Physical feature shared online transfer learning framework for cross-engineering rock mass quality perception during TBM excavation

• A physical shared online transfer learning framework, including an online data stream processor, online learning module, and transfer learning module, is proposed for real-time rock mass quality perception. • The rock-cutting physical process-related parameters are extracted in shared feature space of different tunnels for transfer availability. • An online learning module employing LSTM and the probabilistic state machine methods is developed. • Online transfer learning application is carried out, and performance is validated against offline machine-learning methods. Real-time and accurate perception of rock mass quality during Tunnel Boring Machine (TBM) excavation is crucial for ensuring effective and safe tunnelling operations. This is particularly challenging during the early stages of newly constructed tunnels, where limited prior geological investigations hinder the prompt evaluation of rock mass quality. Therefore, utilizing monitored data from completed tunnels to develop transferable intelligent models remains an urgent task. This study proposes a physical feature shared online transfer learning framework for cross-engineering rock mass quality prediction. First, an online data stream processor, incorporating real-time data acquisition, segmentation, and stream processing, is developed to extract physical shared features. Next, an online learning module based on the long short-term memory and probabilistic state machine (LSTM-PSM) is developed for classification probability evolution estimation. Finally, a case study is conducted to validate the framework using shared features from two tunnels. Field data from the Songhua River Tunnel (YS, 19.77 km) is used as the source domain, while data from the Chaoer to Xiliao River Tunnel (YC, 2.27 km) serves as the target domain. Results demonstrate that (1) The proposed online transfer learning method can accurately identify rock mass classification and state transition processes, (2) Physical shared features from the rock-cutting process exhibit superior performance in similarity measures and provide ease of transfer due to their accessibility in different tunnels, and (3) The online learning module outperforms state-of-the-art offline transfer learning methods in terms of accuracy and model reliability. This study offers a novel approach to cross-engineering real-time rock mass quality perception, facilitating rock mass stability assessment and supporting measure selection.