Research On Tunnel Surface Structure Defects Recognition Based On Convolutional Neural Network

With the rapid development of tunnel construction in recent years,numerous tunnels have been put into operation.It has been found that tunnel linings may deteriorate due to poor hydrogeological conditions,harsh climate and environment,long-term loading,extended usage,and inadequate follow-up maintenance practices,and therefore it is prone to tunnel lining defects such as cracks,leakage,and spalling,etc.These tunnel lining defects reduce structural strength,shorten tunnel lifespans,and seriously threaten the safety of people's lives and property.Therefore,it is of great significance to timely and accurately assess the tunnel health condition during the operation period.At present,the tunnel condition assessment mainly relies on the manual visual inspection method which is inefficient and prone to false detection and missed detection.It is urgent to develop an automatic tunnel defects recognition method to identify defects efficiently and accurately.Convolutional neural network-based computer vision recognition technology can overcome the shortcomings of manual inspection to identify tunnel defects efficiently and accurately.Therefore,it is of important engineering significance and scientific research value to carry out research on automatic tunnel defects recognition based on convolutional neural networks and establishing a complete set of methods for intelligent identification of tunnel surface defects.With the main purpose of accurately and automatically identifying tunnel surface defects,this paper proposes a set of automatic tunnel defects recognition methods which include a tunnel defects classification method based on convolutional neural networks,a tunnel defects data augmentation algorithm based on DCGAN,and a tunnel defects segmentation method based on an improved U-net model.The main research contents and results are as follows:(1)The traditional tunnel defects classification methods rely on hand-designed features.These methods have poor universality,are easy to be disturbed in complex tunnel visual environments,and have low accuracy.This paper designs a convolutional neural network-based tunnel defects classification method.Compared with traditional method,this method exhibits the ability to automatically learn strong characterization capabilities and robustness from defects data,and improves the classification accuracy of tunnel defects.In the experiment,the K-CV is utilized to verify that the tunnel defects classification model based on convolutional neural network does not fall into overfitting and performs well.(2)Aiming at the problems of few tunnel defect data samples and unbalanced categories,this paper analyzes the applicability of several typical algorithm solutions,and establishes a DCGAN-based tunnel defects data augmentation method.DCGAN can learn the data distribution of minority categories which are crack and leakage.The crack and leakage defects data which is generated by the generator of DCGAN is utilized to expend and balance the tunnel defects data set.Through experimental verification,it is found that this method overcomes the problems of small sample and category imbalance,and significantly improves the performance of the tunnel defects recognition model.(3)Aiming at the problems that the spatial output accuracy of tunnel defects segmentation model is not high and the defect boundary is blurred,a tunnel defects segmentation model based on the improved U-net model is proposed.This model combines the low-level detail features and high-level semantic features of the tunnel defects,and retains the detailed information of the defects.The SE-ResNet block in improved U-net can address vanishing gradients,deepen the network depth,fit perturbations,and recalibrate features,thereby improving the accuracy of the tunnel defects segmentation model.The effectiveness of this method is verified through experiments.