To solve the problem of feature localization with multi-feature intersection and improve the performance of machining feature recognition for complex parts, this paper proposes Brep3pNet, a machining feature recognition method under the framework of instance segmentation. Firstly, based on the boundary representation (B-rep) of 3D models, Brep3pNet extracts geometric and topological data such as face point clouds and face adjacency graphs to construct a graph representation of the 3D model. We utilize point cloud learning networks and graph neural networks to learn surface-level embedding representations of the 3D model. Secondly, a probabilistic positional embedding method is proposed, which introduces spatial position prior information to encode the face into ternary Gaussian distribution, and measures the similarity among those face embeddings by Bhattacharyya kernel for the purpose of locating machining features and generating candidate machining feature instances. Finally, a score network is designed to predict the quality of the instance generated, so as to guide the non-maximum suppression between instances to remove redundant feature instances, thereby obtaining the final machining features. Brep3pNet is evaluated on four multi-feature datasets, including MFCAD, MFCAD++, MFInstSeg and a synthetic dataset of rotary parts. The research results indicate that Brep3pNet outperforms other state-of-the-art methods on feature localization accuracy, and can achieve optimal feature recognition accuracy with lightweight model parameters, demonstrating its potential application in intersecting features recognition.