Classifying Rock Fragments Produced by Tunnel Boring Machine Using Optimized Convolutional Neural Network

Grain size and uniformity index of fragments produced by tunnel boring machine (TBM) reflect tunneling states. Traditional manual fragment identification methods will reduce tunneling safety and efficiency. Therefore, an automatic fragment classification method is significant. However, the fragments classification accuracy of existing deep learning methods will decrease under different tunnel projects with various kinds of fragments. In this paper, a rock fragment classification method is proposed based on convolutional neural network (CNN). To distinguish fragment and background with very similar grey level in the image, contrast limited adaptive histogram equalization is used to enhance fragment edge features. In addition, image regeneration is used in data preparation process to balance data set. During hyperparameter tuning in model training, this research applies Bayesian optimization to acquire optimum model. Experiments show that proposed CNN model can acquire an accuracy of 91.88% in classifying various types of fragments, improving 39.21%, 11.64%, 13.64% and 9.45%, respectively, compared with LeNet, ResNet, VGG and AlexNet. Batch normalization, DropBlock and global average pooling skills are used to alleviate overfitting of CNN model. Based on proposed model pre-trained on a single project data set and a small amount of new data, the migrated model can achieve 89.86% accuracy on a new tunnel data set. The experiment results demonstrate a great generalization ability of proposed model dealing with various kinds of fragments. Developing a fragments image-based deep learning architecture to classify rock fragments produced by tunnel boring machine.Trained by Bayesian optimization, proposed method is tested on rock fragment images collected from tunnel projects and acquire better accuracy, precision and recall on the test set than the LeNet, ResNet, VGG and AlexNet based fragment classification models.By integrating the deep learning theories to alleviate model overfitting, proposed model has higher accuracy than existing image classification models when dealing with rock fragments under different tunnel projects.

Automated summary

This paper proposes a rock fragment classification method based on convolutional neural network (CNN), which uses contrast limited adaptive histogram equalization to enhance fragment edge features and balances the data set through image regeneration. The experiments show that the proposed CNN model achieves high accuracy in classifying various types of fragments and outperforms other classic fragment classification models.