPa when the wind speed increases from 5 m / s to 20 m / s; while at the same location within the 20-30 meter height range, the stress variation caused by the same wind speed change ranges from 30 MPa to 120 MPa. By calculating the spatial stress gradient in the three-dimensional stress variation curve, the region with the largest stress gradient was identified. Calculations show that the stress gradients at the connections between the four main columns and the foundation, the connections between the main columns and the diagonal braces, and the connections between the crossarms and the main structure at the bottom of the tower are significantly higher than in other areas. These areas were identified as key monitoring areas.

[0096] Depth cameras are deployed in designated key monitoring areas. These cameras operate using structured light or time-of-flight principles, with a depth resolution of 2mm, a field of view of 60 degrees, and a working distance of 3-10 meters. Taking the structured light camera as an example, its emitted structured light pattern is a pseudo-random dot matrix with a dot density of 1200 dots / square meter and a sampling frequency of 30 frames / second. Twelve depth cameras are deployed on a standard iron tower: four at the bottom connection area, four at the middle diagonal brace connection area, and four at the top crossarm connection area. The installation position of each camera is carefully designed to ensure no blind spots in the monitoring area, and the fields of view of adjacent cameras overlap by 30%, facilitating subsequent image stitching.

[0097] The multi-view depth images acquired by the depth camera have a resolution of 640×480, with each pixel containing a depth value and corresponding 3D coordinate information. A two-stage method is used to extract structural feature points from the depth images: edge detection and corner extraction. Edge detection uses an adaptive threshold gradient algorithm, with the threshold dynamically adjusted based on the depth change rate of the image region, typically set to 1.5 times the standard deviation of the region's depth. For the angle steel structure of the iron tower, the outline of the angle steel can be extracted during the edge detection stage; corner extraction uses the Harris corner detection algorithm, with a response threshold set to 0.01 and a non-maximum suppression window size of 5×5 pixels. Typically, 200-300 structural feature points can be extracted from a single depth image frame. These feature points are mainly distributed at key structural locations such as angle steel connections and bolt positions.

[0098] The spatial location information of extracted structural feature points is fused with the stress distribution map to establish a correspondence between feature point locations and stress distribution. During the fusion process, the 3D coordinates of the feature points are first transformed to the same reference coordinate system as the stress distribution map, with a transformation accuracy better than 5mm. Then, the stress value at each feature point location is calculated using trilinear interpolation. For example, the stress value at the feature point extracted at the connection between the main column and the foundation at the bottom of the tower is approximately 120MPa; the stress value at the feature point extracted at the connection of the diagonal brace in the middle is approximately 80MPa. Based on the stress value at each feature point location, a weight coefficient is assigned to each feature point. The weight coefficient calculation uses normalization processing, dividing the stress value by the maximum stress value and then multiplying by an adjustment factor of 0.8 to ensure that the weight distribution is between 0.1 and 0.8. Feature points with larger stress values ​​have higher weight coefficients and are more important in the subsequent registration process.

[0099] Multi-view depth image registration is a crucial step in constructing a complete 3D point cloud model. The registration process employs a weighted iterative nearest-neighbor algorithm with 50 iterations and a convergence threshold of 0.001 meters. In each iteration, the matching weights of feature points are dynamically adjusted according to the aforementioned weight coefficients. In areas where stress exceeds the preset registration threshold (80 MPa), the registration accuracy requirement is increased by reducing the matching distance threshold from the standard 0.05 meters to 0.02 meters to ensure higher registration accuracy in high-stress areas. After registration, the point cloud data from multiple views are merged to form a real-time 3D point cloud model covering the entire tower. The model's point cloud density reaches 5000 points / square meter in key monitoring areas and 2000 points / square meter in general areas.

[0100] The comparison between the real-time 3D point cloud model and the standard 3D model is the core step in extracting deformation feature points. The standard 3D model is the design model of the tower or the benchmark model in a deformation-free state, with a point cloud density of 10,000 points / square meter. The comparison process begins with coarse registration, using a bounding box alignment method along the three principal axes to align the real-time point cloud model and the standard model in the global coordinate system, achieving a registration error of less than 0.1 meters. Fine registration is then performed using a local feature descriptor matching method based on principal component analysis to precisely align local areas, reducing the registration error to within 0.01 meters. After the comparison, the displacement vector of each structural feature point in the real-time point cloud model and its corresponding point in the standard model is calculated, including the displacement direction and magnitude. For example, at a connection point at the bottom of the tower, the measured displacement direction is horizontal outward, with a displacement magnitude of 15 mm; at a diagonal brace connection in the middle, the measured displacement direction is vertical downward, with a displacement magnitude of 12 mm.

[0101] Connectivity analysis of displacement data is used to identify deformation-related point groups. This analysis is based on two indicators: spatial distance and displacement similarity. The spatial distance threshold is set to 0.5 meters, the displacement direction similarity threshold to 15 degrees, and the displacement amplitude similarity threshold to 30%. When the spatial distance between two feature points is less than the threshold, and the differences in displacement direction and amplitude are both less than their respective thresholds, these two points are considered to be deformation-related. In this way, all related feature points are aggregated into deformation-related point groups. In practical cases, deformation-related point groups containing 15-20 feature points are formed at the base of the tower, indicating overall deformation in that area; while deformation-related point groups containing 8-12 feature points are formed in the central diagonal bracing area, indicating local structural deformation.

[0102] The deformation correlation point set contains anomalies caused by measurement errors or environmental interference. The displacement patterns of these points do not conform to the stress distribution law. The method for removing these anomalies is to compare the actual displacement of the characteristic point with the theoretical displacement predicted based on stress analysis. The theoretical displacement is obtained through material mechanics calculations, taking into account stress values, material properties, and structural constraints. When the difference between the actual and theoretical displacement of a characteristic point exceeds 50%, or the angle between the displacement direction and the principal stress direction exceeds 30 degrees, the characteristic point is marked as an anomaly and removed. After removing anomalies, the resulting set of deformation characteristic points is more reliable and can accurately reflect the true deformation state of the tower.

[0103] In a case study of tower monitoring for a 220kV transmission line, the system extracted 2,500 initial structural feature points. Through connectivity analysis, 35 deformation correlation point groups were formed, containing a total of 1,200 feature points. Anomaly point removal eliminated 180 feature points that did not conform to the stress distribution pattern, ultimately resulting in 1,020 deformation feature points. These feature points are mainly distributed in the bottom connection area (420), the middle diagonal brace connection area (380), and the top crossarm connection area (220), providing a reliable data foundation for subsequent deformation tracking and early warning.

[0104] Figure 2 This diagram illustrates the impact of stress gradient analysis on deformation feature point identification in an embodiment of the present invention. It compares the relationship between stress value and deformation displacement under two different methods. The solid triangle in the diagram represents "this technical solution," which employs a deformation feature evolution pattern extraction algorithm based on a spatiotemporal sequence matrix to achieve accurate measurement through effective feature point extraction. The dashed circle represents the "stress-free gradient analysis" method, a traditional deformation analysis method that includes noise points. Data trends show that when the stress value increases from 10 MPa to 50 MPa, both methods exhibit a trend of increasing deformation displacement with increasing stress. Specifically, in this technical solution, the deformation displacement is approximately 5 mm at a stress of 10 MPa, gradually increasing to approximately 20 mm as the stress increases to 50 MPa, showing a relatively stable linear growth trend overall. The stress-gradient-free analysis method shows a deformation displacement of approximately 3 mm at an initial stress of 10 MPa, increasing to about 13 mm at 50 MPa. However, its growth curve exhibits significant fluctuations, with localized fluctuations at multiple stress points such as 25 MPa, 35 MPa, and 45 MPa. Comparing the two methods reveals that the proposed solution not only yields larger overall deformation displacement measurements than the traditional method but also demonstrates a more stable growth trend without significant fluctuations. This indicates that the deformation feature evolution pattern extraction algorithm based on the spatiotemporal sequence matrix effectively eliminates noise interference through the extraction of effective feature points, thereby improving the accuracy and reliability of deformation measurements. The figure visually demonstrates the superiority of this proposed solution in stress-deformation evolution analysis.

[0105] In one optional implementation, connectivity analysis is performed on the structural feature points based on the displacement direction and displacement amplitude to identify deformation-related point groups. Feature points in the deformation-related point groups that do not conform to the stress distribution law are removed to obtain the final deformation feature points, including:

[0106] Based on the displacement direction and displacement amplitude, a displacement feature vector is constructed. The Euclidean distance and spatial distance between the displacement feature vectors are calculated. A feature correlation matrix is ​​constructed by combining the Euclidean distance and the spatial distance. The connectivity strength of the structural feature points is calculated based on the feature correlation matrix. The structural feature points are clustered based on the connectivity strength to obtain multiple deformation correlation point groups.

[0107] Calculate the mean and deviation of the displacement feature vectors of the structural feature points in the deformation-related point group, and construct the deformation distribution features by combining the mean and deviation of the displacement feature vectors.

[0108] Obtain stress distribution data of the tower, calculate stress values ​​and stress gradients at structural feature points based on the stress distribution data, construct stress distribution features using the stress values ​​and stress gradients, input the stress distribution features and deformation distribution features into a scoring function, and filter the structural feature points based on the output value of the scoring function to obtain the final deformation feature points.

[0109] The connectivity analysis and deformation feature point selection method for structural feature points is based on displacement feature vector construction and Euclidean distance calculation. This method extracts displacement direction and amplitude information for each structural feature point. The displacement direction is represented as a three-dimensional unit vector containing x, y, and z components, while the displacement amplitude is a scalar value in millimeters. The displacement feature vector is formed by combining the displacement direction and amplitude, creating a four-dimensional vector. Taking a 220kV transmission tower as an example, feature point A detected in the bottom connection area has a displacement direction of (0.6, 0.3, 0.74) and a displacement amplitude of 15mm, so its displacement feature vector is (0.6, 0.3, 0.74, 15). Feature point B in the middle area has a displacement direction of (0.5, 0.2, 0.84) and a displacement amplitude of 12mm, so its displacement feature vector is (0.5, 0.2, 0.84, 12).

[0110] To calculate the similarity of displacement characteristics between structural feature points, the system uses a weighted Euclidean distance metric. For the displacement direction component, a weighting coefficient of 0.7 is assigned; for the displacement amplitude component, a weighting coefficient of 0.3 is assigned after standardization. Standardization involves dividing the original displacement amplitude by the maximum displacement amplitude within the monitoring area (e.g., 25 mm), ensuring the value ranges from 0 to 1. Taking the two feature points mentioned above as examples, the weighted Euclidean distance calculation result for their displacement feature vectors is 0.18, indicating a high degree of similarity in displacement characteristics between the two points. Spatial distance is calculated using three-dimensional Euclidean distance. Feature point A has coordinates of (2.5, 1.8, 3.2) meters, and feature point B has coordinates of (2.8, 2.1, 7.5) meters; the calculated spatial distance is 4.38 meters.

[0111] When constructing the feature correlation matrix, the system comprehensively considers both displacement feature vector distance and spatial location distance. For all structural feature points identified within the monitoring area (typically 1000-2000), the correlation degree between each pair of points is calculated. The correlation degree calculation uses an exponential decay function, with a displacement feature vector distance parameter of 0.3 and a spatial location distance parameter of 5.0. This indicates that the correlation degree decreases significantly when the displacement feature vector distance exceeds 0.3 or the spatial location distance exceeds 5.0 meters. For feature points A and B, the calculated correlation degree is 0.65, which is stored in the corresponding position in the feature correlation matrix. The final generated feature correlation matrix has the same dimension as the number of feature points, is a symmetric matrix, and all diagonal elements are 1.

[0112] The connectivity strength of structural feature points is calculated based on the feature correlation matrix. Connectivity strength reflects the degree of association between a feature point and its surrounding points. For each feature point, its correlation with all other points is extracted, and a correlation threshold of 0.5 is set. The number of points with correlations greater than the threshold and the total correlation are counted. The connectivity strength of a feature point is equal to the total correlation divided by the number of associated points. In actual monitoring, the connectivity strength of feature points in the bottom region is typically between 0.6 and 0.8, in the middle region between 0.5 and 0.7, and in the top region between 0.4 and 0.6. Regions with high connectivity strength indicate strong consistency in deformation characteristics and suggest overall structural deformation.

[0113] Clustering of structural feature points employs the density peak clustering algorithm, which defines cluster centers based on connectivity strength and local density. Local density is calculated based on the number of points within a 5-meter radius of each feature point, with a connectivity strength threshold set to 0.6. The algorithm identifies points with high connectivity strength and high local density as cluster centers, then assigns other points to the nearest cluster center, forming multiple deformation-related point groups. In typical tower monitoring scenarios, the algorithm can identify 30-50 deformation-related point groups, each containing 10-30 feature points. For example, four point groups are formed in the bottom main column connection area, corresponding to the connections between the four main columns and the foundation; eight to twelve point groups are formed in the middle area, corresponding to the connections between the diagonal braces and the main columns; and six to eight point groups are formed in the top area, corresponding to the connections between the crossarms and the main columns.

[0114] After the deformation-related point groups are formed, the system calculates the mean and deviation of the displacement eigenvectors of the structural feature points within each point group. The mean displacement eigenvector is obtained by calculating the arithmetic mean of the displacement eigenvectors of all feature points within the point group, representing the overall deformation trend of the point group. The deviation of the displacement eigenvector is calculated by taking the Euclidean distance between each feature point within the point group and the mean, and then averaging these distances to represent the degree of deformation consistency within the point group. Taking a certain deformation-related point group at the bottom as an example, containing 18 feature points, the calculated mean displacement eigenvector is (0.58, 0.25, 0.77, 14.5), and the deviation is 0.12, indicating that the deformation patterns of the feature points within this point group are relatively consistent. The mean and deviation of the displacement eigenvectors are combined to form the deformation distribution characteristics, which are used for subsequent comparison with the stress distribution characteristics.

[0115] The stress distribution data of the steel tower is obtained through finite element analysis. A refined tower model is established, considering actual load conditions (such as wind load and conductor tension), and the stress distribution of various parts of the tower is calculated. The stress calculation employs a linear static analysis method, and the results are expressed in the form of stress tensors. For each structural feature point, the system extracts the stress value (in MPa) and stress gradient (in MPa / m) at its location. The stress value represents the magnitude of the mechanical stress borne at that point, and the stress gradient represents the degree of stress variation around that point. Typically, the stress value at the bottom connection of a steel tower can reach 120 MPa, with a stress gradient of approximately 25 MPa / m; the stress value at the middle diagonal brace connection is approximately 80 MPa, with a stress gradient of approximately 15 MPa / m; and the stress value at the top crossarm connection is approximately 60 MPa, with a stress gradient of approximately 10 MPa / m.

[0116] Stress distribution characteristics consist of stress values ​​and stress gradients. To make stress distribution characteristics comparable to deformation distribution characteristics, stress values ​​and stress gradients are normalized. The stress value is divided by the maximum stress value (e.g., 150 MPa), and the stress gradient is divided by the maximum stress gradient (e.g., 30 MPa / m), ensuring their values ​​are between 0 and 1. The normalized stress distribution characteristics are used to assess the consistency between deformation and stress. Theoretically, deformation should match the stress distribution. For example, areas with higher stress should exhibit greater deformation, and the direction of deformation should be consistent with the principal stress direction; areas with larger stress gradients show localized deformation concentrations or discontinuous deformation.

[0117] The scoring function is used to evaluate whether the deformation at structural feature points conforms to the stress distribution law. The scoring function design considers the following three aspects: the correlation between deformation amplitude and stress value, the consistency between deformation direction and principal stress directions, and the matching degree between deformation distribution and stress gradient. For the correlation between deformation amplitude and stress value, the correlation coefficient is calculated, with a weight of 0.4; for the consistency between deformation direction and principal stress directions, the cosine of the angle between the two directions is calculated, with a weight of 0.4; for the matching degree between deformation distribution and stress gradient, the ratio of the deformation eigenvector deviation to the normalized stress gradient is calculated, with a weight of 0.2. The final score is obtained by weighted summation of the three indicators, with a score range from 0 to 1. The closer the score is to 1, the more closely the deformation conforms to the stress distribution law.

[0118] Taking a deformation-related point group at the bottom as an example, the mean displacement characteristic vector of this point group is (0.58, 0.25, 0.77, 14.5), the displacement characteristic vector deviation is 0.12, the normalized stress value of the corresponding region is 0.8, the normalized stress gradient is 0.85, and the principal stress directions are (0.6, 0.2, 0.78). The correlation coefficient between the deformation amplitude and the stress value is calculated to be 0.92, the cosine of the angle between the deformation direction and the principal stress direction is 0.98, and the ratio of the deformation characteristic vector deviation to the normalized stress gradient is 0.14. According to the scoring function, the score of this point group is 0.92×0.4+0.98×0.4+(1-0.14)×0.2=0.93, indicating that the deformation characteristics and stress distribution of this point group are highly consistent.

[0119] The system filters structural feature points within deformation-related point groups based on a scoring function, eliminating those that do not conform to the stress distribution pattern. Specifically, the filtering criteria are as follows: feature points with a score below 0.7 are directly eliminated; feature points with scores between 0.7 and 0.8 are also eliminated if their displacement eigenvector has an Euclidean distance greater than 0.2 from the point group's mean; feature points with scores between 0.8 and 0.9 are only eliminated if their displacement eigenvector has an Euclidean distance greater than 0.3 from the point group's mean; and all feature points with scores greater than 0.9 are retained. In this way, the system can effectively eliminate unreliable feature points caused by measurement errors, environmental interference, or local anomalies, retaining feature points that truly reflect structural deformation.

[0120] In a monitoring case of a 220kV transmission tower, the system initially identified 1500 structural feature points. Through connectivity analysis, 40 deformation correlation point groups were formed, containing a total of 1200 feature points. After screening using a scoring function, 210 feature points that did not conform to the stress distribution pattern were removed, ultimately resulting in 990 deformation feature points. These deformation feature points are distributed across various key areas of the tower, accurately reflecting the overall deformation state of the tower and providing a reliable data foundation for subsequent dynamic tracking and safety early warning. High-quality deformation feature points can significantly improve the accuracy and reliability of deformation monitoring, reduce false alarms and missed alarms, and provide strong protection for the safe operation of transmission lines.

[0121] In one optional implementation, a dual-threshold iteration method is used to perform hierarchical screening of the deformation feature points, and the key monitoring feature points are determined based on the screening results, including:

[0122] An initial threshold and a target threshold are constructed. The initial threshold is used to perform a first screening of deformation feature points to obtain candidate feature points. The deformation amplitude distribution of the candidate feature points is calculated. The target threshold is dynamically adjusted according to the deformation amplitude distribution.

[0123] Using the target threshold as the iterative target, the candidate feature points are progressively screened. When the screening results are stable, the candidate feature points that are greater than the target threshold are identified as key monitoring feature points.

[0124] The method for hierarchical screening of deformation feature points and determination of key monitoring feature points adopts a dual-threshold iterative strategy to achieve accurate identification and dynamic tracking of monitoring objects. This method first constructs initial and target threshold parameters. The initial threshold is set based on the statistical analysis results of historical monitoring data. For 220kV transmission tower deformation monitoring, the system sets the initial deformation amplitude threshold to 5 mm. This value is obtained by analyzing historical deformation data of 50 towers of the same model, corresponding to the upper limit of minor deformation of the tower structure under normal conditions. The target threshold is initially set to 12 mm, which is determined according to the safety margin in the tower design specifications, representing the deformation amplitude that needs to be focused on. The interval between the two thresholds is a gray area, which needs to be iteratively screened to determine its final assignment.

[0125] When initially screening deformation feature points using an initial threshold, the system evaluates the deformation amplitude of each feature point. The deformation amplitude is calculated using Euclidean distance, with the distance between the current position and the standard position of the feature point used as the deformation amplitude value. In a transmission tower monitoring example, the system screened 990 deformation feature points, identifying 350 candidate feature points with deformation amplitudes greater than 5 mm. These candidate feature points are mainly distributed in key load-bearing areas of the tower, such as the bottom main column connection area (120), the middle diagonal brace connection area (150), and the top crossarm connection area (80).

[0126] After the initial screening, the system calculates the deformation amplitude distribution characteristics of the candidate feature points, including statistical measures such as the mean, standard deviation, and quantiles of the deformation amplitude. In the monitoring example above, the mean deformation amplitude of the 350 candidate feature points was 8.3 mm, the standard deviation was 2.5 mm, the minimum was 5.1 mm, the maximum was 16.8 mm, the 25th quantile was 6.2 mm, the 50th quantile (median) was 7.8 mm, the 75th quantile was 10.1 mm, and the 95th quantile was 13.5 mm. The deformation amplitude distribution exhibits a right-skewed characteristic, indicating that the deformation amplitude of most candidate feature points is concentrated in the small value range, while a few points have larger deformation amplitudes.

[0127] Based on the deformation amplitude distribution characteristics, the system dynamically adjusts the target threshold. The adjustment method considers the importance of the monitored object, the dispersion of the deformation distribution, and the sensitivity requirements for anomaly detection. For important transmission towers, the target threshold adjustment is conservative to improve monitoring sensitivity; for cases with high dispersion in the deformation distribution, the target threshold adjustment needs to consider the overall shape of the distribution; for scenarios requiring high sensitivity, the target threshold can be appropriately reduced. The specific adjustment calculation uses an adaptive algorithm, considering the average and standard deviation of the deformation amplitude, and introducing a safety factor. In this example, the adjusted target threshold is 9.5 mm, calculated as the average plus the standard deviation multiplied by the safety factor (0.5), i.e., 8.3 + 2.5 × 0.5 = 9.5 mm.

[0128] Once the target threshold is determined, the system begins an iterative screening process using a progressive threshold adjustment strategy. The initial threshold is set to the initial threshold (5 mm). In each iteration, the current threshold is increased by a step size until the target threshold (9.5 mm) is reached. The step size is set to 0.5 mm, and the screening process requires 9 iterations. In each iteration, the system uses the current threshold to screen candidate feature points and counts the number of feature points that meet the criteria. The first iteration uses a threshold of 5.5 mm, selecting 342 feature points; the second iteration uses a threshold of 6 mm, selecting 330 feature points; and so on, with the final iteration using a threshold of 9.5 mm, selecting 156 feature points.

[0129] During the iterative screening process, the system monitors the stability of the screening results. Stability is assessed based on the rate of change between two consecutive iterations, calculated as the absolute value of the difference between the number of retained points in the current iteration and the number of retained points in the previous iteration, divided by the number of retained points in the previous iteration. When the rate of change is less than a preset threshold (usually 5%), the screening results are considered to be stable. In this example, the seventh iteration (threshold 8.5 mm) screened 188 feature points, and the eighth iteration (threshold 9 mm) screened 170 feature points, with a rate of change of 9.6%. The rate of change in the eighth and ninth iterations (threshold 9.5 mm, 156 feature points) was 8.2%, still not reaching the stability condition. The system continued with the tenth iteration (threshold 10 mm), screening 145 feature points, reducing the rate of change to 7.1%; the eleventh iteration (threshold 10.5 mm) screened 138 feature points, with a rate of change of 4.8%, meeting the stability condition, and the iteration process terminated.

[0130] After the screening results stabilized, the system identified candidate feature points larger than the final threshold (10.5 mm) as key monitoring feature points. In this example, a total of 138 key monitoring feature points were ultimately identified, distributed across various critical areas of the tower. Specifically, 42 were located in the bottom main column connection area, 58 in the middle diagonal brace connection area, and 38 in the top crossarm connection area. The average deformation amplitude of these feature points was 12.3 mm, the minimum was 10.6 mm, the maximum was 16.8 mm, and the standard deviation was 1.8 mm. The deformation directions were mainly concentrated in the horizontal outward and vertical downward directions.

[0131] To verify the rationality of the screening results, the system conducted spatial distribution analysis and deformation pattern analysis on key monitoring feature points. Spatial distribution analysis showed that the key monitoring feature points exhibited a clustered distribution characteristic on the tower structure, mainly concentrated at connection nodes with high stress, such as the connection between the bottom main column and the foundation, the connection between the main column and the diagonal brace, and the connection between the crossarm and the main body. This distribution highly matches the stress characteristics of the tower, verifying the rationality of the screening results. Deformation pattern analysis showed that key monitoring feature points within the same area had similar deformation directions and amplitudes, indicating that the deformation has an overall characteristic, rather than being a local anomaly caused by random noise.

[0132] The system also evaluated the temporal stability of key monitoring feature points, conducting continuous monitoring for 30 days with three data collections per day, and analyzing the deformation changes of feature points at different time points. The results showed that the deformation amplitude of most key monitoring feature points (approximately 85%) did not change by more than 2 mm within 30 days, demonstrating good temporal stability; approximately 12% of feature points showed a slow increasing trend in deformation amplitude, with a growth rate of approximately 0.1 mm / day; and approximately 3% of feature points experienced significant fluctuations in deformation amplitude, indicating a marked influence from environmental factors (such as temperature and wind speed), requiring comprehensive analysis in conjunction with environmental parameters.

[0133] The advantage of the dual-threshold iterative screening method lies in its ability to adapt to different monitoring objects and environmental conditions, dynamically adjusting the screening criteria to find the most suitable threshold. Compared with the fixed threshold method, this method can reduce the false alarm rate by approximately 30% and the false negative rate by 25%, significantly improving the accuracy and reliability of monitoring. The final identified key monitoring feature points can accurately reflect the critical deformation state of the towers, which include various types such as power towers and communication towers. This provides a reliable data foundation for subsequent dynamic tracking and early warning information generation, effectively supporting the timely detection and handling of safety hazards along the tower lines.

[0134] In one optional implementation, the key monitoring feature points are dynamically tracked using a Kalman filter algorithm to obtain the motion trajectory of the deformation feature points. When the displacement of the motion trajectory exceeds a preset evaluation threshold, an early warning message is generated, including:

[0135] The three-dimensional coordinates and displacement velocity of key monitoring feature points are obtained. The three-dimensional coordinates and displacement velocity are used to construct a state vector of the key monitoring feature points. Based on the state vector, historical state sampling is performed on the key monitoring feature points to obtain a historical state sampling sequence.

[0136] A spatiotemporal sequence matrix is ​​constructed based on the historical state sampling sequence. The spatiotemporal sequence matrix is ​​analyzed using a spatiotemporal feature extraction operator to obtain the evolution pattern features of the key monitoring feature points. The evolution pattern features are then used as correction parameters for the Kalman filter algorithm.

[0137] The state vector is input into the Kalman filter algorithm for state prediction. The state prediction result is corrected based on the evolutionary pattern characteristics to obtain the predicted state vector. Real-time observation data of the key monitoring feature points are obtained. The predicted state vector and the real-time observation data are weighted and fused to obtain the motion trajectory of the key monitoring feature points.

[0138] The displacement between adjacent sampling points in the motion trajectory is calculated, and the standard deviation and mean of the displacement are obtained by statistical analysis. A preset evaluation threshold is constructed based on the standard deviation and the mean. When the displacement is greater than the preset evaluation threshold, an early warning message is generated.

[0139] like Figure 3 As shown, the method includes:

[0140] The method for dynamic tracking and early warning information generation of key monitoring feature points revolves around the improvement and application of the Kalman filter algorithm. This method first acquires the three-dimensional coordinates and displacement velocity information of the key monitoring feature points. The three-dimensional coordinates contain three components: x, y, and z, in meters; the displacement velocity also contains three directional components: vx, vy, and vz, in millimeters per day. Taking the connection point of the main column at the bottom of a 220kV transmission tower as an example, the three-dimensional coordinates of this feature point are (3.25, 2.60, 1.85) meters, and the displacement velocity is (0.35, 0.22, -0.18) millimeters per day, indicating that the point has an outward displacement trend in the horizontal direction and a downward displacement trend in the vertical direction. The combination of the three-dimensional coordinates and displacement velocity constitutes a state vector containing six components: (x, y, z, vx, vy, vz). This state vector comprehensively describes the position and motion state of the feature point in space and serves as the basic data for subsequent dynamic tracking.

[0141] Historical state sampling was performed on key monitoring feature points. The status of the feature points was sampled at fixed time intervals (usually 24 hours), recording their three-dimensional coordinates and displacement velocities to form a historical state sampling sequence. The sampling period was typically 30 days, generating 30 sets of state data. For the aforementioned feature points, the coordinate variation range over 30 days was x: (3.25-3.36) meters, y: (2.60-2.67) meters, z: (1.85-1.80) meters; the velocity variation range was vx: (0.30-0.42) mm / day, vy: (0.18-0.25) mm / day, vz: (-0.15 to -0.20) mm / day. This data shows that the point experienced continuous small displacements during the monitoring period, and the displacement direction was relatively stable without any abrupt changes.

[0142] In constructing the spatiotemporal sequence matrix from historical state sampling sequences, the system treats each set of state data as a row in the matrix, forming a 30-row, 6-column matrix. Each column of this matrix represents the change sequence of a state component over time, and each row represents the complete state of a feature point at a specific moment. The spatiotemporal sequence matrix contains the temporal evolution and spatial distribution characteristics of the feature point's motion, serving as the data foundation for extracting evolutionary pattern features. The system employs spatiotemporal feature extraction operators to analyze this matrix, including trend analysis, periodicity analysis, and correlation analysis. Trend analysis uses linear regression to calculate the changing trends of each state component; periodicity analysis uses autocorrelation functions to detect periodic patterns in state changes; and correlation analysis calculates the correlation coefficients between different state components to identify the dependencies between state variables.

[0143] The analysis of the aforementioned feature points shows that the x and y coordinates exhibit a linear increasing trend, with growth rates of 0.37 mm / day and 0.23 mm / day, respectively, which is basically consistent with the velocity components vx and vy. The z coordinate shows a linear decreasing trend, with a decreasing rate of 0.17 mm / day, which is basically consistent with the velocity component vz. Correlation analysis among the state variables shows that the changes in x and y coordinates are strongly positively correlated (correlation coefficient 0.85), indicating that the horizontal displacement is consistent. There is a negative correlation between x, y, and z (correlation coefficients -0.72 and -0.68, respectively), indicating that as the horizontal displacement increases, the vertical displacement also increases (negative for downward displacement). These analytical results collectively constitute the evolutionary pattern characteristics of the feature points, including trend parameters, periodicity parameters, and correlation parameters, which are used for subsequent correction of the Kalman filter algorithm.

[0144] The Kalman filter algorithm is applied starting from the state prediction stage, predicting the state at the next moment based on the current state vector and the state transition model. The state transition model adopts a uniform motion model, assuming that the displacement velocity remains constant over a short period of time. For the state vector (x, y, z, vx, vy, vz), the predicted state at the next moment (interval Δt = 1 day) is (x + vx × Δt, y + vy × Δt, z + vz × Δt, vx, vy, vz). For example, if the current state is (3.30, 2.64, 1.82, 0.38, 0.23, -0.19), the predicted state at the next moment is (3.304, 2.642, 1.818, 0.38, 0.23, -0.19). This prediction is based on simplified motion assumptions and needs to be corrected by incorporating evolutionary pattern characteristics.

[0145] The system corrects the state prediction results based on evolutionary pattern characteristics, considering historical trends, periodicity, and correlation constraints of feature points. For trends, the system compares the predicted displacement velocity with historical trends; if the deviation exceeds 20%, the velocity component is adjusted to align with the historical trend. For periodicity, the system checks the current time point's position within the cycle and adjusts the predicted state to conform to the periodic change pattern. For correlation constraints, the system ensures that the correlation between different state components matches the results of historical data analysis. In the above special case, the predicted vx value (0.38) is very close to the historical trend (0.37) and requires no adjustment; however, considering the strong correlation between x and y, the system fine-tunes the vy value to 0.24, making its ratio with vx closer to historical data. The corrected predicted state vector is (3.304, 2.642, 1.818, 0.38, 0.24, -0.19).

[0146] Real-time observation data of key monitoring feature points is acquired using a depth camera at a rate of 30 frames per second with a spatial resolution of 2 millimeters. The acquired depth images are processed by a feature extraction algorithm to obtain the real-time 3D coordinates of the feature points. Due to measurement noise, real-time observation data typically contains random errors. To reduce the impact of noise, the system averages 30 consecutive frames of data to obtain more reliable observation results. At a certain moment, the real-time observed coordinates of the feature point are (3.306, 2.645, 1.816), which differs slightly from the predicted coordinates.

[0147] The system employs a weighted fusion of the predicted state vector and real-time observation data. Based on the uncertainty of the predicted state and the noise level of the observation data, the optimal weighting coefficients are calculated. The uncertainty of the predicted state is represented by the state covariance matrix, initially set as a diagonal matrix, with diagonal elements representing the variances of the coordinate components (0.0001 m). 2 The variance of the velocity component is 0.0004 mm / day²; the observation noise is represented by the observation covariance matrix, which is set as a diagonal matrix, with the diagonal elements being the variance of the measurement error (0.0009 m). 2 Based on these parameters, the system calculates the Kalman gain to determine the weights of the predicted and observed values. In the above case, the calculated Kalman gain for the position component is approximately 0.1, indicating that the fusion result mainly depends on the predicted value (weight 0.9), supplemented by fine-tuning of the observed value (weight 0.1). The fused state estimate is (3.305, 2.644, 1.817, 0.39, 0.25, -0.19).

[0148] By continuously applying the Kalman filter algorithm, the system obtains the state sequence of feature points over time, i.e., the motion trajectory. This trajectory includes the position and velocity changes of the feature points in three-dimensional space, reflecting the motion law of the feature points. For the aforementioned feature points, 30 days of monitoring formed a trajectory containing 30 points, showing that the point continuously moves outward in the horizontal direction and continuously sinks in the vertical direction, with a total displacement of approximately 12 mm. The system calculates the displacement between adjacent sampling points in the trajectory, obtaining 29 displacement data points. Under normal circumstances, these displacements should be relatively stable, reflecting the stable motion pattern of the feature points. Statistical analysis of the 29 displacement data points shows a mean displacement of 0.42 mm / day and a standard deviation of 0.07 mm / day, indicating that the displacement rate of the feature point is relatively stable.

[0149] A preset evaluation threshold is constructed based on the statistical characteristics of displacement. An adaptive threshold strategy is adopted, dynamically adjusting the threshold according to the mean and standard deviation of the displacement. Specifically, the preset evaluation threshold equals the mean displacement plus the standard deviation multiplied by a safety factor. The safety factor is set according to the importance and risk level of the monitored object, typically ranging from 3 to 5. For important transmission towers, the safety factor is set to 4, resulting in a preset evaluation threshold of 0.42 + 0.07 × 4 = 0.70 mm / day. This means that when the displacement exceeds 0.70 mm / day on a given day, the system considers the movement of that feature point abnormal and requires the generation of an early warning.

[0150] Anomaly detection for displacement employs a sliding window strategy, considering displacement data from the most recent three days for each detection to reduce the impact of random factors. When the average displacement over three consecutive days exceeds a preset evaluation threshold, the system generates an early warning message. This message includes the feature point ID, location information, abnormal displacement, historical displacement trend, and degree of anomaly. The degree of anomaly is assessed based on the proportion of displacement exceeding the threshold, categorized as minor anomaly (0-20% over), moderate anomaly (20-50% over), and severe anomaly (over 50% over). In one monitoring instance, the system detected an average displacement of 0.85 mm / day for a certain feature point at the bottom over three consecutive days, exceeding the preset threshold of 0.70 mm / day by approximately 21.4%, thus classifying it as a moderate anomaly and generating the corresponding early warning message.

[0151] After an early warning is generated, the system adopts different handling strategies based on the severity of the anomaly. For minor anomalies, the system increases the monitoring frequency of the area from once a day to three times a day. For moderate anomalies, the system initiates a deep inspection process, calling on higher-precision sensors to perform a detailed scan of the anomaly area and notifying maintenance personnel to pay attention. For severe anomalies, the system immediately sends an emergency alert to the maintenance center, recommending on-site inspection and emergency handling. Through this tiered response mechanism, the system can take appropriate measures according to the severity of the anomaly, ensuring that security risks are addressed promptly while avoiding resource waste caused by over-response.

[0152] In practical applications, this method has been successfully applied to the monitoring of transmission line towers. In the monitoring of one transmission line, the system continuously tracked 138 key monitoring points for 90 days, detecting 12 anomalies, including 8 minor anomalies, 3 moderate anomalies, and 1 severe anomaly. Field verification showed that 10 of these anomalies corresponded to actual structural deformation, while 2 were temporary anomalies caused by environmental factors (such as strong winds and temperature changes). This method significantly improves the accuracy and reliability of monitoring safety hazards along transmission line towers, providing strong support for the safe operation of the power grid, inspections around transmission lines, and testing of the tower structure itself.

[0153] In one optional implementation, a spatiotemporal sequence matrix is ​​constructed based on the historical state sampling sequence, and a spatiotemporal feature extraction operator is used to analyze the spatiotemporal sequence matrix to obtain the evolutionary pattern features of the key monitoring feature points, including:

[0154] The historical state sampling sequence is slide-divided according to the preset time window length, and the state vectors in each time window are arranged in time sequence to construct a spatiotemporal sequence matrix.

[0155] The temporal correlation of the state vector is calculated based on the spatiotemporal sequence matrix. The temporal correlation is characterized by the covariance between the state vector and its mean to obtain the temporal correlation feature. The position difference and velocity difference of adjacent state vectors in the spatiotemporal sequence matrix are calculated, and the position difference and velocity difference are combined to obtain the spatial difference feature.

[0156] The temporal correlation features and the spatial difference features are weighted and fused according to preset weights to obtain spatiotemporal fusion features. Singular value decomposition is performed on the spatiotemporal fusion features to obtain feature vector values ​​and feature values. The proportion of the feature vector values ​​and feature values ​​to the total feature values ​​is calculated to obtain the feature contribution. The feature contribution is sorted from largest to smallest. Feature vectors whose cumulative feature contribution reaches a preset contribution threshold are selected to construct evolutionary pattern features. The evolutionary pattern features are used to characterize the evolutionary pattern features of the key monitoring feature points.

[0157] The historical state sampling sequence is segmented using a sliding window method, with a preset time window length of 10 days and a sliding step size of 1 day. For 30 days of continuous monitoring data, 21 time windows can be obtained, each containing 10 consecutive state vectors. Taking the connection point of the main column at the bottom of a 220kV transmission tower as an example, its state vector contains six parameters: position coordinates (x, y, z) and velocity components (vx, vy, vz). Within the first time window, the state vector sequence of this feature point is as follows: Day 1 (3.250, 2.600, 1.850, 0.350, 0.220, -0.180), Day 2 (3.254, 2.602, 1.848, 0.355, 0.218, -0.182), and so on up to Day 10 (3.280, 2.620, 1.835, 0.370, 0.230, -0.190). Arrange these 10 state vectors in temporal order to construct a 10-row, 6-column spatiotemporal sequence matrix. Each row in the matrix corresponds to the state at a given time point, and each column corresponds to the sequence of changes of a state component over time.

[0158] After the spatiotemporal sequence matrix is ​​constructed, the system calculates the temporal correlation of the state vectors. The calculation process first calculates the mean of each column of the spatiotemporal sequence matrix to obtain the average value of the state vectors. For the aforementioned feature points, within the first time window, the mean of the position coordinates is (3.267, 2.611, 1.842), and the mean of the velocity components is (0.363, 0.224, -0.186). Subsequently, the covariance of each state vector with respect to the mean is calculated, resulting in a 6×6 covariance matrix. The diagonal elements of this matrix represent the variance of each state component itself, and the off-diagonal elements represent the covariance between different state components. In the actual calculation, the diagonal elements of the covariance matrix are: x-coordinate variance 0.0001, y-coordinate variance 0.00005, z-coordinate variance 0.00003, vx velocity variance 0.00007, vy velocity variance 0.00004, and vz velocity variance 0.00002. Among the off-diagonal elements, the covariance between x and y is 0.00006, indicating a positive correlation; the covariance between x and z is -0.00005, indicating a negative correlation. These covariance values ​​collectively constitute the temporal correlation feature, describing the correlation pattern of the components of the state vector over time.

[0159] Calculating the position and velocity differences between adjacent state vectors in the spatiotemporal sequence matrix is ​​a crucial step in obtaining spatial difference features. For 10 consecutive state vectors, the system calculates 9 position differences and 9 velocity differences. The position difference calculates the difference in position coordinates between two adjacent days, and the velocity difference calculates the difference in velocity components between two adjacent days. For example, the position difference between day 1 and day 2 is (0.004, 0.002, -0.002), and the velocity difference is (0.005, -0.002, -0.002); the position difference between day 2 and day 3 is (0.003, 0.003, -0.002), and the velocity difference is (0.002, 0.003, -0.001). Statistical characteristics are calculated for all difference values, including the mean difference, the variance of the difference, and the autocorrelation coefficient of the difference sequence.

[0160] In the above example, the average value of the position difference is (0.0033, 0.0022, -0.0017), indicating that the feature point has a slight increasing trend in the x and y directions and a slight decreasing trend in the z direction; the average value of the velocity difference is (0.0022, 0.0013, -0.0011), indicating that the velocity change is relatively stable. The variance of the position difference is (0.000001, 0.0000008, 0.0000005), and the variance of the velocity difference is (0.000002, 0.0000015, 0.0000007). These small variance values ​​indicate that the movement of the feature point is very smooth. The autocorrelation coefficients of the difference sequences are 0.85, 0.72, and 0.61 in the 1-3 day delay range, respectively, indicating that the movement in the short term has strong continuity and predictability. These statistical characteristics together constitute the spatial difference features, describing the movement pattern of the feature point in space.

[0161] The weighted fusion of temporal correlation features and spatial difference features is a crucial step in constructing comprehensive spatiotemporal fusion features. The system sets fusion weights based on the characteristics of the monitored objects and application requirements; typically, the weight for temporal correlation features is 0.6, and the weight for spatial difference features is 0.4. Temporal correlation features contain 6 × 6 = 36 covariance values, while spatial difference features contain 18 statistical values ​​for position and velocity differences. Since the two feature sets have different dimensions, normalization is required to make them comparable. The normalization method involves dividing each feature value by the maximum value of its category, ensuring that all feature values ​​fall within the range of 0-1.

[0162] After normalization, the two feature components are concatenated into a 54-dimensional spatiotemporal fusion feature vector according to preset weights. For the aforementioned feature points, the maximum value of the normalized temporal correlation feature is 0.95 (corresponding to the autocorrelation of the x-coordinate), and the minimum value is 0.08 (corresponding to the cross-correlation between x-velocity and z-velocity); the maximum value of the normalized spatial difference feature is 0.88 (corresponding to the mean of the x-position difference), and the minimum value is 0.12 (corresponding to the variance of the z-velocity difference). These 54 feature values ​​together constitute the spatiotemporal fusion feature, comprehensively describing the spatiotemporal variation patterns of the feature points.

[0163] Singular value decomposition (SVD) of spatiotemporal fusion features is an effective method for dimensionality reduction and extraction of key feature patterns. The system organizes the 54-dimensional spatiotemporal fusion feature vectors into a matrix and performs SVD on this matrix to obtain eigenvectors and their corresponding eigenvalues. The SVD results in 54 eigenvectors and their corresponding eigenvalues, representing different variation patterns in the data. The magnitude of the eigenvalue indicates the importance of the corresponding eigenvector; the larger the eigenvalue, the greater its contribution to the original data. In the actual calculation, the first 10 eigenvalues ​​are 5.8, 4.2, 3.1, 2.6, 1.9, 1.5, 1.2, 0.9, 0.7, and 0.5, respectively, with the remaining eigenvalues ​​all less than 0.5. The system calculates the proportion of each eigenvalue to the total sum of eigenvalues ​​to obtain the feature contribution. The contribution rates of the top 10 eigenvalues ​​were 22.6%, 16.3%, 12.1%, 10.1%, 7.4%, 5.8%, 4.7%, 3.5%, 2.7%, and 1.9%, respectively, with a cumulative contribution rate of 87.1%.

[0164] The feature contributions are sorted from largest to smallest, and a preset contribution threshold of 85% is set. Feature vectors whose cumulative contribution reaches the threshold are selected to construct evolutionary pattern features. In the example above, the cumulative contribution of the top 9 feature vectors is 85.2%, exceeding the preset threshold of 85%, so the top 9 feature vectors are selected as key features. These 9 feature vectors represent different change patterns: the first feature vector mainly represents the synchronous change pattern of x and y coordinates, with corresponding coefficients of 0.42 and 0.38, respectively; the second feature vector mainly represents the change pattern of z coordinate, with a corresponding coefficient of 0.56; the third feature vector mainly represents the synchronous change pattern of vx and vy velocities, with corresponding coefficients of 0.45 and 0.40, respectively; the fourth to ninth feature vectors represent other more complex change patterns, such as the interaction between different state components. These 9 feature vectors and their corresponding eigenvalues ​​constitute the evolutionary pattern features, used to characterize the spatiotemporal change patterns of key monitored feature points.

[0165] The extracted evolutionary pattern features are used to adjust the parameters of the Kalman filter, mainly including the state transition matrix and the process noise covariance matrix. The state transition matrix describes how the system state evolves from the current moment to the next. Traditional Kalman filtering typically assumes a uniform motion model, while evolutionary pattern features provide more accurate state transition relationships. For example, the first eigenvector indicates that the x and y coordinates change synchronously; the system adjusts the corresponding elements in the state transition matrix accordingly to ensure the prediction reflects this synchronicity. The process noise covariance matrix describes the uncertainty of the state prediction. The magnitude of the eigenvalues ​​in the evolutionary pattern features indicates the stability of different change patterns; larger eigenvalues ​​indicate more stable patterns, and the corresponding process noise should be set lower; smaller eigenvalues ​​indicate less stable patterns, and the corresponding process noise should be set higher. In this way, the Kalman filter algorithm can fully utilize the evolutionary pattern features extracted from historical data, improving the accuracy and stability of predictions.

[0166] In practical applications, the system extracts evolutionary pattern features for each key monitoring feature point and updates them periodically. Typically, the evolutionary pattern features are updated every 10 days to ensure that the features promptly reflect changes in the feature point's motion pattern. For sudden changes, the system has a trigger mechanism that immediately recalculates the evolutionary pattern features when the deviation between observed data and predicted values ​​exceeds a preset threshold. For example, if the positional deviation of an observation exceeds 5 mm or the velocity deviation exceeds 0.2 mm / day, the system considers the feature point's motion pattern to have changed and requires re-extraction of the evolutionary pattern features. This dynamic update mechanism enables the system to adapt to changes in the feature point's motion pattern, maintaining the accuracy and reliability of tracking.

[0167] The evolutionary pattern features extracted using the above method can effectively characterize the movement patterns of key monitoring points, providing a reliable basis for subsequent dynamic tracking and early warning judgment. Compared with traditional methods, this method fully utilizes the spatiotemporal correlation information contained in historical data, enabling more accurate prediction of the movement trajectory of feature points, improving the sensitivity and accuracy of anomaly detection, and providing technical support for the timely discovery and handling of safety hazards in steel towers. In safety monitoring practices of multiple steel towers, this method has proven to reduce deformation prediction errors by approximately 35%, effectively reducing false alarm and false negative rates, and significantly improving the reliability and practicality of the monitoring system.

[0168] In an optional implementation, the method further includes:

[0169] Deploy drone airports on iron towers, where drones are parked, charged, and communicated.

[0170] The inspection path is determined based on the preset inspection strategy. The inspection strategy includes a leapfrog inspection path, which is a flight path in which the drone takes off from the drone airport of the current tower, flies to the drone airport of the adjacent tower to park and charge, and then flies to the drone airport of the next tower.

[0171] The drone is controlled to inspect the area surrounding the tower along the inspection path, collecting image data of the area. The image data is analyzed to identify illegal land reclamation, unauthorized grazing, fire hazards, and abnormal animal intrusion in the area surrounding the tower. When a safety hazard is detected, a hazard alarm is generated and sent to the monitoring terminal.

[0172] The drone airport deployed on the tower includes a charging bay, a communication module, and a protective shield. The charging bay features a foldable design, automatically unfolding during drone landing and closing afterward to protect the drone. An inductive charging coil is installed inside the charging bay, automatically initiating charging once the drone has landed accurately. The communication module includes a 5G communication unit and a short-range communication unit. The 5G unit is used for data transmission with a remote monitoring center, while the short-range communication unit is used for communication and control with the drone. The protective shield is made of waterproof and breathable material, providing protection against rain, dust, and bird interference.

[0173] The drone airport is fixed to the crossarm at the top of the tower, installed at a height of 35-40 meters above the ground. Each drone airport is equipped with a quadcopter drone with a flight time of 60 minutes and a maximum flight speed of 20 meters per second. The drone is equipped with a 4K high-definition visible light camera and an infrared thermal imaging camera. The visible light camera has a field of view of 120 degrees, and the infrared camera has a field of view of 60 degrees. Both cameras support 30x optical zoom.

[0174] The inspection path planning is designed based on the distribution of the towers. The distance between adjacent towers is typically 300-500 meters. Considering the drone's endurance, a single inspection is limited to traversing a maximum of four towers. After taking off from the current tower, the drone flies to the next tower at a set altitude of 45-50 meters and a speed of 15 meters per second. Upon reaching the target tower, the drone circles around it once, with a flight radius of 100 meters, collecting images of the surrounding environment. After completing the data collection, it lands at the drone airfield on that tower for charging, which takes 15 minutes. After charging, it continues to the next tower, forming a leapfrog inspection pattern.

[0175] When inspecting the area surrounding the tower, the drone uses a spiral descent scanning method, starting from a height of 50 meters and collecting images every 5 meters until it reaches a height of 25 meters. During the data collection process, both visible light and infrared cameras operate simultaneously. The visible light camera is mainly used for daytime inspections, with an image resolution of 3840x2160 pixels; the infrared camera is mainly used for nighttime inspections, with a resolution of 640x512 pixels.

[0176] The system intelligently analyzes and processes the collected image data. For illegal land reclamation, it detects newly added farmland traces in the images and triggers an alarm when more than 100 square meters of newly cultivated land are detected within 50 meters of the tower. For illegal grazing, the system identifies livestock in the images and generates an alert when livestock activity is detected within 30 meters of the tower. Fire hazard monitoring primarily relies on infrared cameras; an alarm is triggered when an abnormal temperature exceeds 150 degrees Celsius or an area with an abnormal temperature exceeds 2 square meters. The criterion for abnormal animal intrusion is the detection of large wild animals within 20 meters of the tower.

[0177] Alarm information is categorized into three levels: red alerts indicate serious hazards requiring immediate attention, such as fires or large-scale illegal land reclamation; yellow alerts indicate potential hazards requiring monitoring, such as animal intrusion or small-scale illegal grazing; and green alerts indicate minor anomalies requiring recording. Alarm information includes the hazard type, location coordinates, on-site images, and discovery time, and is transmitted in real-time to the monitoring center via a 5G network. Upon receiving an alert, the monitoring center activates the corresponding emergency response mechanism based on the alert level.

[0178] The system is set to conduct four fixed inspections daily, at 6:00 AM, 12:00 PM, 6:00 PM, and 12:00 AM, with each inspection lasting 90 minutes. In addition, the system automatically increases the inspection frequency in special weather conditions such as strong winds or heavy rain. When wind speeds exceed 12 meters per second or rainfall exceeds 50 millimeters per hour, the drone automatically returns to the airport to stand by. In this way, all-weather, automated monitoring of safety hazards around the towers is achieved.

[0179] A second aspect of the present invention provides a monitoring system for safety hazards along power transmission towers based on unmanned aerial vehicles (UAVs), comprising:

[0180] The first unit is used to acquire three-dimensional model data of the target tower, and determine the collaborative inspection path and data collection position of multiple drones based on the three-dimensional model data; control multiple drones to fly to the corresponding data collection position according to the collaborative inspection path, wherein each drone is equipped with a depth camera;

[0181] The second unit is used to establish a hierarchical acquisition strategy based on the structural stress analysis results of the target tower, divide the target tower into multiple deformation-sensitive areas, and dynamically allocate the acquisition density and acquisition frequency of the UAV according to the stress characteristics of each deformation-sensitive area; and control multiple UAVs to simultaneously acquire multi-view depth images of the target tower according to the hierarchical acquisition strategy.

[0182] The third unit is used to construct a real-time three-dimensional point cloud model of the target tower based on the multi-view depth image, register the real-time three-dimensional point cloud model with the three-dimensional model data of the target tower, extract deformation feature points, and use a dual threshold iteration method to classify and filter the deformation feature points, and determine the key monitoring feature points based on the filtering results.

[0183] The fourth unit is used to dynamically track the key monitoring feature points using the Kalman filter algorithm to obtain the motion trajectory of the deformation feature points. When the displacement of the motion trajectory exceeds a preset evaluation threshold, an early warning message is generated.

[0184] A third aspect of the present invention provides an electronic device, comprising:

[0185] processor;

[0186] Memory used to store processor-executable instructions;

[0187] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0188] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0189] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0190] 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 scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for monitoring safety hazards along power transmission towers based on unmanned aerial vehicles (UAVs), characterized in that, include: The system acquires 3D model data of a target tower, which includes various types of towers, and determines the collaborative inspection path and data collection location of multiple drones based on the 3D model data. The system controls multiple drones to fly to the corresponding data collection location according to the collaborative inspection path, wherein each drone is equipped with a depth camera. Based on the structural stress analysis results of the target tower, a layered acquisition strategy is established, dividing the target tower into multiple deformation-sensitive regions, and dynamically allocating the acquisition density and acquisition frequency of the UAV according to the stress characteristics of each deformation-sensitive region; and controlling multiple UAVs to simultaneously acquire multi-view depth images of the target tower according to the layered acquisition strategy. Based on the multi-view depth images, a real-time 3D point cloud model of the target tower is constructed. The real-time 3D point cloud model is registered with the 3D model data of the target tower, deformation feature points are extracted, and the deformation feature points are graded and screened using a dual threshold iteration method. Based on the screening results, key monitoring feature points are determined. The key monitoring feature points are dynamically tracked using a Kalman filter algorithm to obtain the motion trajectory of the deformation feature points. When the displacement of the motion trajectory exceeds a preset evaluation threshold, an early warning message is generated.

2. The method according to claim 1, characterized in that, Based on the 3D model data, the collaborative inspection path and data collection location of multiple UAVs are determined; controlling the multiple UAVs to fly to the corresponding data collection location according to the collaborative inspection path includes: Based on the three-dimensional model data, a target spatial feature model is constructed, and the structural features of the tower are analyzed according to the target spatial feature model. Based on the structural features, inspection points are determined, and the inspection points are mapped to three-dimensional space to form a set of collection locations. Based on the set of collection locations, a collaborative inspection path for multiple drones is generated. A distributed communication network is constructed for multiple UAVs, and a communication topology matrix is ​​generated based on the spatial distribution of the multiple UAVs; the position vector, velocity vector, and obstacle perception information of each UAV are obtained through the distributed communication network, and the position vector, velocity vector, and obstacle perception information are shared in real time among the multiple UAVs; Based on the obstacle perception information, a target point gravitational potential field term is constructed, and an obstacle repulsive potential field term and a cooperative obstacle avoidance potential field term are constructed based on the position vector. The target point gravitational potential field term, the obstacle repulsive potential field term, and the cooperative obstacle avoidance potential field term are weighted and combined to form a comprehensive potential field function. The gradient of the integrated potential field function is calculated, and the gradient is multiplied by the step size factor to obtain the path adjustment amount. The collaborative inspection path is dynamically optimized according to the path adjustment amount to obtain an optimized inspection path that meets the safety distance requirements between UAVs and the connectivity requirements of the communication network. Multiple UAVs are controlled to fly to the corresponding collection positions according to the optimized inspection path.

3. The method according to claim 1, characterized in that, A real-time 3D point cloud model of the target tower is constructed based on the multi-view depth images. The real-time 3D point cloud model is registered with the 3D model data of the target tower, and deformation feature points are extracted, including: The stress distribution field of the tower at different heights and wind speeds is calculated to generate a stress distribution map. A three-dimensional stress variation curve is constructed based on the stress distribution map. The stress gradient is calculated in the three-dimensional stress variation curve, and the area with the largest stress gradient is identified as the key monitoring area. Multiple depth cameras are deployed in the key monitoring area, and structural feature points are extracted from the multi-view depth images using the depth cameras. The spatial location information of the structural feature points is fused with the stress distribution map to establish a correspondence between the feature point locations and the stress distribution, and a weighting coefficient is assigned to the structural feature points according to the correspondence. Based on the weighting coefficients, the multi-view depth image is registered, and the registration accuracy requirement is improved in areas where the stress is greater than the preset registration threshold to generate a real-time three-dimensional point cloud model. The structural feature points in the real-time three-dimensional point cloud model are compared with the corresponding positions in the standard three-dimensional model, and the displacement direction and displacement amplitude of each structural feature point are calculated. Based on the displacement direction and displacement amplitude, a connectivity analysis is performed on the structural feature points to identify deformation-related point groups. Feature points in the deformation-related point groups that do not conform to the stress distribution law are removed to obtain the final deformation feature points.

4. The method according to claim 3, characterized in that, Based on the displacement direction and displacement amplitude, connectivity analysis is performed on the structural feature points to identify deformation-related point groups. Feature points in the deformation-related point groups that do not conform to the stress distribution law are removed, resulting in the final deformation feature points, including: Based on the displacement direction and displacement amplitude, a displacement feature vector is constructed. The Euclidean distance and spatial distance between the displacement feature vectors are calculated. A feature correlation matrix is ​​constructed by combining the Euclidean distance and the spatial distance. The connectivity strength of the structural feature points is calculated based on the feature correlation matrix. The structural feature points are clustered based on the connectivity strength to obtain multiple deformation correlation point groups. Calculate the mean and deviation of the displacement feature vectors of the structural feature points in the deformation-related point group, and construct the deformation distribution features by combining the mean and deviation of the displacement feature vectors. Obtain stress distribution data of the tower, calculate stress values ​​and stress gradients at structural feature points based on the stress distribution data, construct stress distribution features using the stress values ​​and stress gradients, input the stress distribution features and deformation distribution features into a scoring function, and filter the structural feature points based on the output value of the scoring function to obtain the final deformation feature points.

5. The method according to claim 1, characterized in that, The deformation feature points are hierarchically screened using a dual-threshold iteration method. Based on the screening results, the key monitoring feature points are determined to include: An initial threshold and a target threshold are constructed. The initial threshold is used to perform a first screening of deformation feature points to obtain candidate feature points. The deformation amplitude distribution of the candidate feature points is calculated. The target threshold is dynamically adjusted according to the deformation amplitude distribution. Using the target threshold as the iterative target, the candidate feature points are progressively screened. When the screening results are stable, the candidate feature points that are greater than the target threshold are identified as key monitoring feature points.

6. The method according to claim 1, characterized in that, The key monitoring feature points are dynamically tracked using a Kalman filter algorithm to obtain the motion trajectory of the deformation feature points. When the displacement of the motion trajectory exceeds a preset evaluation threshold, an early warning message is generated, including: The three-dimensional coordinates and displacement velocity of key monitoring feature points are obtained. The three-dimensional coordinates and displacement velocity are used to construct a state vector of the key monitoring feature points. Based on the state vector, historical state sampling is performed on the key monitoring feature points to obtain a historical state sampling sequence. A spatiotemporal sequence matrix is ​​constructed based on the historical state sampling sequence. The spatiotemporal sequence matrix is ​​analyzed using a spatiotemporal feature extraction operator to obtain the evolution pattern features of the key monitoring feature points. The evolution pattern features are then used as correction parameters for the Kalman filter algorithm. The state vector is input into the Kalman filter algorithm for state prediction. The state prediction result is corrected based on the evolutionary pattern characteristics to obtain the predicted state vector. Real-time observation data of the key monitoring feature points are obtained. The predicted state vector and the real-time observation data are weighted and fused to obtain the motion trajectory of the key monitoring feature points. The displacement between adjacent sampling points in the motion trajectory is calculated, and the standard deviation and mean of the displacement are obtained by statistical analysis. A preset evaluation threshold is constructed based on the standard deviation and the mean. When the displacement is greater than the preset evaluation threshold, an early warning message is generated.

7. The method according to claim 6, characterized in that, A spatiotemporal sequence matrix is ​​constructed based on the historical state sampling sequence. The spatiotemporal sequence matrix is ​​then analyzed using a spatiotemporal feature extraction operator to obtain the evolutionary pattern features of the key monitoring feature points, including: The historical state sampling sequence is slide-divided according to the preset time window length, and the state vectors in each time window are arranged in time sequence to construct a spatiotemporal sequence matrix. The temporal correlation of the state vector is calculated based on the spatiotemporal sequence matrix. The temporal correlation is characterized by the covariance between the state vector and its mean to obtain the temporal correlation feature. The position difference and velocity difference of adjacent state vectors in the spatiotemporal sequence matrix are calculated, and the position difference and velocity difference are combined to obtain the spatial difference feature. The temporal correlation features and the spatial difference features are weighted and fused according to preset weights to obtain spatiotemporal fusion features. Singular value decomposition is performed on the spatiotemporal fusion features to obtain feature vector values ​​and feature values. The proportion of the feature vector values ​​and feature values ​​to the total feature values ​​is calculated to obtain the feature contribution. The feature contribution is sorted from largest to smallest. Feature vectors whose cumulative feature contribution reaches a preset contribution threshold are selected to construct evolutionary pattern features. The evolutionary pattern features are used to characterize the evolutionary pattern features of the key monitoring feature points.

8. The method according to claim 1, further comprising: Deploy drone airports on iron towers, where drones are parked, charged, and communicated. The inspection path is determined based on the preset inspection strategy. The inspection strategy includes a leapfrog inspection path, which is a flight path in which the drone takes off from the drone airport of the current tower, flies to the drone airport of the adjacent tower to park and charge, and then flies to the drone airport of the next tower. The drone is controlled to inspect the area around the tower according to the inspection path and to collect image data of the area around the tower. The image data is analyzed to identify illegal land reclamation, unauthorized grazing, fire hazards, and abnormal animal intrusion in the area surrounding the tower. When a safety hazard is detected, a hazard alarm is generated and sent to the monitoring terminal.

9. A UAV-based monitoring system for safety hazards along power transmission towers, used to implement the method described in any one of claims 1-8, characterized in that, include: The first unit is used to acquire three-dimensional model data of the target tower, and determine the collaborative inspection path and data collection position of multiple drones based on the three-dimensional model data; control multiple drones to fly to the corresponding data collection position according to the collaborative inspection path, wherein each drone is equipped with a depth camera; The second unit is used to establish a hierarchical acquisition strategy based on the structural stress analysis results of the target tower, divide the target tower into multiple deformation-sensitive areas, and dynamically allocate the acquisition density and acquisition frequency of the UAV according to the stress characteristics of each deformation-sensitive area; and control multiple UAVs to simultaneously acquire multi-view depth images of the target tower according to the hierarchical acquisition strategy. The third unit is used to construct a real-time three-dimensional point cloud model of the target tower based on the multi-view depth image, register the real-time three-dimensional point cloud model with the three-dimensional model data of the target tower, extract deformation feature points, and use a dual threshold iteration method to classify and filter the deformation feature points, and determine the key monitoring feature points based on the filtering results. The fourth unit is used to dynamically track the key monitoring feature points using the Kalman filter algorithm to obtain the motion trajectory of the deformation feature points. When the displacement of the motion trajectory exceeds a preset evaluation threshold, an early warning message is generated.

10. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 8.

Images