Knowledge extraction and insertion to deep belief network for gearbox fault diagnosis

Deep neural network (DNN) with a complex structure and multiple nonlinear processing units has achieved great success for feature learning in machinery fault diagnosis. Due to the "black box" problem in DNNs, there are still many obstacles to the application of DNNs in fault diagnosis. This paper proposes a new DNN model, knowledge-based deep belief network (KBDBN), which inserts confidence and classification rules into the deep network structure. This not only enables the model to have good pattern recognition performance but also to adaptively determine the network structure and obtain a good understanding of the features learned by the deep network. The knowledge extraction algorithm is proposed to offer a good representation of layerwise networks (i.e., restricted Boltzmann machines (RBMs)). The layerwise extraction can produce an improvement in feature learning of RBMs. Moreover, the extracted confidence rules that characterize the deep network offers a novel method for insertion of prior knowledge in the deep RBM. The classification knowledge extracted from the data is further inserted into the classification layer of DBN. KBDBN is used to generate the discriminant features from the data and then construct a complex mapping between vibration signals and gearbox defects. The testing results of KBDBN on a gearbox test rig not only effectively extracts knowledge from the deep network, but also shows better classification performance than the typical classifiers and DBNs. Moreover, the interpretable network model helps us understand what DBN has learned from vibration signals and then makes it be applied easily in real-world cases. (C) 2020 Elsevier B.V. All rights reserved.