Bi-LSTM neural network for remaining useful life prediction of bearings

Rolling bearing is a key part of rotating machine and its healthy condition is of significance on safety in production. The prediction for operating condition and residual lifetime of the rolling bearing is one of main challenges in intelligent diagnosis field. In order to attain the whole process of rolling bearing degradation, a method of Convolution Autoencode with improved loss function is proposed in this paper. The proposed method can obtain the condition of rolling bearing from vibration signals collected by multi-sensors avoiding the loss of local information as well as achieving fault character in deeper layer. Then a cyclic neural network structure based on bi-directional long and short time memory (Bi-LSTM) is suggested in this paper to learn the principle of rolling bearing degradation in practical work by means of its ability to process the time series data, which realizes the residual lifetime prediction of the rolling bearing. In addition, with the aim of improving the prediction accuracy and ability to be used widely of model, the Bi-LSTM network is trained by receiving the sample with random length to make the model accept continuous data instead of segmented data. Finally, the IMS data set from NASA is utilized to operate experiment and comparative test. The result shows that the proposed prediction model of rolling bearing lifetime based on CE-Bi-LSTM exhibits higher precision than that of other methods. © 2021, Editorial Board of Journal of Vibration Engineering. All right reserved.