Uncertainty Measurement of the Prediction of the Remaining Useful Life of Rolling Bearings

In the study of the remaining useful life (RUL) prediction of neural networks based on deep learning, most of the RUL prediction models use point estimation models. However, due to the influence of the measurement noise and the parameters in the deep learning model, the prediction results will be quite different, which makes the point prediction meaningless. For this reason, this paper proposes a multi-scale convolutional neural network based on approximate Bayesian inference to realize the credibility measurement of bearing RUL prediction results. First, in order to avoid the problem of insufficient single-scale feature representation, parallel multiple dilated convolutions are used to extract multiple features. At the same time, the channel attention mechanism is used to allocate its importance, which can avoid the redundancy of multi-dimensional information. Then, Monte Carlo Dropout can be used to describe the probability characteristics of the results, so as to achieve the measurement of the uncertainty of the RUL prediction results. Finally, the prediction and health management data set is used to verify that the method has less volatility compared with the traditional point estimation prediction results, which provides a more valuable reference for predictive maintenance.