An unmanned aerial vehicle-based building outer wall disease intelligent identification system and method

Abstract

The application provides a kind of building outer wall disease intelligent identification system and method based on unmanned aerial vehicle, including dynamic task area division and collaborative path planning to building outer wall, obtain optimized coverage shooting task set;Multiple synchronous shooting and point information collection are carried out to building outer wall, obtain visible light image set, infrared image set and corresponding pose point information set;Multi-scale cross-modal feature fusion and disease identification are carried out to the texture features and thermal anomaly features of visible light image set and infrared image set, obtain primary identification result;Generate the digital twin model of building outer wall, map the primary identification result to the digital twin model, and obtain the environmental semantic features of the corresponding position;Based on environmental semantic features, context logic constraint and false alarm filtering are carried out to primary identification result, obtain disease identification result, carry out disease cause tracing analysis to disease identification result, and encapsulate the tracing analysis result into outer wall disease report.

Description

Technical Field

[0001] This invention relates to the field of building inspection technology, and in particular to an intelligent identification system and method for building exterior wall defects based on unmanned aerial vehicles (UAVs). Background Technology

[0002] With the acceleration of urban modernization, regular inspection and defect identification of high-rise building facades have become crucial for ensuring public safety and extending building lifespan. However, traditional building facade inspection methods heavily rely on manual suspended platforms or spider-man operations, which suffer from significant drawbacks such as high-risk high-altitude work, numerous blind spots, and low efficiency.

[0003] Although drones have begun to be used for photographic inspections, they still face the following technical bottlenecks: First, traditional drone flight path planning often uses static preset modes, which are difficult to adapt to complex lighting changes, sudden gusts of wind, and obstacle distribution among buildings, resulting in inconsistent data collection quality and difficulty in guaranteeing coverage. Second, single visible light inspections are easily interfered with by environmental noise such as wall shadows, water stains, and oil stains, leading to high misdiagnosis rates and difficulty in detecting hidden defects such as hollow areas. Furthermore, existing identification schemes mostly focus on feature extraction at the two-dimensional image level, lacking spatial correlation with the building's physical structure and environmental semantics. This prevents the system from scientifically tracing the causes of defects and generating in-depth diagnostic reports with causal inference capabilities, ultimately resulting in maintenance recommendations lacking specificity and objectivity. Summary of the Invention

[0004] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides an intelligent identification system and method for building exterior wall defects based on unmanned aerial vehicles (UAVs). It has the advantages of adaptive collaborative perception, accurate identification through multimodal feature fusion, and defect mapping driven by digital twins, solving the problems of low detection efficiency, inaccurate defect identification, and reliance on subjective experience in traditional exterior wall defect detection.

[0005] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: This invention provides an intelligent identification system for building exterior wall defects based on unmanned aerial vehicles (UAVs), comprising a path planning module, a data acquisition module, a primary identification module, a result mapping module, and a source tracing analysis module, wherein: The path planning module dynamically divides the building exterior wall into task areas based on the environmental data perceived in real time by the UAV swarm, generates an initial area task set, and performs collaborative path planning on the initial area task set based on the improved particle swarm algorithm to obtain an optimized coverage shooting task set. The data acquisition module controls the drone swarm to perform multi-dimensional synchronous shooting and point information acquisition on the building exterior wall according to the optimized coverage shooting task set, and obtains a set of visible light images, an infrared image set and a corresponding set of pose and point information. The primary identification module performs bi-branch feature extraction on the visible light image set and the infrared image set, performs cross-modal feature fusion on the extracted multi-scale texture feature map set and multi-scale thermal anomaly feature map set to obtain a fused disease feature map, and performs disease identification on the fused disease feature map to obtain the primary identification result. The result mapping module generates a digital twin model of the building exterior wall based on the visible light image set, infrared image set, and shooting point information set, maps the primary recognition result to the digital twin model, and obtains the environmental semantic features of the corresponding position of the primary recognition result in the digital twin model. The source tracing analysis module performs contextual logic constraints and false alarm filtering on the initial identification results based on the environmental semantic features to obtain the disease identification results. It then combines the environmental semantic features to perform source tracing analysis on the disease identification results and encapsulates the source tracing analysis results into an exterior wall disease report.

[0006] According to a preferred embodiment of the present invention, when the path planning module performs dynamic task area division of the building exterior wall based on real-time environmental data perceived by the UAV swarm and generates an initial area task set, it includes: A swarm of drones is used to pre-scan the exterior walls of buildings to obtain environmental data, including the distribution of light intensity, the complexity of wall texture, the distribution of instantaneous gust wind speed, and the spatial distribution of obstacles on each exterior wall facade. Based on the environmental data, the shooting cost corresponding to each exterior wall facade is calculated, and a global shooting cost heatmap is generated. Cost clustering and spatial merging are performed on the global captured cost heatmap to obtain the external wall cost region set, and redundant overlapping regions are set for the external wall cost region set to obtain the standard cost region set. The performance status parameters of each UAV in the UAV swarm are obtained to form a performance status parameter set. A task mapping relationship between the performance status parameter set and the standard cost region set is established based on a greedy algorithm to obtain an initial region task set. The performance status parameters include the remaining effective payload, remaining battery power, real-time position, zoom capability, and resolution of each UAV.

[0007] According to another preferred embodiment of the present invention, when the path planning module performs cooperative path planning on the initial region task set based on an improved particle swarm optimization algorithm to obtain an optimized coverage shooting task set, it includes: The standard cost region set corresponding to the initial regional task set is rasterized to obtain a regional waypoint network set. The regional waypoint network set is then quantized based on the flight constraints of each UAV to obtain a constrained waypoint network set. Based on the constrained waypoint network set, particle encoding and population initialization are performed on the UAV swarm to obtain an initial path particle swarm. A path fitness function is constructed based on the global shooting cost heatmap corresponding to the initial regional task set and the regional coverage of the initial path particle swarm. Based on the pheromone mechanism of the ant colony algorithm and the path fitness function, the initial path particle swarm is heuristically updated to obtain the iterative shooting path set; The iterative shooting path set is subjected to path conflict detection and conflict resolution to obtain an obstacle avoidance shooting path set. The obstacle avoidance shooting path set is then bound to the initial area task set to obtain an optimized coverage shooting task set.

[0008] According to another preferred embodiment of the present invention, the primary recognition module, when performing dual-branch feature extraction on the visible light image set and the infrared image set, includes: Histogram equalization and pixel value normalization are performed on the visible light image set and the infrared image set respectively to obtain the equalized visible light image set and the equalized infrared image set. The key feature point set of the visible light image set and the key feature point set of the infrared image set are extracted respectively. Feature point matching is performed on the key feature point set of the visible light image set and the key feature point set of the infrared image set. Based on the feature point matching result, the image space of the balanced visible light image set and the balanced infrared image set is aligned to obtain the visible light-infrared image pair set. Contour information and texture features are extracted from the balanced visible light image set of the visible light-infrared image pair to obtain a multi-scale texture feature map set; Thermal radiation brightness distribution and thermal field features are extracted from the set of uniform infrared images in the visible-infrared image pair to obtain a multi-scale thermal anomaly feature map.

[0009] According to another preferred embodiment of the present invention, when the primary identification module performs cross-modal feature fusion on the extracted multi-scale texture feature map set and multi-scale thermal anomaly feature map set to obtain a fused disease feature map, it includes: The multi-scale texture feature map set and the multi-scale thermal anomaly feature map set are subjected to feature alignment and scale matching to obtain a texture-thermal anomaly feature map pair set. Based on the cross-attention mechanism, cross-modal attention weights are calculated for each texture-thermal anomaly feature map pair in the texture-thermal anomaly feature map pair set to obtain a modal attention weight matrix set. Based on the modal attention weight matrix set, feature enhancement and dynamic aggregation are performed on each texture-thermal anomaly feature map pair to obtain a multi-scale fused feature map set; The multi-scale fused feature map is fused across scales according to the feature pyramid structure to obtain a primary disease feature map. Then, the primary disease feature map is subjected to global context pooling and weight calibration based on an efficient channel attention mechanism to obtain a fused disease feature map.

[0010] According to another preferred embodiment of the present invention, when the result mapping module generates a digital twin model of the building exterior wall based on the visible light image set, the infrared image set, and the shooting point information set, it includes: A set of visible light-infrared image pairs is generated based on the visible light image set and the infrared image set. Visible light-infrared image pairs in the set of visible light-infrared image pairs are selected one by one as target image pairs. The shooting point information corresponding to the target image pairs is filtered out from the shooting point information set as target point information. Based on the target point information, geometric inversion and physical scale projection are performed on the target image pair to obtain the shooting point point cloud, and all the shooting point point clouds corresponding to the shooting point information set are aggregated into a building sparse point cloud model. The sparse point cloud model of the building is smoothed and denoised. The smoothed and denoised sparse point cloud model of the building is then extracted into a plane according to the random sampling consensus algorithm. The extracted sparse point cloud model of the building is then reconstructed into a three-dimensional model to obtain a three-dimensional mesh model of the building. The pixel information of the visible light-infrared image pair set is mapped onto the 3D mesh model of the building to obtain the textured 3D building model; The textured 3D building model is subjected to environmental semantic segmentation and semantic annotation to obtain a digital twin model.

[0011] According to another preferred embodiment of the present invention, when the result mapping module performs environmental semantic segmentation and semantic annotation on the textured architectural 3D model to obtain a digital twin model, it includes: The textured 3D building model is segmented by model boundaries and its components are identified to obtain a semantic set of building components; Based on the semantic set of building components, the textured 3D building model is transformed into a solid entity to obtain a set of building component entities. Based on the positional relationship of the textured 3D building model, a component topology map corresponding to the set of building component entities is generated. By combining the geographical location information of the building's exterior walls, environmental attribute tags are injected into each building component entity in the component topology map to obtain a digital twin model. The environmental attribute tags include ambient lighting attributes and building function attributes.

[0012] According to another preferred embodiment of the present invention, when the source tracing analysis module performs contextual logic constraints and false alarm filtering on the primary identification results based on the environmental semantic features to obtain the disease identification results, it includes: Each disease bounding box in the initial identification results is selected as the target disease bounding box, and the disease category and confidence level corresponding to the target disease bounding box are extracted. The building component entities and environmental attribute labels corresponding to the target disease bounding box are extracted from the environmental semantic features, and the disease category is logically conflict corrected based on the building component semantics of the building component entities to obtain the corrected disease category. Based on the positional relationship corresponding to the environmental attribute labels, spatial positional logic constraints are applied to the corrected disease categories, and functional logic constraints are applied to the corrected disease categories based on the building function attributes in the environmental attribute labels, to obtain constrained disease categories. The confidence level is dynamically compensated based on the ambient light attribute in the environmental attribute label to obtain a calibration confidence level. The initial identification result is then updated using all calibration confidence levels and all constrained disease categories to obtain the disease identification result.

[0013] According to another preferred embodiment of the present invention, when the source tracing analysis module performs source tracing analysis of the disease identification results by combining the environmental semantic features, it includes: The geometric parameters and disease types corresponding to each disease are extracted from the disease identification results, and the disease types and geometric parameters are characterized and encoded to obtain the external wall disease feature set. Extract the environmental feature set corresponding to the external wall disease feature set from the environmental semantic features, and extract the structural feature set corresponding to the environmental feature set of the disease from the component topology map; Based on the aforementioned external wall disease feature set, disease environment feature set, and component structural feature set, a Bayesian causal network is constructed, and a set of disease cause hypotheses is generated in the Bayesian causal network. The Bayesian causal network is used to perform causal reasoning and consistency constraint verification on the set of disease cause hypotheses to obtain the disease cause set. The disease cause set is subjected to consistency verification and multi-source confidence fusion to obtain a fused disease cause set, and the source tracing analysis results are generated based on the causal relationship between the fused disease cause set and the disease identification results.

[0014] This invention provides a method for intelligent identification of building exterior wall defects based on unmanned aerial vehicles (UAVs), including: Based on the real-time environmental data perceived by the drone swarm, the building exterior wall is dynamically divided into task areas to generate an initial task set. Then, based on the improved particle swarm algorithm, the initial task set is used for collaborative path planning to obtain an optimized coverage shooting task set. The drone swarm is controlled to perform multi-dimensional synchronous shooting and point information collection on the building exterior wall according to the optimized coverage shooting task set, so as to obtain a set of visible light images, an infrared image set and a corresponding set of pose and point information. The visible light image set and the infrared image set are subjected to bi-branch feature extraction. The extracted multi-scale texture feature map set and multi-scale thermal anomaly feature map set are subjected to cross-modal feature fusion to obtain a fused disease feature map. The fused disease feature map is then used to identify diseases to obtain a preliminary identification result. A digital twin model of the building's exterior wall is generated based on the visible light image set, infrared image set, and shooting point information set. The initial recognition result is mapped into the digital twin model, and the environmental semantic features of the corresponding position of the initial recognition result in the digital twin model are obtained. Based on the environmental semantic features, the initial identification results are subjected to contextual logic constraints and false alarm filtering to obtain the disease identification results. The disease identification results are then combined with the environmental semantic features to conduct a disease cause tracing analysis, and the tracing analysis results are packaged into an external wall disease report.

[0015] (III) Beneficial Effects Compared with existing technologies, the present invention provides an intelligent identification system for building exterior wall defects based on unmanned aerial vehicles (UAVs), which has the following beneficial effects: This UAV-based intelligent identification system for building exterior wall defects avoids the problem of insufficient adaptability to changes in lighting, wind disturbance, and obstacles in traditional static flight path planning by directly incorporating real-time environmental perception results into the path planning decision-making process. By constructing a global shooting cost heatmap and performing cost clustering and redundant region design, it achieves adaptive division of the building exterior wall task area, improving the continuity of shooting and the integrity of digital twin reconstruction. Through greedy task mapping based on UAV performance status, high-cost and high-difficulty areas are prioritized for execution by UAVs with better performance, improving the overall system reliability. By introducing an ant colony pheromone mechanism to heuristically guide the particle swarm algorithm in the collaborative path planning stage, it effectively alleviates the problems of premature convergence and local optima in the particle swarm algorithm. Furthermore, by using conflict detection and resolution, it ensures the safety of parallel operation of multiple UAVs, improving the efficiency of path planning for UAV shooting tasks and the quality of captured images.

[0016] This UAV-based intelligent identification system for building exterior wall defects significantly improves the robustness and accuracy of defect identification in complex environments through multi-source synchronous acquisition and cross-modal deep fusion. By simultaneously acquiring visible light and infrared images and combining them with four-system positioning, laser ranging, and attitude information, it achieves high-precision correlation between defect images and real spatial locations, providing a reliable foundation for subsequent digital twin modeling and defect tracing. By utilizing a dual-branch structure to enhance texture details and thermal anomaly response respectively, and by achieving modal complementarity through cross-attention mechanism and multi-scale feature pyramid, it effectively suppresses interference factors such as illumination changes and surface contamination. By using a decoupled head for defect detection, it makes positioning, classification, and confidence assessment independent of each other, thereby improving the accuracy of defect identification.

[0017] This UAV-based intelligent identification system for building exterior wall defects introduces environmental semantic features to constrain the contextual logic of the initial defect identification results, effectively eliminating false alarms that do not match the semantics, spatial location, and building function of the components. This significantly improves the accuracy and reliability of the defect identification results. By using a multimodal reliability rebalancing mechanism based on the lighting environment, the system mitigates the interference of complex lighting conditions on visual detection results, enhancing its robustness in real-world scenarios. Furthermore, by combining the corrected defect identification results with component topology and environmental semantic information, and by implementing a Bayesian causal network to trace the causes of defects, the system further improves the accuracy of exterior wall defect identification. Attached Figure Description

[0018] Figure 1 The diagram shown is a structural diagram of an intelligent identification system for building exterior wall defects based on unmanned aerial vehicles (UAVs) according to the present invention.

[0019] Figure 2 The diagram shown is a flowchart of an intelligent identification method for building exterior wall defects based on unmanned aerial vehicles (UAVs) according to the present invention. Detailed Implementation

[0020] The following description is intended to disclose the present invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious modifications will occur to those skilled in the art. The basic principles of the invention defined in the following description can be applied to other embodiments, modifications, improvements, equivalents, and other technical solutions that do not depart from the spirit and scope of the invention.

[0021] It is understood that the term "a" should be understood as "at least one" or "one or more", that is, in one embodiment, the number of an element can be one, while in another embodiment, the number of the element can be multiple, and the term "a" should not be understood as a limitation on the number.

[0022] Example 1: Please refer to Figure 1 This invention discloses an intelligent identification system for building exterior wall defects based on unmanned aerial vehicles (UAVs). The system mainly includes a path planning module, a data acquisition module, a primary identification module, a result mapping module, and a source tracing analysis module, wherein: The path planning module dynamically divides the building exterior wall into task areas based on real-time environmental data perceived by the UAV swarm, generates an initial task set, and performs collaborative path planning on the initial task set based on an improved particle swarm algorithm to obtain an optimized coverage shooting task set.

[0023] In detail, when the path planning module performs dynamic task area division of the building exterior based on real-time environmental data perceived by the drone swarm and generates an initial area task set, it includes: A swarm of drones is used to pre-scan the exterior walls of buildings to obtain environmental data, including the distribution of light intensity, the complexity of wall texture, the distribution of instantaneous gust wind speed, and the spatial distribution of obstacles on each exterior wall facade. Based on the environmental data, the shooting cost corresponding to each exterior wall facade is calculated, and a global shooting cost heatmap is generated. Cost clustering and spatial merging are performed on the global captured cost heatmap to obtain the external wall cost region set, and redundant overlapping regions are set for the external wall cost region set to obtain the standard cost region set. The performance status parameters of each UAV in the UAV swarm are obtained to form a performance status parameter set. A task mapping relationship between the performance status parameter set and the standard cost region set is established based on a greedy algorithm to obtain an initial region task set. The performance status parameters include the remaining effective payload, remaining battery power, real-time position, zoom capability, and resolution of each UAV.

[0024] Pre-scanning refers to using a swarm of drones to perform a preliminary scanning of the building's exterior walls, determining the area of ​​each facade and simultaneously collecting environmental data. Light intensity distribution can be obtained from the photosensors of each drone during shooting, based on the camera's exposure sensitivity parameters. Wall texture complexity can be obtained through image gradient amplitude statistics. Instantaneous gust wind speed distribution can be calculated based on the attitude disturbances of the drone's inertial measurement unit (IMU). Obstacle spatial distribution can be calculated using the drone's onboard laser rangefinder or through image component detection. The shooting cost can be a multivariate weighted function of light intensity distribution, wall texture complexity, instantaneous gust wind speed distribution, and obstacle spatial distribution. For example, in strong winds and shaded areas, drones need to consume more power to maintain their attitude and require longer exposures, thus resulting in a higher shooting cost.

[0025] Specifically, the global shooting cost heatmap refers to mapping the shooting cost of each unit at each location on each exterior facade onto a two-dimensional unfolded plane of the exterior facade area, resulting in a global shooting heatmap representing the distribution of shooting costs. The cost clustering and spatial merging refer to using anisotropic density-based clustering algorithms... Noise (ADCN) performs density clustering on the shooting costs of each location unit in the global shooting cost heatmap, merging location units with similar costs and adjacent geographical locations to obtain the outer wall cost region. Setting redundant overlapping regions means that at the boundary of two adjacent outer wall cost regions, a fixed-width overlapping region is extended outward for each outer wall cost region to ensure that important features of adjacent regions are not missed, and to facilitate alignment during image stitching and panoramic reconstruction of digital twin models. Establishing task mapping relationships based on a greedy algorithm means that standard cost regions with high shooting costs are preferentially assigned to the drone with the best performance status parameters. At the same time, the Euclidean distance between the drone and each standard cost region is calculated based on the real-time position in the performance status parameters, and the nearest matching is performed. Each initial region task in the initial region task set includes the corresponding drone ID, the regional coordinates of the corresponding standard cost region, the suggested flight speed, and the camera preset parameters.

[0026] Specifically, when the path planning module performs cooperative path planning on the initial region task set based on the improved particle swarm optimization algorithm to obtain an optimized coverage shooting task set, it includes: The standard cost region set corresponding to the initial regional task set is rasterized to obtain a regional waypoint network set. The regional waypoint network set is then quantized based on the flight constraints of each UAV to obtain a constrained waypoint network set. Based on the constrained waypoint network set, particle encoding and population initialization are performed on the UAV swarm to obtain an initial path particle swarm. A path fitness function is constructed based on the global shooting cost heatmap corresponding to the initial regional task set and the regional coverage of the initial path particle swarm. Based on the pheromone mechanism of the ant colony algorithm and the path fitness function, the initial path particle swarm is heuristically updated to obtain the iterative shooting path set; The iterative shooting path set is subjected to path conflict detection and conflict resolution to obtain an obstacle avoidance shooting path set. The obstacle avoidance shooting path set is then bound to the initial area task set to obtain an optimized coverage shooting task set.

[0027] Rasterization refers to dividing the standard cost region into a uniform spatial grid using a raster method, thereby obtaining a regional waypoint network. Each regional waypoint network in the regional waypoint network set is a graph-structured regional representation. Each node in the regional waypoint network represents a possible waypoint of the UAV, and the lines connecting each waypoint constitute the edges of the regional waypoint network. Flight constraints include the minimum turning radius, maximum climb speed, battery endurance limit, and obstacle safety distance for each UAV. Constraint quantization refers to quantifying the flight constraints into mathematical constraint terms. For example, for hard constraints such as the minimum turning radius, edges that do not meet the conditions can be directly extracted from the regional waypoint network. For soft constraints such as obstacle safety distance, they are transformed into penalty terms of the path fitness function. For example, a high cost is added when the path is too close to an obstacle.

[0028] Specifically, particle encoding refers to mapping a complex path composed of multiple waypoints into a high-dimensional vector in the particle swarm algorithm to obtain initial path particles. Population initialization refers to forming an initial path particle swarm from multiple encoded particles. Constructing a path fitness function based on the global shooting cost heatmap refers to establishing a function to calculate the path fitness corresponding to the initial path particles. The path fitness function is a standard function used to evaluate the quality of a path. For example, the sum of the shooting costs of each waypoint corresponding to the initial path particle in the global shooting cost heatmap is used as the cumulative cost value to calculate the area coverage corresponding to the initial path particle. The weighted sum of the cumulative cost value and the penalty term corresponding to the initial path particle is subtracted from the weighted area coverage to obtain the path fitness. The area coverage refers to the shooting coverage of the path corresponding to the initial path particle in the corresponding standard cost area, which can be calculated by shooting simulation projection. The initial path particle swarm is heuristically updated based on the ant colony algorithm's pheromone mechanism and the path fitness function to obtain the iterative shooting path set. This involves introducing the ant colony algorithm's pheromone mechanism during the particle swarm iteration process, mapping the path fitness of high-quality paths to pheromone intensity, and embedding the pheromone as a heuristic factor in the particle position and velocity update process to guide particles to converge towards high-coverage, low-cost paths. This heuristically updates the initial path particle swarm to obtain the iterative shooting path set. Conflict detection refers to detecting whether multiple drones appear in the same waypoint location range at the same time. Conflict resolution can be achieved through time-shifting adjustments, path fine-tuning, or local replanning. Task binding refers to binding each obstacle avoidance shooting path with the drone ID, the corresponding standard cost area's coordinates, flight speed, and camera preset parameters.

[0029] By directly incorporating real-time environmental perception results into the path planning decision-making process, the problem of insufficient adaptability to changes in lighting, wind disturbance, and obstacles in traditional static flight path planning is avoided. By constructing a global shooting cost heatmap and performing cost clustering and redundant region design, adaptive division of the task area for building exterior walls is achieved, improving the continuity of shooting and the integrity of digital twin reconstruction. Through greedy task mapping based on UAV performance status, high-cost and high-difficulty areas are prioritized for execution by UAVs with better performance, improving the overall system reliability. By introducing an ant colony pheromone mechanism to heuristically guide the particle swarm algorithm in the collaborative path planning stage, the problems of premature convergence and local optima in the particle swarm algorithm are effectively alleviated. Conflict detection and resolution ensure the safety of parallel operation of multiple UAVs, improving the efficiency of path planning for UAV shooting tasks and the quality of captured images.

[0030] The data acquisition module controls the drone swarm to perform multi-dimensional synchronous shooting and point information acquisition on the building exterior wall according to the optimized coverage shooting task set, and obtains a set of visible light images, an infrared image set, and a corresponding set of pose and point information.

[0031] The aforementioned multi-mode synchronous shooting refers to using the visible light cameras and thermal imaging cameras of various drones to simultaneously capture visible light and infrared images of the same area. Combined with the drone's pose and orientation, the normal distance from the drone to the outer wall surface is measured by the drone's laser rangefinder. Then, the pose and position information of each shooting point is obtained by combining the real-time positioning system of each drone. The pose and position information includes the drone's three-dimensional position, normal distance, and other position information; pitch angle, roll angle, and heading angle, and other attitude information. The real-time positioning system is a four-system real-time kinematic (RTK) positioning system, which refers to the GPS positioning system, GLONASS positioning system, Galileo positioning system, and Beidou positioning system.

[0032] The primary identification module performs bi-branch feature extraction on the visible light image set and the infrared image set, performs cross-modal feature fusion on the extracted multi-scale texture feature map set and multi-scale thermal anomaly feature map set to obtain a fused disease feature map, and performs disease identification on the fused disease feature map to obtain the primary identification result.

[0033] Specifically, when the primary recognition module performs dual-branch feature extraction on the visible light image set and the infrared image set, it includes: Histogram equalization and pixel value normalization are performed on the visible light image set and the infrared image set respectively to obtain the equalized visible light image set and the equalized infrared image set. The key feature point set of the visible light image set and the key feature point set of the infrared image set are extracted respectively. Feature point matching is performed on the key feature point set of the visible light image set and the key feature point set of the infrared image set. Based on the feature point matching result, the image space of the balanced visible light image set and the balanced infrared image set is aligned to obtain the visible light-infrared image pair set. Contour information and texture features are extracted from the balanced visible light image set of the visible light-infrared image pair to obtain a multi-scale texture feature map set; Thermal radiation brightness distribution and thermal field features are extracted from the set of uniform infrared images in the visible-infrared image pair to obtain a multi-scale thermal anomaly feature map.

[0034] Pixel value normalization refers to normalizing the pixel values ​​of the histogram-equalized visible light and infrared images to a uniform range, thereby improving the efficiency of subsequent feature extraction. Scale-Invariant Feature Transform (SIFT) or Oriented Fast and Rotated Brief (ORB) feature point matching algorithms can be used to extract key feature point sets for both visible and infrared images. Feature point matching is then performed using methods such as FLANN or Hamming distance, followed by geometric correction and scale remapping using homography matrices to achieve pixel-level alignment between the equalized visible light image and the corresponding equalized infrared image. The aligned image pairs are then used as visible light-infrared image pairs. Contour information extraction involves capturing low-level edge and boundary information using shallow convolutions. The Canny or Sobel operators are used to highlight boundary information of edge structures such as cracks in the equalized visible light image. Finally, a parallel spatial attention convolutional block (Convolutional block with Parallel) incorporating residual connections and attention mechanisms is used. SpatialAttention (C2PSA) is used for feature extraction; texture feature extraction refers to extracting fine-grained textures using convolution with a large receptive field, Gabor filter response, and LBP-inspired local patterns, and extracting features at different scales using the multi-kernel parallel approach of the Inception structure, and using the features at each scale containing texture features and contour information as a multi-scale texture feature set; thermal radiation brightness distribution extraction refers to capturing the pixel intensity distribution of the equal infrared image using convolution; thermal field feature extraction refers to capturing thermal anomaly patterns, such as temperature gradients caused by cracks, hot spots or cold spots, through pooling and attention mechanisms, and detecting hidden anomalies through high contrast to highlight the abnormal features of the temperature field, thus obtaining a multi-scale thermal anomaly feature set.

[0035] In detail, when the primary identification module performs cross-modal feature fusion on the extracted multi-scale texture feature map and multi-scale thermal anomaly feature map to obtain the fused disease feature map, it includes: The multi-scale texture feature map set and the multi-scale thermal anomaly feature map set are subjected to feature alignment and scale matching to obtain a texture-thermal anomaly feature map pair set. Based on the cross-attention mechanism, cross-modal attention weights are calculated for each texture-thermal anomaly feature map pair in the texture-thermal anomaly feature map pair set to obtain a modal attention weight matrix set. Based on the modal attention weight matrix set, feature enhancement and dynamic aggregation are performed on each texture-thermal anomaly feature map pair to obtain a multi-scale fused feature map set; The multi-scale fused feature map is fused across scales according to the feature pyramid structure to obtain a primary disease feature map. Then, the primary disease feature map is subjected to global context pooling and weight calibration based on an efficient channel attention mechanism to obtain a fused disease feature map.

[0036] Feature alignment and scale matching refer to aligning the feature channels and matching the scales of texture feature maps at various scales in the multi-scale texture feature map set with the corresponding scales of thermal anomaly feature maps in the multi-scale thermal anomaly feature map set. This can be achieved through latent feature space projection. Cross-modal attention weight calculation involves using the texture feature map in the texture-thermal anomaly feature map pair as the query vector and the thermal anomaly feature map as both the key and value features. Dot product attention is used to calculate the spatial correspondence between the two, resulting in a modal attention weight matrix. Feature enhancement and dynamic aggregation refer to weighted fusion of texture feature maps and thermal anomaly feature maps at the same scale based on the modal attention weight matrix. Cross-scale fusion refers to bidirectional propagation between fused feature maps at different scales using upsampling and downsampling feature fusion algorithms, following a feature pyramid structure. This allows for full fusion of high-level and shallow features to obtain a primary disease feature map. Efficient Channel Attention (ECA) for global context pooling and weight calibration involves applying global average pooling to the primary disease feature map. Pooling (GAP) compresses spatial dimensions and performs weight calibration by learning channel attention weights. The calibrated weights are then multiplied back into the primary disease feature map channel by channel to obtain the fused disease feature map.

[0037] In detail, the initial identification result of the fused disease feature map refers to the use of the decoupled head module of the pre-trained YOLO series model for disease identification. The initial identification result is output through three branches of the decoupled head module: the regression branch outputs the position coordinates of the disease bounding box, the classification branch outputs the disease category within the disease bounding box, and the confidence branch outputs the confidence of the category classification and location within the disease bounding box. The disease categories include leakage, cracking and peeling, and hollowing.

[0038] By employing multi-dimensional synchronous acquisition and cross-modal deep fusion, the robustness and accuracy of building exterior wall defect identification in complex environments are significantly improved. Through the synchronous acquisition of visible light and infrared images, combined with four-system positioning, laser ranging, and attitude information, high-precision correlation between defect images and real spatial locations is achieved, providing a reliable foundation for subsequent digital twin modeling and defect tracing. By utilizing a dual-branch structure to enhance texture details and thermal anomaly response respectively, and by achieving modal complementarity through cross-attention mechanism and multi-scale feature pyramid, interference factors such as illumination changes and surface contamination are effectively suppressed. By using a decoupled head for defect detection, positioning, classification, and confidence assessment are made independent of each other, improving the accuracy of defect identification.

[0039] The result mapping module generates a digital twin model of the building's exterior wall based on the visible light image set, infrared image set, and shooting point information set. It then maps the initial recognition result to the digital twin model and obtains the environmental semantic features of the corresponding position of the initial recognition result in the digital twin model.

[0040] Specifically, when the result mapping module generates a digital twin model of the building's exterior wall based on the visible light image set, infrared image set, and shooting point information set, it includes: A set of visible light-infrared image pairs is generated based on the visible light image set and the infrared image set. Visible light-infrared image pairs in the set of visible light-infrared image pairs are selected one by one as target image pairs. The shooting point information corresponding to the target image pairs is filtered out from the shooting point information set as target point information. Based on the target point information, geometric inversion and physical scale projection are performed on the target image pair to obtain the shooting point point cloud, and all the shooting point point clouds corresponding to the shooting point information set are aggregated into a building sparse point cloud model. The sparse point cloud model of the building is smoothed and denoised. The smoothed and denoised sparse point cloud model of the building is then extracted into a plane according to the random sampling consensus algorithm. The extracted sparse point cloud model of the building is then reconstructed into a three-dimensional model to obtain a three-dimensional mesh model of the building. The pixel information of the visible light-infrared image pair set is mapped onto the 3D mesh model of the building to obtain the textured 3D building model; The textured 3D building model is subjected to environmental semantic segmentation and semantic annotation to obtain a digital twin model.

[0041] Among them, geometric inversion refers to constructing collinearity equations based on the camera intrinsic parameter matrix and the extrinsic parameter matrix formed by the target image pair and the target point information, thereby calculating the point cloud corresponding to each matching key feature point in the target image pair. Geometric inversion can be performed using the Structure from Motion (SfM) algorithm based on pose constraints. Physical scale projection refers to using the position information of the laser rangefinder in the target point information as a constraint to determine the true physical scale of the point cloud model and realize the three-dimensional scaling of the point cloud model. Bilateral filtering or Gaussian filtering methods can be used for smoothing and noise reduction. Pixel information mapping refers to using texture mapping technology to map pixel information to the three-dimensional mesh model of the building according to the projection relationship corresponding to geometric inversion.

[0042] In detail, when the result mapping module performs environmental semantic segmentation and semantic annotation on the textured building 3D model to obtain a digital twin model, it includes: The textured 3D building model is segmented by model boundaries and its components are identified to obtain a semantic set of building components; Based on the semantic set of building components, the textured 3D building model is transformed into a solid entity to obtain a set of building component entities. Based on the positional relationship of the textured 3D building model, a component topology map corresponding to the set of building component entities is generated. By combining the geographical location information of the building's exterior walls, environmental attribute tags are injected into each building component entity in the component topology map to obtain a digital twin model. The environmental attribute tags include ambient lighting attributes and building function attributes.

[0043] The process involves using pre-trained 3D U-Net networks or semantic segmentation models such as Mesh-Segmenter for model boundary segmentation and component recognition. The semantics of the building components in the semantic set include windows, balconies, smoke vents, air conditioning units, rainwater pipes, and decorative lines. Entity transformation refers to converting the components corresponding to the segmented model boundaries into independent entity objects. A component topology map is generated by calculating the three-dimensional center coordinates, bounding boxes, and relative depth relationships with the walls of each building component entity in the model. The ambient lighting attributes are determined based on geographical location information to determine the lighting orientation of each wall, such as east-facing or west-facing. The building function attributes include kitchens, corridor windows, and balconies, which can be determined based on the building drawings in the geographical location information.

[0044] In detail, mapping the primary identification result to the digital twin model means mapping the bounding box position information corresponding to the primary identification result to the digital twin model according to the pixel information mapping method, and labeling the confidence level and disease category information in the primary identification result to the corresponding position. Obtaining the environmental semantic features of the corresponding position of the primary identification result in the digital twin model means determining the positional relationship between each disease bounding box and the surrounding building component entities, including distance and direction, according to the position of the primary identification result in the digital twin model and the Euclidean distance algorithm, and aggregating the environmental attribute labels of the building component entities located within the neighborhood distance and the corresponding positional relationship into environmental semantic features.

[0045] By unifying the modeling of multi-source visible light images, infrared images, and high-precision shooting point information, a high-precision mapping of building exterior walls from two-dimensional image space to three-dimensional semantic space was achieved. A real-scale point cloud model was constructed through geometric inversion and physical scale projection, and a textured building model was generated through planar extraction and three-dimensional reconstruction, giving the exterior wall spatial structure real physical constraints. Through three-dimensional semantic segmentation and building component materialization, the building exterior wall was upgraded from a pure geometric model to a digital twin model with component hierarchy, topological relationships, and environmental attributes. By accurately mapping the primary disease identification results into the digital twin model and combining the semantic features of the component's neighborhood environment, spatial correlation analysis between diseases and environmental factors was achieved, improving the accuracy and interpretability of disease location.

[0046] The source tracing analysis module performs contextual logic constraints and false alarm filtering on the initial identification results based on the environmental semantic features to obtain the disease identification results. It then combines the environmental semantic features to perform source tracing analysis on the disease identification results and encapsulates the source tracing analysis results into an exterior wall disease report.

[0047] In detail, when the source tracing analysis module performs contextual logic constraints and false alarm filtering on the initial identification results based on the environmental semantic features to obtain the disease identification results, it includes: Each disease bounding box in the initial identification results is selected as the target disease bounding box, and the disease category and confidence level corresponding to the target disease bounding box are extracted. The building component entities and environmental attribute labels corresponding to the target disease bounding box are extracted from the environmental semantic features, and the disease category is logically conflict corrected based on the building component semantics of the building component entities to obtain the corrected disease category. Based on the positional relationship corresponding to the environmental attribute labels, spatial positional logic constraints are applied to the corrected disease categories, and functional logic constraints are applied to the corrected disease categories based on the building function attributes in the environmental attribute labels, to obtain constrained disease categories. The confidence level is dynamically compensated based on the ambient light attribute in the environmental attribute label to obtain a calibration confidence level. The initial identification result is then updated using all calibration confidence levels and all constrained disease categories to obtain the disease identification result.

[0048] The logical conflict correction refers to using a pre-defined physical semantic rule base or knowledge graph to determine whether the disease category conforms to the semantics of the corresponding building component. For example, when the building component entity corresponding to the target disease boundary box is a glass window, the disease category "hollow" should be removed. Spatial location logical constraints refer to determining whether the disease category conforms to the location constraints of the building component entity. For example, when the semantics of the building component include window frames and walls, the disease category "crack" can be a structural settlement crack. Functional logical constraints refer to determining whether the disease category conforms to the functional constraints of the building component entity. For example, when the target disease boundary box of the disease category "leakage" is close to the building component... When the defect is located below the physical "air conditioner outdoor unit", the defect is updated to "air conditioner outdoor unit leakage". When the defect category "peeling" is located at the physical "kitchen exhaust vent" of the building component, the defect is judged as a false alarm and filtered. In actual use, spatial location logic constraints and functional logic constraints are used simultaneously, and the constraints are also implemented based on the physical semantic rule base or knowledge graph. Dynamic compensation refers to dynamically adjusting the confidence level of defect detection based on the ambient light attributes. For example, for the bounding box of the target defect located in the "strong light reflection area" or "deep shadow area", its confidence level is reduced according to the light intensity distribution, and the discrimination weight of infrared features is increased to achieve multimodal confidence rebalancing.

[0049] Specifically, when the source tracing analysis module performs source tracing analysis on the disease identification results by combining the environmental semantic features, it includes: The geometric parameters and disease types corresponding to each disease are extracted from the disease identification results, and the disease types and geometric parameters are characterized and encoded to obtain the external wall disease feature set. Extract the environmental feature set corresponding to the external wall disease feature set from the environmental semantic features, and extract the structural feature set corresponding to the environmental feature set of the disease from the component topology map; Based on the aforementioned external wall disease feature set, disease environment feature set, and component structural feature set, a Bayesian causal network is constructed, and a set of disease cause hypotheses is generated in the Bayesian causal network. The Bayesian causal network is used to perform causal reasoning and consistency constraint verification on the set of disease cause hypotheses to obtain the disease cause set. The disease cause set is subjected to consistency verification and multi-source confidence fusion to obtain a fused disease cause set, and the source tracing analysis results are generated based on the causal relationship between the fused disease cause set and the disease identification results.

[0050] The geometric parameters include the area, length, width, crack direction (determined through main direction analysis), angle, depth, and location coordinates of the defects. Extracting the structural feature set of components refers to extracting the attributes and relationships of each defective building component entity within the structural system of the digital twin model. Component structural features include the semantics of the building component, component hierarchy, and topological relationships between adjacent components, such as being 5cm from a drain pipe or located below an air conditioner condensate outlet. The prior probabilities of the Bayesian causal network originate from historical maintenance big data or an expert knowledge base. A set of hypothesis sets for the causes of defects can be generated through reverse retrieval. For example, for the defect "hollowing out," the corresponding hypothesis can be automatically retrieved from the expert knowledge base: {H1: thermal expansion and contraction, H2: construction quality, H3: base layer dampness}. These hypotheses constitute the nodes of the Bayesian network. Bayesian inference formulas can be used to... Causal reasoning is performed using approximate reasoning methods; consistency constraint verification refers to checking whether the inferred causes violate basic physical laws or spatial logic and making corrections based on logical conflict correction methods; source consistency verification can be performed through causal chain analysis or counterfactual reasoning, and confidence fusion can be performed through Bayesian updates or DS evidence theory. For example, the confidence of a disease in the disease identification result is only 70%, but because environmental characteristics show that it is below the air conditioner leak point, the correlation is 95%, and the confidence of the final source tracing result after weighted fusion is 90%; encapsulating the source analysis results into an exterior wall disease report means that the source analysis results are encapsulated according to a preset text format, and treatment suggestions are generated for each disease type through regular expression matching or database retrieval methods, resulting in a disease identification result that includes disease type, disease location, possible causes of disease, and treatment suggestions.

[0051] By introducing environmental semantic features to constrain the contextual logic of the primary defect identification results, false alarms that do not match the semantics, spatial location, and building function of the components are effectively eliminated, significantly improving the accuracy and reliability of the defect identification results. Through a multimodal reliability rebalancing mechanism based on the lighting environment, the interference of complex lighting conditions on visual detection results is mitigated, enhancing the robustness of the system in real-world scenarios. By combining the corrected defect identification results with component topology and environmental semantic information, and realizing the causal analysis of defects through Bayesian causal networks, the accuracy of exterior wall defect identification is improved.

[0052] Example 2: Please refer to Figure 2 This invention discloses an intelligent identification method for building exterior wall defects based on unmanned aerial vehicles (UAVs), comprising the following steps: Based on the real-time environmental data perceived by the drone swarm, the building exterior wall is dynamically divided into task areas to generate an initial task set. Then, based on the improved particle swarm algorithm, the initial task set is used for collaborative path planning to obtain an optimized coverage shooting task set. The drone swarm is controlled to perform multi-dimensional synchronous shooting and point information collection on the building exterior wall according to the optimized coverage shooting task set, so as to obtain a set of visible light images, an infrared image set and a corresponding set of pose and point information. The visible light image set and the infrared image set are subjected to bi-branch feature extraction. The extracted multi-scale texture feature map set and multi-scale thermal anomaly feature map set are subjected to cross-modal feature fusion to obtain a fused disease feature map. The fused disease feature map is then used to identify diseases to obtain a preliminary identification result. A digital twin model of the building's exterior wall is generated based on the visible light image set, infrared image set, and shooting point information set. The initial recognition result is mapped into the digital twin model, and the environmental semantic features of the corresponding position of the initial recognition result in the digital twin model are obtained. Based on the environmental semantic features, the initial identification results are subjected to contextual logic constraints and false alarm filtering to obtain the disease identification results. The disease identification results are then combined with the environmental semantic features to conduct a disease cause tracing analysis, and the tracing analysis results are packaged into an external wall disease report.

[0053] The processes described above with reference to the flowcharts in the embodiments disclosed in this invention can be implemented as computer software programs. The embodiments disclosed in this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication component, and / or installed from a removable medium. When the computer program is executed by a central processing unit (CPU), it performs the functions defined in the methods of this application. It should be noted that the computer-readable medium described above in this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wire segments, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on a computer-readable medium may be transmitted using any suitable medium, including but not limited to: wireless segments, wire segments, optical fibers, RF, etc., or any suitable combination thereof.

[0054] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0055] Those skilled in the art should understand that the embodiments of the present invention described above and shown in the accompanying drawings are merely examples and do not limit the present invention. The purpose of the present invention has been fully and effectively achieved. The functions and structural principles of the present invention have been shown and explained in the embodiments. Without departing from the stated principles, the implementation of the present invention may have any variations or modifications.

Claims

1. An unmanned aerial vehicle (UAV)-based intelligent building exterior wall disease identification system, characterized in that, The system includes a path planning module, a data acquisition module, a primary identification module, a result mapping module, and a source tracing analysis module, wherein: The path planning module dynamically divides the building exterior wall into task areas based on the environmental data perceived in real time by the UAV swarm, generates an initial area task set, and performs collaborative path planning on the initial area task set based on the improved particle swarm algorithm to obtain an optimized coverage shooting task set. The data acquisition module controls the drone swarm to perform multi-dimensional synchronous shooting and point information acquisition on the building exterior wall according to the optimized coverage shooting task set, and obtains a set of visible light images, an infrared image set and a corresponding set of pose and point information. The primary identification module performs bi-branch feature extraction on the visible light image set and the infrared image set, performs cross-modal feature fusion on the extracted multi-scale texture feature map set and multi-scale thermal anomaly feature map set to obtain a fused disease feature map, and performs disease identification on the fused disease feature map to obtain the primary identification result. The result mapping module generates a digital twin model of the building exterior wall based on the visible light image set, infrared image set, and shooting point information set, maps the primary recognition result to the digital twin model, and obtains the environmental semantic features of the corresponding position of the primary recognition result in the digital twin model. The source tracing analysis module performs contextual logic constraints and false alarm filtering on the initial identification results based on the environmental semantic features to obtain the disease identification results. It then combines the environmental semantic features to perform source tracing analysis on the disease identification results and encapsulates the source tracing analysis results into an exterior wall disease report.

2. The intelligent identification system for building exterior wall defects based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, When the path planning module dynamically divides the building exterior into task areas based on real-time environmental data perceived by the drone swarm and generates an initial task set, it includes: A swarm of drones is used to pre-scan the exterior walls of buildings to obtain environmental data, including the distribution of light intensity, the complexity of wall texture, the distribution of instantaneous gust wind speed, and the spatial distribution of obstacles on each exterior wall facade. Based on the environmental data, the shooting cost corresponding to each exterior wall facade is calculated, and a global shooting cost heatmap is generated. Cost clustering and spatial merging are performed on the global captured cost heatmap to obtain the external wall cost region set, and redundant overlapping regions are set for the external wall cost region set to obtain the standard cost region set. The performance status parameters of each UAV in the UAV swarm are obtained to form a performance status parameter set. A task mapping relationship between the performance status parameter set and the standard cost region set is established based on a greedy algorithm to obtain an initial region task set. The performance status parameters include the remaining effective payload, remaining battery power, real-time position, zoom capability, and resolution of each UAV.

3. The intelligent identification system for building exterior wall defects based on unmanned aerial vehicles (UAVs) according to claim 2, characterized in that, When the path planning module performs cooperative path planning on the initial region task set based on the improved particle swarm optimization algorithm to obtain an optimized coverage shooting task set, it includes: The standard cost region set corresponding to the initial regional task set is rasterized to obtain a regional waypoint network set. The regional waypoint network set is then quantized based on the flight constraints of each UAV to obtain a constrained waypoint network set. Based on the constrained waypoint network set, particle encoding and population initialization are performed on the UAV swarm to obtain an initial path particle swarm. A path fitness function is constructed based on the global shooting cost heatmap corresponding to the initial regional task set and the regional coverage of the initial path particle swarm. Based on the pheromone mechanism of the ant colony algorithm and the path fitness function, the initial path particle swarm is heuristically updated to obtain the iterative shooting path set; The iterative shooting path set is subjected to path conflict detection and conflict resolution to obtain an obstacle avoidance shooting path set. The obstacle avoidance shooting path set is then bound to the initial area task set to obtain an optimized coverage shooting task set.

4. The intelligent identification system for building exterior wall defects based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, The primary recognition module, when performing dual-branch feature extraction on the visible light image set and the infrared image set, includes: Histogram equalization and pixel value normalization are performed on the visible light image set and the infrared image set respectively to obtain the equalized visible light image set and the equalized infrared image set. The key feature point set of the visible light image set and the key feature point set of the infrared image set are extracted respectively. Feature point matching is performed on the key feature point set of the visible light image set and the key feature point set of the infrared image set. Based on the feature point matching result, the image space of the balanced visible light image set and the balanced infrared image set is aligned to obtain the visible light-infrared image pair set. Contour information and texture features are extracted from the balanced visible light image set of the visible light-infrared image pair to obtain a multi-scale texture feature map set; Thermal radiation brightness distribution and thermal field features are extracted from the set of uniform infrared images in the visible-infrared image pair to obtain a multi-scale thermal anomaly feature map.

5. The intelligent identification system for building exterior wall defects based on unmanned aerial vehicles (UAVs) according to claim 4, characterized in that, The primary identification module, when performing cross-modal feature fusion on the extracted multi-scale texture feature maps and multi-scale thermal anomaly feature maps to obtain a fused disease feature map, includes: The multi-scale texture feature map set and the multi-scale thermal anomaly feature map set are subjected to feature alignment and scale matching to obtain a texture-thermal anomaly feature map pair set. Based on the cross-attention mechanism, cross-modal attention weights are calculated for each texture-thermal anomaly feature map pair in the texture-thermal anomaly feature map pair set to obtain a modal attention weight matrix set. Based on the modal attention weight matrix set, feature enhancement and dynamic aggregation are performed on each texture-thermal anomaly feature map pair to obtain a multi-scale fused feature map set; The multi-scale fused feature map is fused across scales according to the feature pyramid structure to obtain a primary disease feature map. Then, the primary disease feature map is subjected to global context pooling and weight calibration based on an efficient channel attention mechanism to obtain a fused disease feature map.

6. The intelligent identification system for building exterior wall defects based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, When the result mapping module generates a digital twin model of the building's exterior wall based on the visible light image set, infrared image set, and shooting point information set, it includes: A set of visible light-infrared image pairs is generated based on the visible light image set and the infrared image set. Visible light-infrared image pairs in the set of visible light-infrared image pairs are selected one by one as target image pairs. The shooting point information corresponding to the target image pairs is filtered out from the shooting point information set as target point information. Based on the target point information, geometric inversion and physical scale projection are performed on the target image pair to obtain the shooting point point cloud, and all the shooting point point clouds corresponding to the shooting point information set are aggregated into a building sparse point cloud model. The sparse point cloud model of the building is smoothed and denoised. The smoothed and denoised sparse point cloud model of the building is then extracted into a plane according to the random sampling consensus algorithm. The extracted sparse point cloud model of the building is then reconstructed into a three-dimensional model to obtain a three-dimensional mesh model of the building. The pixel information of the visible light-infrared image pair set is mapped onto the 3D mesh model of the building to obtain the textured 3D building model; The textured 3D building model is subjected to environmental semantic segmentation and semantic annotation to obtain a digital twin model.

7. The intelligent identification system for building exterior wall defects based on unmanned aerial vehicles (UAVs) according to claim 6, characterized in that, When the result mapping module performs environmental semantic segmentation and semantic annotation on the textured 3D building model to obtain a digital twin model, it includes: The textured 3D building model is segmented by model boundaries and its components are identified to obtain a semantic set of building components; Based on the semantic set of building components, the textured 3D building model is transformed into a solid entity to obtain a set of building component entities. Based on the positional relationship of the textured 3D building model, a component topology map corresponding to the set of building component entities is generated. By combining the geographical location information of the building's exterior walls, environmental attribute tags are injected into each building component entity in the component topology map to obtain a digital twin model. The environmental attribute tags include ambient lighting attributes and building function attributes.

8. The intelligent identification system for building exterior wall defects based on unmanned aerial vehicles (UAVs) according to claim 7, characterized in that, When the source tracing analysis module performs contextual logic constraints and false alarm filtering on the initial identification results based on the environmental semantic features to obtain the disease identification results, it includes: Each disease bounding box in the initial identification results is selected as the target disease bounding box, and the disease category and confidence level corresponding to the target disease bounding box are extracted. The building component entities and environmental attribute labels corresponding to the target disease bounding box are extracted from the environmental semantic features, and the disease category is logically conflict corrected based on the building component semantics of the building component entities to obtain the corrected disease category. Based on the positional relationship corresponding to the environmental attribute labels, spatial positional logic constraints are applied to the corrected disease categories, and functional logic constraints are applied to the corrected disease categories based on the building function attributes in the environmental attribute labels, to obtain constrained disease categories. The confidence level is dynamically compensated based on the ambient light attribute in the environmental attribute label to obtain a calibration confidence level. The initial identification result is then updated using all calibration confidence levels and all constrained disease categories to obtain the disease identification result.

9. A building exterior wall defect intelligent identification system based on unmanned aerial vehicles (UAVs) according to claim 8, characterized in that, When the source tracing analysis module performs source tracing analysis on the disease identification results by combining the environmental semantic features, it includes: The geometric parameters and disease types corresponding to each disease are extracted from the disease identification results, and the disease types and geometric parameters are characterized and encoded to obtain the external wall disease feature set. Extract the environmental feature set corresponding to the external wall disease feature set from the environmental semantic features, and extract the structural feature set corresponding to the environmental feature set of the disease from the component topology map; Based on the aforementioned external wall disease feature set, disease environment feature set, and component structural feature set, a Bayesian causal network is constructed, and a set of disease cause hypotheses is generated in the Bayesian causal network. The Bayesian causal network is used to perform causal reasoning and consistency constraint verification on the set of disease cause hypotheses to obtain the disease cause set. The disease cause set is subjected to consistency verification and multi-source confidence fusion to obtain a fused disease cause set, and the source tracing analysis results are generated based on the causal relationship between the fused disease cause set and the disease identification results.

10. A method for intelligent identification of building exterior wall defects based on unmanned aerial vehicles (UAVs), characterized in that, The method includes: Based on the real-time environmental data perceived by the drone swarm, the building exterior wall is dynamically divided into task areas to generate an initial task set. Then, based on the improved particle swarm algorithm, the initial task set is used for collaborative path planning to obtain an optimized coverage shooting task set. The drone swarm is controlled to perform multi-dimensional synchronous shooting and point information collection on the building exterior wall according to the optimized coverage shooting task set, so as to obtain a set of visible light images, an infrared image set and a corresponding set of pose and point information. The visible light image set and the infrared image set are subjected to bi-branch feature extraction. The extracted multi-scale texture feature map set and multi-scale thermal anomaly feature map set are subjected to cross-modal feature fusion to obtain a fused disease feature map. The fused disease feature map is then used to identify diseases to obtain a preliminary identification result. A digital twin model of the building's exterior wall is generated based on the visible light image set, infrared image set, and shooting point information set. The initial recognition result is mapped into the digital twin model, and the environmental semantic features of the corresponding position of the initial recognition result in the digital twin model are obtained. Based on the environmental semantic features, the initial identification results are subjected to contextual logic constraints and false alarm filtering to obtain the disease identification results. The disease identification results are then combined with the environmental semantic features to conduct a disease cause tracing analysis, and the tracing analysis results are packaged into an external wall disease report.

Images