Research On Deep Learning Algorithm For Feature Recognition Of Tunnel Leakage

Currently,conventional tunnel leakage detection mainly relies on engineers in China,which has disadvantages of low detection efficiency,strong subjectivity,low accuracy,high labor cost and harsh environment and inability to adapt to the rapid development of tunnel construction.This thesis proposed to combine the deep learning algorithms with conventional image processing technologies to realize the identification of tunnel leakage.Initially,the causes of three types of point,line and large area leakage phenomena,and corresponding treatment measurements were analyzed.According to the temperature difference between the seepage and non-seepage area on the cable tunnel lining,the temperature characteristics distribution matrix was captured by the infrared imager and then converted into a digital image by normalization.Considering that it is difficult to collect a large number of images,traditional image processing techniques of rotation,histogram equalization,bilateral filtering etc.and current popular generative adversarial nets were proposed to enrich the datasets.Secondly,several typical convolutional neural network structures in deep learning were deeply studied from the aspects of depth,width and attention.The three network structures of AlexNet5,VGG16 and ResNet34 examples in depth aspect,the GoogLeNet21 example in width aspect and attention module example in attention aspect were mainly studied.The structures of AM-VGG16 and AM-ResNet34 were designed based on experimental equipment conditions.Comparing the test error rate on testing datasets and training loss on training datasets of AM-VGG16 to AM-ResNet34,the AM-VGG16 with an error rate of 1.94% and a loss value of 0.06 was selected to identify the characteristics of tunnel leakage.Finally,the model parameters with an epoch of 30 and an AUC value of 0.932 were selected for testing by comparing ROC curves and the values of AUC.The selected model parameters were tested on 4634 pictures,whose accuracy reached up to 98.4% and average inference time was 12 ms.This detection technology greatly improves recognition accuracy and recognition efficiency compared to human eye detection.