An unsupervised deep learning surrounding rock perception method for TBM operational parameter multiobjective optimization

In the tunnel boring machine (TBM) excavation process, accurately assessing the surrounding rock conditions in real time is critical for effectively controlling the TBM operational parameters. This study introduces a novel unsupervised deep learning rock perception method for TBM operational parameter multiobjective optimization. The TBM operational data were preprocessed using various techniques, including excavation phase segmentation, dimensionality reduction via the adaptive piecewise constant approximation (APCA), and combined outlier detection. A database was established by using the processed data to formulate a technique for grading rocks on the basis of the deep clustering model stacked sparse autoencoder-fuzzy c-means (SSA-FC). On the basis of the clustered results, a comprehensive multiobjective optimization (MOO) framework for TBM operational parameters has been developed using the NSGA-III algorithm, which targets objectives such as the penetration rate, cutter disc wear, and energy consumption in rock breaking. This framework and associated models, along with a predictive model for excavation parameters, were tested against 80 TBM excavation cases, confirming their efficacy. The results of the optimization showed a 50.19 % improvement in the penetration rate on average, an 88.23 % decrease in energy consumption for rock breaking, and a 79.52 % reduction in wear on the cutter disc. These methodologies and models offer supportive insights for decision-making in the optimization of operational parameters for comparable TBM tunnel projects.