<p>The precise identification of rock mass discontinuities is a crucial part of the stability analysis of rock mass slopes. Aiming at the problems such as the low efficiency of traditional machine learning methods relying on manually annotated data, and the limitations of the existing PointNet + + network in the extraction of point cloud neighborhood features, this study proposes an innovative deep learning method named Attention-PointNet + + that integrates the forward modeling of discontinuities and attention enhancement. Its specific process includes: (1) Using the point cloud coordinates and normal vectors as input features, adopting a local area random sampling strategy, and constructing a training set (accounting for 80%) and a test set (accounting for 20%) based on the forward modeling of discontinuities; (2) Introducing the Attention-PointNet + + network architecture improved by integrating the self-attention mechanism, and strengthening the ability to capture complex geometric features through multi-scale feature fusion; (3) Based on the trained deep learning model, realizing the intelligent identification of rock mass discontinuities, and completing the extraction of a single discontinuity and the calculation of orientation parameters by combining the density-based spatial clustering of applications with noise (DBSCAN) algorithm and random sample consensus (RANSAC) algorithms. Experiments show that in the standard dataset, the identification results of the three groups of discontinuities are basically consistent with the known data, and for the measured point clouds of two typical engineering cases, this method also shows good robustness. This study provides a new technical solution for the automatic identification of discontinuities in complex geological environments.</p>