A remaining useful life prediction method for bearing based on deep neural networks

With the purpose of improving the prediction accuracy and generalization ability of remaining useful life (RUL) prediction models, this paper proposes a new method to predict the RUL of bearings based on the convolutional neural network (CNN). First, the 3 sigma criterion is applied to denoise the original data and remove gross errors. Subsequently, the frequency features are obtained from the original data by the fast Fourier transform (FFT), and the root mean square is employed as the tracking metric. Then, stratified sampling, which differs from the traditional time series data partitioning method, is applied to data partitioning to completely learn the data features. A deep convolutional neural network (DCNN) model without a pooling layer, which consists of three convolutional layers and two fully connected layers, is constructed to avoid feature loss. Finally, the NASA IMS dataset is utilized to assess the preprocessing method, DCNN accuracy and generalization ability. The experimental results show the effectiveness of the proposed method.