| As an essential infrastructure to ensure safe aircraft operation,airport runways serve as the area for aircraft takeoff and landing.Environmental factors and repeated takeoffs and landings inevitably lead to subsurface defects to the runway structure,such as typical voids,cracks,subsidence,and water-rich defects.These issues can severely impact the runway’s service life and safety performance.However,early detection of such subsurface defects can significantly reduce maintenance costs and ensure safety.Ground Penetrating Radar(GPR)has been widely used in runway inspection tasks as an important tool for detecting underground targets.However,the interpretation of GPR data still relies on human experts,and the manual interpretation process often suffers from subjectivity,error-proneness,high cost,and time-consuming issues.In recent years,deep learning technology has made significant progress in fields such as image recognition and object detection,providing new solutions for automated detection of runway structural diseases.Therefore,developing effective automatic detection methods for runway structural diseases can improve the efficiency of runway inspection while increasing the accuracy and reliability of the detection,which is of great significance for ensuring airport operation safety and promoting the application of artificial intelligence technology in the engineering field.Due to the non-uniformity and strong attenuation characteristics of the airport pavement structure,the electromagnetic wave propagation environment of the ground-penetrating radar(GPR)is complex and varied,leading to severe interference in the radar data features of underground defects.Automatic detection of various types of underground defects from complex radar images is the primary problem that needs to be solved for airport pavement structure detection.Different types of defects require different repair measures,and achieving fine-grained identification of airport pavement structure defects is of great importance for later airport runway maintenance work.However,due to differences in defect size,severity,and development stage,different high-incidence underground defects exhibit similar features,subtle differences,and are difficult to distinguish in GPR profile images.In addition,occasional underground defects containing water are scarce in sample data,making it difficult to use machine learning methods for defect identification.Moreover,the tailing effect caused by antenna ringing is very similar to the water-rich feature in radar images,making it difficult to distinguish the two under interference conditions,and automatic detection of small-sample defects still remains a key challenge.Finally,different airport environments differ greatly,and it is necessary to design model fusion methods to give full play to the respective advantages of the research algorithms and improve the practical application effect of disease detection in unknown airport pavement scenarios.In summary,this thesis focuses on the detection of hidden defects in airport pavement structures,which face many challenges such as strong noise interference,difficult differentiation of morphologies,and scarce sample data.The research is mainly based on deep learning and ground-penetrating radar(GPR)technology.The main contributions of this study are as follows:(1)Research and implementation of an automatic detection method for airport pavement structure defects in complex backgrounds.Unlike existing 2D object detection methods,the proposed 2D-3D hybrid convolutional neural network takes into account the correlation between GPR data in adjacent channels to extract features from different dimensions.Based on the fusion features,a bottom-up proposal generation strategy is used to filter out noise points and reduce the search space for defects.Then,a fully connected layer is constructed to refine the initial proposals in terms of location and classification,resulting in precise detection of underground defects.This method achieves automatic detection of common underground targets from complex GPR images,overcoming the interference issues caused by non-uniformity and strong attenuation of airport pavement structural layers,significantly enhancing the efficiency of radar data interpretation.(2)Based on multi-view learning,fine-grained detection of runway structural diseases at airports.In response to the feature differences of different types of high-frequency underground diseases in different views,a multi-view convolutional neural network is designed to extract GPR side and top views separately,and the two are fused according to their actual corresponding positions.At the same time,a dense connection network is used to repeatedly utilize the extracted features to enhance feature transfer efficiency.In addition,a radar channel attention mechanism is introduced to guide the network to focus on clues that are useful for fine-grained feature recognition.A large number of experimental results indicate that this method can provide more accurate fine-grained disease type and location information,enabling precise preventative maintenance for airport pavement surfaces.(3)Small sample detection of runway structural defects at airports.By constructing a multimodal fusion network,the correlation and complementarity between different modalities of radar data are utilized to extract features of radar raw signals,image data tail signals,and underground target echoes.A reasonable joint mechanism between modalities is established to maximize the information content.A forward model is established between small sample defects and radar echoes to obtain a large amount of annotated simulation data for model pre-training.Finally,real small sample defect data is used for transfer learning to transfer the defect detection capability from simulated data to actual airport runway data.Experimental results demonstrate that this method can address the challenges of scarce data on sporadic diseases and difficulty in distinguishing them from trailing interference,providing effective early warning for severe sporadic underground diseases at airports.(4)Airport runway structural damage detection based on multi-model fusion.To enhance the generalization performance of the proposed methods in real-world application scenarios,automatic augmentation techniques are employed to expand the GPR dataset.The existing learning models are trained with this expanded dataset,extracting parameters and features from multiple models and using the predicted results as training samples to establish a new prediction model.By conducting tests on complete runway projects at domestic airports,it was found that the consistency between test results and human-annotated results reached over90%.Faced with unknown GPR data from diverse airport runway structures,this method allows each model to leverage its strengths,enhancing the detection accuracy of all types of underground diseases,resulting in a more comprehensive and reliable integrated model. |
PDF Full Text Request