Surface Defect Detection Of Diversion Tunnel Based On Deep Neural Network Model Compression

As the key structure of water conservancy and hydropower project,diversion tunnel has the functions of hydropower generation,agricultural irrigation,domestic water scheduling.Due to the difficult construction technology,complex surrounding rock geology and uneven pipeline water pressure,the tunnel surface is easy to produce high-risk defects such as cracks,exposed reinforcement and leakage.If the surface defects of the tunnel can not be detected timely and accurately,the diversion tunnel may stop operation,and even serious consequences such as personal safety accidents may occur.At present,the surface defect detection method of diversion tunnel is mainly manual detection,which has many problems,such as large hidden danger of personal safety,long defect detection cycle,too subjective inspection results and low utilization rate of defect features.Therefore,it is urgent to take into account the high accuracy and low delay of surface defect autonomous detection technology.Based on the deep neural network model compression technology,this paper studies the autonomous detection technology of surface defects of diversion tunnel.(1)Aiming at the problems of turbid water quality,dim light and deposition in the diversion tunnel environment,the deep curve estimation network was designed to optimize the defect images collected by the tunnel robot,which could effectively improve the image quality in the low illumination environment.At the same time,combined with the defect characteristic mechanism,the data set of surface defects of diversion tunnel including cracks,pitted surface,shedding,exposed reinforcement,calcification and leakage was constructed.(2)Aiming at the problem that most of the existing deep neural network can not effectively extract image features of surface defects of diversion tunnel,a dynamic convolution module with attention mechanism was constructed to replace the traditional static convolution,and the dynamic features obtained could dynamically adjust the model parameters according to different defect features for a single sample.(3)Aiming at the problems of huge parameters and long model reasoning time in deep neural network,a model compression method based on dynamic feature distillation was proposed.A dynamic feature distillation loss was constructed by fusing discriminator structure in knowledge distillation framework.The dynamic feature knowledge was transferred from a deep teacher network to a shallow student network,which greatly reduced the model reasoning time while six kinds of defects could be detected with high precision simultaneously.Finally,compared with the original residual network on the constructed data set of surface defects of diversion tunnel,the detection accuracy of this method can reach 96.15%,and the amount of model parameters and reasoning time are reduced to 1/2 and 1/6 of the original respectively.At the same time,the integrated development of diversion tunnel surface defect detection system was used to test the data of a diversion tunnel in Sichuan.Experimental results show that fusing the compressed information of the dynamic feature model of defect images into the deep neural network can effectively improve the efficiency of the surface defect detection of diversion tunnel and has engineering significance for the long-term safe operation and maintenance of the diversion tunnel and the intelligent management of water conservancy projects.