Communication facility data automatic acquisition and three-dimensional modeling system based on unmanned aerial vehicle

Abstract

The invention provides a communication facility data automatic acquisition and three-dimensional modeling system based on an unmanned aerial vehicle. Multi-dimensional data of a communication facility is acquired through a data acquisition module; the data fusion module extracts features from the multi-dimensional data, captures association among cross-modal data, and dynamically allocates feature weights in combination with an attention mechanism to obtain fusion features; the target detection module performs target detection based on multi-scale detection of a YOLO-SCDI algorithm and a dynamic attention mechanism according to the fusion features so as to match the image, the point cloud and the position data of the same communication facility; the model generation module identifies facility components in the matched data through a built-in communication facility feature library, calls a feature library model to be spliced and optimized with the identified communication facility data, and generates an editable three-dimensional model; automatic acquisition and three-dimensional modeling of communication facility data are realized, and the precision of three-dimensional modeling is improved while the accuracy and integrity of the data are guaranteed.

Description

Technical Field

[0001] This invention relates to the field of digital operation and maintenance technology for communication facilities, and more particularly to a system for automatic data acquisition and 3D modeling of communication facilities based on unmanned aerial vehicles (UAVs). Background Technology

[0002] In today's era of rapid digital and information technology development, communication facilities, as key carriers of information transmission, require efficient management and maintenance. Accurate and comprehensive data collection from communication facilities, and the construction of high-precision 3D models, are of paramount importance for all aspects of the communication industry, including planning, construction, operation, and management. However, traditional methods of automatic data collection and modeling for communication facilities have many limitations. Traditional data collection relies on manual climbing or handheld devices, which is not only inefficient but also dangerous, making it difficult to guarantee the accuracy and integrity of the data. When using fixed-point 3D coordinate measuring machines and 3D laser scanners to collect point cloud data from equipment rooms, the data is easily affected by factors such as object occlusion and uneven lighting, resulting in blind spots and holes in areas with complex shapes, low modeling accuracy, and a lack of temporal and engineering attributes. Traditional methods of communication facility data collection and monitoring not only consume significant manpower and resources but are also inefficient and inaccurate, failing to meet the needs of the rapidly developing communication industry.

[0003] Therefore, there is an urgent need for a system that integrates automatic data acquisition and 3D modeling. This system can automatically acquire, analyze, and model fused multimodal data, combining the advantages of multimodal data and 3D modeling to address the shortcomings of existing technologies. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides an automatic data acquisition and 3D modeling system for communication facilities based on unmanned aerial vehicles (UAVs). The system automatically acquires multimodal data of communication facilities using data acquisition equipment mounted on a UAV, captures and fuses the correlations between the multimodal data, performs target detection based on the YOLO-SCDI algorithm's multi-scale detection and dynamic attention mechanism, and combines this with a built-in communication facility feature library for 3D modeling. This approach improves the accuracy and completeness of the 3D modeling while ensuring data accuracy and completeness.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: This invention provides an automatic data acquisition and 3D modeling system for communication facilities based on unmanned aerial vehicles (UAVs), including a data acquisition module for communication connection, a data fusion module, a target detection module, and a model generation module; Data acquisition module: Based on 5G and Beidou drone scheduling algorithm, it controls drones to operate in zones according to preset routes, collects multi-dimensional data of communication facilities in real time, and preliminarily filters invalid data through edge computing; the multi-dimensional data includes the appearance texture, three-dimensional coordinates and equipment temperature of communication facilities; Data fusion module: used to extract features from multi-dimensional data acquired by the data acquisition module, capture the correlation between cross-modal data, and dynamically allocate feature weights by combining an attention mechanism to obtain fused features; Target detection module: Based on fused features, target detection is performed using multi-scale detection and dynamic attention mechanism based on YOLO-SCDI algorithm to match image, point cloud and location data of the same communication facility and generate standardized data format; Model generation module: It is used to identify facility components in the matching data through the built-in communication facility feature library, call the feature library model and the identified communication facility data to stitch and optimize, generate an editable 3D model, and annotate key parameters.

[0006] Preferably, edge computing devices are deployed on drones to perform real-time preliminary analysis of the collected data based on pre-trained deep learning models and data filtering algorithms, filtering out invalid data.

[0007] Preferably, the ground station uses a 5G and BeiDou-based drone scheduling algorithm to control drones to operate in zones according to preset flight paths, and collects multi-dimensional information from communication facilities in real time, including the following steps: The ground station receives core data transmitted in real time from each drone via the 5G network and the BeiDou short message communication link, analyzes the core data, and updates the drone dynamic status database. Based on the facility distribution list of the preset route, compare the facility IDs and component data collected by each drone to determine the current task completion rate, and dynamically adjust the task weight in combination with the priority of communication facilities. Based on the current location of the drone, the remaining workload, and geographical environment data, the optimal flight path is replanned using the A* algorithm. Real-time calculation of the relative position, flight direction and flight speed of all airborne drones; setting two levels of monitoring thresholds to assess collision risk. Based on the UAV's flight speed and turning response speed, the minimum safe flight interval is calculated, and avoidance is carried out according to the priority avoidance strategy and the partition scheduling mode. The drone receives adjusted flight routes, task allocation instructions, and avoidance instructions in real time via an encrypted communication link, and then executes them after parsing.

[0008] Preferably, the core data includes the real-time position output by the BeiDou positioning module, the flight attitude fed back by the IMU inertial measurement unit, the mission progress, the sensor working status, and the remaining battery power.

[0009] Preferably, the execution process of the data fusion module includes the following steps: The acquired image, point cloud, and temperature data are standardized, noise filtered, and missing values ​​repaired. Visual features, point cloud features, and temperature features are extracted separately, and then mapped to a shared embedding space through a fully connected layer to construct feature vectors. Constraint loss is introduced to initially establish the correlation between modal data. Based on the contrastive learning mechanism, positive and negative sample pairs are constructed to capture the correlation between data from different modalities, and the feature differences between cross-modal data are eliminated by training the loss function. Modal features are aligned and fused using a three-layer attention mechanism consisting of intra-modal attention, inter-modal attention, and global fusion attention to obtain fused features.

[0010] Preferably, the InfoNCE loss function is used to enhance the feature similarity of positive sample pairs and suppress the association of negative sample pairs. The formula is as follows: ; Where sim(·) represents the mean cosine similarity of the three-modal features, τ=0.1 represents the temperature coefficient, N represents the number of positive samples, M represents the number of negative samples, i represents the index variable of the positive sample pairs, used to traverse the set of all positive sample pairs, and j represents the index variable of the negative sample pairs, used to traverse the set of all negative sample pairs.

[0011] Preferably, the execution process of the target detection module includes the following steps: The fused features are preprocessed and optimized, and the optimized feature map is divided into different detection scales; Construct a YOLO-SCDI multi-scale detection network, extract features from feature maps of different scales, and output candidate detection boxes; The Inner-ShapeIoU metric method is used in conjunction with the geometric features of the communication facilities to calibrate the bounding boxes, and the optimal candidate boxes are matched based on the Hungarian algorithm. Based on the spatial coordinates of the optimal candidate box, combined with the timestamp of BeiDou time synchronization, the spatiotemporal consistency is verified by associating the original multimodal data, and the credibility is verified by a dual verification mechanism of size and temperature. In accordance with the digital operation and maintenance standards for communication facilities, the associated multimodal data and detection results are encapsulated into a standardized format, stored, and output.

[0012] Preferably, the C2f-SCSA module is used in the YOLO-SCDI multi-scale detection network to simultaneously extract local detail features and global structural features, accurately focusing on key component areas of communication facilities; a CCFM-Neck structure is constructed to simultaneously complete channel compression and cross-channel information fusion, and an occlusion perception submodule is embedded in the CCFM to assign higher weights to occluded region features based on the spatial depth information in the fused feature map; a Dynamic Head is used as the detection head, and the feature response is dynamically adjusted through a scale-space-task three-dimensional attention mechanism; a dynamic convolution kernel generation network is introduced to adaptively switch the convolution kernel size according to the characteristics of the input fused features to optimize the detection effect of different types of components; based on the Anchor-Free mechanism, the bounding box coordinates, class probability, and confidence of communication facility components are directly regressed.

[0013] Preferably, the execution process of the model generation module includes the following steps: It receives standardized data and parses it, then calls the built-in communication facility feature library to compare the features of the parsed data. Based on the current 3D modeling objectives, select component data that are highly relevant to the requirements and call the corresponding model templates; Based on the physical installation logic of communication facilities and the component connection rules preset in the feature library, the splicing order and spatial constraint relationship are determined, and splicing errors are corrected. The geometric shape of the assembled preliminary model is optimized, and the geometric accuracy of the model is calibrated. Key parameters required for the operation and maintenance of communication facilities are selected, categorized and labeled according to geometric parameters, electrical parameters, and performance parameters. A linkage mechanism between parameters and model geometry is established, and editing permissions are configured.

[0014] Preferably, it also includes a data encryption module: by generating independent keys for each drone, each communication link, and each storage node, distributing the keys based on BeiDou short messages or encrypted command channels, and periodically and automatically updating the encryption keys, the data is encrypted throughout the entire lifecycle. Risk monitoring module: Based on the optimized 3D model and labeled parameters, combined with real-time collected multimodal data, the module monitors communication facilities in real time and provides risk warnings.

[0015] The beneficial effects of this invention are as follows: By equipping drones with multiple sensors to achieve simultaneous multi-source data acquisition, data collection efficiency is improved. Combined with BeiDou timing and inertial navigation, the data timestamp synchronization accuracy reaches ±1ms, and the spatial coordinate error is ≤5cm, completely eliminating the limitations of traditional manual data collection, which is characterized by low efficiency, poor accuracy, and incomplete coverage. The real-time scheduling system ensures the safety of drones through dynamic route planning and collision avoidance algorithms. At the same time, dynamic optimization of task allocation increases the drone's endurance utilization rate by 20%, improving operational efficiency.

[0016] By employing a cross-modal contrastive learning mechanism and geometric constraint loss, semantic and spatial alignment of image, point cloud, and temperature data is achieved. The dynamic attention mechanism and multi-scale detection architecture of the YOLO-SCDI algorithm adapt to the differentiated detection needs of communication facility components, particularly for components occluded by trees or auxiliary facilities. Through an occlusion perception module that integrates point cloud depth information, the effective recognition rate is improved to over 90%, overcoming detection bottlenecks in complex scenes. The built-in communication facility feature library and automated stitching algorithm significantly shorten the time required for the entire "data-model-parameter" process. Furthermore, the model supports parametric editing, and the high-precision 3D model and standardized parameter annotations form a digital asset library for communication facilities, directly serving the entire process of planning and design, construction and acceptance, and emergency repair. Attached Figure Description

[0017] Figure 1 This is a block diagram of the automatic data acquisition and 3D modeling system for communication facilities based on unmanned aerial vehicles (UAVs) according to the present invention. Detailed Implementation

[0018] Please see Figure 1 As shown, the present invention provides an automatic data acquisition and 3D modeling system for communication facilities based on unmanned aerial vehicles (UAVs), including a data acquisition module, a data fusion module, a target detection module, and a model generation module for communication connection; Data acquisition module: Based on 5G and Beidou drone scheduling algorithm, it controls drones to operate in zones according to preset routes, collects multi-dimensional data of communication facilities in real time, and preliminarily filters invalid data through edge computing; the multi-dimensional data includes the appearance texture, three-dimensional coordinates and equipment temperature of communication facilities; The drone is equipped with a high-definition industrial camera to capture extremely fine textures and structural features on the surface of communication facilities. To ensure high-quality appearance and texture images are acquired under different flight conditions and operating environments, the camera automatically triggers the acquisition frequency based on factors such as the drone's flight speed and the distance to the communication facilities. In addition, the camera is equipped with an intelligent image stabilization algorithm that can compensate for vibrations and jitters during drone flight in real time, further improving image quality. The high-definition industrial camera undergoes intrinsic parameter calibration (using the Zhang Zhengyou calibration method to determine the camera's internal parameters, such as focal length, principal point position, and distortion coefficient) and extrinsic parameter calibration (determining the camera's position and attitude in the world coordinate system).

[0019] The three-dimensional coordinate information of the communication facility is acquired by lidar, and combined with assisted positioning technologies such as inertial measurement unit (IMU) and global navigation satellite system (GNSS), the facility's position and orientation information are obtained in real time. Using this information, the distance data of each lidar reflection point is converted into three-dimensional coordinates (x, y, z) in a geographic coordinate system and integrated into point cloud data. This point cloud data accurately depicts the three-dimensional shape and spatial location of the communication facility, and performs distance and angle calibration on the lidar data.

[0020] Infrared thermal imagers use infrared radiation emitted by communication facilities to acquire temperature data, thereby enabling effective monitoring of equipment operating status. During the data acquisition process, they have an automatic trigger acquisition mechanism to perform temperature calibration on the infrared thermal imager.

[0021] Edge computing devices are deployed on drones to perform real-time preliminary analysis of collected data and filter invalid data based on pre-trained deep learning models and data filtering algorithms.

[0022] Communication between the UAV and the ground station employs redundant communication links, combining 5G networks and satellite communication to ensure uninterrupted communication. To improve communication efficiency and data transmission accuracy, efficient data compression and encryption algorithms are used to compress data collected by the UAV in real time before transmission, reducing data transmission volume and time. Simultaneously, the transmitted data is encrypted to prevent theft or tampering during transmission, ensuring data security.

[0023] The ground station uses a 5G and BeiDou-based drone scheduling algorithm to control drones to operate in zones according to preset routes, and collects multi-dimensional information from communication facilities in real time, including the following steps: The ground station receives core data transmitted in real time from each drone via the 5G network and the BeiDou short message communication link, analyzes the core data, and updates the drone dynamic status database. The core data includes the real-time position (latitude and longitude, altitude, accuracy ±2cm) output by the BeiDou positioning module, the flight attitude (roll angle, pitch angle, yaw angle, error <0.5°) fed back by the IMU inertial measurement unit, the mission progress (number of facilities collected, remaining flight path length, current type of collected components), the sensor working status (collection status and data integrity of camera / LiDAR / infrared thermal imager), and the remaining battery power (accuracy ±1%).

[0024] Based on the facility distribution list of the preset route, compare the facility IDs and component data collected by each drone to determine the current task completion rate, and dynamically adjust the task weight in combination with the priority of communication facilities. For example, when the point cloud void ratio is less than 5% and the Brenner gradient value of the image clarity is greater than 80, the component data is considered complete (qualified). For facility components with unqualified data, they are marked as needing to be supplemented. For areas where all qualified data collection has been completed, the task is considered to be completed. For uncovered facility areas, the remaining data collection workload (including the number of components to be collected and the estimated flight distance) is calculated in real time.

[0025] Based on the drone's current location, remaining workload, and geographical environment data, the following methods are adopted: The algorithm re-plans the optimal route; If a drone completes its current area task and has sufficient battery power (≥30%), the data collection task for the adjacent unassigned area will be assigned to that drone, generating a new zone flight path to ensure that the flight path coverage is non-overlapping and the shortest possible. If the drone has insufficient battery power (<20%), a return charging route will be planned, and its unfinished tasks will be assigned to drones with slower current task progress and sufficient battery power. For components marked for supplementary data collection, targeted supplementary data collection routes will be planned, and flight altitude and angle will be adjusted to ensure that the supplementary data collection is qualified.

[0026] Real-time calculation of the relative position, flight direction and flight speed of all airborne drones; setting two levels of monitoring thresholds to assess collision risk. Two levels of monitoring thresholds are set: a safety warning threshold of 50m and a collision risk threshold of 20m. When the relative distance between two drones is ≤50m, the system enters a warning state and continuously tracks the relative motion trend; when the relative distance is ≤20m or it is predicted that they will enter the 20m range within the next 3 seconds, a collision risk is determined, and the avoidance mechanism is triggered.

[0027] Based on the UAV's flight speed and turning response speed, the minimum safe flight interval is calculated, and avoidance is carried out according to the priority avoidance strategy and the partition scheduling mode. To address collision risks, a priority avoidance strategy is employed. Drones with higher mission priority maintain their original flight path and speed, while drones with lower mission priority adjust their flight direction or reduce speed until the relative distance between them is ≥30m. If two drones have the same priority, the drone that must avoid collision is determined by its flight direction: drones flying clockwise deflect to the left, and drones flying counter-clockwise deflect to the right, ensuring no intersections in the avoidance paths. For areas with dense drone traffic, a zoned scheduling model is used, dividing temporary flight channels and limiting each channel to only one drone at a time to avoid collisions.

[0028] The drone receives adjusted flight routes, task allocation instructions, and avoidance instructions in real time via an encrypted communication link, and then executes them after parsing.

[0029] The flight path includes a sequence of latitude and longitude coordinates, flight altitude, and speed parameters; the task allocation command includes the ID of the facility to be collected and the collection requirements; the avoidance command includes the deflection angle and speed adjustment value. After receiving the command, the UAV is parsed and executed by the flight control system, which simultaneously provides feedback on the command execution status once per second (e.g., flight path following error ≤ 1m, speed adjustment response time ≤ 0.3s). The ground station verifies the execution effect in real time. If the UAV does not adjust according to the command, the command is reissued with a larger adjustment range to ensure flight safety and mission execution accuracy.

[0030] Data fusion module: used to extract features from multi-dimensional data acquired by the data acquisition module, capture the correlation between cross-modal data, and dynamically allocate feature weights by combining an attention mechanism to obtain fused features; The execution process of the data fusion module includes the following steps: The acquired image, point cloud, and temperature data are standardized, noise filtered, and missing values ​​repaired. For high-definition images (RGB format), lidar point clouds (PLY format), and infrared thermal imager temperature matrices (TIF format) acquired by the UAV, the sampling frequency of each sensor's data was synchronized to 20Hz, and the time error was calibrated to ±1ms, using the BeiDou time synchronization system as a reference. Through the UAV's onboard IMU inertial measurement unit and calibration board, a transformation matrix (rotation matrix R, translation vector T) was established from the sensor coordinate system to the UAV's body coordinate system. Combined with BeiDou positioning data, all data was uniformly mapped to the WGS-84 geodetic coordinate system, ensuring a three-dimensional spatial coordinate error ≤5cm.

[0031] Image data is processed by bilateral filtering to remove Gaussian noise, and Otsu thresholding is used to segment invalid background areas. Point cloud data is processed by statistical filtering to remove outliers (points that are more than twice the standard deviation from the mean), and K-nearest neighbor interpolation (K=5) is used to repair missing coordinates in occluded areas. Temperature data is processed by median filtering to smooth the data, and missing temperature values ​​in occluded areas are filled by interpolation based on the temperature gradient of adjacent pixels.

[0032] Visual features, point cloud features, and temperature features are extracted separately, and then mapped to a shared embedding space through a fully connected layer to construct feature vectors. Constraint loss is introduced to initially establish the correlation between modal data. Visual features: An improved ResNet-50 network is used. Based on ImageNet pre-training, fine-tuning is performed on communication facility components (tower angle steel, antenna radiating units, etc.) to extract 2048-dimensional image semantic features, with a focus on enhancing edge texture and geometric shape feature responses.

[0033] Point cloud features: Using the PointNet++ network, local geometric features are extracted through multi-scale grouping (radius 0.1m, 0.3m, 0.5m), and combined with global feature aggregation to generate a 1024-dimensional point cloud shape descriptor, which effectively represents the three-dimensional structural information of the component.

[0034] Temperature characteristics: The temperature matrix is ​​convolved in spatial dimensions (kernel size 3×3) by a 1D convolutional neural network to extract 512-dimensional temperature distribution features, highlighting the temperature gradient changes in the abnormal heating area of ​​the equipment.

[0035] A fully connected layer is used to uniformly map the features of the three modalities to a 512-dimensional shared embedding space, constructing a visual feature vector f. v Point cloud feature vector f p Temperature eigenvector f t Furthermore, a shared space constraint loss is introduced to force the multimodal characteristics of similar communication facility components to be similar in the shared space, thus initially establishing the correlation between modal data. The formula is as follows: .

[0036] Based on the contrastive learning mechanism, positive and negative sample pairs are constructed to capture the correlation between data from different modalities, and the feature differences between cross-modal data are eliminated by training the loss function. Communication facilities are divided into basic semantic units (such as tower standard sections, antenna arrays, and feeder connectors), with each unit corresponding to a multi-mode data set (I). i ,P i ,T iFor each semantic unit, data is collected from different perspectives by the drone, generating 3 sets of positive samples (data from different perspectives of the same unit) and 5 sets of negative samples (data from different units or background data), and constructing a set of contrastive learning sample pairs S={(s+,s-),i=1,2,…,N}.

[0037] The InfoNCE loss function is used to enhance the feature similarity of positive sample pairs and suppress the association of negative sample pairs. The formula is as follows: ; Where sim(·) represents the mean cosine similarity of the three-modal features, τ=0.1 represents the temperature coefficient, N represents the number of positive samples, M represents the number of negative samples, i represents the index variable of the positive sample pairs, used to traverse all positive sample pair sets, and j represents the index variable of the negative sample pairs, used to traverse all negative sample pair sets. Through this loss function, the multimodal features of the same semantic unit are forced to cluster in the shared space, and the features of different units are separated.

[0038] Modal features are aligned and fused using a three-layer attention mechanism consisting of intra-modal attention, inter-modal attention, and global fusion attention to obtain fused features.

[0039] Intramodal attention: Self-attention calculation is performed on single-modal features to generate weighted features. (Preserve spatial dimensions, such as visual feature maps containing pixel positions and texture distribution, point cloud feature maps containing voxelized spatial coordinates and density information, and temperature feature maps containing temperature distribution corresponding to spatial positions), highlighting key component areas (such as antenna interfaces and tower connection points).

[0040] Intermodal cross-attention: Using the visual feature vector as the query vector and the point cloud feature vector and temperature feature vector as the key vectors, calculate the cross-modal attention weights. ; ; W1 and W2 are learnable weight matrices used to align visual features with the spatial positions of other modalities.

[0041] Global Fusion Attention: Integrates attention features from three modalities to generate aligned fusion features. This ensures that multimodal features are aligned both semantically and spatially.

[0042] To address the spatial differences between point clouds and images, a reprojection error constraint is constructed using the projection relationship between the 3D coordinates of the LiDAR point cloud and image pixels. The formula is as follows: ; Where π represents the projection function from the point cloud to the image, P kLet u represent the k-th point in the 3D point cloud. k This represents the corresponding pixel coordinates in the image, K represents the total number of samples, i.e. the total number of point cloud points participating in geometric alignment, and k represents the kth data point. Through gradient descent optimization, the reprojection error is controlled within ±2 pixels to ensure the accuracy of geometric alignment.

[0043] Target detection module: Based on fused features, target detection is performed using multi-scale detection and dynamic attention mechanism based on YOLO-SCDI algorithm to match image, point cloud and location data of the same communication facility and generate standardized data format; The execution process of the target detection module includes the following steps: The fused features are preprocessed and optimized, and the optimized feature map is divided into different detection scales; An adaptive contrast enhancement algorithm is applied to the fused feature map to enhance the feature response of edge details and abnormal temperature areas of communication facility components, thereby improving the recognition of small components and occluded components; a 3×3 Gaussian filter is used to eliminate residual noise interference during the fusion process.

[0044] Based on the size differences of communication facility components (large components: towers, base station cabinets; medium components: antennas, brackets; small components: feeder connectors, fixing bolts), the optimized fused feature map is divided into three detection scales proportionally to adapt to the detection needs of targets of different sizes: Large-scale feature map (1 / 4 of the original size): Focusing on large main components such as tower body and base station cabinet, and capturing global spatial morphology and geographic coordinate information; Mid-scale feature map (1 / 2 of the original size): For medium-sized components such as antennas and brackets, it balances texture details and spatial positioning accuracy, taking into account both detection efficiency and accuracy; Small-scale feature maps (original size): specially adapted for small components such as feeder connectors and fixing bolts, highlighting local texture features and temperature gradient changes (such as overheated areas of connectors), improving the detection accuracy of small components.

[0045] Construct a YOLO-SCDI multi-scale detection network, extract features from feature maps of different scales, and output candidate detection boxes; Specifically, the C2f-SCSA module is used in the YOLO-SCDI multi-scale detection network to simultaneously extract local detail features and global structural features, accurately focusing on key component areas of communication facilities; a CCFM-Neck structure is constructed to simultaneously complete channel compression and cross-channel information fusion, and an occlusion perception submodule is embedded in the CCFM to assign higher weights to occluded region features based on the spatial depth information in the fused feature map; a Dynamic Head is used as the detection head, and the feature response is dynamically adjusted through a scale-space-task three-dimensional attention mechanism; a dynamic convolution kernel generation network is introduced to adaptively switch the convolution kernel size according to the characteristics of the input fused features to optimize the detection effect of different types of components; based on the Anchor-Free mechanism, the bounding box coordinates (x, y, w, h), class probability, and confidence of communication facility components are directly regressed.

[0046] Specifically, the improved C2f-SCSA module is adopted as the backbone network core of YOLO-SCDI. This module dynamically fuses 3×3, 5×5, 7×7, and 9×9 multi-scale convolutional kernels to simultaneously extract local detail features and global structural features. Combined with the Spatial-Channel Collaborative Attention (SCSA) mechanism, it accurately focuses on key component areas of communication facilities, suppresses background interference such as sky and vegetation, and improves the relevance and effectiveness of feature representation.

[0047] A CCFM (Cross-Scale Feature Fusion Module) is introduced to construct the CCFM-Neck structure. Four 1×1 convolutions simultaneously perform channel compression (reducing the number of feature channels from 256 to 128) and cross-channel information fusion. For scenarios with densely distributed communication facility components and occlusion, an occlusion perception submodule is embedded in the CCFM. Based on the spatial depth information (derived from LiDAR point clouds) in the fused feature map, higher weights are assigned to features in occluded areas, reducing missed detections caused by occlusion and enhancing the model's adaptability to complex scenarios.

[0048] A Dynamic Head is used as the detection head, and the feature response is dynamically adjusted through a scale-space-task 3D attention mechanism: Scale dimension: Assign higher feature weights to small-scale components (such as feeder connectors with a diameter of <10cm) to compensate for the weak feature signals of small targets; Spatial dimension: Focus on key structural areas of components (such as the radiating elements of an antenna and the interface panels of a cabinet), enhance the characteristic response of the target area, and suppress interference in non-target areas; Task dimension: Distinguish between component classification (such as tower, antenna, feeder) and location regression tasks, and adaptively adjust the focus of feature extraction (enhance semantic features for classification tasks and coordinate features for regression tasks).

[0049] Meanwhile, a dynamic convolution kernel generation network is introduced, which adaptively switches between 3×3 and 5×5 convolution kernels based on the texture richness, point cloud density, and other characteristics of the input fusion features, thereby optimizing the detection effect of different types of parts (5×5 convolution kernels are used for parts with complex textures, and 3×3 convolution kernels are used for parts with simple structures).

[0050] Based on the Anchor-Free mechanism, the bounding box coordinates (x, y, w, h), class probability, and confidence of communication facility components are directly regressed, avoiding the adaptation problem caused by traditional anchor box matching, so as to adapt to the diverse geometric shapes of communication facility components.

[0051] The Inner-ShapeIoU metric method is used in conjunction with the geometric features of the communication facilities to calibrate the bounding boxes, and the optimal candidate boxes are matched based on the Hungarian algorithm. The Inner-Shape IoU metric method is adopted, combining the geometric features of communication facility components. A shape consistency constraint and an auxiliary bounding box mechanism (ratio=0.75) are introduced into the IoU calculation to reduce bounding box offsets caused by viewpoint changes and occlusion, improving the shape matching accuracy between the predicted bounding box and the actual component by more than 15%, while controlling the positioning error within ±2cm. An IoU threshold of 0.3 is set to filter candidate boxes at each scale, eliminating duplicate detection boxes for the same component to adapt to scenarios with densely distributed communication facility components.

[0052] The Hungarian algorithm is used to associate candidate boxes output from feature maps at different scales. If multiple candidate boxes at different scales correspond to the same component (spatial coordinate overlap ≥70%, consistent class probability), their feature information is fused, and the coordinate data with the highest accuracy and the class determination result are taken to generate a unified optimal candidate box, ensuring the uniqueness and accuracy of the detection results.

[0053] Based on the spatial coordinates of the optimal candidate box, combined with the timestamp of BeiDou time synchronization, the spatiotemporal consistency is verified by associating the original multimodal data, and the credibility is verified by a dual verification mechanism of size and temperature. High-resolution texture fragments are extracted from the image data corresponding to the candidate bounding boxes, ensuring a texture coverage of ≥95% and preserving the surface details of the component. Point cloud data is cropped from the point cloud fragments within the spatial range of the candidate bounding boxes, ensuring a point cloud void rate of <5% and fully preserving the three-dimensional geometric structure of the component. Location and temperature data are bound to the precise geographical location (latitude and longitude ±2cm, altitude ±5cm) and temperature distribution matrix (local average temperature, temperature gradient) corresponding to the candidate bounding boxes. The timestamp verification error is ≤0.5s to ensure that all associated data comes from the same acquisition time and avoid data confusion across time periods.

[0054] A dual verification mechanism is adopted for the test results: On the one hand, the actual size of the lidar point cloud computing is compared with the size of the component marked in the test results. If the error is less than 3%, it is judged as qualified to ensure the accuracy of the size data; on the other hand, the temperature data of the infrared thermal imager is combined to verify whether the temperature of the area marked as the equipment component is within the reasonable working range. If it exceeds the range, it is marked as pending verification, and subsequent manual review confirms whether it is an equipment failure or a test error.

[0055] In accordance with the digital operation and maintenance standards for communication facilities, the associated multimodal data and detection results are encapsulated into a standardized format, stored, and output.

[0056] The core fields include: Basic information: Facility ID, detection time (accurate to milliseconds), geographical location (latitude and longitude, altitude); Component information: Component category (e.g., tower standard section, parabolic antenna, feeder), dimensional parameters (length / width / height / diameter), spatial coordinate range (xmin / ymin / zmin ~ xmax / ymax / zmax); Feature information: texture feature vector (2048-dimensional), point cloud data fragment (PLY format), temperature distribution matrix (TIF format); Quality information: detection accuracy (coordinate error, size error), data integrity (texture coverage, point cloud integrity), temperature status (normal / pending verification).

[0057] Standardized data is carried in JSON format and associated with corresponding image, point cloud, and temperature data files (named using UUIDs to ensure uniqueness). After being encrypted by the data encryption module, it is synchronously stored in the local database and cloud server, and simultaneously output to the model generation module, providing structured and highly reliable input data for subsequent 3D modeling.

[0058] Model generation module: It is used to identify facility components in the matching data through the built-in communication facility feature library, call the feature library model and the identified communication facility data to stitch and optimize, generate an editable 3D model, and annotate key parameters.

[0059] The execution process of the model generation module includes the following steps: It receives standardized data and parses it, then calls the built-in communication facility feature library to compare the features of the parsed data. The system receives standardized data output from the target detection module, parses the core information, including facility ID, component category, spatial coordinate range, multimodal features, and data quality indicators (texture coverage, point cloud void rate), filters valid component data with data integrity ≥95%, and removes invalid data caused by occlusion or noise.

[0060] Activate the built-in communication facility feature library, which is built based on massive amounts of actual communication facility data and industry standards. It stores multi-dimensional feature templates for various components, including physical features (such as the height, cross-sectional dimensions, and flange hole positions of standard tower sections, antenna aperture, number of radiating elements, and mounting hole spacing), electrical features (such as the antenna's operating frequency, gain, polarization, and the impedance and attenuation coefficient of the feeder), and functional features (such as the component's role in the communication system and its connection relationship with other components).

[0061] The data processing algorithm compares the parsed component data with the feature library templates one by one. For complex components, a multi-dimensional weighted comparison is required, taking into account electrical parameters, functional connections, and spatial installation location, to ensure the accuracy of component identification.

[0062] Based on the current 3D modeling objectives, select component data that are highly relevant to the requirements and call the corresponding model templates; For example, if the target is to model a steel tower, prioritize retaining the data of the main components such as standard tower sections, connecting flanges, and ladders, temporarily store the data of auxiliary components such as antennas and feeders, and call the steel tower model of the same specification in the feature library.

[0063] Based on the physical installation logic of communication facilities and the component connection rules preset in the feature library, the splicing order and spatial constraint relationship are determined, and splicing errors are corrected. The main components are assembled first. For example, when modeling a steel tower, they are assembled from top to bottom in the order of "base section → standard section → top platform". Each standard section is aligned using the flange hole coordinates, and the z-axis height of each section is determined using the spatial coordinates in the point cloud data. Auxiliary components are assembled in association. For example, when assembling an antenna, the spatial relationship between the antenna and the support frame (the z-coordinates of the antenna bottom and the support frame top are consistent, and the x / y coordinate deviation is ≤5cm) is first used to call the support model from the feature library and assemble it with the antenna model. Then, the assembled "antenna-support" assembly is assembled with the main body of the steel tower according to the preset installation position, ensuring that the connection logic between components conforms to actual engineering specifications. Errors generated during the assembly process are corrected by comparing the spatial relationship between the point cloud data and the assembled model. For example, if a flange gap appears after assembling a standard section of the steel tower (the gap width shown in the point cloud is >2mm), the x / y coordinates of the next standard section are adjusted (based on the actual position of the flange hole in the point cloud) to make the gap ≤1mm.

[0064] The geometric shape of the assembled preliminary model is optimized, and the geometric accuracy of the model is calibrated. The initial model after assembly is smoothed to eliminate unnatural areas such as sharp edges and gaps at component connections. For example, the angle steel connections of the standard tower sections are treated with "beveling + surface fitting" to ensure the edge roundness conforms to actual manufacturing processes (beveling radius 2-5mm). Based on high-definition texture fragments and industry standards, engineering details of the components are added, such as equipment nameplates and ventilation holes on the cabinet model, enhancing the model's realism and practicality.

[0065] Key parameters required for the operation and maintenance of communication facilities are selected, categorized and labeled according to geometric parameters, electrical parameters, and performance parameters. A linkage mechanism between parameters and model geometry is established, and editing permissions are configured.

[0066] Geometric parameters are directly related to the model's geometry and are labeled as text tags on the corresponding component surfaces. Electrical parameters are related to the component's electrical properties and are labeled in the component's functional identification area. Performance parameters are related to the component's operational capabilities and are labeled in a prominent position on the front of the cabinet for easy viewing by maintenance personnel. A linkage mechanism between parameters and model geometry is established. For example, if a user modifies the height of a standard tower section while editing the model, the system automatically updates the corresponding labeled height parameters and simultaneously adjusts the spatial positions of the upper and lower components connected to that standard section, ensuring consistency between parameters and the model. Editing permissions are configured for different components of the model, and the optimized 3D model is exported to commonly used formats in the communications industry, including editable formats (such as FBX and OBJ, supporting subsequent modifications in 3ds Max and Blender) and lightweight display formats (such as glTF, supporting fast loading on web and mobile devices). The model's layer structure is preserved during export, allowing users to edit specific components by layer.

[0067] It also includes a data encryption module: by generating independent keys for each drone, each communication link, and each storage node, distributing keys based on BeiDou short messages or encrypted command channels, and periodically and automatically updating encryption keys, it encrypts data throughout the entire lifecycle. Risk monitoring module: Based on the optimized 3D model and labeled parameters, combined with real-time collected multimodal data, the module monitors communication facilities in real time and provides risk warnings.

[0068] Specifically, the system loads the editable 3D model and annotation parameters output by the model generation module, extracts the geometric datum, operational parameter datum, and performance parameter datum for each component of the facility, and constructs a datum parameter library. It receives multimodal real-time data (high-definition images, LiDAR point clouds, and infrared temperature data) collected during subsequent UAV inspections, as well as standardized data updated by the target detection module. Through BeiDou timestamps and the spatial coordinates of the 3D model, the real-time data is associated with the corresponding components in the model, ensuring spatiotemporal consistency between the real-time data and the datum model.

[0069] Based on the comparison of real-time point cloud data with the geometric benchmarks of the 3D model, the spatial morphology and positional changes of facility components are monitored. For example, the verticality of the tower body is calculated through point cloud matching; if the actual verticality exceeds the benchmark value of 0.1°, it is judged as an abnormal geometric deformation. Combining real-time temperature data with labeled electrical / performance parameters, the operating status of equipment is monitored. For example, the temperature of the cabinet power module collected by an infrared thermal imager is read; if it exceeds the benchmark range of -10~55℃, it is marked as a temperature anomaly. Based on the comparison of real-time high-definition images with the component list of the 3D model, the system monitors for missing components or added foreign objects.

[0070] Anomaly data from multi-dimensional monitoring is aggregated and categorized into structural, operational, and performance risks. Structural risks include geometric deformation and component damage / missing parts; operational risks include abnormal temperatures and electrical parameter deviations; and performance risks include communication capacity attenuation and insufficient signal strength. A built-in communication facility risk rule base (built based on industry operation and maintenance standards and historical fault data) is invoked to match risk weights and scoring criteria for each type of anomaly. The impact of anomalies on the overall facility is analyzed by combining the component relationships in the 3D model, ultimately outputting the risk impact range and potential failure consequences. Anomaly locations and risk levels are visually marked in the 3D model, triggering corresponding early warning mechanisms based on the risk level.

[0071] Furthermore, in communication network planning and design, planners can simulate and analyze different communication facility layout schemes in a virtual environment. By adjusting parameters such as the location of base stations, the orientation and height of antennas in the model, and combining them with the key parameters marked on the model, they can intuitively observe the coverage effect of communication signals under different schemes, thereby quickly finding the optimal network planning scheme and greatly saving the time and cost of on-site surveys and experiments.

[0072] In terms of communication facility maintenance and management, maintenance personnel can clearly understand the internal structure and component composition of the communication facility through the model. Combined with the marked key parameters, they can more accurately determine the cause of the fault, quickly select the appropriate maintenance tools and replacement parts, improve maintenance efficiency, and reduce communication interruption time when performing equipment maintenance and troubleshooting.

[0073] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A system for automatic data acquisition and 3D modeling of communication facilities based on unmanned aerial vehicles (UAVs), characterized in that, It includes a data acquisition module for communication connection, a data fusion module, a target detection module, and a model generation module; Data acquisition module: Based on 5G and Beidou drone scheduling algorithm, it controls drones to operate in zones according to preset routes, collects multi-dimensional data from communication facilities in real time, and preliminarily filters invalid data through edge computing; The multi-dimensional data includes the appearance texture, three-dimensional coordinates, and equipment temperature of the communication facilities; Data fusion module: used to extract features from multi-dimensional data acquired by the data acquisition module, capture the correlation between cross-modal data, and dynamically allocate feature weights by combining an attention mechanism to obtain fused features; Target detection module: Based on fused features, target detection is performed using multi-scale detection and dynamic attention mechanism based on YOLO-SCDI algorithm to match image, point cloud and location data of the same communication facility and generate standardized data format; Model generation module: It is used to identify facility components in the matching data through the built-in communication facility feature library, call the feature library model and the identified communication facility data to stitch and optimize, generate an editable 3D model, and annotate key parameters.

2. The UAV-based automatic data acquisition and 3D modeling system for communication facilities according to claim 1, characterized in that, Edge computing devices are deployed on drones to perform real-time preliminary analysis of collected data and filter invalid data based on pre-trained deep learning models and data filtering algorithms.

3. The automatic data acquisition and 3D modeling system for communication facilities based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, The ground station uses a 5G and BeiDou-based drone scheduling algorithm to control drones to operate in zones according to preset routes, and collects multi-dimensional information from communication facilities in real time, including the following steps: The ground station receives core data transmitted in real time from each drone via the 5G network and the BeiDou short message communication link, analyzes the core data, and updates the drone dynamic status database. Based on the facility distribution list of the preset route, compare the facility IDs and component data collected by each drone to determine the current task completion rate, and dynamically adjust the task weight in combination with the priority of communication facilities. Based on the drone's current location, remaining workload, and geographical environment data, the following methods are adopted: The algorithm re-plans the optimal route; Real-time calculation of the relative position, flight direction and flight speed of all airborne drones; setting two levels of monitoring thresholds to assess collision risk. Based on the UAV's flight speed and turning response speed, the minimum safe flight interval is calculated, and avoidance is carried out according to the priority avoidance strategy and the partition scheduling mode. The drone receives adjusted flight routes, task allocation instructions, and avoidance instructions in real time via an encrypted communication link, and then executes them after parsing.

4. The UAV-based automatic data acquisition and 3D modeling system for communication facilities according to claim 3, characterized in that, The core data includes the real-time location output by the BeiDou positioning module, the flight attitude fed back by the IMU inertial measurement unit, the mission progress, the sensor working status, and the remaining battery power.

5. The UAV-based automatic data acquisition and 3D modeling system for communication facilities according to claim 1, characterized in that, The execution process of the data fusion module includes the following steps: The acquired image, point cloud, and temperature data are standardized, noise filtered, and missing values ​​repaired. Visual features, point cloud features, and temperature features are extracted separately, and then mapped to a shared embedding space through a fully connected layer to construct feature vectors. Constraint loss is introduced to initially establish the correlation between modal data. Based on the contrastive learning mechanism, positive and negative sample pairs are constructed to capture the correlation between data from different modalities, and the feature differences between cross-modal data are eliminated by training the loss function. Modal features are aligned and fused using a three-layer attention mechanism consisting of intra-modal attention, inter-modal attention, and global fusion attention to obtain fused features.

6. The UAV-based automatic data acquisition and 3D modeling system for communication facilities according to claim 5, characterized in that, The InfoNCE loss function is used to enhance the feature similarity of positive sample pairs and suppress the association of negative sample pairs. The formula is as follows: ； Where sim(·) represents the mean cosine similarity of the three-modal features, τ=0.1 represents the temperature coefficient, N represents the number of positive samples, M represents the number of negative samples, i represents the index variable of the positive sample pairs, used to traverse the set of all positive sample pairs, and j represents the index variable of the negative sample pairs, used to traverse the set of all negative sample pairs.

7. The automatic data acquisition and 3D modeling system for communication facilities based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, The execution process of the target detection module includes the following steps: The fused features are preprocessed and optimized, and the optimized feature map is divided into different detection scales; Construct a YOLO-SCDI multi-scale detection network, extract features from feature maps of different scales, and output candidate detection boxes; The Inner-ShapeIoU metric method is used in conjunction with the geometric features of the communication facilities to calibrate the bounding boxes, and the optimal candidate boxes are matched based on the Hungarian algorithm. Based on the spatial coordinates of the optimal candidate box, combined with the timestamp of BeiDou time synchronization, the spatiotemporal consistency is verified by associating the original multimodal data, and the credibility is verified by a dual verification mechanism of size and temperature. In accordance with the digital operation and maintenance standards for communication facilities, the associated multimodal data and detection results are encapsulated into a standardized format, stored, and output.

8. The automatic data acquisition and 3D modeling system for communication facilities based on unmanned aerial vehicles (UAVs) according to claim 7, characterized in that, In the YOLO-SCDI multi-scale detection network, the C2f-SCSA module is used to simultaneously extract local detail features and global structural features, accurately focusing on key component areas of communication facilities; a CCFM-Neck structure is constructed to simultaneously complete channel compression and cross-channel information fusion, and an occlusion perception submodule is embedded in the CCFM, giving higher weight to occlusion area features based on the spatial depth information in the fused feature map. A Dynamic Head is used as the detection head, and the feature response is dynamically adjusted through a scale-space-task three-dimensional attention mechanism. A dynamic convolution kernel generation network is introduced to adaptively switch the convolution kernel size according to the characteristics of the input fused features, thereby optimizing the detection effect of different types of parts. Based on the Anchor-Free mechanism, the bounding box coordinates, class probability and confidence of communication facility parts are directly regressed.

9. The automatic data acquisition and 3D modeling system for communication facilities based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, The execution process of the model generation module includes the following steps: It receives standardized data and parses it, then calls the built-in communication facility feature library to compare the features of the parsed data. Based on the current 3D modeling objectives, select component data that are highly relevant to the requirements and call the corresponding model templates; Based on the physical installation logic of communication facilities and the component connection rules preset in the feature library, the splicing order and spatial constraint relationship are determined, and splicing errors are corrected. The geometric shape of the assembled preliminary model is optimized, and the geometric accuracy of the model is calibrated. Key parameters required for the operation and maintenance of communication facilities are selected, categorized and labeled according to geometric parameters, electrical parameters, and performance parameters. A linkage mechanism between parameters and model geometry is established, and editing permissions are configured.

10. The automatic data acquisition and 3D modeling system for communication facilities based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, It also includes a data encryption module: by generating independent keys for each drone, each communication link, and each storage node, distributing keys based on BeiDou short messages or encrypted command channels, and periodically and automatically updating encryption keys, it encrypts data throughout the entire lifecycle. Risk monitoring module: Based on the optimized 3D model and labeled parameters, combined with real-time collected multimodal data, the module monitors communication facilities in real time and provides risk warnings.

Images