Deep Learning–Based Inspection Data Mining and Derived Information Fusion for Enhanced Bridge Deterioration Assessment
Pengyong Miao, Guohua Xing, Shengchi Ma, Teeranai Srimahachota
Abstract
Inspection data are usually utilized to assess bridge situations for directing further maintenance and preservation. However, due to the complexity of inspection data, mining and fusing valuable information to assess bridge situations remains challenging. To address these issues, a novel inspection data analysis framework was proposed in this study. The framework integrated a gated recursive unit (GRU) model, a semantic segmentation (Seg) model, and a Yolo V4 object detector to analyze both time-series data and images. Seg and Yolo were used to detect defective pixels, which were then evaluated using refined fuzzy inference systems (RFISs) to determine the deterioration grade. The GRU and RFIS models were employed used to infer the probability of bridge deterioration grades. These probabilities were then fused by the novel fusion technique to determine the final deterioration grade. A verification showed GRU, Seg, and Yolo detectors to have 0.9299, 0.9580, and 0.7967 accuracy values for analyzing time-series data and images, respectively. RFISs also performed well in determining concrete and steel deterioration grades with R-values of 0.9968 and 0.9962. Compared with Dempster–Shafer and its two variants, the proposed fusion technique improved the accuracy rates by 11.65%, 2.19%, and 3.38%, respectively. Prototype models also demonstrated abilities to clearly understand deterioration grades and the spatial relationship of defects. Overall, the proposed method could sufficiently mine inspection data and more reasonably assess bridge situations.