A Deep Feature Optimization Fusion Method for Extracting Bearing Degradation Features

Abstract

Centrifugal pumps with bearings are widely used in nuclear power plants. A challenge in data-driven prognostic technologies for centrifugal pump bearings is to evaluate their degradation features, which can be combined into degradation trajectory; however, the massive data gathered continuously from condition monitoring systems has created challenges to extracting degradation features effectively. The traditional degradation feature extraction methods are highly reliant on prior knowledge and diagnostic expertise, have limited capacities for learning the complex relationships between the degradation features and the massive amounts of measurement data, and their feature design processes are not automated and require intensive human labor. Deep neural networks offer remarkable abilities to extract features from massive data, and they can be used to automatically extract highly abstracted features that correlate well with bearing degradation. Therefore, to extract centrifugal pump bearing degradation features from massive amounts of vibration data, this paper proposes a deep feature optimization fusion method. First, data from the vibration-frequency-domain are mapped to a nonlinear spatial domain using an enhanced autoencoder. Second, the neural nodes in the last hidden layer of the enhanced autoencoder are divided into several child modules. The minimum quantization errors of each child module are used as the candidate degradation features. Finally, the optimal degradation trajectory is obtained via a weighted fusion of the candidate degradation features. The optimal weighting coefficients are calculated using the grey wolf optimizer algorithm. In this paper, experiments were performed using the IEEE PHM2012 bearing prognostic data set and a centrifugal pump bearing condition-monitoring data set. The results demonstrate that the degradation trajectory obtained using the proposed method offers stronger predictive capabilities than do those obtained using other methods, thereby improving the accuracy of predictions of the bearings’ remaining useful life.