Intelligent Bearing Fault Diagnosis Based on Adaptive Deep Belief Network Under Variable Working Conditions

Abstract

In engineering, working environment and operating state are constantly changing, which decreases the correct rate of equipment fault diagnosis, resulting in the loss of time and cost. The structure of the deep belief network is investigated for the time-varying factors in the mechanical system. In combination with the signal decomposition technology of fixed learning step size, the original characteristics of the sensor data are retained. In addition, the deep key information of the signal is repeatedly extracted layer by layer. The data loss technology is integrated to optimize the network structure to avoid over-fitting problems. Further, considering the domain adaptive method in transfer learning, the memory characteristics of different levels of deep belief networks are solidified. Therefore, a domain adaptive deep belief network with shift-invariant features (SIF-DADBN) is proposed for rolling bearing fault diagnosis. By identifying the characteristic information of similar fault signals with variable working conditions, the accuracy and generalization of bearing intelligent fault diagnosis are both improved. Based on the public data set of rolling bearings, the average correct rate of the fault diagnosis technology is found to be as high as 95.65%. Compared with five other methods, the effectiveness and accuracy of SIF-DADBN under variable working conditions are verified.

Keywords

• variable working conditions
• fault diagnosis
• deep belief network
• domain adaptation
• dropout