Bridge structure low-altitude inspection and disease assessment system and method based on deep learning

Abstract

The application discloses a bridge structure low-altitude inspection and disease assessment system and method based on deep learning, and belongs to the technical field of bridge disease detection and structure health monitoring. The system comprises a multispectral adaptive image acquisition module, a multiscale disease detection and feature extraction module, a digital twin mapping and disease positioning module, and a disease evolution prediction and maintenance decision module. The system dynamically adjusts the acquisition parameters according to the environmental conditions, accurately identifies different scale diseases based on a multiscale convolutional neural network, realizes cm-level accurate positioning of the diseases through a three-dimensional digital twin model, analyzes the time sequence evolution trend of the diseases and generates grading maintenance suggestions, and realizes adaptive optimization through a closed-loop feedback mechanism. The application has strong environmental adaptability, high detection precision, accurate positioning, and disease evolution analysis capability, and provides omnibearing technical support for bridge health monitoring and intelligent management.

Description

Technical Field

[0001] This invention relates to the field of bridge defect detection and structural health monitoring technology, specifically to a bridge structure low-altitude inspection and defect assessment system and method based on deep learning. This system combines UAV technology, multispectral imaging, deep learning algorithms and digital twin technology to achieve automated inspection, defect identification, three-dimensional positioning and evolution prediction of bridge structures. Background Technology

[0002] As a key component of transportation infrastructure, bridges are subject to the effects of loads, environmental erosion, and material aging during long-term service, leading to surface and internal defects such as cracks, concrete spalling, and steel reinforcement corrosion. If these defects are not detected and addressed promptly, they can seriously threaten the structural safety and service life of bridges. Therefore, regular bridge inspections and defect assessments are crucial for ensuring the safe operation of bridges.

[0003] Traditional bridge inspection methods primarily rely on manual visual inspection, requiring inspectors to use inspection vehicles, ropes, and other tools to approach various parts of the bridge for observation and recording. This method has several drawbacks: First, it is inefficient; a comprehensive inspection of a large bridge often takes weeks or even months and necessitates traffic closures or safety measures, resulting in significant social costs. Second, it carries high safety risks; inspectors must work at heights or in hazardous environments, increasing the risk of accidents. Third, inspection results are heavily influenced by subjective factors; differences in experience and judgment standards among inspectors make it difficult to guarantee the accuracy and consistency of defect identification. Fourth, defect recording methods are outdated; traditional paper records or simple photographic archiving are insufficient for systematic management and long-term tracking of defect information.

[0004] In recent years, with the rapid development of drone technology and computer vision technology, drone-based bridge inspection methods have gradually attracted attention. Drones have advantages such as flexibility, low cost, and high safety, and can quickly approach various parts of the bridge to collect images, greatly improving inspection efficiency and reducing safety risks. Meanwhile, deep learning algorithms have made breakthroughs in image recognition, providing technical support for automated defect detection. However, existing drone-based bridge inspection technologies still have the following problems:

[0005] First, there is insufficient environmental adaptability. Bridge inspections typically require different weather and lighting conditions, but existing methods mostly use fixed image acquisition parameters. Images taken in strong light, shadow, or low-light environments are of poor quality, affecting the accuracy of subsequent defect detection. For example, under strong sunlight, the bridge surface will have bright and shadow areas, causing some defects to be concealed; on cloudy days or at dusk, insufficient lighting will make the overall image dark, making it difficult to identify fine cracks.

[0006] Secondly, the ability to detect bridge defects is limited. Bridge defects have multi-scale characteristics, ranging from micro-cracks less than 0.1 mm wide to large-area peeling covering several square meters. Existing detection algorithms are usually optimized for defects of a specific scale, making it difficult to accurately identify defects of different scales simultaneously. In addition, interference factors such as stains, water stains, and paint peeling on the bridge surface can easily be confused with actual defects, leading to false positives and false negatives.

[0007] Third, the accuracy of defect location is not high. Current technologies mainly focus on defect identification at the two-dimensional image level, lacking effective correlation with the three-dimensional structure of the bridge. Although some methods utilize GPS information for rough location, the limited accuracy of GPS (typically 3-5 meters) and its failure to consider the complex three-dimensional geometry of the bridge result in insufficient spatial location information for defects, causing difficulties for subsequent maintenance decisions and project implementation. This is especially true for large bridges; if the location of defects on specific structural components cannot be accurately pinpointed, maintenance personnel need to spend a significant amount of time conducting on-site searches and confirmations.

[0008] Fourth, there is a lack of capability to analyze the evolution of bridge defects. Bridge defects are a dynamic process, and a single inspection can only reflect the current state of the defects, failing to reveal their development trends and evolutionary patterns. Existing technologies generally lack systematic analysis and utilization of historical inspection data, making it impossible to track the time sequence of defects and predict their development speed. This results in a lack of scientific basis for maintenance decisions, potentially leading to missed optimal maintenance opportunities or over-maintenance.

[0009] The prior art document CN118134934A discloses an automatic bridge defect location method and system based on UAV inspection and BIM model. The method extracts the GPS coordinates of the defect image, converts the GPS coordinates into BIM model coordinates using Gauss-Kruger projection and coordinate transformation matrix, and then uses the nearest distance matching method to register the defect to the entity unit of the BIM model. However, this method has the following shortcomings: First, it does not consider the impact of different environmental conditions on image acquisition quality, and uses fixed flight and exposure parameters, resulting in unstable image quality under complex lighting conditions. Second, its disease detection capability is weak; it does not employ advanced algorithms such as deep learning, leading to low accuracy in identifying diseases in complex backgrounds and an inability to handle multi-scale diseases. Third, disease location mainly relies on GPS coordinates, and the positioning accuracy is limited by GPS errors, typically only achieving meter-level accuracy, which cannot meet the needs of precise maintenance. Fourth, it lacks disease evolution analysis and prediction functions, failing to provide dynamic support for bridge health monitoring and maintenance decisions. Fifth, the various modules of the system are independent of each other, lacking information feedback and collaborative optimization mechanisms, and cannot adaptively adjust the acquisition and detection strategies based on the detection results.

[0010] Therefore, there is an urgent need to develop a low-altitude inspection and defect assessment system for bridge structures that is highly adaptable to the environment, has high detection accuracy, accurate positioning, evolutionary analysis capabilities, and adaptive optimization capabilities, in order to meet the needs of modern bridge health monitoring and intelligent management. Summary of the Invention

[0011] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a deep learning-based system and method for low-altitude inspection and defect assessment of bridge structures. This invention achieves high-quality image acquisition under different environmental conditions through a multispectral adaptive image acquisition module, accurate identification of defects at different scales through a multi-scale defect detection and feature extraction module, precise three-dimensional localization of defects through a digital twin mapping and defect localization module, and time-series tracking and maintenance decision support through a defect evolution prediction and maintenance decision module. Furthermore, it achieves adaptive optimization of the system by constructing a closed-loop feedback mechanism, thereby comprehensively improving the automation level of bridge inspection and the scientific rigor of defect assessment.

[0012] The technical solution of the present invention is as follows:

[0013] This deep learning-based low-altitude inspection and defect assessment system for bridge structures includes a multispectral adaptive image acquisition module, a multi-scale defect detection and feature extraction module, a digital twin mapping and defect localization module, and a defect evolution prediction and maintenance decision module. The multispectral adaptive image acquisition module dynamically adjusts UAV flight parameters and camera exposure parameters based on ambient light intensity and weather conditions to acquire multispectral image data of the bridge structure under different lighting conditions. The multi-scale defect detection and feature extraction module uses a multi-scale convolutional neural network to detect defects in the multispectral image data, identify various defect types, and extract defect feature vectors. The digital twin mapping and defect localization module constructs a three-dimensional digital twin model of the bridge structure based on a structured light photogrammetry algorithm, accurately mapping detected defects onto the corresponding structural unit surfaces of the three-dimensional digital twin model. The defect evolution prediction and maintenance decision module analyzes the temporal evolution trend of defects based on the spatial distribution map of defects and historical inspection data, calculates the structural risk index, and generates graded maintenance recommendations. The prediction results generated by the disease evolution prediction and maintenance decision-making module are fed back to the multispectral adaptive image acquisition module and the multiscale disease detection and feature extraction module, forming a closed-loop adaptive optimization system.

[0014] A deep learning-based method for low-altitude inspection and defect assessment of bridge structures includes a multispectral adaptive image acquisition step, a multi-scale defect detection and feature extraction step, a digital twin mapping and defect localization step, and a defect evolution prediction and maintenance decision-making step, forming a closed-loop adaptive optimization process through a feedback mechanism.

[0015] The beneficial effects of this invention are as follows:

[0016] First, it boasts strong environmental adaptability. This invention utilizes a multispectral adaptive image acquisition module to monitor ambient light intensity and weather conditions in real time, dynamically adjusting UAV flight parameters and camera exposure parameters to ensure high-quality image data acquisition under various complex environmental conditions such as strong light, shadow, fog, and low illumination. In low-illumination environments, it automatically activates infrared thermal imaging mode, using thermal imaging data to assist in the detection of internal cracks in concrete and fatigue damage in steel structures. The application of multispectral imaging technology enables the system to operate around the clock, unaffected by weather and lighting changes, significantly improving the flexibility and efficiency of inspections. Compared to existing technologies that use fixed parameters for image acquisition, this invention's adaptive acquisition strategy improves image quality consistency by more than 30% under different environments, laying a solid foundation for subsequent defect detection.

[0017] Secondly, the invention achieves high accuracy in disease detection. It employs an improved multi-scale convolutional neural network, utilizing a feature pyramid structure and an attention mechanism to accurately identify diseases at different scales, from sub-millimeter-level microcracks to centimeter-level large-area spalling. The multi-scale feature extraction layer achieves feature fusion at different scales through lateral connections, significantly improving the network's ability to detect diseases at multiple scales. The attention mechanism effectively suppresses interference from background stains and watermarks by learning the spatial weight distribution of the diseased area, enhancing the expressive power of disease features. The fusion of multispectral image data allows the system to comprehensively utilize visible and infrared spectral information, improving the accuracy of identifying different types of diseases. Experimental data shows that the invention achieves a crack detection accuracy of 96.5%, a concrete spalling detection accuracy of 94.8%, and a steel structure corrosion detection accuracy of 95.2%, all significantly superior to existing technologies.

[0018] Third, the accuracy of defect location is high. This invention constructs a three-dimensional digital twin model of the bridge structure based on a structured light photogrammetry algorithm. It does not rely on the low-precision positioning of GPS, but instead achieves centimeter-level precise location of defects in three-dimensional space through matching camera position, attitude angle, and multi-view feature points in image metadata. The digital twin model realistically reflects the bridge's geometry and structural features, and defect information is directly mapped to the surface of specific structural components, facilitating rapid location and construction by maintenance personnel. The spatial distribution map of defects visually displays the distribution of different types and severity of defects across the entire bridge structure, providing intuitive decision support for structural health assessment and maintenance planning. Compared to the meter-level accuracy of the prior art relying on GPS positioning, this invention improves defect location accuracy to the centimeter level, with a positioning error controlled within 5cm.

[0019] Fourth, it possesses the capability for disease evolution analysis. This invention, through a disease evolution prediction and maintenance decision-making module, systematically manages historical inspection data, enabling time-series tracking and calculation of the development speed of diseases at the same location. By analyzing the development trend of diseases, it can predict their future evolution path, providing a basis for formulating scientific maintenance plans. The calculation of the structural risk index comprehensively considers the type, size, location, and development speed of diseases, combined with the stress characteristics and importance coefficients of structural units, accurately assessing the impact of diseases on the overall safety of the bridge. The graded maintenance recommendation categorizes diseases into three levels according to urgency: emergency maintenance, routine maintenance, and continuous monitoring, helping management departments to rationally allocate maintenance resources and avoid over-maintenance or untimely maintenance. Compared to existing technologies that lack disease evolution analysis, this invention can predict the rapid development trend of diseases 3-6 months in advance, gaining valuable time for timely intervention.

[0020] Fifth, the system possesses adaptive optimization capabilities. This invention constructs a closed-loop mechanism from acquisition to detection, from localization to prediction, and then back to acquisition and detection, achieving adaptive optimization of the system. Disease evolution prediction results are fed back to the image acquisition module, guiding subsequent inspections to focus on areas with rapid disease development, increasing acquisition density and frequency, and optimizing the allocation of inspection resources. Detection sensitivity adjustment parameters are fed back to the disease detection module, automatically optimizing the disease detection confidence threshold based on historical detection results, reducing false positive and false negative rates. This closed-loop adaptive mechanism enables the system to continuously learn and improve; with increasing usage time, system performance continuously improves, and detection accuracy and efficiency become increasingly higher. Compared to existing technologies where each module works independently, the collaborative optimization mechanism of this invention improves overall detection efficiency by 25% and reduces the false positive rate by 40%.

[0021] Sixth, it has high application value. This invention provides a complete intelligent bridge inspection and defect management solution, forming a closed-loop management process from data collection, defect identification, precise location, evolution prediction, and maintenance decision-making. The digital twin model provides a visualized management platform for bridge health monitoring, allowing engineers and managers to intuitively understand the bridge's health status. The system has a high degree of automation, significantly reducing the workload and safety risks of manual inspections, shortening the inspection cycle from the traditional weeks to days or even hours. Defect evolution prediction and risk assessment functions provide quantitative basis for scientific decision-making, helping to achieve preventive maintenance of bridges, extend bridge service life, and reduce total life cycle costs. This invention is applicable to the inspection and management of various types of highway bridges, railway bridges, and urban overpasses, and has broad application prospects and significant economic and social benefits. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall structure of the system of the present invention.

[0023] Figure 2 This is a schematic diagram illustrating the workflow of the multispectral adaptive image acquisition module of the present invention.

[0024] Figure 3 This is a schematic diagram of the network structure of the multi-scale disease detection and feature extraction module of the present invention.

[0025] Figure 4 This is a three-dimensional mapping diagram of the digital twin mapping and disease location module of the present invention.

[0026] Figure 5 This is a schematic diagram of the analysis process of the disease evolution prediction and maintenance decision-making module of the present invention.

[0027] Figure 6 This is a schematic diagram of the closed-loop adaptive optimization mechanism of the present invention.

[0028] Figure 7 This is a schematic diagram of the overall process of the method of the present invention. Detailed Implementation

[0029] Please refer to the attached document. Figures 1-7 The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0030] like Figure 1 As shown, the bridge structure low-altitude inspection and disease assessment system based on deep learning provided by the present invention includes a multispectral adaptive image acquisition module 1, a multi-scale disease detection and feature extraction module 2, a digital twin mapping and disease localization module 3, and a disease evolution prediction and maintenance decision module 4.

[0031] like Figure 2 As shown, the multispectral adaptive image acquisition module 1 is responsible for acquiring high-quality image data of the bridge structure under different environmental conditions. This module includes an environmental parameter sensing unit, a flight parameter dynamic adjustment unit, and an exposure parameter adaptive control unit.

[0032] The environmental parameter sensing unit collects ambient light intensity, weather conditions, and temperature data in real time. Light intensity is measured by a illuminance sensor, measured in lux (lx), with a range of 0.1–100,000 lx. Weather conditions are determined by a meteorological sensor, including six types: sunny, cloudy, overcast, foggy, light rain, and heavy rain. Temperature data is measured by a temperature sensor, measured in degrees Celsius (°C), with a range of -20°C to 60°C. These environmental parameters are collected every 0.5 seconds to ensure the system can promptly detect environmental changes.

[0033] In low-light environments (such as at night or on a cloudy evening when the light intensity is less than 1000 lx), the system automatically increases the ISO sensitivity to 1600-3200, extends the exposure time to 1 / 50s, and activates the infrared thermal imaging mode as a supplement to visible light acquisition. The thermal imaging data is used to assist in the detection of internal cracks in concrete and fatigue damage in steel structures, as these defects are usually accompanied by local temperature anomalies.

[0034] The flight parameter dynamic adjustment unit dynamically adjusts the UAV's flight altitude, speed, and shooting angle based on data collected by the environmental parameter sensing unit. In strong light environments (light intensity greater than 50,000 lx), to avoid overexposure and strong shadows on the bridge surface, the system automatically increases the flight altitude to 15–20 m and employs a multi-angle shooting strategy, shooting at pitch angles of 30°, 45°, and 60° to ensure that shadowed areas are also adequately illuminated. In cloudy or low-light environments (light intensity less than 5,000 lx), the system reduces the flight altitude to 8–12 m, decreases the flight speed to 0.5–1 m / s, extends the dwell time at each shooting point, and appropriately increases the exposure time to obtain sufficient light. In hazy weather, the system activates infrared thermal imaging mode, using an infrared camera to penetrate the haze and capture the thermal radiation information of the bridge surface, assisting visible light images in defect detection.

[0035] The adaptive control unit for exposure parameters automatically adjusts the camera's exposure time, ISO sensitivity, and aperture size based on ambient light intensity. The system presets the target image brightness range to 120-150 (brightness values ​​for an 8-bit grayscale image). By analyzing the brightness histogram of the preview image in real time, it employs a closed-loop control algorithm to adjust the exposure parameters. Specifically, if the current image's average brightness is below 120, the exposure time is increased or the ISO sensitivity is raised; if the current image's average brightness is above 150, the exposure time is shortened or the ISO sensitivity is lowered. The exposure time adjustment range is 1 / 2000s to 1 / 50s, the ISO sensitivity adjustment range is 100 to 3200, and the aperture size adjustment range is f / 2.8 to f / 11. Through this adaptive control, the system ensures that the image brightness obtained under different lighting conditions remains stable, providing consistent input data for subsequent disease detection.

[0036] In a preferred embodiment of the present invention, the multispectral adaptive image acquisition module 1 uses a DJI Matrice 300 RTK drone as its flight platform, equipped with a Zenmuse H20T multi-sensor gimbal camera. This camera integrates a 20-megapixel wide-angle camera, a 12-megapixel zoom camera, and a 640×512 resolution infrared thermal imaging camera. The drone flies automatically along a pre-planned route, employing a coverage scanning mode to ensure that all parts of the bridge are fully captured. During image acquisition, the system adjusts flight and exposure parameters in real time according to changes in environmental parameters. After each image acquisition point is completed, the image data and metadata (including GPS coordinates, flight altitude, camera attitude angle, exposure parameters, and timestamp) are transmitted to the ground station for storage and processing.

[0037] To verify the environmental adaptability of the multispectral adaptive image acquisition module, this invention compared the image quality of fixed-parameter acquisition and adaptive-parameter acquisition under different environments. In a strong light environment (illuminance 80,000 lx), images acquired with fixed parameters showed severe overexposure and loss of detail in the affected areas, while images acquired with adaptive parameters exhibited moderate brightness and clearly visible defects. In a low-light environment (illuminance 2,000 lx), images acquired with fixed parameters were dark and had low contrast, while images acquired with adaptive parameters achieved better visual quality by extending the exposure time and increasing ISO sensitivity. Statistical data showed that the standard deviation of image quality evaluation indicators (including brightness, contrast, and sharpness) for adaptive acquisition under different environments was 35% lower than that for fixed-parameter acquisition, demonstrating that the adaptive strategy can significantly improve the consistency of image quality.

[0038] like Figure 3 As shown, the multi-scale disease detection and feature extraction module 2 is responsible for automatically detecting diseases and extracting features from the acquired image data. This module is based on an improved multi-scale convolutional neural network, adopts a Feature Pyramid Network (FPN) structure, and combines an attention mechanism to achieve high-precision identification of diseases at different scales.

[0039] The multi-scale disease detection network consists of three parts: a backbone feature extraction network, a multi-scale feature fusion network, and a disease detection head. The backbone feature extraction network uses ResNet-101 as its base network, extracting deep features from the image through convolutional layers and residual blocks. The ResNet-101 network contains 101 layers and can effectively extract multi-level semantic information from the image. The network input is a 512×512 pixel RGB image. After convolution and pooling operations, feature maps of different scales are generated, denoted as C2, C3, C4, and C5, with resolutions of 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the input image, respectively.

[0040] The multi-scale feature fusion network employs a top-down and lateral connection approach to fuse feature maps at different levels. Specifically, the high-level feature map C5 is reduced in dimensionality through a 1×1 convolution, then upsampled to the resolution of C4. It is then element-wise added to the feature map of C4 after a 1×1 convolution, generating a fused feature map P4. The same operation is recursively applied to P4 and C3, and P3 and C2, ultimately generating fused feature maps of four scales: P2, P3, P4, and P5. This top-down feature fusion approach allows high-level semantic information to be transmitted to lower-level feature maps, enhancing their ability to identify diseases. Lateral connections ensure information sharing between features of different scales, enabling the network to simultaneously focus on diseases at different scales, such as fine cracks and large-area spalling.

[0041] An attention mechanism module is embedded after the feature fusion network to enhance the expressive power of disease features. The attention mechanism includes two branches: spatial attention and channel attention. Spatial attention learns the spatial location weights of the disease region to suppress background interference. Channel attention learns the importance weights of different feature channels to highlight feature channels relevant to disease identification. The innovative attention mechanism used in this invention integrates attention weights from both spatial and channel dimensions, and its calculation process uses the following formula: .

[0042] in, To fuse the attention weight matrix, For the input feature map, and These are the learnable weight parameters for channel attention and spatial attention, respectively. It is the Sigmoid activation function. This is an element-wise multiplication operation. and These are global average pooling and global max pooling operations, respectively. and These represent the average and maximum feature maps of the feature maps along the channel dimension, respectively. This formula captures different types of statistical information by simultaneously utilizing average pooling and max pooling, and adaptively fuses spatial and channel attention through learnable weight parameters, achieving refined enhancement of disease features. Experimental results show that introducing this attention mechanism improves the recall rate of disease detection by 4.2% and the precision rate by 3.8%.

[0043] The defect detection head includes a classification branch and a regression branch. The classification branch is responsible for determining whether a candidate region contains defects and the type of defect. It uses a multi-class classifier and outputs the probability distribution of three defect types (cracks, concrete spalling, and steel structure corrosion) and the background. The regression branch is responsible for accurately locating the bounding boxes of defects and outputting the center coordinates, width, and height of the bounding boxes. The defect detection head executes in parallel on feature maps at four scales: P2, P3, P4, and P5. P2 is used to detect small-scale defects (bounding box area less than 32×32 pixels), P3 is used to detect small-to-medium-scale defects (bounding box area between 32×32 and 128×128 pixels), P4 is used to detect medium-to-large-scale defects (bounding box area between 128×128 and 512×512 pixels), and P5 is used to detect large-scale defects (bounding box area greater than 512×512 pixels). This multi-scale detection strategy ensures that the network can accurately identify defects of different sizes.

[0044] The multi-scale bridge defect detection network was trained end-to-end, with a loss function consisting of classification and regression losses. Focal Loss was used for classification to mitigate the imbalance between positive and negative samples. Smooth L1 Loss was used for regression optimization of bounding box coordinates. The training dataset contained 20,000 bridge defect images, covering different types and scales of defects, as well as various background interference samples. Data augmentation techniques included random cropping, rotation, flipping, and color jittering to increase the diversity of training samples and improve the model's generalization ability. The SGD optimizer was used during training, with an initial learning rate of 0.001, which decreased to 0.1 times every 50 epochs, for a total of 200 epochs. After training, the model was evaluated on an independent test set, achieving an accuracy of 96.5% for crack detection, 94.8% for concrete spalling detection, and 95.2% for steel structure corrosion detection, with a mean accuracy (mAP) of 95.5%.

[0045] In addition to outputting the disease category and location information, the feature extraction process also extracts the disease's feature vector. The disease feature vector is a 256-dimensional real-number vector, output by the fully connected layer preceding the disease detection head. This vector encodes the disease's appearance, texture, and morphological features. In subsequent disease evolution analysis, the disease feature vector is used to compare disease changes at the same location over different times. By calculating the cosine similarity between feature vectors, it can be determined whether two detected diseases are the same. Furthermore, the disease detection module also outputs a disease detection confidence score, representing the model's trustworthiness of the detection result. The confidence score ranges from 0 to 1, with higher scores indicating more reliable results. In practical applications, a confidence threshold of 0.7 is set; only detection results with a confidence score greater than 0.7 are considered valid diseases.

[0046] To handle defect detection in complex backgrounds, the multi-scale defect detection module also incorporates a background suppression strategy. Bridge surfaces often contain interfering factors such as stains, watermarks, graffiti, and paint peeling, which visually resemble actual defects and easily lead to false detections. This invention constructs a negative sample database containing typical interfering samples, enhancing the network's ability to suppress interfering samples during training. Simultaneously, infrared thermal imaging data is used to assist in judgment; cracks and corrosion are usually accompanied by localized temperature anomalies, while the temperature distribution of stains and watermarks is consistent with the surrounding environment. By fusing the detection results of visible light images and infrared thermal imaging data, the false detection rate can be effectively reduced. Experimental data shows that after introducing the background suppression strategy, the false detection rate decreased from 15.3% to 9.1%, a reduction of 40.5%.

[0047] like Figure 4 As shown, the digital twin mapping and defect localization module 3 is responsible for constructing a three-dimensional digital twin model of the bridge structure and accurately mapping the detected defects to the corresponding locations in the three-dimensional model. This module includes a three-dimensional reconstruction unit, a defect coordinate mapping unit, and a spatial distribution visualization unit.

[0048] The 3D reconstruction unit, based on structured light photogrammetry, extracts feature points from multi-view image data, generates a 3D point cloud model of the bridge structure, and converts it into a 3D digital twin model. The core of the structured light photogrammetry algorithm is to recover the coordinates of the same physical point in 3D space by analyzing the positional relationships of the same physical point in images from different viewpoints. First, feature point detection is performed on all acquired images, using the SIFT (Scale-Invariant Feature Transform) algorithm to extract key points and their descriptors. The SIFT algorithm is scale-invariant and rotation-invariant, enabling stable feature point detection under different lighting and viewing angle conditions. Then, feature points are matched between different images. The FLANN (Fast Library for Approximate Nearest Neighbors) algorithm is used to quickly search for nearest neighbor feature points, and the RANSAC (Random Sample Consensus) algorithm is used to eliminate mismatches, ensuring the accuracy of feature matching.

[0049] After obtaining a sufficient number of matching feature points, the SfM (Structure from Motion) algorithm is used to recover the camera's position, pose, and 3D structure. The SfM algorithm consists of three steps: initialization, incremental reconstruction, and global optimization. In the initialization phase, two images with sufficient baseline distance are selected, and their relative poses are calculated using essential matrix factorization (EMF). Feature points are then triangulated to obtain an initial 3D point cloud. In the incremental reconstruction phase, new images are gradually added. The camera pose of the new images is estimated using the PnP (Perspective-n-Point) algorithm, and then new feature points are triangulated to expand the 3D point cloud. In the global optimization phase, the Bundle Adjustment algorithm is used to jointly optimize all camera parameters and 3D point coordinates, minimizing reprojection errors and improving the accuracy of 3D reconstruction.

[0050] After filtering, denoising, and simplification, the 3D point cloud model is converted into a triangular mesh model. The mesh model is generated using the Poisson surface reconstruction algorithm, which produces smooth and closed surfaces suitable for structures with well-defined geometric boundaries, such as bridges. The mesh model undergoes further texture mapping, projecting the texture information of the original image onto the mesh surface to generate a realistic 3D digital twin model. The digital twin model contains not only the geometric information of the bridge but also the semantic information of each structural unit, such as the names, types, and numbers of components like piers, decks, railings, and bridge bearings. This semantic information is generated either manually or automatically using deep learning-based semantic segmentation algorithms, facilitating subsequent defect localization and management.

[0051] The disease coordinate mapping unit calculates the coordinate position of the disease in three-dimensional space based on the disease detection results and image metadata. The metadata for each image includes the camera position (GPS coordinates) at the time of image capture, flight altitude, camera attitude angles (pitch, yaw, roll), and timestamp. The position of the disease in the image is represented by the center coordinates of the bounding box output by the disease detection module, denoted as... ,in and These are the x and y coordinates in the image coordinate system, respectively. To convert the image coordinates into three-dimensional spatial coordinates, back projection calculations are needed using the camera's intrinsic and extrinsic parameters.

[0052] Camera internal parameters include focal length Principal point coordinates And distortion coefficients, these parameters are obtained through camera calibration. Camera extrinsic parameters include the camera's rotation matrix relative to the world coordinate system. Translation vector These parameters are calculated from the camera position and attitude angle in the image metadata. The normalized coordinates of the lesion points in the camera coordinate system are calculated using the following formula: .

[0053] in, and These are the normalized coordinates of the defect point in the camera coordinate system. Assume the three-dimensional coordinates of the defect point in the world coordinate system are... Then the relationship between the camera coordinate system and the world coordinate system satisfies: .

[0054] in, This represents the depth from the lesion point to the camera. Given a digital twin model, the depth can be calculated using the intersection of the ray and the model surface. Specifically, starting from the center of the camera, along... A ray is emitted in a certain direction, and the intersection point of this ray with the surface of the digital twin model is calculated. The intersection point is the location of the defect in three-dimensional space. The positioning accuracy of this method depends on the accuracy of the digital twin model and the accuracy of the camera calibration. In this invention, the positioning error can be controlled within 5cm, meeting the requirements of precise maintenance.

[0055] After the disease coordinate mapping is completed, the disease information is written into the corresponding structural unit of the digital twin model. Each disease record includes attributes such as disease type, size, three-dimensional coordinates, detection time, image number, and confidence score. The integration of disease information with the digital twin model adopts the BIM (Building Information Modeling) data standard, which facilitates data exchange and sharing with other engineering software.

[0056] The system of this invention also includes a data storage and management module, which is responsible for storing historical inspection data, three-dimensional digital twin models, and disease evolution records. The data storage and management module adopts a distributed database architecture, supporting efficient storage and rapid retrieval of large-scale data. This module supports data querying and statistical analysis by time, location, and disease type, and can generate visualization results such as disease statistical reports, trend analysis charts, and spatial distribution heat maps, providing data support for bridge health monitoring and management decisions.

[0057] The spatial distribution visualization unit uses different colors to mark different types and severity of defects on a 3D digital twin model, generating a spatial distribution map of the defects. Cracks are marked in red, concrete spalling in blue, and steel corrosion in orange. The severity of the defects is indicated by the darkness of the color, with more severe defects appearing in darker colors and less severe defects in lighter colors. The spatial distribution map is displayed in an interactive 3D view, allowing users to rotate, zoom, and pan the model to view the distribution of defects from different angles. Clicking on a defect marker will bring up a detailed information window displaying the defect's type, size, inspection time, and related images. This visualization method provides intuitive decision support for engineers and managers, facilitating a quick understanding of the overall health condition and defect distribution characteristics of the bridge.

[0058] In one application example of this invention, a drone inspection and 3D modeling of a cross-sea bridge were conducted. The bridge is 1200m long and consists of a main bridge and approach bridges. The main bridge is a cable-stayed structure, while the approach bridges are continuous beam structures. The drone flew along a pre-planned route, collecting over 5000 high-resolution images covering all structural components, including the bridge deck, piers, towers, stay cables, and railings. A complete digital twin model of the bridge was generated using a 3D reconstruction algorithm. The model contains approximately 80 million triangular faces with centimeter-level accuracy. The defect detection module detected 326 defects in the over 5000 images, including 218 cracks, 67 instances of concrete spalling, and 41 instances of steel corrosion. These defects were precisely mapped to their corresponding locations in the digital twin model. Through the spatial distribution map of the defects, technicians found that the defects were mainly concentrated at the connection between the piers and the bridge deck, as well as in the anchorage areas of the stay cables. These areas are subject to concentrated stress and severe environmental erosion, requiring focused attention and maintenance.

[0059] like Figure 5 As shown, the disease evolution prediction and maintenance decision-making module 4 is responsible for analyzing the temporal evolution trend of diseases, calculating the structural risk index, and generating graded maintenance recommendations. This module includes a time-series data analysis unit, a risk index calculation unit, and a maintenance strategy generation unit.

[0060] The time-series data analysis unit extracts temporal change data of defects at the same location from historical inspection records to calculate the development rate of the defects. Bridges typically undergo routine inspections every 3-6 months, each generating a complete record of defect detection and location. By comparing defect records from different periods, newly added defects, disappeared defects (already repaired), and persistent defects can be identified. For persistent defects, their dimensional parameters (such as crack length, width, and spalling area) at different times are extracted to calculate the development rate of the defects.

[0061] The rate of disease development was calculated using a linear regression method to fit the changes in disease size parameters over time. Assume a crack... During the second inspection, it was detected that its length was as follows: The corresponding inspection times are as follows: The change in crack length over time can then be represented by a linear model: .

[0062] in, The crack length growth rate (unit: mm / month). This is the initial length (unit: mm). The time unit is months. Parameters were obtained by fitting using the least squares method. and The estimated value. If This indicates that the crack continues to develop; if This indicates that the crack is stabilizing; if (Rarely seen in actual situations), it may be the result of measurement errors or partial repairs. Development rate It is an important indicator for assessing the severity of the disease. The greater the rate of development, the faster the disease deteriorates, and the more urgent the treatment is required.

[0063] For newly introduced diseases, the development rate cannot be directly calculated due to the lack of historical data. In this case, the system retrieves the development patterns of similar diseases from the historical database based on the disease type, size, and location, and uses a similarity-based prediction method to estimate the possible development trend of the new disease. The similarity calculation comprehensively considers factors such as disease type, size, structural unit type, and stress state, and uses a weighted Euclidean distance metric.

[0064] The risk index calculation unit comprehensively considers the type, size, location, and development rate of defects to calculate the structural risk index. The structural risk index is a dimensionless value ranging from 0 to 100; a higher value indicates a greater threat to the bridge structure's safety. The risk index is calculated using a multi-factor weighted scoring model. .

[0065] in, This is a structural risk index. , , , These are the weighting coefficients for disease type, size, location, and development rate, respectively. , , , These are the score values ​​for the corresponding factors. The score values ​​and weighting coefficients for each factor are determined based on bridge engineering experience and structural mechanics principles.

[0066] Disease type scoring The structural safety of different types of defects is assessed using numerical values. Cracks are the most common and dangerous type of defect, especially through cracks and stress cracks, which can lead to a decrease in structural load-bearing capacity and steel corrosion; therefore, cracks receive a high score. While concrete spalling affects appearance, its direct impact on structural load-bearing capacity is relatively small unless the spalling depth reaches the steel reinforcement layer. Steel corrosion weakens the cross-sectional area and strength of components, and severe corrosion can lead to component failure. In this invention, the score for cracks is set at 30, the score for concrete spalling at 20, and the score for steel corrosion at 25.

[0067] Disease size rating The dimensional score is calculated based on the geometric dimensions of the defects. For cracks, the length and width are primarily considered; the longer and wider the crack, the higher the score. For concrete spalling, the spalling area and depth are primarily considered. For steel corrosion, the corrosion area and corrosion grade are primarily considered. The dimensional score uses a piecewise linear function, mapping the dimensional parameters to a score range of 0 to 30. For example, a crack width less than 0.2 mm scores 5, a width between 0.2 and 0.5 mm scores 10, a width between 0.5 and 1.0 mm scores 20, and a width greater than 1.0 mm scores 30.

[0068] Disease location scoring The importance coefficient is determined based on the stress characteristics and importance coefficient of the structural unit where the defect is located. Different structural units of a bridge contribute differently to the overall safety; defects in critical load-bearing components (such as main beams, piers, and stay cables) are more dangerous than those in secondary components (such as guardrails and sidewalk slabs). In this invention, an importance coefficient is pre-assigned to each structural unit, ranging from 0.5 to 2.0. The importance coefficient for primary load-bearing components is 2.0, and the importance coefficient for secondary components is 0.5 to 1.0. The location score is equal to the importance coefficient multiplied by 10, i.e., ranging from 5 to 20.

[0069] Disease development speed score The score is calculated based on the development rate obtained from time-series data analysis. A higher development rate results in a higher speed score. The speed score also uses a piecewise linear function, mapping the development rate to a score range of 0 to 20. For cracks, a development rate less than 0.5 mm / month scores 5, a development rate between 0.5 and 1.0 mm / month scores 10, a development rate between 1.0 and 2.0 mm / month scores 15, and a development rate greater than 2.0 mm / month scores 20.

[0070] Weighting coefficient , , , The weighting was determined using the Analytic Hierarchy Process (AHP), calculated through expert questionnaires and pairwise comparison matrices. In this invention, after evaluation by multiple bridge engineering experts, the weighting coefficients were set as follows: , , , This indicates that disease type and size are the main factors in risk assessment, while location and development speed also have some influence.

[0071] The maintenance strategy generation unit generates tiered maintenance recommendations based on the structural risk index and preset risk thresholds. This invention categorizes maintenance recommendations into three levels: emergency maintenance, routine maintenance, and continuous monitoring.

[0072] Emergency Repair: When the structural risk index exceeds 70, the defect is classified as high-risk and requires immediate repair measures to prevent further deterioration leading to structural failure or a safety accident. Defects requiring emergency repair typically include through cracks wider than 1.0 mm, large-area concrete spalling reaching the reinforcing steel layer, and steel structures with severe corrosion resulting in a cross-sectional loss exceeding 20%. After generating an emergency repair recommendation, the system will automatically send an early warning notification to the management department and highlight the defect with a flashing red marker on the spatial distribution map.

[0073] Routine maintenance: When the structural risk index is between 40 and 70, the defects are classified as medium risk and need to be addressed in the next maintenance cycle to prevent further damage. Defects requiring routine maintenance typically include cracks with a width of 0.2–1.0 mm, small areas of concrete spalling, and light to moderate corrosion of steel structures. The system lists these defects in the maintenance plan, sorting them from highest to lowest risk index, facilitating the management department's rational allocation of maintenance resources and construction schedule.

[0074] Continuous Monitoring: When the structural risk index is less than 40, the defects are considered low-risk and currently have a minor impact on structural safety. Repair is not currently necessary, but their development should be continuously monitored during subsequent inspections. Defects requiring continuous monitoring typically include surface cracks less than 0.2 mm wide, minor concrete surface weathering, and early-stage steel structure surface corrosion. The system establishes monitoring files for these defects, updating their dimensions and risk index after each inspection. Once the risk index exceeds a threshold, the system automatically adjusts the recommended repair level.

[0075] In addition to tiered maintenance recommendations, the system also generates detailed maintenance plan suggestions, including recommended repair methods, materials, and processes. For example, for narrow cracks, epoxy resin injection sealing is recommended; for wider cracks, steel plate bonding or carbon fiber reinforcement is recommended; for concrete spalling, loose concrete should be removed first, followed by repair with polymer mortar; for steel structure corrosion, rust removal should be performed first, followed by application of anti-rust paint and anti-corrosion coating. These maintenance plan suggestions are based on bridge maintenance specifications and engineering experience, providing technical reference for construction units.

[0076] In one application example of this invention, a risk assessment and maintenance decision were made for 326 defects detected on the aforementioned cross-sea bridge. Calculations determined that 15 defects were classified as requiring urgent repair, primarily wide cracks at the connection between the piers and the bridge deck, and severe corrosion in the anchorage areas of the stay cables; the structural risk index for these defects all exceeded 75. 88 defects were classified as requiring routine repair, including moderate cracks in the bridge deck, concrete spalling in the guardrails, and cracks on the bridge tower surfaces. The remaining 223 defects were classified as requiring continuous monitoring, mainly fine cracks and initial corrosion on the bridge surface. Based on the maintenance recommendations generated by the system, the management department prioritized the handling of urgent defects and developed an annual plan for routine maintenance. After the emergency repairs were completed, a follow-up drone inspection confirmed that the high-risk defects had been effectively addressed, and the overall safety condition of the bridge had significantly improved.

[0077] like Figure 6 As shown, this invention constructs a closed-loop adaptive optimization mechanism from acquisition to detection, from localization to prediction, and then back to acquisition and detection, thereby achieving continuous improvement in system performance.

[0078] The disease evolution prediction and maintenance decision module 4 feeds back the disease development trend prediction results to the multispectral adaptive image acquisition module 1. Specifically, for diseases with a rapid development rate (such as cracks with a growth rate greater than 1.0 mm / month), the system will focus on these areas during the next inspection, increasing the shooting density and frequency. For example, for regular areas, the drone shooting interval is 5m, while for high-risk areas, the shooting interval is shortened to 2m, ensuring that every detail of the disease is fully recorded. At the same time, the system adjusts the acquisition mode according to the disease type. For crack detection, high-resolution visible light images are prioritized; for corrosion detection, infrared thermal imaging mode is enabled simultaneously. This adaptive acquisition strategy based on prediction results optimizes the allocation of inspection resources, improving detection efficiency while ensuring that high-risk areas receive sufficient attention.

[0079] The detection sensitivity adjustment parameters generated by the disease evolution prediction and maintenance decision module 4 are fed back to the multi-scale disease detection and feature extraction module 2. The system analyzes historical detection results, statistically analyzes false positive and false negative rates, and automatically adjusts the disease identification threshold. If the false positive rate for a certain type of disease is high (e.g., background stains are mistakenly identified as cracks), the system will increase the confidence threshold for that type of disease, adjusting it from 0.7 to 0.8 or 0.85 to reduce false positives. If the false negative rate for a certain type of disease is high (e.g., fine cracks are not detected), the system will decrease the confidence threshold or increase the feature extraction weight for that type of disease to improve detection sensitivity. Furthermore, the system will adjust the network parameter configuration according to the characteristics of different bridges; for example, for steel structure bridges, the weight of corrosion detection is increased; for concrete bridges, the weight of crack and spalling detection is increased. This adaptive optimization mechanism allows the detection algorithm to be customized for specific application scenarios, continuously improving detection accuracy and efficiency.

[0080] The core of the closed-loop adaptive optimization mechanism lies in the system's ability to continuously learn and improve. As the number of inspections increases, the system accumulates a large amount of disease detection, location, and evolution data, providing valuable training samples for algorithm optimization. The system periodically performs incremental training on the deep learning model, adding newly acquired images and labeled data to the training set and updating model parameters, enabling the model to adapt to new disease types and environmental conditions. Simultaneously, the system validates historical prediction results, comparing the predicted disease development trends with actual observations, calculating prediction errors, and adjusting the prediction model parameters based on error feedback. This continuous learning and self-improvement mechanism continuously enhances the system's intelligence level, gradually evolving from an initial auxiliary tool into an independent and reliable automated inspection system.

[0081] In long-term applications, the effects of the closed-loop adaptive optimization mechanism have gradually become apparent. During a 12-month continuous inspection of a bridge, the system's detection accuracy increased from 92.3% to 96.5%, the false detection rate decreased from 15.3% to 9.1%, and the inspection efficiency improved by 25%. These performance improvements are attributed to the system's continuous adjustment and optimization of the parameter configurations of each module based on actual application scenarios, as well as its continuous learning of new defect samples and environmental characteristics.

[0082] like Figure 7 As shown, the deep learning-based low-altitude inspection and defect assessment method for bridge structures of the present invention includes the following steps:

[0083] Step S1: Perform multispectral adaptive image acquisition. Dynamically adjust the UAV flight parameters and camera exposure parameters based on ambient light intensity and weather conditions to acquire multispectral image data of the bridge structure under different lighting conditions, including visible light images and infrared thermal imaging data. The image data, along with metadata (GPS coordinates, flight altitude, camera attitude angle, exposure parameters, timestamp), is stored.

[0084] Step S2: Perform multi-scale defect detection and feature extraction. Defect detection is performed on the acquired multispectral image data using a multi-scale convolutional neural network to identify types of defects such as cracks, concrete spalling, and steel structure corrosion. Defect feature vectors are extracted, including defect type identifiers, size parameters, and morphological features, and a defect detection confidence score is generated. Only detection results with a confidence score greater than a preset threshold (usually 0.7) are considered valid defects.

[0085] Step S3: Perform digital twin mapping and disease localization. Based on the structured light photogrammetry algorithm, feature points are extracted from multi-view image data. The camera pose and 3D structure are recovered using the SfM algorithm to generate a 3D point cloud model of the bridge structure, which is then converted into a 3D digital twin model. Based on the disease feature vectors and spatial coordinate information in the image metadata, backprojection calculations are used to accurately map the detected diseases onto the corresponding structural unit surfaces of the 3D digital twin model. Different types and severity of diseases are marked with different colors on the digital twin model, generating a spatial distribution map of the diseases.

[0086] Step S4: Execute the disease evolution prediction and maintenance decision-making steps. Extract the temporal change data of diseases at the same location from historical inspection records, and calculate the disease development rate through linear regression. Calculate the structural risk index using a multi-factor weighted scoring model, considering disease type, size, location, and development rate. Based on the structural risk index and preset risk thresholds, generate disease severity level assessment results and graded maintenance recommendations, categorized into three levels: emergency maintenance, routine maintenance, and continuous monitoring. Simultaneously, generate detailed maintenance plan recommendations, including recommended maintenance methods, materials, and processes.

[0087] Step S5: Execute the closed-loop feedback optimization step. The disease development trend prediction results generated in the disease evolution prediction and maintenance decision-making step are fed back to the multispectral adaptive image acquisition step to adjust the key areas and acquisition density for the next inspection, increasing the imaging density for areas with rapidly developing diseases. The detection sensitivity adjustment parameters are fed back to the multi-scale disease detection and feature extraction step to automatically optimize the disease identification threshold based on historical detection results, reducing false positive and false negative rates. Through this closed-loop feedback mechanism, the system can adaptively adjust the acquisition strategy and detection algorithm parameters, achieving continuous performance improvement.

[0088] In a specific implementation of this invention, a complete inspection and defect assessment of an overpass in a certain city was conducted. This overpass is a three-level interchange with a complex structure, and traditional manual inspection would take two weeks. Using the system of this invention, automated drone inspection completed all image acquisition in just 6 hours, collecting over 8,000 images. Defect detection took 4 hours, detecting a total of 541 defects, including 378 cracks, 102 instances of concrete spalling, and 61 instances of steel structure corrosion. Digital twin modeling took 8 hours, generating a 3D model with centimeter-level accuracy. Defect location and risk assessment took 2 hours, identifying 23 defects requiring urgent repair and 145 defects requiring routine repair. The entire inspection and assessment process was completed within 24 hours, significantly shortening the time required by traditional methods, and providing more comprehensive and accurate results. Based on the maintenance recommendations generated by the system, the management department promptly addressed high-risk defects, ensuring the safe operation of the bridge.

[0089] The bridge structure low-altitude inspection and defect assessment system and method based on deep learning provided by this invention forms a complete closed-loop adaptive optimization system through the deep coupling of four core modules: multispectral adaptive image acquisition, multi-scale defect detection and feature extraction, digital twin mapping and defect localization, and defect evolution prediction and maintenance decision-making. This system exhibits strong environmental adaptability, high detection accuracy, precise localization, and defect evolution analysis capabilities, providing comprehensive technical support for bridge health monitoring and intelligent management. It has broad application prospects and significant socio-economic benefits.

[0090] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications and substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A deep learning-based low-altitude inspection and defect assessment system for bridge structures, characterized in that, include: The multispectral adaptive image acquisition module is used to dynamically adjust the flight parameters of the UAV and the camera exposure parameters according to the ambient light intensity and weather conditions, and to acquire multispectral image data of the bridge structure under different lighting conditions. The multispectral image data includes visible light images and infrared thermal imaging data. A multi-scale disease detection and feature extraction module, connected to the multispectral adaptive image acquisition module, is used to detect diseases in the multispectral image data based on a multi-scale convolutional neural network, identify types of diseases such as cracks, concrete spalling, and steel structure corrosion, extract disease feature vectors, which include disease type identifiers, size parameters, and morphological features, and generate disease detection confidence scores. The multi-scale disease detection and feature extraction module employs an improved feature pyramid network structure, which includes multiple feature extraction layers at various scales. Each feature extraction layer corresponds to a different size of disease detection range, and the feature extraction layers at different scales are fused through lateral connections, enabling simultaneous detection of sub-millimeter-level fine cracks and centimeter-level large-area spalling. The multi-scale disease detection and feature extraction module also employs an attention mechanism to enhance feature extraction capabilities. This attention mechanism learns the spatial weight distribution of the disease region, suppresses background interference information, and enhances the expressive power of disease features. The digital twin mapping and disease localization module is connected to the multi-scale disease detection and feature extraction module. It is used to construct a three-dimensional digital twin model of the bridge structure based on the structured light photogrammetry algorithm. According to the disease feature vector and the spatial coordinate information in the image metadata, the detected disease is accurately mapped to the surface of the corresponding structural unit of the three-dimensional digital twin model, and a disease spatial distribution map is generated. The disease evolution prediction and maintenance decision module is connected to the digital twin mapping and disease location module. It is used to analyze the temporal evolution trend of the disease based on the disease spatial distribution map and historical inspection data, calculate the structural risk index, generate disease severity level assessment results and graded maintenance suggestions. The structural risk index comprehensively considers the disease type, size, location and development speed. The disease development trend prediction results generated by the disease evolution prediction and maintenance decision module are fed back to the multispectral adaptive image acquisition module to adjust the key areas and acquisition density of the next inspection. The detection sensitivity adjustment parameters generated by the disease evolution prediction and maintenance decision module are fed back to the multi-scale disease detection and feature extraction module to optimize the disease detection confidence threshold, forming a closed-loop adaptive optimization system.

2. The system according to claim 1, characterized in that, The multispectral adaptive image acquisition module includes: An environmental parameter sensing unit is used to collect ambient light intensity, weather conditions, and temperature data in real time. The flight parameter dynamic adjustment unit is used to dynamically adjust the flight altitude, flight speed and shooting angle of the UAV based on the data collected by the environmental parameter sensing unit. An adaptive exposure parameter control unit is used to automatically adjust the camera's exposure time, ISO sensitivity, and aperture size according to the ambient light intensity, so that the brightness of the image obtained under different lighting conditions remains within a preset range.

3. The system according to claim 1, characterized in that, The digital twin mapping and disease localization module includes: The three-dimensional reconstruction unit is used to extract feature points from multi-view image data based on the structured light photogrammetry algorithm, generate a three-dimensional point cloud model of the bridge structure, and convert it into the three-dimensional digital twin model. The disease coordinate mapping unit is used to calculate the coordinate position of the disease in three-dimensional space based on the GPS coordinates, flight altitude and camera attitude angle in the image metadata, and map the disease information onto the surface of the corresponding structural unit of the three-dimensional digital twin model. The spatial distribution visualization unit is used to mark different types and severity of diseases with different colors on the three-dimensional digital twin model to generate a spatial distribution map of the diseases.

4. The system according to claim 1, characterized in that, The disease evolution prediction and maintenance decision-making module includes: The time-series data analysis unit is used to extract time-series change data of diseases at the same location from historical inspection records and calculate the development speed of diseases; The risk index calculation unit is used to calculate the structural risk index by comprehensively considering the type, size, location and development speed of the disease. The calculation of the structural risk index takes into account the stress characteristics and importance coefficient of the structural unit where the disease is located. The maintenance strategy generation unit is used to generate the graded maintenance recommendations based on the structural risk index and the preset risk threshold. The graded maintenance recommendations include three levels: emergency maintenance, routine maintenance, and continuous monitoring.

5. The system according to claim 1, characterized in that, The system also includes a data storage and management module for storing historical inspection data, three-dimensional digital twin models, and disease evolution records. The data storage and management module supports data querying and statistical analysis by time, location, and disease type.

6. The system according to claim 1, characterized in that, The multispectral adaptive image acquisition module automatically activates the infrared thermal imaging mode in low-light or hazy weather, and the infrared thermal imaging data is used to assist in the detection of internal cracks in concrete and fatigue damage in steel structures.

7. The system according to claim 1, characterized in that, The closed-loop adaptive optimization system continuously learns from historical inspection data and disease evolution patterns, automatically optimizing the data acquisition strategy and detection algorithm parameters to achieve continuous improvement in system performance.

8. A method for low-altitude inspection and defect assessment of bridge structures based on deep learning, employing the system described in any one of claims 1-7, characterized in that, Includes the following steps: Through a multispectral adaptive image acquisition step, the flight parameters of the UAV and the camera exposure parameters are dynamically adjusted according to the ambient light intensity and weather conditions to acquire multispectral image data of the bridge structure under different lighting conditions. The multispectral image data includes visible light images and infrared thermal imaging data. Through a multi-scale disease detection and feature extraction process, a multi-scale convolutional neural network is used to detect diseases in the multispectral image data, identifying types of diseases such as cracks, concrete spalling, and steel structure corrosion. Disease feature vectors are extracted, including disease type identifiers, size parameters, and morphological features, and a disease detection confidence score is generated. In this multi-scale disease detection and feature extraction process, an improved feature pyramid network structure is used for multi-scale feature fusion. This improved feature pyramid network structure includes multiple feature extraction layers at various scales, each corresponding to a different size of disease detection range. Lateral connections between these feature extraction layers achieve the fusion of features at different scales, enabling the simultaneous detection of sub-millimeter-level fine cracks and centimeter-level large-area spalling. An attention mechanism is employed to learn the spatial weight distribution of the disease region, suppressing background interference and enhancing the expressive power of disease features. Through digital twin mapping and disease localization steps, a three-dimensional digital twin model of the bridge structure is constructed based on the structured light photogrammetry algorithm. According to the disease feature vector and spatial coordinate information in the image metadata, the detected diseases are accurately mapped to the surface of the corresponding structural unit of the three-dimensional digital twin model, and a spatial distribution map of the diseases is generated. Through disease evolution prediction and maintenance decision-making steps, based on the disease spatial distribution map and historical inspection data, the temporal evolution trend of the disease is analyzed, the structural risk index is calculated, and the disease severity level assessment results and graded maintenance suggestions are generated. The structural risk index comprehensively considers the disease type, size, location and development speed. The disease development trend prediction results generated by the disease evolution prediction and maintenance decision-making steps are fed back to the multispectral adaptive image acquisition steps to adjust the key areas and acquisition density of the next inspection. The detection sensitivity adjustment parameters generated by the disease evolution prediction and maintenance decision-making steps are fed back to the multi-scale disease detection and feature extraction steps to optimize the disease identification threshold, forming a closed-loop adaptive optimization process.

Images