Stacked Convolutional LSTM Models for Prognosis of Bearing Performance Degradation

Bearing fault is one of leading causes among industry machine failures. Detection and diagnosis of bearing faults is therefore pretty significant for health monitoring of mechanical equipment. The latest years has witnessed more and more attention to deep learning techniques for prediction and health management (PHM) of bearings. The degradation performance induced by vibration features can be predicted with less requirements on prior knowledge. However, many existing bearing RUL prediction methods have the following two shortcomings: (i) the effects of the entire bearings running time and continuous variation are often ignored; (ii) gradient vanishing or gradient explosion is often involved in learning long term dependency in traditional neural networks. To overcome these problems, a deep structure of bearing degradation prognosis based on staked convolutional and long-short-term memory (CNN-LSTM) has been proposed in the paper. Firstly, convolution neural networks are applied to extract features from the original vibration signals, and then the extracted features are fed to the stacked long-short-term memory networks through the connection layer. Finally, the bearing performance degradation trend is estimated. In this experiment, the bearing dataset of NASA is utilized to verify the prognosis performance of the proposed method. The experiment results show the prognosis effectiveness of the proposed method is prior to traditional CNNs and plain LSTM networks.