Shield Tunneling Speed Prediction Method Based on the GWO-CNN-Transformer Model

Digging speed is a key parameter to characterize the efficient and safe digging of shield machine, in order to realize the accurate prediction of shield digging speed, a prediction method of GWO-CNN-Transformer model is proposed. Based on the CNN model, local features are extracted from the input data, and the Transformer model is further employed to capture the global dependencies, thus constructing a hybrid CNN-Transformer deep learning model. By introducing the GWO algorithm to optimize the CNN-Transformer network model, the prediction model of digging speed is established based on the actual engineering dataset, and its prediction results are compared and analyzed with the prediction results of recurrent neural network. The results show that: the shield tunneling speed prediction model constructed based on the GWO-CNN-Transformer algorithm has a MAE (mean absolute error) of about 3 mm and a MAPE (mean relative error) of less than $10 \%$ on the training and validation sets, which indicates that the model has a high prediction accuracy; the GWO-optimized CNN-Transformer network The CNN-Transformer network model optimized by GWO only needs 12 iterations to achieve convergence, indicating that the model can efficiently adjust the hyper-parameters; applying the trained model in the subsequent excavation lining ring, the MAE value of the model is $\mathbf{3. 1 8 5}$ mm, and the MAPE value is $8.406 \%$, indicating that the model’s prediction accuracy can satisfy the engineering requirements; compared with the recurrent neural network model, the CNN-Transformer model has a higher prediction accuracy. Compared with the optimal recurrent neural network model, the MAE and MAPE values of CNN-Transformer model are the lowest, in which the MAE of CNN-Transformer model is reduced by $16.735 \%$ and the MAPE is reduced by $19.129 \%$, which indicates that the model has significant advantages in shield tunneling speed prediction.