Explainable Transfer Learning for Modeling and Assessing Risks in Tunnel Construction
Hanbin Luo, Jialin Chen, Peter E.D. Love, Weili Fang
Abstract
Deep learning models are black boxes. Thus, determining the source domain data contributing to transfer learning for ground settlement prediction is impossible. The research presented in this article aims to determine the source domain data (i.e., the dataset or domain used for model pre-training) that contributes most to transfer learning for risk prediction in tunnel construction and quantify its contribution to improving prediction accuracy. We propose a novel explainable transfer learning approach to quantify the selection of degraded knowledge from source and sub-source domains. Our approach comprises: (1) feature selection and space point clustering; (2) construction of a similarity metric between the target domain and each sub-source domain; and (3) construction of a stacked Deep Neural Network model with selective transfer learning. We apply our model to a real-life tunnel project to demonstrate its feasibility and effectiveness. The results indicate that: (1) our proposed explainable transfer learning approach outperforms other transparent and opaque analysis models on risk prediction with <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">R</i> <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> above 0.5 by adjusting the clustering, transferring, and freezing strategy; (2) the optimal number of freezing layers should be less than half of the total number of layers, and the best number of freezing layers is 1. We show that explaining transfer learning enables transparency in training and understanding the source domain data, contributing to ground settlement prediction.