Geological information ahead of the tunnel face is vital for ensuring safe and efficient tunnel construction. This paper presents a semi-supervised approach that combines a constrained dense convolutional autoencoder (CDAE) with a hybrid deep neural network (HDNN) to predict rock mass discontinuities. The factors related to rock mass discontinuity are categorized into four main classes: geometry, geology, construction, and measurement while drilling (MWD). The geological similarity between neighbouring blastholes is taken into consideration to constrain the training process of CDAE. These four categories of data serve as input parameters for CDAE-HDNN to predict rock mass discontinuity. Data from the Yangjiawopu tunnel is employed to assess the performance and generalization capability of the proposed method. The results show that CDAE can extract features related to rock mass discontinuity, and the proposed model exhibits superior performance and robustness compared to conventional deep learning and machine learning models.