An ai-based road maintenance patrol disease analysis system and method

Abstract

The application discloses the technical field of road maintenance and artificial intelligence, in particular to a road maintenance patrol disease analysis system and method based on AI, first, the mobile inspection vehicle, unmanned aerial vehicle aerial photography and fixed sensor are used to realize multi-dimensional road data collection; then, the improved Transform-CNN hybrid model is combined with the road field knowledge graph to realize accurate identification; based on this, the multi-modal data collection mode of the mobile inspection vehicle collection, unmanned aerial vehicle aerial photography collection and fixed sensor real-time collection is integrated, multi-dimensional data such as road surface images, three-dimensional profiles, position attitudes, macro scenes and environmental parameters are synchronously acquired, and the collected data is pretreated by image denoising, point cloud deduplication and sensor data filtering, so that the advantages of comprehensive and effective acquisition of road data in complex environments and avoidance of data distortion or loss caused by single visible light collection are realized.

Description

Technical Field

[0001] This invention relates to the fields of road maintenance technology and artificial intelligence technology, specifically to an AI-based road maintenance inspection and disease analysis system and method. Background Technology

[0002] As a core component of transportation infrastructure, the quality of roads directly impacts traffic safety, transportation efficiency, and the public's travel experience. With the continuous growth of road mileage in my country (the total length of highways nationwide exceeded 5.5 million kilometers by the end of 2024), the traditional road maintenance model relying on manual inspections is no longer sufficient to meet the demands. Manual inspections are not only time-consuming and labor-intensive (a single inspection of a highway requires 2-3 people, covering less than 200 kilometers per day), but also suffer from problems such as strong subjectivity in defect identification, high rates of missed or incorrect diagnoses (the accuracy rate for identifying defects such as minor cracks and hidden spalling is less than 60%), and delayed maintenance decisions. In recent years, the rapid development of artificial intelligence (AI) technology in computer vision and data analysis has driven the transformation of road maintenance towards "intelligent inspection." Existing AI-based road defect analysis systems mostly rely on single cameras mounted on inspection vehicles to collect images and use pre-trained deep learning models (such as CNNs) to identify common defects such as cracks and potholes. However, such systems still have significant limitations in practical applications, making it difficult to adapt to complex and ever-changing road environments and the needs of refined maintenance. Specific problems are as follows: (i) Data collection is limited to a single dimension, resulting in poor data validity in complex environments. Existing systems mostly rely on visible light cameras mounted on inspection vehicles to collect road images, without integrating multiple types of perception data. This leads to data distortion or loss in special environments, affecting the accuracy of subsequent AI analysis.

[0003] (ii) Disease identification is limited to a single type, and the rate of missed diagnosis for compound diseases is high. Existing AI models are mostly trained for single defects (such as cracks, potholes, and spalling), and cannot identify complex scenarios with "multiple defects overlapping". They also do not associate with road structural attributes (such as surface material and base layer thickness), leading to missed or incorrect judgments.

[0004] (iii) Lack of ability to trace the source and predict the development of diseases, resulting in passive maintenance decisions. The existing system can only locate and mark diseases, but it cannot analyze the causes of diseases or predict the development trend of diseases. This results in the maintenance work being in a passive state of "solving problems after they are discovered", making it difficult to prevent the spread of diseases in advance.

[0005] Based on the above, an AI-based road maintenance inspection and disease analysis system and method are invented. Summary of the Invention

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: An AI-based road maintenance and inspection system for analyzing road defects includes: The multimodal data acquisition module is used to collect multi-dimensional road data by utilizing mobile inspection vehicles, drone aerial photography, and fixed sensors. The AI ​​intelligent analysis module is used to achieve accurate identification based on an improved Transformer-CNN hybrid model and combined with a road domain knowledge graph. First, the model is trained with a large number of single and complex disease samples, and a road structure attention mechanism and a multi-label classification head are added to simultaneously identify multiple diseases and their corresponding structural layer locations. Then, the knowledge graph is used to perform a second verification of the identification results to improve the identification accuracy to a practical level, providing reliable identification results for disease tracing and maintenance decisions. The disease source tracing and prediction module first uses graph neural networks to mine the heterogeneous graph relationships constructed from disease data and multimodal data to trace the causes of diseases; then it uses long short-term memory networks to combine historical disease data with future environmental data to predict the extent and risk level of disease expansion within 1-7 days. The maintenance decision push module is used to automatically generate personalized maintenance plans based on the types of defects, risk levels, and road grades identified by AI, combined with industry standards and historical maintenance case databases. It clarifies the process steps, material and equipment lists, and schedule requirements, and enables the plans to be accurately pushed to maintenance units through multiple channels. The multi-department collaborative scheduling module is used to first establish a mapping relationship between disease type and responsible department, so that collaborative tasks can be automatically pushed to relevant departments and response time limits can be set; then, the progress of each department is synchronized in real time through the collaborative task dashboard to avoid construction waiting; at the same time, it can identify and warn of task time and space conflicts, and assist departments in negotiating and adjusting. The data feedback optimization module is responsible for receiving maintenance effect data from maintenance units, marking the feedback data as training samples, updating the model parameters of the AI ​​intelligent analysis module, and improving recognition accuracy. At the same time, it will also calculate the cost-effectiveness of different maintenance solutions, optimize the solution generation rules of the maintenance decision push module, and prioritize recommending efficient processes.

[0007] As a preferred embodiment of the AI-based road maintenance and inspection defect analysis system described in this invention, the multimodal data acquisition module includes: The inspection vehicle's data acquisition unit is equipped with a high-definition industrial camera, LiDAR, and inertial measurement unit to simultaneously acquire road surface images, road surface 3D contour data, and inspection vehicle position and attitude data. The drone data acquisition unit is used to take aerial photos of areas that are difficult to cover by inspection vehicles using multi-rotor drones in order to obtain macroscopic images of the road surface and temperature field data. Fixed sensor units are used to deploy pressure sensors, humidity sensors, and displacement sensors at key locations on the road to detect road surface load, base layer moisture content, and subgrade settlement data. The processing unit is used to transmit the collected data to the cloud database through the 5G / BeiDou dual-mode communication module, and at the same time perform data preprocessing to ensure data validity.

[0008] As a preferred embodiment of the AI-based road maintenance and inspection defect analysis system described in this invention, the AI ​​intelligent analysis module includes: The model training optimization unit is used to train a combination of public road defect datasets and actual engineering defect data. It adds a road structure attention mechanism to the Transformer encoder layer to make the model prioritize structural regions, and adds a multi-label classification head to the CNN decoder layer to support the simultaneous output of multiple defect types and their corresponding structural layer positions. The domain knowledge graph fusion unit is used to construct a road defect knowledge graph, so that after the model identifies defects, the knowledge graph can be called for secondary verification.

[0009] As a preferred embodiment of the AI-based road maintenance and inspection disease analysis system described in this invention, the disease source tracing and prediction module includes: The disease tracing unit is used to construct a heterogeneous graph by combining multimodal data with disease data identified by AI, and analyzes the node relationships through graph neural networks to output the causes of diseases and generate a tracing report. The development prediction unit is used to predict the development trend of diseases by taking historical disease data and real-time environmental data as input and using a long short-term memory network model. It outputs the extent of disease expansion in the next 1-7 days and marks the risk level.

[0010] As a preferred embodiment of the AI-based road maintenance inspection and defect analysis system described in this invention, the maintenance decision push module includes: The scheme generation unit is used to generate schemes for different diseases and risk levels by combining highway maintenance technical specifications and historical maintenance case databases. The push unit is used to push solutions to maintenance units via system APP, SMS, and office OA system, including the location of the disease, disease details, recommended maintenance process, list of required materials and equipment, and estimated construction period.

[0011] As a preferred embodiment of the AI-based road maintenance and inspection defect analysis system described in this invention, the multi-department collaborative scheduling module includes: The departmental association mapping unit is used to preset the associated departments involved in road maintenance in order to establish a mapping relationship between disease type and responsible department; The collaborative task allocation unit is used to automatically push collaborative tasks to relevant departments after the maintenance plan is generated, and to set task response time limits; The real-time progress synchronization unit is used to establish a collaborative task dashboard, allowing each department to update task progress in real time and automatically synchronize progress with the maintenance unit, avoiding construction waiting. The conflict early warning unit is used to automatically identify and push conflict reminders when there are time and space conflicts in the tasks of multiple departments, and to assist departments in negotiating and adjusting the plan.

[0012] As a preferred embodiment of the AI-based road maintenance and inspection defect analysis system described in this invention, the data feedback optimization module includes: The model optimization unit is used to mark maintenance feedback data as correctly identified, missed, or misidentified samples, and retrain the Transformer-CNN model of the AI ​​intelligent analysis module. The model parameters are updated every quarter to continuously improve the recognition accuracy. The scheme optimization unit is used to statistically analyze the disease recurrence cycle of different maintenance schemes in order to optimize the scheme generation rules and prioritize the recommendation of cost-effective maintenance processes.

[0013] An AI-based method for analyzing road maintenance defects includes the following specific steps: S1: Utilize mobile inspection vehicles, drone aerial photography, and fixed sensors to achieve multi-dimensional road data collection; S2: Accurate identification is achieved by using an improved Transformer-CNN hybrid model combined with a road domain knowledge graph. First, the model is trained with a large number of single and composite disease samples, and a road structure attention mechanism and a multi-label classification head are added to simultaneously identify multiple diseases and their corresponding structural layer locations. Then, the knowledge graph is used to perform a second verification of the identification results to improve the identification accuracy to a practical level, providing reliable identification results for disease tracing and maintenance decisions. S3: First, use graph neural networks to mine the heterogeneous graph relationships constructed from disease data and multimodal data to trace the causes of diseases; then use long short-term memory networks to combine historical disease data and future environmental data to predict the extent and risk level of disease expansion within 1-7 days. S4: Based on the types of defects, risk levels, and road grades identified by AI, combined with industry standards and historical maintenance case databases, it automatically generates personalized maintenance plans, clearly defining process steps, material and equipment lists, and schedule requirements, and enables the plans to be accurately pushed to maintenance units through multiple channels. S5: First, establish a mapping relationship between disease type and responsible department so that collaborative tasks can be automatically pushed to relevant departments and response deadlines can be set; then, the progress of each department can be synchronized in real time through the collaborative task dashboard to avoid construction waiting; at the same time, it can identify and warn of time and space conflicts of tasks and assist departments in negotiating and adjusting. S6: First, it is responsible for receiving maintenance effect data from maintenance units. Then, it marks the feedback data as training samples, updates the model parameters of the AI ​​intelligent analysis module, and improves the recognition accuracy. At the same time, it will also calculate the cost-effectiveness of different maintenance solutions, optimize the solution generation rules of the maintenance decision push module, and prioritize recommending efficient processes.

[0014] Compared with existing technologies: 1. By integrating multimodal data acquisition methods such as mobile data acquisition by inspection vehicles, aerial data acquisition by drones, and real-time data acquisition by fixed sensors, multi-dimensional data such as road surface images, 3D contours, position and attitude, macroscopic scene and environmental parameters are acquired simultaneously. The acquired data is preprocessed by image denoising, point cloud deduplication, and sensor data filtering. It has the advantages of achieving comprehensive and effective acquisition of road data in complex environments and avoiding data distortion or loss caused by single visible light acquisition.

[0015] 2. By adopting an improved Transformer-CNN hybrid model, a road structure attention mechanism is added to the model structure to focus on key structural regions. A multi-label classification head is added to support the simultaneous identification of multiple disease types. At the same time, a road domain knowledge graph containing disease types, road attributes, and environmental factors is integrated to perform secondary verification of the identification results. This has the advantages of achieving accurate identification of single diseases and composite diseases with multiple diseases superimposed, and avoiding missed or wrong judgments caused by model training limitations or lack of association with road attributes.

[0016] 3. By constructing heterogeneous graphs based on graph neural networks and multi-source data such as underground pipelines, meteorology, and traffic volume, the system mines data correlations to analyze the causes of diseases. Furthermore, by combining long short-term memory networks with historical disease data and future environmental data, it predicts disease development trends and labels risk levels. This approach has the advantages of enabling clear tracing of disease causes and early prediction of disease development trends, providing a scientific basis for maintenance work, and promoting a shift in maintenance decision-making from a passive mode of discovering and solving problems to proactive prevention and planning. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall framework of the present invention; Figure 2 This is a schematic diagram of the multimodal data acquisition module framework of the present invention; Figure 3 This is a schematic diagram of the AI ​​intelligent analysis module framework of the present invention; Figure 4This is a schematic diagram of the disease source tracing and prediction module framework of the present invention; Figure 5 This is a schematic diagram of the maintenance decision push module framework of the present invention; Figure 6 This is a schematic diagram of the multi-department collaborative scheduling module framework of the present invention; Figure 7 This is a schematic diagram of the data feedback optimization module framework of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0019] This invention provides an AI-based road maintenance inspection and defect analysis system. Please refer to [link / reference]. Figures 1-7 ,include: The multimodal data acquisition module is used to collect multi-dimensional road data by utilizing mobile inspection vehicles, drone aerial photography, and fixed sensors. The AI ​​intelligent analysis module is used to achieve accurate identification based on an improved Transformer-CNN hybrid model and combined with a road domain knowledge graph. First, the model is trained with a large number of single and complex disease samples, and a road structure attention mechanism and a multi-label classification head are added to simultaneously identify multiple diseases and their corresponding structural layer locations. Then, the knowledge graph is used to perform a second verification of the identification results to improve the identification accuracy to a practical level, providing reliable identification results for disease tracing and maintenance decisions. The disease source tracing and prediction module first uses graph neural networks to mine the heterogeneous graph relationships constructed from disease data and multimodal data to trace the causes of diseases; then it uses long short-term memory networks to combine historical disease data with future environmental data to predict the extent and risk level of disease expansion within 1-7 days. The maintenance decision push module is used to automatically generate personalized maintenance plans based on the types of defects, risk levels, and road grades identified by AI, combined with industry standards and historical maintenance case databases. It clarifies the process steps, material and equipment lists, and schedule requirements, and enables the plans to be accurately pushed to maintenance units through multiple channels. The multi-department collaborative scheduling module is used to first establish a mapping relationship between disease type and responsible department, so that collaborative tasks can be automatically pushed to relevant departments and response time limits can be set; then, the progress of each department is synchronized in real time through the collaborative task dashboard to avoid construction waiting; at the same time, it can identify and warn of task time and space conflicts, and assist departments in negotiating and adjusting. The data feedback optimization module is responsible for receiving maintenance effect data from maintenance units, marking the feedback data as training samples, updating the model parameters of the AI ​​intelligent analysis module, and improving recognition accuracy. At the same time, it will also calculate the cost-effectiveness of different maintenance solutions, optimize the solution generation rules of the maintenance decision push module, and prioritize recommending efficient processes.

[0020] The multimodal data acquisition module includes: The inspection vehicle's data acquisition unit is equipped with a high-definition industrial camera (4K resolution, 30fps frame rate, with anti-reflective filter and automatic fill light function), a lidar (range accuracy ±2mm, scanning frequency 100Hz), and an inertial measurement unit (IMU, sampling rate 1000Hz) to simultaneously acquire road surface images, road surface 3D contour data (such as pothole depth and crack width), and inspection vehicle position and attitude data. The drone data acquisition unit is used to take aerial photos of areas that are difficult to cover by inspection vehicles, such as bridges, steep slopes, and tunnel entrances and exits, using multi-rotor drones (with a flight time of 40 minutes and equipped with a 50-megapixel camera and thermal imager) to obtain macroscopic images of the road surface and temperature field data (such as road icing warnings and thermal anomalies in permeable areas of the base layer). Fixed sensor units are used to deploy pressure sensors, humidity sensors, and displacement sensors at key locations on roads (such as sections with heavy vehicle traffic or sections where underground pipelines cross) to detect road surface load, base layer moisture content, and subgrade settlement data. The processing unit is used to transmit the collected data to the cloud database through the 5G / BeiDou dual-mode communication module, and at the same time perform data preprocessing (such as image denoising, lidar point cloud deduplication, and sensor data filtering) to ensure data validity.

[0021] The AI ​​intelligent analysis module includes: The model training optimization unit uses publicly available road defect datasets (such as Crack500 and Pothole600) and composite defect data collected from actual engineering projects (such as 200,000 sets of samples of "cracks + spalling" and "potholes + settlement") as training sets. A road structure attention mechanism is added to the Transformer encoder layer to make the model prioritize structural regions such as the surface layer and the base layer. A multi-label classification head is added to the CNN decoder layer to support the simultaneous output of multiple defect types and their corresponding structural layer locations (such as "surface spalling + base layer cracks"). The domain knowledge graph fusion unit is used to construct a road defect knowledge graph, which includes entities and relationships such as defect types (15 common defects + 8 complex defects), road attributes (surface material, base layer thickness, design load), and environmental factors (rainfall, temperature, traffic volume). After the model identifies defects, the knowledge graph is called for secondary verification (e.g., "cracks that appear on asphalt pavement in winter are preferentially associated with the 'low temperature shrinkage' attribute"), thereby improving the recognition accuracy.

[0022] The disease source tracing and prediction module includes: The disease tracing unit is used to construct a heterogeneous graph by combining multimodal data (such as underground pipeline data, meteorological data, and traffic volume data) with disease data identified by AI. It also analyzes the node correlation relationship (such as the correlation strength of "pothole location → underground pipeline leakage point → rainfall") through graph neural network, outputs the cause of the disease (such as "rainwater leakage leads to softening of the base layer → heavy vehicle rolling forms potholes"), and generates a tracing report. The development prediction unit is used to predict the development trend of diseases by taking historical disease data (the changes in diseases of the same road section in the past 3 years) and real-time environmental data (rainfall and temperature in the next 7 days) as input, and outputting the extent of disease expansion in the next 1-7 days (e.g., "the current 0.5㎡ peeling area will expand to 1.8㎡ in 7 days, and the depth will increase by 3mm"), and marking the risk level (high / medium / low, high risk means that a safety accident may occur within 7 days).

[0023] The maintenance decision push module includes: The scheme generation unit is used to combine the "Technical Specifications for Highway Maintenance" (JTGH10-2009) with the historical maintenance case library to generate schemes for different defects and risk levels (e.g., "High-risk potholes (depth > 10cm): adopt the process of 'cutting → cleaning the bottom → spraying tack coat → asphalt concrete repair', which is recommended to be completed within 24 hours; medium-risk cracks (width 3-5mm): use crack sealant to fill the cracks, which is recommended to be completed within 7 days"). The push unit is used to push solutions to maintenance units via system APP, SMS, and office OA system, including the location of the disease (accurate to latitude and longitude ±1m), disease details (type, size, risk level), recommended maintenance process, list of required materials and equipment, and estimated construction period.

[0024] The multi-department collaborative scheduling module includes: The departmental association mapping unit is used to pre-set the associated departments involved in road maintenance (such as traffic management departments, municipal pipeline departments, and emergency rescue departments) to establish a mapping relationship between the type of road damage and the responsible department (e.g., "road damage caused by underground pipeline leakage" is automatically associated with the municipal pipeline department, and "high-risk road damage causing traffic congestion" is automatically associated with the traffic management department). The collaborative task allocation unit is used to automatically push collaborative tasks to relevant departments after the maintenance plan is generated (such as pushing a road closure application for "K150+200 section to be closed for construction within 48 hours" to the traffic management department, and pushing a cooperation request for "K150+200 underground sewage pipe leakage repair" to the municipal department), and setting task response time limits (response within 1 hour for emergency tasks, and within 4 hours for routine tasks). The real-time progress synchronization unit is used to establish a collaborative task dashboard, where each department can update the task progress in real time (such as "Road closure notice has been issued" and "Pipeline repair has been completed"), and will automatically synchronize the progress with the maintenance unit to avoid construction waiting. The conflict early warning unit is used to automatically identify and push conflict reminders when there are time and space conflicts in the tasks of multiple departments (such as overlapping road section maintenance and pipeline repair plans), and assist departments in negotiating and adjusting the plan.

[0025] The data feedback optimization module includes: The model optimization unit is used to mark maintenance feedback data as correctly identified, missed, or misidentified samples, and retrain the Transformer-CNN model of the AI ​​intelligent analysis module. The model parameters are updated every quarter to continuously improve the recognition accuracy. The scheme optimization unit is used to statistically analyze the disease recurrence cycle of different maintenance schemes (e.g., the average recurrence cycle of pothole repair by process A is 6 months, while that of process B is 12 months) in order to optimize the scheme generation rules and prioritize the recommendation of cost-effective maintenance processes.

[0026] An AI-based method for analyzing road maintenance defects includes the following specific steps: S1: Utilize mobile inspection vehicles, drone aerial photography, and fixed sensors to achieve multi-dimensional road data collection; S2: Accurate identification is achieved by using an improved Transformer-CNN hybrid model combined with a road domain knowledge graph. First, the model is trained with a large number of single and composite disease samples, and a road structure attention mechanism and a multi-label classification head are added to simultaneously identify multiple diseases and their corresponding structural layer locations. Then, the knowledge graph is used to perform a second verification of the identification results to improve the identification accuracy to a practical level, providing reliable identification results for disease tracing and maintenance decisions. S3: First, use graph neural networks to mine the heterogeneous graph relationships constructed from disease data and multimodal data to trace the causes of diseases; then use long short-term memory networks to combine historical disease data and future environmental data to predict the extent and risk level of disease expansion within 1-7 days. S4: Based on the types of defects, risk levels, and road grades identified by AI, combined with industry standards and historical maintenance case databases, it automatically generates personalized maintenance plans, clearly defining process steps, material and equipment lists, and schedule requirements, and enables the plans to be accurately pushed to maintenance units through multiple channels. S5: First, establish a mapping relationship between disease type and responsible department so that collaborative tasks can be automatically pushed to relevant departments and response time limits can be set; then, the progress of each department can be synchronized in real time through the collaborative task dashboard to avoid construction waiting; at the same time, it can identify and warn of task time and space conflicts, and assist departments in negotiating and adjusting. S6: First, it is responsible for receiving maintenance effect data from maintenance units. Then, it marks the feedback data as training samples, updates the model parameters of the AI ​​intelligent analysis module, and improves the recognition accuracy. At the same time, it will also calculate the cost-effectiveness of different maintenance solutions, optimize the solution generation rules of the maintenance decision push module, and prioritize recommending efficient processes.

[0027] Although the present invention has been described above with reference to embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, as long as there is no structural conflict, the features in the disclosed embodiments can be combined with each other in any manner. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. An AI-based road maintenance and inspection disease analysis system, characterized in that, include: The multimodal data acquisition module is used to collect multi-dimensional road data by utilizing mobile inspection vehicles, drone aerial photography, and fixed sensors. The AI ​​intelligent analysis module is used to achieve accurate identification based on an improved Transformer-CNN hybrid model and combined with a road domain knowledge graph. First, the model is trained with a large number of single and complex disease samples, and a road structure attention mechanism and a multi-label classification head are added to simultaneously identify multiple diseases and their corresponding structural layer locations. Then, the knowledge graph is used to perform a second verification of the identification results to improve the identification accuracy to a practical level, providing reliable identification results for disease tracing and maintenance decisions. The disease source tracing and prediction module first uses graph neural networks to mine the heterogeneous graph relationships constructed from disease data and multimodal data to trace the causes of diseases; then it uses long short-term memory networks to combine historical disease data with future environmental data to predict the extent and risk level of disease expansion within 1-7 days. The maintenance decision push module is used to automatically generate personalized maintenance plans based on the types of defects, risk levels, and road grades identified by AI, combined with industry standards and historical maintenance case databases. It clarifies the process steps, material and equipment lists, and schedule requirements, and enables the plans to be accurately pushed to maintenance units through multiple channels. The multi-department collaborative scheduling module is used to first establish a mapping relationship between disease type and responsible department, so that collaborative tasks can be automatically pushed to relevant departments and response time limits can be set; then, the progress of each department is synchronized in real time through the collaborative task dashboard to avoid construction waiting; at the same time, it can identify and warn of task time and space conflicts, and assist departments in negotiating and adjusting. The data feedback optimization module is responsible for receiving maintenance effect data from maintenance units, marking the feedback data as training samples, updating the model parameters of the AI ​​intelligent analysis module, and improving recognition accuracy. At the same time, it will also calculate the cost-effectiveness of different maintenance solutions, optimize the solution generation rules of the maintenance decision push module, and prioritize recommending efficient processes.

2. The road maintenance and inspection defect analysis system based on AI according to claim 1, characterized in that, The multimodal data acquisition module includes: The inspection vehicle's data acquisition unit is equipped with a high-definition industrial camera, LiDAR, and inertial measurement unit to simultaneously acquire road surface images, road surface 3D contour data, and inspection vehicle position and attitude data. The drone data acquisition unit is used to take aerial photos of areas that are difficult to cover by inspection vehicles using multi-rotor drones in order to obtain macroscopic images of the road surface and temperature field data. Fixed sensor units are used to deploy pressure sensors, humidity sensors, and displacement sensors at key locations on the road to detect road surface load, base layer moisture content, and subgrade settlement data. The processing unit is used to transmit the collected data to the cloud database through the 5G / BeiDou dual-mode communication module, and at the same time perform data preprocessing to ensure data validity.

3. The road maintenance and inspection defect analysis system based on AI according to claim 1, characterized in that, The AI ​​intelligent analysis module includes: The model training optimization unit is used to train a combination of public road defect datasets and actual engineering defect data. It adds a road structure attention mechanism to the Transformer encoder layer to make the model prioritize structural regions, and adds a multi-label classification head to the CNN decoder layer to support the simultaneous output of multiple defect types and their corresponding structural layer positions. The domain knowledge graph fusion unit is used to construct a road defect knowledge graph, so that after the model identifies defects, the knowledge graph can be called for secondary verification.

4. The road maintenance and inspection defect analysis system based on AI according to claim 1, characterized in that, The disease source tracing and prediction module includes: The disease tracing unit is used to construct a heterogeneous graph by combining multimodal data with disease data identified by AI, and analyzes the node relationships through graph neural networks to output the causes of diseases and generate a tracing report. The development prediction unit is used to predict the development trend of diseases by taking historical disease data and real-time environmental data as input and using a long short-term memory network model. It outputs the extent of disease expansion in the next 1-7 days and marks the risk level.

5. The road maintenance and inspection defect analysis system based on AI according to claim 1, characterized in that, The maintenance decision push module includes: The scheme generation unit is used to generate schemes for different diseases and risk levels by combining highway maintenance technical specifications and historical maintenance case databases. The push unit is used to push solutions to maintenance units via system APP, SMS, and office OA system, including the location of the disease, disease details, recommended maintenance process, list of required materials and equipment, and estimated construction period.

6. The road maintenance and inspection defect analysis system based on AI according to claim 1, characterized in that, The multi-department collaborative scheduling module includes: The departmental association mapping unit is used to preset the associated departments involved in road maintenance in order to establish a mapping relationship between disease type and responsible department; The collaborative task allocation unit is used to automatically push collaborative tasks to relevant departments after the maintenance plan is generated, and to set task response time limits; The real-time progress synchronization unit is used to establish a collaborative task dashboard, allowing each department to update task progress in real time and automatically synchronize progress with the maintenance unit, avoiding construction waiting. The conflict early warning unit is used to automatically identify and push conflict reminders when there are time and space conflicts in the tasks of multiple departments, and to assist departments in negotiating and adjusting the plan.

7. The road maintenance and inspection defect analysis system based on AI according to claim 1, characterized in that, The data feedback optimization module includes: The model optimization unit is used to mark maintenance feedback data as correctly identified, missed, or misidentified samples, and retrain the Transformer-CNN model of the AI ​​intelligent analysis module. The model parameters are updated every quarter to continuously improve the recognition accuracy. The scheme optimization unit is used to statistically analyze the disease recurrence cycle of different maintenance schemes in order to optimize the scheme generation rules and prioritize the recommendation of cost-effective maintenance processes.

8. A road maintenance inspection and defect analysis method based on AI, characterized in that, The specific steps are as follows: S1: Utilize mobile inspection vehicles, drone aerial photography, and fixed sensors to achieve multi-dimensional road data collection; S2: Accurate identification is achieved by using an improved Transformer-CNN hybrid model combined with a road domain knowledge graph. First, the model is trained with a large number of single and composite disease samples, and a road structure attention mechanism and a multi-label classification head are added to simultaneously identify multiple diseases and their corresponding structural layer locations. Then, the knowledge graph is used to perform a second verification of the identification results to improve the identification accuracy to a practical level, providing reliable identification results for disease tracing and maintenance decisions. S3: First, use graph neural networks to mine the heterogeneous graph relationships constructed from disease data and multimodal data to trace the causes of diseases; then use long short-term memory networks to combine historical disease data and future environmental data to predict the extent and risk level of disease expansion within 1-7 days. S4: Based on the types of defects, risk levels, and road grades identified by AI, combined with industry standards and historical maintenance case databases, it automatically generates personalized maintenance plans, clearly defining process steps, material and equipment lists, and schedule requirements, and enables the plans to be accurately pushed to maintenance units through multiple channels. S5: First, establish a mapping relationship between disease type and responsible department so that collaborative tasks can be automatically pushed to relevant departments and response time limits can be set; then, the progress of each department can be synchronized in real time through the collaborative task dashboard to avoid construction waiting; at the same time, it can identify and warn of task time and space conflicts, and assist departments in negotiating and adjusting. S6: First, it is responsible for receiving maintenance effect data from maintenance units. Then, it marks the feedback data as training samples, updates the model parameters of the AI ​​intelligent analysis module, and improves the recognition accuracy. At the same time, it will also calculate the cost-effectiveness of different maintenance solutions, optimize the solution generation rules of the maintenance decision push module, and prioritize recommending efficient processes.

Images