An optimization strategy to improve the deep learning‐based recognition model of leakage in shield tunnels

Abstract

AbstractDue to the interference problems of complex on‐site installations attached to shield tunnel lining surface, deep learning models, developed for leakage datasets of shield tunnels, are not prepared to meet engineering requirements. Therefore, it is of utmost importance to optimize the original model based on the characteristics of leakage datasets. For this purpose, the present study adopted Mask R‐CNN as the baseline and improved its performance from two aspects, including the properties of shield tunnel leakage datasets and detection errors of the original model in the testing set. With reference to the properties of leakage datasets, the model compression technique was implemented to remove the redundant parameters in the training stage and enhance the detection accuracy and speed of the original model. Besides, three error types were grouped to explore the optimization direction in practical application. Accordingly, four different optimization measurements were fulfilled step‐by‐step to improve the model performance. It was concluded that the compressed model with 62% sparsity reached average precision (AP) of 0.399 at the detection speed of 7 frames per second (FPS), which was 0.144 higher and 2 FPS faster than that of the original model, respectively. The improvements in the model concerned were also quite different with the three error types moderated, indicating that the optimization direction of the original model needed to be focused on the error types with higher AP promotion. In addition, the AP value of the original model significantly augmented from 25.5% to 51.1% upon minimizing the detection errors on shield tunnel leakage datasets.