Design And Implementation Of A Tunnel Inner Wall Leakage Detection System Based On Deep Learning

The metro is one of the commonest ways of transportation for modern cities.To minimize the impact on the surface urban landscape,most metro lines are located in underground tunnels.Due to the presence of groundwater,leakage problems are a common disease in underground tunnels.The detection of leakage in tunnels still relies on the cruise by linemen at present,which is time-consuming.With heavy workloads,the daily maintenance of the metro is arduous,and there is an urgent need for an automated method to detect leaks and improve the efficiency of maintenance tasks.Object detection tasks are a common problem in computer vision.Traditional object detection relies on feature engineering methods and machine learning methods,and there are proven methods with high accuracy only in some popular areas.Deep learning technology can automatically extract features from samples and optimize prediction accuracy.Therefore,in recent years,feature engineering methods and traditional machine learning algorithms in object detection tasks have gradually started to be replaced by deep learning models.We take the Shanghai Metro tunnel as the actual application scenario and compares the performance of different deep learning models in the tunnel inner wall leak detection using the images of the inner wall of the metro tunnel.We also use the improved neural network model to design and implement a tunnel inner wall leakage detection system.The main work of the thesis includes:(1)Collecting images of the inner walls of tunnels on different lines of the Shanghai Metro,pre-processing the dataset by screening labeling and so on.Using image augmentation methods to expand the dataset.proposing an image augmentation method based on background replacement.Comparing and Analyzing the effects of different image augmentation methods on model performance.(2)Comparing the recognition accuracy of different object detection models on the leakage dataset.Selecting Faster R-CNN with the best effect as the target detection model,and optimizing the parameters of the model.(3)Solving the problem that the object detection model lacks scale invariance by proposing a procedure of firstly performing scale detection,then scaling the image and finally performing object detection.Comparing the performance of different models as scale detection models.Selecting the modified ResNet model as the scale detection model,and optimizing the parameters of the model.Proposing an averaging method by removing extremum based on the median,which improves the accuracy of prediction.(4)Designing and implementing the front-end web page and back-end system of the tunnel inner wall leakage detection system using frameworks such as Flask and Vue.js.This thesis proposes image augmentation methods and model training and optimization methods for a dataset for specific application scenarios.By applying the data augmentation method and the optimized neural network model,the m AP performance of the system on the test set is 84.7%.We also design and implement a leakage detection system based on our detection method.