Towards a fault diagnosis method for rolling bearing with Bi-directional deep belief network

Intelligent fault diagnosis model based on machine learning algorithm has extensive application, while the unsatisfied training data quality leads to the lower accuracy of intelligent fault diagnosis and inferior generalization performance. To address this problem, a method called Bi-directional deep belief network (Bi-DBN) is developed for fault diagnosis of rolling bearings. The forward training part of Bi-DBN can learn the fault features from the original vibration signals first, and then the reverse generation part synthesizes samples according to the weight sharing by the forward training part. The noise time-shift layer is innovatively introduced to reduce the similarity between the synthesized sample and the original sample. Finally, Quantum genetic algorithm (QGA) is applied to optimize the key parameters of Bi-DBN to improve the feature learning efficiency. The experimental results of rolling bearing signals indicate that the developed method significantly limits the effect of training data quality on the diagnosis accuracy.