Abstract
Abstract
The remaining useful life (RUL) of bearings in rotating machinery is continuously affected by time. To address this concern, an improved model based on gated recurrent unit is proposed by taking full advantage of the characteristics of recurrent neural networks to efficiently process sequence data. This model is then applied to different prediction scenarios. First, to construct training and test sets, the required feature data are extracted from the vibration signals of rolling bearings. A health indicator (HI) is required to be constructed as a label for indirect prediction, whereas RUL is directly used as a label for direct prediction. The model is then allowed to learn through training sets to determine its optimal parameters. Finally, test sets are used to predict HI or RUL step by step. The effectiveness and superiority of the novel model in indirect and direct predictions is demonstrated by the comparison of evaluation indexes for prediction results with lower prediction deviations than conventional methods.
Funder
National Natural Science Foundation of China