A new subset based deep feature learning method for intelligent fault diagnosis of bearing

Abstract Intelligent fault diagnosis has attracted considerable attention due to its ability in effectively processing massive data and rapidly providing diagnosis results. However, in the traditional intelligent diagnosis methods of bearing, features are extracted manually. Such process is not only a grueling and time-consuming work but also greatly affects the diagnosis results. In this study, we propose a new intelligent diagnosis method of bearing, which can learn features automatically. First, a new subset approach is developed and it is helpful to learn the discriminative features from different fault patterns. Second, a subset based deep auto-encoder (SBTDA) model is proposed to realize the automatic feature extraction. Additionally, a new self-adaptive fine-tuning operation is designed to ensure the good convergence performance of SBTDA. Finally, to obtain the appropriate configuration, several key parameters are optimized with particle swarm optimization algorithm. The proposed method is evaluated on three public bearing datasets, and achieves the average testing accuracies of 99.65%, 99.66% and 99.60% respectively. The comparisons with 13 intelligent diagnosis methods demonstrate that SBTDA can obtain higher diagnosis accuracy.