A Novel Bearing Health Prognostic Method Based on Time-frequency Analysis and LSTM

Abstract

Bearing plays an enormous role in modern factories. It is of great value to monitor the bearing health conditions. With the rapid development of data acquisition and transmission technology, it has been available to collect, transmit and store enormous operating data of the bearings, which will reflect the health conditions thoroughly. Deep learning is a promising method to completely extract hidden information from the big data and Recurrent Neural Network (RNN) is designed to process time relations of the input signals. However, there are still some shortcomings on the vanilla RNN, and Long Short-Term Memory (LSTM) neural network is proposed to optimize the model. In this paper, a new method based on time-frequency analysis and LSTM is proposed for bearing prognosis. Firstly, vibration raw signals are collected. Then, a time-frequency analysis will be applied on these raw signals. LSTM neural network is established and the time-frequency information is fed into the target network to train the target one or diagnose the bearing health state. Selection of some vital parameters in our model is discussed in detail, and then the proposed method is compared with other methods, such as vanilla RNN, MLP and SVM. Result shows that the proposed method achieves the highest accuracy rate over other methods.