A novel constrained dense convolutional autoencoder and DNN-based semi-supervised method for shield machine tunnel geological formation recognition

Abstract Accurately acquiring the geological information of the tunnel face will help to set the optimal operational parameters, so that the shield machine can achieve better tunneling performance. The design of the shield machine prevents the operator from observing the surrounding environment directly, and the soft methods which can utilize machine parameters to recognize geological conditions indirectly are becoming a research hotspot. However, current soft methods are all supervised methods which can only use the few machine data with geological type labels, and the unlabeled machine data that is much more than the labeled machine data is wasted. To make the most of all the collected in-situ data to boost the performance of the geological formation recognition model, a novel constrained dense convolutional autoencoder and DNN-based semi-supervised method is proposed. To begin with, 177 machine parameters related to geological conditions are selected and preprocessed. Then, a novel geological feature extractor is obtained via the proposed constrained dense convolutional autoencoder and the unlabeled data. Eventually, a DNN-based geological feature classifier is trained on the basis of the established feature extractor and the labeled machine data, which is capable of recognizing geological formation of the tunnel face. In-situ data collected from a Singapore project (stacked twin bored tunnels) was used to prove the superiority of the proposed method. The results show the constrained dense convolutional autoencoder can extract geological-related features accurately, and the proposed method outperforms other supervised soft methods. Its classification performance in one tunnel is 23.98%, 17.47%, 1.93%, and 18.52% higher than the random forest-based, decision tree-based, KNN-based, and SVM-based methods, respectively. Its classification performance in another tunnel is 33.54%, 33.75%, 42.87%, 43.58%, 33.75%, 49.91%, 37.77%, and 27.04% higher than the random forest-based, SVM-based, decision tree-based, DNN-based, KNN-based, CNN-based, ResNet-based, and DenseNet-based methods, respectively. Thus, the novel semi-supervised method has significantly better generalizability than the currently adopted supervised soft methods.