Unmanned aerial vehicle bridge disease inspection method and system based on DeepSeek-YOLO target detection

Abstract

The invention discloses an unmanned aerial vehicle bridge disease inspection method and system based on DeepSeek-YOLO target detection. The method comprises the following steps: S100, planning a flight path of an unmanned aerial vehicle; s200, a unified time reference is provided for all the sensors, and hardware trigger signal synchronization is carried out; s300, the unmanned aerial vehicle flies according to the planned route, and the multi-sensor collaborative acquisition module acquires data at the same time; s400, processing the collected data, fusing the processed data, and constructing a bridge disease database; s500, establishing a YOLO target detection model, and optimizing the YOLO target detection model through the convolutional neural network in combination with DeepSeek; s600, performing disease identification by using a YOLO target detection model to obtain bridge disease information; s700, selecting multi-dimensional data, and generating a three-dimensional model of the bridge; s800, outputting an image in the digital twinborn visualization platform; and S900, updating the three-dimensional model in real time according to bridge disease information, comparing an identified disease result with an expert label, and feeding back the result to the YOLO target detection model to further train the YOLO target detection model.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of unmanned aerial vehicle inspection, and more particularly relates to an unmanned aerial vehicle bridge disease inspection system and method based on DeepSeek-YOLO target detection. BACKGROUND

[0002] As the core hub of the transportation network, the safety and durability of the bridge are directly related to public safety and the stability of economic operation. With the increase of the service life of the bridge, environmental erosion (such as salt spray, freeze-thaw cycle), material aging (such as concrete carbonization, steel corrosion) and long-term overload operation (heavy vehicles passing far beyond the design load) and other factors jointly act, leading to continuous degradation of structural performance. Some early built bridges have low design standards, and there are significant safety hazards under the current traffic load. If the diseases of the bridge are not detected in time, local collapse may occur, causing major accidents. Therefore, it is necessary to regularly inspect the bridge.

[0003] The existing bridge detection technology system is divided into three categories: traditional manual detection, non-destructive detection and intelligent monitoring technology. The conventional visual inspection is still the basic means, and the detection personnel use crack observation instruments, steel tapes and other tools to record the surface diseases of the bridge deck pavement, expansion joints and beam cracks. Non-destructive detection technology penetrates the internal structure through physical signals, such as ultrasonic detection of internal defects of concrete, X-ray detection of steel member weld quality, and infrared thermal imaging method to identify bridge leakage area. The intelligent monitoring system realizes real-time monitoring of strain and temperature change through fiber Bragg grating sensors; through unmanned aerial vehicle carrying high-definition camera and laser radar, accurate three-dimensional modeling of bridge tower, cable and other high-altitude parts is realized; deep learning algorithm extracts features from collected crack images, improves recognition accuracy.

[0004] However, the method of inspection by unmanned aerial vehicle carrying high-definition camera and laser radar has the following problems: it is difficult to obtain disease data in actual engineering, which leads to insufficient model generalization ability, it is difficult to comprehensively identify various diseases, and the detection accuracy in complex environment needs to be improved. Therefore, an unmanned aerial vehicle bridge disease inspection system and method based on DeepSeek-YOLO target detection is needed to improve the multi-modal data processing capability, accurately locate the spatial distribution of diseases, and improve the efficiency of inspection. SUMMARY

[0005] In view of the above defects or improvement needs of the prior art, the present application provides a UAV bridge disease inspection method and system based on DeepSeek-YOLO target detection, which combines the lightweight model at the UAV end with the fine analysis model at the cloud end to realize real-time processing and deep analysis of data. The UAV end is responsible for preliminary disease identification and data compression, and the cloud end is responsible for complex data analysis and model training. The digital twin visualization platform can run on various devices, including desktop computers, mobile devices, etc., to facilitate users to check the health status of the bridge at any time and anywhere. The safety design of the system is strengthened to ensure the safety of data transmission and storage. At the same time, the performance of the system is optimized to ensure its stable and reliable operation.

[0006] To achieve the above object, according to a first aspect of an embodiment of the present application, a UAV bridge disease inspection method based on DeepSeek-YOLO target detection is provided, which specifically comprises the following steps:

[0007] S100, the flight path of the UAV is planned by controlling the overlap rate of the detection angle and the detection area;

[0008] S200, a unified time reference is provided for all sensors in the multi-sensor cooperative acquisition module using GPS, Beidou or atomic clock, and hardware trigger signal synchronization is performed through FPGA to ensure time alignment;

[0009] S300, the UAV flies according to the planned route, the multi-sensor cooperative acquisition module simultaneously acquires data, and the data is marked and aligned through time stamp and spatial positioning information;

[0010] S400, the data acquired by the multi-sensor cooperative acquisition module is processed and fused, and a bridge disease database is constructed, the acquired data is labeled and included in the bridge disease database;

[0011] S500, a YOLO target detection model is established, the YOLO target detection model is optimized through a convolutional neural network combined with DeepSeek, and the YOLO target detection model is trained using existing data;

[0012] S600, the fused data in the bridge disease database is used to identify diseases through the YOLO target detection model to obtain bridge disease information;

[0013] S700, multi-dimensional data is selected from the data acquired by the multi-sensor cooperative acquisition module, and a three-dimensional model of the bridge is generated after the multi-dimensional data is fused and processed;

[0014] S800, the bridge disease information identified by the YOLO target detection model is integrated with the three-dimensional model, and an image is output in the digital twin visualization platform.

[0015] S900, updating the three-dimensional model in real time according to the bridge disease information, and comparing the identified disease result with the expert annotation to feed back to the YOLO target detection model for further training.

[0016] Further, in step S400, the data collected by the multi-sensor cooperative collection module is processed and fused, specifically: combining visible light images, infrared images and laser radar data to more comprehensively evaluate the health status of the bridge.

[0017] Further, the fusion includes early fusion and late fusion.

[0018] The early fusion refers to fusion in the data collection or feature extraction stage, fusing the features of optical images and infrared images to generate more rich feature representations.

[0019] The late fusion refers to fusion in the decision-making stage, comprehensively analyzing the detection results of different sensors to obtain the final disease judgment.

[0020] When fusing, a multi-branch network structure is used to process data of different modalities respectively, and then fuse at an appropriate stage.

[0021] Further, in step S500, when optimizing the YOLO target detection model through the convolutional neural network combined with DeepSeek, a technical closed loop needs to be constructed from three dimensions of model optimization, data enhancement and data integration.

[0022] When optimizing the model, an attention mechanism needs to be added to improve detection accuracy; when data enhancement, the multi-modal data processing capability of DeepSeek needs to be combined; when data integration, edge computing and API calling need to be used to guide the deployment of the YOLO target detection model to the actual environment, and a three-level architecture of UAV-edge server-DeepSeek cloud is constructed.

[0023] Finally, a feedback data model is also needed for continuous learning and optimization.

[0024] Further, when training the YOLO target detection model, distributed training and hybrid training are used, and a multi-data labeling system is used to develop a hybrid labeling tool chain, and the YOLO native rectangular frame labeling is used to label bridge diseases including cracks and peeling; the DeepSeek point cloud segmentation module is used to process laser radar data, and the detection report is automatically mapped to the image label.

[0025] Further, the step of calling the API is: first, register an account and log in to the console to obtain an API key starting with `sk-`, and then refer to the official documentation to confirm the request endpoint, authentication method, model name, and parameter restrictions of the target API;

[0026] Send JSON format data to the specified URL through HTTP POST request, need to carry `Authorization:Bearer YOUR_API_KEY` in the request header, the content includes dialogue message, temperature value;

[0027] Parse the JSON result and extract `choices[0].message.content` after receiving the response, and need to check the status code to handle errors;

[0028] Through streaming, adjusting the advanced parameters of `temperature` / `max_tokens`, when calling, attention should be paid to the permission billing, data security and model token length limit.

[0029] Further, in step S600, the fused data is transmitted in real time to the ground control station or cloud server through wireless communication, and the operator can view the data in the bridge disease database in real time and perform preliminary analysis;

[0030] When a crack is identified, it is analyzed whether it is a transverse crack or a longitudinal crack, and its severity is evaluated, combined with the disease detection results and the historical data of the bridge, the overall health condition of the bridge is evaluated, and a health evaluation report is generated;

[0031] According to the severity and distribution of the disease, a maintenance priority ranking is generated to provide scientific maintenance recommendations for the bridge maintenance department.

[0032] Further, in step S700, a three-dimensional model of the bridge is generated, specifically: using the point cloud data generated by the laser radar, combined with the real-time positioning and map building algorithm to generate a three-dimensional model of the bridge, which truly reflects the geometric shape, structural details and spatial relationship of the bridge;

[0033] And map the high-resolution images taken by the optical camera onto the three-dimensional model to give the model a real appearance texture, making the digital twin model more realistic in vision.

[0034] According to the second aspect of the embodiment of the application, a UAV bridge disease inspection system based on DeepSeek-YOLO target detection is provided, which includes a multi-sensor cooperative acquisition module, an intelligent analysis module, and a digital twin visualization platform. Each module works cooperatively through data flow to ensure the accuracy of the determination result.

[0035] The multi-sensor cooperative acquisition module is integrated on the unmanned aerial vehicle, and the unmanned aerial vehicle drives the multi-sensor cooperative acquisition module to perform bridge inspection to simultaneously acquire various information of the bridge.

[0036] The intelligent analysis module processes the data collected by the multi-sensor cooperative acquisition module, and comprehensively analyzes the detection results after processing to obtain a final disease judgment.

[0037] The digital twin visualization platform fuses and processes the bridge information acquired by the unmanned aerial vehicle, and generates a three-dimensional model of the bridge, and integrates the detected bridge diseases with the three-dimensional model.

[0038] Further, the digital twin visualization platform is provided with an output interface, and the output interface includes a three-dimensional visualization interface, a disease distribution diagram, and real-time monitoring and alarm.

[0039] The three-dimensional visualization interface takes a high-precision three-dimensional digital twin model as a core, updates disease detection data in real time, and marks disease positions through colors or icons.

[0040] Users can view disease details, compare historical data, and view bridge details from different angles through the perspective control through the interactive function.

[0041] The disease distribution diagram provides a two-dimensional plan and a heat map to intuitively display disease distribution and severity, and help users quickly identify disease concentration areas.

[0042] The real-time monitoring and alarm display data collected by the unmanned aerial vehicle in real time, and issue an alarm prompt when serious diseases are detected.

[0043] Overall, the above technical solutions conceived by the present application can achieve the following beneficial effects compared with the prior art:

[0044] 1. The unmanned aerial vehicle bridge disease inspection method of the present application combines the lightweight model at the unmanned aerial vehicle end with the fine analysis model in the cloud to realize real-time processing and deep analysis of data. The unmanned aerial vehicle end is responsible for preliminary disease identification and data compression, and the cloud is responsible for complex data analysis and model training. Ensure that the digital twin visualization platform can run on various devices, including desktop computers, mobile devices, etc., to facilitate users to view the health status of the bridge anytime and anywhere. Strengthen the safety design of the system to ensure the safety of data transmission and storage. At the same time, optimize the performance of the system to ensure its stable and reliable operation.

[0045] 2. The unmanned aerial vehicle bridge disease inspection method of the present application has a simple and easy-to-use output interface, supports multi-device access, and allows users to view the bridge status at any time and anywhere. The disease detection report generated by the system provides an overview of the disease, a detailed disease list, disease pictures and videos, and classifies and quantitatively analyzes the disease to generate intuitive charts. The health assessment report generated by the system is based on the disease detection results and historical data, generates bridge health assessment indicators, analyzes health trends, and provides maintenance recommendations. By comparing disease data in different time periods, the disease evolution trend is analyzed, and comparison analysis with bridges of the same type is supported.

[0046] 3. The unmanned aerial vehicle bridge disease inspection method of the present application accesses the DeepSeek API to the YOLO bridge disease detection system. When YOLO detects bridge cracks, peeling and other diseases, the system will automatically call the DeepSeek API and input the disease image and detection result. DeepSeek generates a structured report through its multi-modal or text API, including disease type, risk level, and repair recommendations, greatly improving the accuracy and efficiency of disease identification.

[0047] 4. The unmanned aerial vehicle bridge disease inspection method of the present application, DeepSeek also supports quantitative analysis of the identified disease, determines the category and severity of the disease, and generates a disease distribution map and a quantitative report to support bridge maintenance decisions. By building a digital twin visualization platform, customers are served more intuitively. Not only does it improve the accuracy of disease detection, but it also provides a solid technical foundation for long-term health monitoring and decision support for bridges. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 A step schematic diagram of an unmanned aerial vehicle bridge disease inspection method based on DeepSeek-YOLO target detection according to an embodiment of the present application;

[0049] Figure 2 A flowchart of an unmanned aerial vehicle bridge disease inspection method based on DeepSeek-YOLO target detection according to an embodiment of the present application;

[0050] Figure 3 A framework diagram of an unmanned aerial vehicle bridge disease inspection system based on DeepSeek-YOLO target detection according to an embodiment of the present application;

[0051] Figure 4 A digital twin visualization platform workflow diagram of an unmanned aerial vehicle bridge disease inspection system based on DeepSeek-YOLO target detection according to an embodiment of the present application;

[0052] Figure 5A user interface diagram of a digital twin visualization platform of an unmanned aerial vehicle bridge disease inspection system based on DeepSeek-YOLO target detection is provided in an embodiment of the present application. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application is further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0054] Embodiment 1

[0055] As shown in Figure 1 , 2 , the present application provides an unmanned aerial vehicle bridge disease inspection method based on DeepSeek-YOLO target detection, which specifically comprises the following steps:

[0056] S100, planning a flight path of the unmanned aerial vehicle by controlling the overlap rate of the detection angle and the detection area;

[0057] S200, using GPS, Beidou or atomic clock to provide a unified time reference for all sensors in the multi-sensor cooperative acquisition module, and synchronizing hardware trigger signals through FPGA to ensure time alignment;

[0058] S300, the unmanned aerial vehicle flies according to the planned route, the multi-sensor cooperative acquisition module simultaneously collects data, and the data is marked and aligned through time stamp and spatial positioning information;

[0059] S400, after processing the data collected by the multi-sensor cooperative acquisition module, the data is fused and a bridge disease database is constructed, the collected data is labeled and included in the bridge disease database;

[0060] S500, establishing a YOLO target detection model, optimizing the YOLO target detection model through convolutional neural network combined with DeepSeek, and training the YOLO target detection model using existing data;

[0061] S600, using the YOLO target detection model to identify the diseases in the fused data in the bridge disease database to obtain bridge disease information;

[0062] S700, selecting multi-dimensional data from the data collected by the multi-sensor cooperative acquisition module, and generating a three-dimensional model of the bridge after fusion processing of the multi-dimensional data;

[0063] S800, integrate the bridge disease information identified by the YOLO target detection model into the three-dimensional model, and output the image in the digital twin visualization platform;

[0064] S900, update the three-dimensional model in real time according to the bridge disease information, and compare the identified disease results with the expert annotation to feed back to the YOLO target detection model for further training.

[0065] In step S100, when planning the flight path of the unmanned aerial vehicle, it is ensured that it can cover all parts of the bridge, especially hidden parts (such as the bridge bottom and high piers). The flight path planning needs to consider the structural characteristics and detection requirements of the bridge. Using the autonomous obstacle avoidance function of the unmanned aerial vehicle, obstacles are avoided to ensure flight safety. At the same time, through the path optimization algorithm, it is ensured that the unmanned aerial vehicle can efficiently cover all key areas to ensure the reliability of subsequent data processing.

[0066] In step S200, according to the requirements of bridge detection, multiple sensors in the multi-sensor cooperative acquisition module are selected, and these sensors are integrated into the unmanned aerial vehicle, and the integrated position needs to ensure that they can work simultaneously and do not interfere with each other, avoiding mutual interference between the optical camera and the infrared thermal imager. Through the cooperative work of multiple sensors, multiple information of the bridge is obtained, specifically: the optical camera provides high-resolution visual images, the infrared thermal imager provides temperature distribution information, and the combination of the two can more comprehensively identify diseases.

[0067] In step S300, during the flight of the unmanned aerial vehicle, multiple sensors simultaneously collect data, the optical camera shoots high-resolution images, the infrared thermal imager records temperature distribution, and the laser radar generates three-dimensional point cloud data. And through image enhancement technology, the image is preprocessed to improve the image quality. The image enhancement technology includes denoising, distortion correction and illumination equalization.

[0068] In step S400, the data collected by the multi-sensor cooperative acquisition module is fused after processing, specifically: combining visible light images, infrared images and laser radar data to more comprehensively evaluate the health status of the bridge. Fusion includes early fusion and late fusion. The early fusion refers to the fusion of the features of the optical image and the infrared image in the data acquisition or feature extraction stage to generate more rich feature representation. The late fusion refers to the fusion of the detection results of different sensors in the decision-making stage for comprehensive analysis to obtain the final disease judgment. When fusing, a multi-branch network structure is adopted to process different modal data respectively, and then fused at an appropriate stage.

[0069] In step S400, the collected data is labeled, specifically: the disease data collected including cracks, corrosion and concrete spalling is labeled combined with expert experience.

[0070] In step S500, when optimizing the YOLO target detection model through the convolutional neural network combined with DeepSeek, a technical closed loop needs to be constructed from three dimensions of model optimization, data enhancement, and data integration.

[0071] When optimizing the model, an attention mechanism needs to be added to improve detection accuracy; when data enhancement is performed, the multi-modal data processing capability of DeepSeek needs to be combined; when data integration is performed, edge computing and API calling are needed to guide the deployment of the YOLO target detection model to the actual environment, and a three-level architecture of UAV-edge server-DeepSeek cloud is constructed. Finally, feedback data models need to be used for continuous learning and optimization.

[0072] When training the YOLO target detection model, distributed training and hybrid training are used, and a multi-data labeling system is adopted to develop a hybrid labeling tool chain. The YOLO native rectangular box labeling is used to include bridge diseases such as cracks and peeling. The DeepSeek point cloud segmentation module is used to process laser radar data, and the detection report is automatically mapped to the image label.

[0073] The steps of calling the API are as follows: first, register an account and log in to the console to obtain an API key starting with `sk-`, then consult the official documentation to confirm the request endpoint (URL) of the target API, the authentication method, the model name (such as `deepseek-chat`), and the parameter limit. Send JSON format data to the specified URL through HTTP POST request, and need to carry `Authorization:Bearer YOUR_API_KEY` in the request header, the content includes conversation messages, temperature values and other parameters. After receiving the response, parse the JSON result and extract `choices[0].message.content`, and need to check the status code to handle errors. Support streaming, adjust `temperature` / `max_tokens` and other advanced parameters, when calling, need to pay attention to the permission billing, data security and model token length limit.

[0074] In step S600, the fused data is transmitted in real time to the ground control station or cloud server through wireless communication. The operator can view the data in the bridge disease database in real time and perform preliminary analysis. When a crack is identified, it is analyzed whether it is a transverse crack or a longitudinal crack, and its severity is evaluated. Combined with the disease detection results and the historical data of the bridge, the overall health status of the bridge is evaluated, and a health assessment report is generated. According to the severity and distribution of the disease, a maintenance priority ranking is generated to provide scientific maintenance recommendations for the bridge maintenance department.

[0075] In step S700, a three-dimensional model of the bridge is generated. Specifically, a three-dimensional model of the bridge is generated using point cloud data generated by the laser radar in combination with a SLAM (simultaneous localization and mapping) algorithm to accurately reflect the geometric shape, structural details and spatial relationships of the bridge. High-resolution images captured by the optical camera are mapped onto the three-dimensional model to give the model a realistic appearance texture, making the digital twin model more visually realistic.

[0076] As a further optimization, after step S900, the functions and performance of the platform need to be continuously optimized based on user feedback and technological development. More advanced deep learning algorithms are introduced to improve the accuracy and efficiency of disease identification. Data fusion and visualization technologies are optimized to enhance user experience.

[0077] It can be understood that the embodiments of the present application combine the lightweight model on the unmanned aerial vehicle side with the fine analysis model on the cloud side to realize real-time processing and deep analysis of data. The unmanned aerial vehicle side is responsible for preliminary disease identification and data compression, and the cloud side is responsible for complex data analysis and model training. The digital twin visualization platform can run on various devices, including desktop computers, mobile devices, etc., allowing users to view the health status of the bridge at any time and anywhere. The security of the system is strengthened to ensure the security of data transmission and storage. At the same time, the performance of the system is optimized to ensure its stable and reliable operation.

[0078] In a preferred embodiment, the architecture of the YOLO target detection model includes:

[0079] Data acquisition layer: DJI M300 RTK unmanned aerial vehicle equipped with Zenmuse L1 laser radar (point frequency 240,000 points / second), FLIR A8580 infrared thermal imager (thermal sensitivity 30 mK), Sony a7R V high-resolution camera (61MP).

[0080] Data processing layer: The collected raw data is processed for denoising, distortion correction, and illumination equalization to improve data quality. Point cloud data is used for registration and three-dimensional reconstruction to generate a high-precision three-dimensional model of the bridge to support disease positioning.

[0081] Model construction layer: YOLO algorithm is used to integrate Deep Seek for accurate positioning of disease areas; convolutional neural network CNN is used for pixel-level disease classification; Deep Seek's multi-modal learning capability combines visible light, infrared, laser radar and other multi-source data to improve the accuracy and robustness of disease identification.

[0082] Simulation analysis layer: Deep Seek calls ANSYS for finite element analysis to identify and classify bridge disease types, supports edge-cloud collaborative computing architecture, and realizes real-time identification and analysis of diseases.

[0083] Digital twin visualization module: Based on the three-dimensional reconstruction results, a digital twin model of the bridge is constructed to realize real-time mapping of the physical bridge and the virtual model. Through high-resolution rendering technology, the disease identification results are displayed in an intuitive visual form on the digital twin model, supporting multi-dimensional data superposition analysis and interactive operation. Disease distribution map and quantitative report are generated to provide scientific basis for bridge maintenance decision-making.

[0084] Application and feedback module: Utilize the efficient inspection capability of the unmanned aerial vehicle to quickly complete the bridge detection task, especially for high-risk areas such as the bottom of the bridge and high piers. Through historical data comparison, analyze the evolution trend of the disease to support long-term health monitoring of the bridge. According to the feedback in actual application, continuously optimize the AI model and detection process to improve the performance and reliability of the system.

[0085] Obtain the disease data of each part of the bridge through the camera, and perform histogram equalization processing and data enhancement processing on the input data to obtain the updated working image.

[0086] Determine the type of disease image and classify it into four categories: cracks, spalling, internal hollowing, and water seepage. Real-time label disease location on the three-dimensional model and provide disease detailed information query function to facilitate users to intuitively view the disease situation.

[0087] The unmanned aerial vehicle + AI detection system can shorten the time of traditional manual detection from several days to several hours, significantly improving the detection efficiency.

[0088] Embodiment 2

[0089] As shown in Figures 3-5 The embodiment of the application provides an unmanned aerial vehicle bridge disease inspection system based on DeepSeek-YOLO target detection, which comprises a multi-sensor cooperative acquisition module, an intelligent analysis module and a digital twin visualization platform. Each module cooperates with each other through data flow to ensure the accuracy of the determination result. The multi-sensor cooperative acquisition module is integrated on the unmanned aerial vehicle, which is driven by the unmanned aerial vehicle to inspect the bridge and obtain various information of the bridge at the same time. The intelligent analysis module processes the data collected by the multi-sensor cooperative acquisition module and comprehensively analyzes the detection results after processing to obtain the final disease judgment. The digital twin visualization platform fuses and processes the bridge information obtained by the unmanned aerial vehicle and generates a three-dimensional model of the bridge, and integrates the detected bridge disease with the three-dimensional model.

[0090] The multi-sensor cooperative acquisition module includes: a high-resolution optical camera for detecting surface diseases including cracks and peeling; an infrared thermal imager for detecting internal problems including hollowing and water seepage; a laser radar for detecting three-dimensional structural deformation; a multi-spectral / hyper-spectral sensor for detecting material aging.

[0091] The intelligent analysis module is a YOLO target detection model optimized using a convolutional neural network combined with DeepSeek, which is used for pixel-level classification in cooperation with a U-Net semantic segmentation enhanced by DeepSeek, and is deployed through a DeepSeek edge computing framework for multi-modal data fusion and real-time analysis to realize real-time disease positioning.

[0092] The digital twin visualization platform is provided with an output interface, which includes a three-dimensional visualization interface, a disease distribution map, and real-time monitoring and alarm. The three-dimensional visualization interface takes a high-precision three-dimensional digital twin model as the core, updates disease detection data in real time, and marks disease positions through color or icons. Users can view disease details, compare historical data, and view bridge details from different angles through perspective control through interactive functions. The disease distribution map provides a two-dimensional plan view and a heat map to intuitively display disease distribution and severity, helping users quickly identify disease concentration areas. The real-time monitoring and alarm display data collected by the unmanned aerial vehicle in real time and issue an alarm prompt when serious diseases are detected.

[0093] It can be understood that the output interface is simple and easy to use, supports multi-device access, and allows users to view the bridge status anytime and anywhere. The disease detection report generated by the system provides an overview of diseases, a detailed disease list, disease pictures and videos, and classifies and quantitatively analyzes diseases to generate intuitive charts. The health assessment report generated by the system is based on disease detection results and historical data to generate bridge health assessment indicators, analyze health trends, and provide maintenance recommendations. By comparing disease data at different time periods, the system analyzes disease evolution trends and supports comparison analysis with similar bridges.

[0094] Those skilled in the art will readily understand that the above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A UAV bridge defect inspection method based on DeepSeek-YOLO target detection, characterized in that: The specific steps include: S100, planning the flight path of the UAV by controlling the detection angle and the overlap rate of the detection area; S200 uses GPS, BeiDou or atomic clocks to provide a unified time base for all sensors in the multi-sensor collaborative acquisition module, and synchronizes hardware trigger signals through FPGA to ensure time alignment; S300, the UAV flies according to the planned route, and the multi-sensor collaborative acquisition module collects data simultaneously, and marks and aligns the data through timestamps and spatial positioning information; S400, processing and fusing the data collected by the multi-sensor collaborative acquisition module, and constructing a bridge disease database, annotating the collected data and including it in the bridge disease database; S500, establishing a YOLO target detection model, optimizing the YOLO target detection model through a convolutional neural network combined with DeepSeek, and training the YOLO target detection model using existing data; S600: Using the YOLO target detection model to identify the fused data in the bridge defect database, and obtaining bridge defect information; S700, selecting multi-dimensional data from the data collected by the multi-sensor collaborative collection module, fusing the multi-dimensional data, and generating a three-dimensional model of the bridge; S800 integrates the bridge damage information identified by the YOLO object detection model into the 3D model and outputs the image on the digital twin visualization platform; S900 updates the 3D model in real time based on bridge damage information, compares the identified damage results with expert annotations, and feeds the feedback to the YOLO target detection model for further training.

2. The method for UAV bridge defect inspection based on DeepSeek-YOLO target detection according to claim 1 is characterized in that: In step S400, the data collected by the multi-sensor collaborative acquisition module is processed and then fused, specifically by combining visible light images, infrared images and lidar data to more comprehensively assess the health status of the bridge.

3. The method for UAV bridge defect inspection based on DeepSeek-YOLO target detection according to claim 2 is characterized in that: The fusion includes early fusion and late fusion; The early fusion refers to the fusion performed during the data acquisition or feature extraction stage, fusing the features of the optical image and the infrared image to generate a richer feature representation; The late fusion refers to the fusion performed at the decision-making stage, which comprehensively analyzes the detection results of different sensors to obtain the final disease judgment; During fusion, a multi-branch network structure is used to process data of different modalities separately, and then fuse them at the appropriate stage.

4. A method for inspecting bridge defects using a drone based on DeepSeek-YOLO target detection according to any one of claims 1 to 3, characterized in that: In step S500, when optimizing the YOLO target detection model by combining the convolutional neural network with DeepSeek, it is necessary to build a technical closed loop from three dimensions: model optimization, data enhancement, and data integration; When optimizing the model, an attention mechanism needs to be added to improve detection accuracy. Data enhancement requires the integration of DeepSeek's multimodal data processing capabilities. Data integration requires edge computing and API calls to guide the deployment of the YOLO object detection model in the actual environment, building a three-level architecture consisting of drones, edge servers, and the DeepSeek cloud. Finally, feedback data models need to be used for continuous learning and optimization.

5. The method for UAV bridge defect inspection based on DeepSeek-YOLO target detection according to claim 4 is characterized in that: When training the YOLO object detection model, distributed training and hybrid training are used, and a multi-data annotation system is adopted to develop a hybrid annotation tool chain. YOLO native rectangular boxes are used to annotate bridge defects such as cracks and spalling. The DeepSeek point cloud segmentation module is used to process lidar data, and the descriptions in the inspection report are automatically mapped to image labels.

6. The method for UAV bridge defect inspection based on DeepSeek-YOLO target detection according to claim 4 is characterized in that: To call an API, first register an account and log in to the console to obtain an API key starting with `sk-`. Then, consult the official documentation to confirm the target API's request endpoint, authentication method, model name, and parameter restrictions. Send JSON format data to the specified URL via HTTP POST request. The request header must include `Authorization:Bearer YOUR_API_KEY`. The content includes the conversation message and temperature value. After receiving the response, parse the JSON result and extract `choices[0].message.content`. At the same time, check the status code and handle errors. By streaming and adjusting the advanced parameters of `temperature` / `max_tokens`, you need to pay attention to permission billing, data security, and model token length limits when calling.

7. A method for UAV bridge defect inspection based on DeepSeek-YOLO target detection according to any one of claims 1-3, characterized in that: In step S600, the fused data is transmitted to a ground control station or a cloud server in real time via wireless communication. The operator views the data in the bridge disease database in real time and performs preliminary analysis. When cracks are identified, they are analyzed to determine whether they are transverse or longitudinal cracks, and their severity is assessed. Combined with the disease detection results and historical bridge data, the overall health of the bridge is evaluated and a health assessment report is generated. According to the severity and distribution of the damage, maintenance priority ranking is generated to provide scientific maintenance recommendations to the bridge maintenance department.

8. A method for UAV bridge defect inspection based on DeepSeek-YOLO target detection according to any one of claims 1-3, characterized in that: In step S700, a three-dimensional model of the bridge is generated, specifically by using point cloud data generated by the LiDAR, combined with real-time positioning and map building algorithms to generate a three-dimensional model of the bridge that realistically reflects the geometric shape, structural details, and spatial relationships of the bridge; The high-resolution images taken by the optical camera are mapped onto the three-dimensional model, giving the model a realistic appearance and texture, making the digital twin model more visually realistic.

9. A UAV bridge defect inspection system based on DeepSeek-YOLO target detection, used to implement the method according to any one of claims 1 to 8, characterized in that: include: The multi-sensor collaborative acquisition module, intelligent analysis module, and digital twin visualization platform work together through data flow to ensure the accuracy of the judgment results; The multi-sensor collaborative acquisition module is integrated into a drone, which is driven by the drone to inspect the bridge to obtain multiple types of bridge information at the same time; The intelligent analysis module processes the data collected by the multi-sensor collaborative acquisition module and conducts a comprehensive analysis of the processed detection results to obtain a final disease judgment; The digital twin visualization platform integrates the bridge information obtained by the drone and generates a three-dimensional model of the bridge, integrating the detected bridge defects with the three-dimensional model.

10. The UAV bridge defect inspection system based on DeepSeek-YOLO target detection according to claim 9 is characterized in that: The digital twin visualization platform is provided with an output interface, which includes a three-dimensional visualization interface, a disease distribution map, and real-time monitoring and alarm; The three-dimensional visualization interface is based on a high-precision three-dimensional digital twin model, which updates disease detection data in real time and marks the location of the disease with colors or icons; Users can use interactive functions to view damage details, compare historical data, and view bridge details from different angles through viewing angle control; The disease distribution map provides a two-dimensional plane map and a heat map to intuitively display the distribution and severity of the disease, helping users to quickly identify areas where the disease is concentrated; The real-time monitoring and alarm display drone collected data in real time, and issued an alarm prompt when a serious disease is detected.

Images