Abstract: As a direct byproduct of TBM rock fragmentation,TBM rock-muck contains a wealth of pertinent geological information.In order to fully harness the informative potential embedded within TBM rock-muck data to enhance the efficacy of surrounding rock classification prediction,this study first preprocessed the raw images and then predicted the classification of surrounding rock using convolutional neural network algorithms including ResNet,VGG,and SqueezeNet.The results show that using convolutional neural networks for TBM rock-muck image recognition is a feasible method for classification prediction of surrounding rock.ResNet 18 and SqueezeNet v1.0 models exhibit high prediction accuracy,with an accuracy (ACC) reaching 0.986.Additionally,the running time for a single image is only about 70ms,which indicates the potential for real-time surrounding rock classification prediction during TBM tunnelling.The surrounding rock classification prediction based on TBM-generated rock-muck images not only facilitates real-time perception and classification of the rock face ahead of the shield but also assists geological engineers in enhancing their work efficiency.This approach thus provides a valuable reference for the automation determinations of surrounding rock classification.