Abstract: Deep neural network (DNN) is a complex structure and multiple nonlinear processing unit models, and has been gradually applied in industrial production processes. Due to the unexplained of "black box" problem, and the huge data demand problem, however, there are still huge obstacles to the application of DNNs in industrial fields. A new DNN model, knowledge-based deep belief network (KBDBN), is proposed. The combination of this logical symbol language and deep neural network not only makes the model have good pattern recognition performance, but also adaptively determines the network model and has interpretable and visual characteristics. Furthermore, the prediction model of workpiece surface roughness processing based on KBDBN is proposed, which realizes accurate prediction and effectively extracts the key knowledge of the manufacturing process. Experimental results show that compared with the traditional machine learner, KBDBN has better network performance, model interpretability and more applicability. Combines symbolic rules with deep learning and establishes a processing roughness prediction model, which can extract process knowledge and guide process optimization under the premise of accurate prediction.

Key words: surface roughness, deep learning, deep belief network, knowledge discovery, pattern recognition