Road construction safety inspection method and system based on unmanned aerial vehicle field

Abstract

The invention provides a highway construction safety inspection method and system based on an unmanned aerial vehicle field, and the method comprises the following steps: collecting the data of a highway construction scene through a collection terminal, and carrying out the preprocessing of the collected real-time data; inputting the preprocessed data into a path planning module, calculating the patrol path weight of the unmanned aerial vehicle, and generating an optimal patrol path; the data is input into an anomaly detection module for edge calculation, and when abnormal information is detected, the abnormal information is fed back to a path planning module to implement path adjustment; and inputting abnormal information into a cloud image processing model, carrying out mutual learning through a knowledge distillation mechanism, training a cloud teacher model by utilizing incremental data, and deploying a trained student model in an anomaly detection module. According to the method, complex environment changes are coped with through reinforcement learning and a self-evaluation algorithm, and efficient detection is realized through an incremental framework and knowledge distillation balance precision and speed; the whole scene is covered through an air-fixed point-moving three-in-one mode, and multi-mode cooperation is achieved.

Description

Technical Field

[0001] The present invention relates to the field of unmanned aerial vehicle inspection, and in particular to a highway construction safety inspection method and system based on an unmanned aerial vehicle airport. Background Art

[0002] The current highway construction safety inspections mainly have the following technical problems: 1. Low inspection efficiency: Traditional manual inspections have limited coverage, and the preset routes of drones lack the ability to dynamically adjust, resulting in risks of repeated inspections and blind spots; 2. Insufficient anomaly detection accuracy: Conventional image recognition algorithms are sensitive to complex lighting and occlusions in construction scenes, with a false alarm rate of more than 30%; 3. Weak model generalization ability: Fixed training sets cannot adapt to morphological changes in different construction stages (such as roadbed paving and asphalt paving); 4. Delayed emergency response: Traditional edge computing devices have limited computing power, and sudden safety hazards cannot be warned in seconds.

[0003] Existing construction inspection systems usually use fixed-route drone inspections, which do not take into account the dynamic changes in the construction area. Although some commercial systems have path planning functions, they use the A* static algorithm and cannot cope with emergencies such as temporary roadblocks. In addition, the detection model relies on offline model updates and cannot achieve real-time incremental learning. Summary of the invention

[0004] In view of the problems of the existing construction safety inspection system lacking dynamic inspection capabilities, insufficient anomaly detection accuracy and system delay, the present invention proposes a highway construction safety inspection method and system based on an unmanned aerial vehicle airport to solve the problems raised in the above background technology. The present invention provides the following technical solutions: In a first aspect, a highway construction safety inspection method based on a drone airport comprises the following steps: Step 1: collect real-time data of the highway construction scene through a collection terminal, and pre-process the collected real-time data; Step 2: Input the preprocessed real-time data into the path planning module, calculate the weight of the drone inspection path through the online DQN network, and generate the optimal inspection path; Step 3: Input the pre-processed real-time data into the anomaly detection module for edge computing. When abnormal information is detected, it is fed back to the path planning module to implement path adjustment. Step 4: Input the abnormal information as incremental data into the cloud image processing model. The cloud image processing model and the anomaly detection module learn from each other through the knowledge distillation mechanism. The cloud teacher model is trained using the incremental data, and the trained lightweight student model is deployed in the anomaly detection module.

[0005] Preferably, step 2 further includes: collecting historical inspection path data of the drone, using an offline PPO model to train inspection strategies for typical construction scenarios, and forming a historical inspection database.

[0006] Preferably, in step 2, the formula for calculating the weight of the drone inspection path is: ; The risk level is determined by the hazardous factors in the construction area, and is judged by real-time data, historical inspection data and construction area type annotations. The judgment factors include geological risk, equipment density and personnel activity frequency. The risk level is proportional to the path weight; the coverage rate is the area that the current path can cover. The grid is divided by the sensor data of the drone, and the ratio of the number of grids covered by the path to the total number of grids is calculated. The coverage rate is proportional to the path weight; the energy consumption is the energy consumption of the drone flight, which is related to the flight parameters, and the energy consumption is inversely proportional to the path weight; the above three indicators are scaled to the [0,1] interval through the Max-Min normalization method; a、b、c is a coefficient determined according to the historical inspection database and used to measure the importance of each factor.

[0007] Preferably, the path planning module in step 2 is trained through a reinforcement learning model, the drone receives external state information obtained by the sensor, executes corresponding actions through the control system, and iterative training is performed by observing the new state and reward after the action is executed. The flight strategy gradually converges to the direction of maximizing the cumulative reward, so that the drone achieves the above-mentioned optimal path.

[0008] Preferably, the formula of the reward function is: ,in a、b、c is the weight coefficient, which is determined by the historical inspection database. The coverage rate and risk level are calculated through sensor data. The proximity to the dangerous area refers to the situation of approaching the high-risk area. The energy consumption is calculated through flight parameters.

[0009] Preferably, the anomaly detection module obtains an anomaly score S through an autoencoder and an isolation forest algorithm to detect abnormal information:

[0010] The autoencoder reconstruction error It is the difference between the input data and the reconstructed output obtained after encoding and decoding by the autoencoder. The calculation formula steps are as follows: The construction area images collected by drones x Input pre-trained autoencoder, the encoder part compresses the image into a low-dimensional latent feature vector , the decoder part restores z to the reconstructed image , for n samples, the mean square error is used to measure the difference between the original image and the reconstructed image:

[0011] Using the mean and standard deviation of historical normal data, the error is mapped to the [0,1] interval:

[0012] Among them, the mean m AE and standard deviation s AE Statistics of normal samples from the training phase; Isolation Forest Anomaly Score Indicates the degree of abnormality of the data point. The calculation formula steps are as follows: Input the latent feature vector z extracted by the autoencoder, randomly select features and segmentation values ​​to recursively divide the data to build an isolated tree until the vector z is isolated to the leaf node. For the vector z, calculate its path length h(z) from the root node to the leaf node in each isolated tree, take the average value E(h(z)) of multiple trees, and calculate the anomaly score:

[0013] in c ( n ) is the normalization factor, which is related to the number of samples n:

[0014] The Euler constant is ≈ 0.5772; When the anomaly score S is less than the threshold T, the anomaly detection module triggers an anomaly information alarm and takes photos for evidence collection; the threshold T is obtained by training historical anomaly data.

[0015] Preferably, the lightweight student model used in the knowledge distillation mechanism is MobileNetV3, and the teacher model is ResNet50.

[0016] In the second aspect, a highway construction safety inspection system based on an unmanned aerial vehicle airport is used to implement the steps of the above method. The system consists of three levels: data acquisition layer, edge computing layer and cloud platform layer; the data acquisition layer collects data through the acquisition terminal and performs preprocessing; the edge computing layer includes a path planning module for generating inspection paths, an anomaly detection module for identifying abnormal information, and an edge-cloud collaborative computing module for allocating computing resources and optimizing data pipelines; the cloud platform layer is composed of a cloud-based GPU cluster, generates a lightweight model through knowledge distillation, and updates the model of the edge computing layer.

[0017] Preferably, the collection terminal includes an aerial inspection module, a fixed-point monitoring module and a mobile evidence collection module.

[0018] Preferably, the computing resource allocation strategy of the edge-cloud collaborative computing module is: real-time obstacle avoidance tasks are allocated to the onboard FPGA, preliminary anomaly identification is allocated to the edge server, and model incremental training is allocated to the cloud GPU cluster.

[0019] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: by adopting a dynamic path planning module, the repeated scanning area is reduced by 42% compared with the traditional method under the same working conditions, and the reinforcement learning and self-evaluation algorithms can cope with complex environmental changes, thereby improving the efficiency of drone path planning; by adopting an incremental anomaly detection module, the incremental framework and knowledge distillation balance accuracy and speed, improve anomaly detection accuracy and model update efficiency, and incremental learning enables the accuracy of new category recognition to reach 80% within 200 samples; by adopting an edge-cloud collaborative computing module, the system response delay is reduced, and the average time from anomaly recognition to alarm issuance is 1.2s; the aerial-fixed-point-mobile trinity mode is adopted to cover the entire scene and realize multimodal collaboration. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings: Figure 1 It is a schematic diagram of the structure of the system of the present invention. DETAILED DESCRIPTION

[0021] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0022] The present invention provides a technical solution: Embodiment 1: A highway construction safety inspection method based on a drone airport comprises the following steps: Step 1: collect real-time data of highway construction scenes through a collection terminal and pre-process the collected real-time data.

[0023] Step 2: Collect historical inspection path data of drones, use offline PPO model to train inspection strategies for typical construction scenarios, and form a historical inspection database. Input preprocessed real-time data into the path planning module. Real-time data includes lidar point cloud data and centimeter-level RTK positioning data. Calculate the weight of drone inspection path through online DQN network to generate the optimal inspection path.

[0024] The formula for calculating the weight of the drone inspection path is: ; The risk level is the degree of danger in the construction area, which is used to measure the priority of path planning. The risk level is determined by the dangerous factors in the construction area (such as landslide area, construction vehicle density, personnel activity frequency, etc.). The random forest classifier is trained through historical inspection data to divide the construction area into high-risk, medium-risk and low-risk areas. The risk level is proportional to the path weight. The specific calculation steps are as follows: The following data are input: real-time sensor data (such as vibration sensors and infrared cameras to monitor landslide risks), historical accident databases (such as landslide records and equipment failure frequencies), and construction area type labels (such as high-risk landslide areas, medium-risk construction areas, and low-risk buffer zones).

[0025] By multi-factor weighting:

[0026] Among them, geological risk is to analyze slope stability through lidar point cloud or use historical landslide probability model; equipment density is the number of construction machinery identified by real-time cameras, and the higher the density, the greater the risk; personnel activity frequency is the personnel distribution statistics through thermal imaging or motion detection; weight Determined based on regression analysis of historical accident data.

[0027] Coverage is the area that the current path can cover. It is used to measure the efficiency of path planning. It is obtained by dividing the area grid through the sensor data of the drone (such as lidar point cloud data) and counting the number of covered grid cells. The coverage is proportional to the path weight. The specific calculation steps are as follows: Divide the construction area into N×M grids (e.g. 1m×1m), mark the grids to be inspected, and after the drone plans the path, calculate the number of grids covered by the path , total number of grids to be inspected ,

[0028] Energy consumption is the energy consumption of the UAV flight, which is related to the flight parameters (wind speed, flight speed and remaining power). Energy consumption is inversely proportional to the path weight. The specific calculation steps are as follows: According to the UAV power model, the basic energy consumption is proportional to the flight time t:

[0029] in is the hover or cruise speed; Energy consumption increases when flying against the wind: ,in is the real-time wind speed, and k is the drag coefficient (determined by the aerodynamic characteristics of the UAV).

[0030] The above data is scaled to the [0,1] interval using the Max-Min normalization method:

[0031] a、b、c is a coefficient, which is determined according to the historical inspection database and is used to measure the importance of each factor. The path planning module will select paths with higher coverage and lower energy consumption for inspection based on real-time data and historical data, reduce repeated scanning, and improve inspection efficiency; at the same time, this non-uniform sampling path optimization can dynamically adjust the sampling density based on risk classification (such as high-risk landslide areas and medium-risk construction areas), making inspections in high-risk areas more frequent and facilitating timely discovery of safety hazards.

[0032] The path planning module is trained through a reinforcement learning model to improve the generalization ability of the module. The reinforcement learning model includes state information, action information and reward function. Specifically: Build a deep neural network and initialize the experience replay pool to store historical interaction data (s t , a t , R t ,s t+1 ), where s t is the current state, a t is the current action, R t is the current reward, s t+1 The new state after executing the action.

[0033] The drone selects action a according to the current strategy t , observe the new state s after execution t+1 and reward R t ; Store the interaction data in the experience replay pool, sample batch data from the replay pool, and calculate the policy gradient (PPO); Update the network parameters to minimize the loss function. Through iterative updates, the strategy gradually converges to the direction of maximizing the cumulative reward.

[0034] The status information is obtained in real time through the drone's sensors (such as GPS, anemometer). The status information includes the three-dimensional coordinates of the area (i.e. the current location coordinates of the drone), the remaining power of the drone (i.e. the current battery power percentage) and the real-time wind speed vector (i.e. the wind speed and direction of the current environment). The status information is used as the state input of the reinforcement learning model.

[0035] Action information refers to the operations that the drone can perform during the inspection process, including the drone's heading angle adjustment, flight speed, and shooting angle. The action information is executed through the drone's control system.

[0036] The formula for the reward function is: ,in a、b、c is the weight coefficient, which is determined according to the historical inspection database. The coverage rate and the proximity to the dangerous area are calculated through sensor data. The proximity to the dangerous area refers to the situation of approaching the high-risk area. The energy consumption is calculated through flight parameters.

[0037] During the training process, the drone receives external state information obtained by the sensor, executes corresponding actions through the control system, and performs iterative training by observing the new state and reward after executing the action, so that the drone can achieve the above-mentioned optimal path (a path with higher coverage, higher risk level, and lower energy consumption). Under the same working conditions, this embodiment reduces the repeated scanning area by 42% compared with the traditional method, thereby improving the path planning efficiency.

[0038] Step 3: Input the preprocessed real-time data into the anomaly detection module for edge computing. When abnormal information is detected, it is fed back to the path planning module to implement path adjustment.

[0039] The anomaly detection module uses the autoencoder and isolation forest algorithm to obtain the anomaly score S to detect abnormal information: the image data of the construction area collected by the drone is input into the autoencoder, the feature code is extracted, and the anomaly score is performed using the following formula:

[0040] The autoencoder reconstruction error It is the difference between the input data and the reconstructed output obtained after encoding and decoding by the autoencoder. It is used to measure the accuracy of the autoencoder in representing the input data. The calculation formula steps are as follows: Input the construction area image (RGB or infrared image) collected by the drone, and use the pre-trained autoencoder encoder part (such as MobileNetV3) to compress the image into a low-dimensional latent feature vector , the decoder restores z to the reconstructed image , for n samples, the mean square error is used to measure the difference between the original image and the reconstructed image:

[0041] Using the mean and standard deviation of historical normal data, the error is mapped to the [0,1] interval:

[0042] Among them, the mean m AE and standard deviation s AE Statistics of normal samples from the training phase; Isolation Forest is a tree-based anomaly detection algorithm with anomaly score Indicates the degree of abnormality of a data point. Isolation Forest constructs multiple isolated trees and calculates the path length of each data point in each tree. The shorter the path length, the easier it is for the data point to be isolated and the more likely it is to be abnormal. The calculation formula steps are as follows: Input the latent feature vector z extracted by the autoencoder, randomly select features and segmentation values ​​to recursively divide the data to build an isolated tree until the sample is isolated to a leaf node. For sample z, calculate the path length h(z) from the root node to the leaf node in each isolated tree, take the average value E(h(z)) of multiple trees, and calculate the anomaly score:

[0043] in c ( n ) is the normalization factor, which is related to the number of samples n:

[0044] The Euler constant is ≈ 0.5772; When the anomaly score S is less than the threshold T, the anomaly detection module triggers an anomaly information alarm and takes photos for evidence collection; the threshold T is obtained by training historical anomaly data.

[0045] Step 4: Input the abnormal information as incremental data into the cloud image processing model. The cloud image processing model and the anomaly detection module learn from each other through the knowledge distillation mechanism. The cloud teacher model is trained using the incremental data, and the trained lightweight student model is deployed in the anomaly detection module.

[0046] The lightweight student model used in the knowledge distillation mechanism is MobileNetV3, and the teacher model is ResNet50.

[0047] After incremental learning, the system's anomaly detection accuracy mAP@0.5 reached 91.7%, an increase of 23.6% compared to the baseline YOLOv5, and the accuracy of new category recognition exceeded 80% within 200 samples.

[0048] Example 2: A highway construction safety inspection system based on a drone airport, such as Figure 1 As shown in the figure, the system consists of three levels: data collection layer, edge computing layer and cloud platform layer; the data collection layer collects data through the collection terminal and performs preprocessing; the edge computing layer includes a path planning module for generating inspection paths, an anomaly detection module for identifying abnormal information, and an edge-cloud collaborative computing module for allocating computing resources and optimizing data pipelines; the cloud platform layer is composed of a cloud-based GPU cluster, which generates a lightweight model through knowledge distillation and updates the model of the edge computing layer.

[0049] The acquisition terminal includes an aerial inspection module, a fixed-point monitoring module and a mobile evidence collection module, specifically: Aerial inspection module: Drones: Drones equipped with 5G units and high-resolution cameras are used to perform dynamic path inspections. Drones have fast response and high maneuverability to adapt to changing aerial environments.

[0050] 5G unit: Utilize the high speed and low latency characteristics of the 5G network to achieve real-time data transmission between the drone and the ground control center, ensuring the real-time and accuracy of inspection data.

[0051] High-resolution camera: A high-resolution camera is used to capture ground details and provide high-quality image data for subsequent analysis.

[0052] Fixed-point monitoring module: Fixed smart cameras: deployed in high-risk areas to focus on monitoring potential danger points. The cameras are equipped with night vision and thermal imaging functions to adapt to different lighting conditions and environments.

[0053] Thermal imaging and laser ranging: Thermal imaging technology is used to detect temperature anomalies, and laser ranging is used to obtain accurate distance data, providing a scientific basis for risk assessment.

[0054] Edge computing nodes: Data is preprocessed through edge computing nodes to reduce data transmission volume and improve processing efficiency. The preprocessed data is uploaded to the cloud for further analysis and decision-making.

[0055] Mobile Forensics Module: Ground inspection robot: It carries a robotic arm and a multi-spectral sensor to conduct close-range review and evidence fixation of the drone alarm area. The robot has autonomous navigation and obstacle avoidance capabilities and can adapt to complex ground environments.

[0056] Robotic arm: used to collect samples or perform on-site operations to improve the flexibility and accuracy of evidence collection.

[0057] Multispectral sensor: Multispectral imaging technology is used to obtain image data in different bands, providing multi-dimensional information for anomaly detection and analysis.

[0058] This three-in-one data acquisition module can achieve all-round and multi-angle monitoring of the construction area, improving the coverage and accuracy of data collection.

[0059] The computing resource allocation strategy of the edge-cloud collaborative computing module is as follows: real-time obstacle avoidance tasks are allocated to the onboard FPGA, initial anomaly identification is allocated to the edge server, and model incremental training is allocated to the cloud GPU cluster.

[0060] This embodiment proposes a specific computing resource allocation strategy: The real-time obstacle avoidance task is assigned to the onboard FPGA, with a latency requirement of less than 50ms and a data volume of 2MB / s. Low-latency processing is achieved through the parallel computing capability of the FPGA.

[0061] The task of preliminary screening of anomalies is assigned to the edge server, with a latency requirement of less than 1s and a data volume of 15MB / s. Fast screening is achieved through GPU acceleration of the edge server.

[0062] The model incremental training task is assigned to the cloud GPU cluster, with a latency requirement of less than 30 minutes and a data volume of 200GB / day. Efficient training is achieved through a distributed computing framework (such as TensorFlow Distributed).

[0063] The data pipeline optimization uses the improved Apache Kafka protocol and is optimized in the following ways: Transmit data in pieces to reduce the amount of data transmitted in a single transmission; Dynamically adjust data priority to ensure that high-priority data is transmitted first; Achieve 98.7% effective data transmission rate at 10Gbps bandwidth, with delay jitter less than 2ms.

[0064] After testing, this embodiment shows that in a 5G SA networking environment, it takes an average of 1.2 seconds from anomaly identification to alarm issuance.

[0065] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A highway construction safety inspection method based on an unmanned aerial vehicle airport, characterized in that: The following steps are involved: Step 1: collect real-time data of the highway construction scene through a collection terminal, and pre-process the collected real-time data; Step 2: Input the preprocessed real-time data into the path planning module, calculate the weight of the drone inspection path through the online DQN network, and generate the optimal inspection path; Step 3: Input the pre-processed real-time data into the anomaly detection module for edge computing. When abnormal information is detected, it is fed back to the path planning module to implement path adjustment. Step 4: Input the abnormal information as incremental data into the cloud image processing model. The cloud image processing model and the anomaly detection module learn from each other through the knowledge distillation mechanism. The cloud teacher model is trained using the incremental data, and the trained lightweight student model is deployed in the anomaly detection module.

2. A highway construction safety inspection method based on an unmanned aerial vehicle airport according to claim 1, characterized in that: Step 2 further includes: collecting historical inspection path data of drones, using the offline PPO model to train inspection strategies for typical construction scenarios, and forming a historical inspection database.

3. A highway construction safety inspection method based on an unmanned aerial vehicle airport according to claim 2, characterized in that: In step 2, the formula for calculating the weight of the drone inspection path is: ; The risk level is determined by the hazardous factors in the construction area, and is judged by real-time data, historical inspection data and construction area type annotations. The judgment factors include geological risk, equipment density and personnel activity frequency. The risk level is proportional to the path weight; the coverage rate is the area that the current path can cover. The grid is divided by the sensor data of the drone, and the ratio of the number of grids covered by the path to the total number of grids is calculated. The coverage rate is proportional to the path weight; the energy consumption is the flight energy consumption of the drone, which is related to the flight parameters, and the energy consumption is inversely proportional to the path weight; the risk level, coverage rate and energy consumption are scaled to the [0,1] interval through the Max-Min normalization method; a, b, c is a coefficient determined according to the historical inspection database and used to measure the importance of each factor.

4. A highway construction safety inspection method based on an unmanned aerial vehicle airport according to claim 3, characterized in that: In step 2, the path planning module is trained through a reinforcement learning model. The drone receives external state information obtained by the sensor, performs corresponding actions through the control system, and iteratively trains by observing the new state and reward after executing the action. The flight strategy gradually converges to the direction of maximizing the cumulative reward.

5. A highway construction safety inspection method based on an unmanned aerial vehicle airport according to claim 4, characterized in that: The formula for the reward is: ,in α, β, γ is the weight coefficient, which is determined by the historical inspection database. The coverage rate and risk level are calculated through sensor data. The proximity to the dangerous area refers to the situation of approaching the high-risk area. The energy consumption is calculated through flight parameters.

6. A highway construction safety inspection method based on an unmanned aerial vehicle airport according to claim 2, characterized in that: The anomaly detection module uses the autoencoder and isolation forest algorithm to obtain the anomaly score S to detect abnormal information: ； The autoencoder reconstruction error It is the difference between the input data and the reconstructed output obtained after encoding and decoding by the autoencoder. The calculation formula steps are as follows: The construction area images collected by drones x Input pre-trained autoencoder, the encoder part compresses the image into a low-dimensional latent feature vector , the decoder part restores z to the reconstructed image , for n samples, the mean square error is used to measure the difference between the original image and the reconstructed image: ； Using the mean and standard deviation of historical normal data, the error is mapped to the [0,1] interval: ； Among them, the mean μ AE and standard deviation σ AE Statistics of normal samples from the training phase; Isolation Forest Anomaly Score Indicates the degree of abnormality of the data point. The calculation formula steps are as follows: Input the latent feature vector z extracted by the autoencoder, randomly select features and segmentation values ​​to recursively divide the data to build an isolated tree until the vector z is isolated to the leaf node. For the vector z, calculate its path length h(z) from the root node to the leaf node in each isolated tree, take the average value E(h(z)) of multiple trees, and calculate the anomaly score: ； in c ( n ) is the normalization factor, which is related to the number of samples n: ； When the anomaly score S is less than the threshold T, the anomaly detection module triggers an anomaly information alarm and takes photos for evidence collection; the threshold T is obtained by training historical anomaly data.

7. A highway construction safety inspection method based on an unmanned aerial vehicle airport according to claim 2, characterized in that: The lightweight student model used in the knowledge distillation mechanism is MobileNetV3, and the teacher model is ResNet50.

8. A highway construction safety inspection system based on an unmanned aerial vehicle airport, characterized in that: The system is used to implement the steps of a highway construction safety inspection method based on an unmanned aerial vehicle airport as described in any one of claims 1-7, and is composed of three-level structures: a data acquisition layer, an edge computing layer, and a cloud platform layer; the data acquisition layer collects data through an acquisition terminal and performs preprocessing; the edge computing layer includes a path planning module for generating an inspection path, an anomaly detection module for identifying abnormal information, and an edge-cloud collaborative computing module for allocating computing resources and optimizing data pipelines; the cloud platform layer is composed of a cloud-side GPU cluster, generates a lightweight model through knowledge distillation, and updates the model of the edge computing layer.

9. A highway construction safety inspection system based on an unmanned aerial vehicle airport according to claim 8, characterized in that: The collection terminal includes an aerial inspection module, a fixed-point monitoring module and a mobile evidence collection module.

10. The highway construction safety inspection system based on the UAV airport according to claim 8 is characterized in that: The computing resource allocation strategy of the edge-cloud collaborative computing module is as follows: real-time obstacle avoidance tasks are allocated to the onboard FPGA, initial anomaly identification is allocated to the edge server, and model incremental training is allocated to the cloud GPU cluster.

Images