A real-time multi-head mixed attention mechanism-based prediction method for tunnel boring machine disc cutter wear

In the process of hard rock tunnel excavation, workers often need to enter the tunnel boring machine (TBM) cutterhead at regular intervals to measure cutter wear. However, this method is time-consuming and labor-intensive. Existing cutter prediction models primarily rely on geological parameters to predict overall cutter wear before construction, making it challenging to monitor real-time wear at different locations of the cutter and obtain accurate geological parameters. To address these challenges, this paper proposes a multi-head mixed attention mechanism-based method for real-time wear prediction of TBM disc cutter. First, a method of cutter wear normalization to eliminate measurement noise is explored. Then, considering the complex correlation of TBM operating parameters in feature and time dimensions, a new multi-head mixed attention mechanism model is designed to establish the dependency between different features and different moments, to better establish the mapping model between operating parameters and cutter wear. Finally, the current cutter wear state can be calculated by accumulating the wear amount of all previous small excavation sections. The effectiveness of the method is verified by using field data from the Mumbai tunnel. The results demonstrate that the method is capable of real-time prediction of front cutter and edge cutter wear on the test set, achieving an average accuracy rate of 95.75%. Moreover, the method can update the cutter wear status after every meter of excavation, which has good real-time performance. In addition, the average accuracy of cutter wear prediction of the proposed method is 11.75%, 10.375%, 3.875%, 3.625%, 1.375%, 3.125%, 2.25%, and 0.75% higher than that of LSTM, CNN, LSTM-CNN, CACNN, SACNN, MMADNN, MTACNN, and MFACNN. In summary, this approach offers an accurate prediction of cutter wear state while reducing inspection time and costs and has high application value.