Abstract: Machining feature recognition can extract valuable manufacturing semantic information from geometric and topological relationships of solid models, and it is an important method to realize seamless connection from design to digital processing. However, the existing machining feature recognition methods can not meet the needs of engineering. In order to improve the recognition performance of complex manufacturing features, a novel feature recognition algorithm based on BP binary classification network was proposed. Firstly, the feature coding rules are defined according to the attributes of the sides of the feature surfaces, and the features are represented as 12-dimensional feature vectors. Then, 6 kinds of multi-BP binary classification networks are designed for 6 kinds of common machining features in milling parts, are trained by self-designed feature data set and obtain the good feature recognition effect. In order to construct a large scale machining feature data set for training, an automatic generation method of feature data set is designed, and the generalized geometric Euler formula is used to verify the validity of the model that automatically generates the feature data set. Finally, based on the concept of ground surface, the intersection features are decomposed into multiple isolated features, and then the network is used to identify the features. Testing on a self-built model and a subset of models from the ESB model library shows that the algorithm can effectively recognize machining features on milling parts.