Enhancing data reuse in tunnelling site investigation through transfer learning-based historical data mining

Vast amounts of valuable historical tunnelling site investigation data remain underutilized due to inefficient content-based archiving and searching tools. This study introduces a novel data-driven framework that integrates transfer learning with reverse image search to revolutionize the utilization of historical data in tunnelling projects. The method indexes excavated tunnel sections with corresponding tunnel face images and identifies similarities between projects based on geological features. Transfer learning with pre-trained deep learning models is employed to compress tunnel face images into compact, lower-dimensional vectors, enabling efficient similarity searches. This transformation converts geological information into comparable vectors, enhancing the efficiency and speed of data searches. An online cloud service is developed to allow engineers to access similar historical projects in real-time. To enhance the quality of the compressed vectors, this study developed a multi-level feature extraction method. This method markedly improves the deep learning models’ ability to accurately identify major features from rock images. When applied to a diverse range of tunnel excavation projects in China, the model exhibited an impressive accuracy of over 90% in retrieving projects with similar geological features. This underscores the model’s potential as a robust tool for enhancing data management and decision-making in tunnelling engineering.