Underground Disasters Recognition Method Research Of Urban Road Based On GPR Data

Road is an important infrastructure.In recent years,urban road collapse accidents caused by underground disasters occur frequently,which seriously endanger people’s life and property safety.Ground penetrating radar(GPR)has been widely used in the recognition of disasters under urban roads because of its advantages of nondestructive,high efficiency and high resolution.However,GPR data contains a large amount of noise,which makes it difficult to interpret.At present,the disasters information in GPR data is mainly interpreted by manual interpretation.Due to the lack of prior knowledge given by physical model and the dependence on subjective judgment of field experts,the efficiency and accuracy of the recognition for underground disasters are low.Therefore,the disasters information in the intelligent diagnosis of radar images has important social value and practical significance.Currently,among the research on recognition of underground disasters at home and abroad,deep learning methods based on artificial intelligence have significant advantages in disasters feature extraction,however their performance depends upon a lot of high quality and labeled data.Due to the difficulty in interpretation of disasters data and the lack of labeled data,efficient and accurate deep learning algorithms are yet to be researched.On the other hand,most of the current researchers focus on the recognition of void disasters and few on other types of underground disasters.In view of the above problems,this thesis focuses on the automatic interpretation of a variety of underground disasters in the case of fewer samples.Through theoretical analysis and forward modelling of underground disasters,a prior guidance based on numerical simulation is given,and a deep learning model of underground disasters classification was proposed to realize intelligent diagnosis of underground disasters.The completed work and research results are as follows:(1)An underground disasters recognition method based on feature distribution calibration(UDC)was proposed.The method constructs a convolutional neural network for GPR data(gpr Net)which can extract the features of the disasters data after trained with a small number of samples.Then we generate feature samples through the transfer of feature statistics to augment the training data set.The trained model can get more accurate feature distribution of underground disasters and realize the calibration of feature distribution of disasters data and further realize the recognition of different kinds of underground disasters.(2)Through the analysis of the characteristics of underground disasters data,forward modelling of underground disasters is carried out based on the FDTD method,and the prior knowledge of underground disasters based on numerical simulation is given.An unsupervised subdomain adaptation method for underground disasters recognition(USAN)was proposed.Through forward modelling of underground disasters,we established the forward modeling data set as the source domain data and training data.And we speed up network training and improving convergence with pretraining strategy.Finally we utilize the local maximum mean discrepancy metric to align the distribution of subdomains in feature space between forward modelling data and real underground disasters data to make the trained model recognize unlabeled data correctly.In conclusion,in this thesis we proposed an underground disasters classification method based on feature distribution calibration for recognition of few samples and an unsupervised subdomain adaptation method for underground disasters recognition.The provided experiments verified the effectiveness of our methods,and suggested that our methods promote the development of underground disasters recognition in GPR data.