Multivariable real-time prediction method of tunnel boring machine operating parameters based on spatio-temporal feature fusion

The tunnel boring machine (TBM) is an important piece of equipment in tunnelling. Accurate prediction of its operating parameters is essential for the operator to adjust the tunnelling strategy in time. Based on the data of a tunnel project in western Sichuan, and considering the easy accessibility of the parameters, this study selects four operational parameters closely related to the tunnelling process as research objects, namely cutterhead speed, total thrust, penetration, and cutterhead torque. A new multi-attention mechanism fusion neural network (TBMformer) based on spatio-temporal feature fusion is proposed. Firstly, based on the establishment of a function to eliminate invalid data to identify different operating states of the TBM. Then the abnormal data were excluded using the isolated forest algorithm, followed by data noise reduction using the Kalman filter, and finally a high-quality TBM dataset was obtained. Secondly, in order to take into account the influence of the TBM real-time running time on the running state of TBM equipment, the correlation between different tunnelling circles and the correlation between different parameters, the time information and ring number information are encoded, and the time attention mechanism and self-attention mechanism are introduced in the time domain and space domain, respectively. In parallel, we employ LSTM to capture the long-term dependencies within TBM sequences. Finally, based on the Informer model, a variety of attention mechanisms are integrated to form the TBMformer model that can deal with the multi-variable real-time prediction of TBM operating parameters. In this study, three datasets with varying spatial resolutions were generated for experimental and analytical purposes, utilising tunnel construction data from two distinct geological contexts in western Sichuan and northern China. The TBMformer model exhibits superior predictive accuracy, with an average accuracy (ACC) of over 94.3% on the three test sets, in comparison to other data-driven methods. The results show that this method can provide real-time guidance to the operator, thereby reducing uncertainty in the control of TBM equipment.