Simulation-based remaining useful life prediction of rolling element bearings under varying operating conditions

Remaining useful life (RUL) prediction of rolling element bearings is a complex task in the frame of condition monitoring which brings cost benefits to the industry by reducing unexpected downtimes and failures. Data-driven approaches based on deep learning have demonstrated exceptional performance in estimating RUL effectively. Nevertheless, challenges such as data scarcity for model training and varying operating conditions add more complexity to prognostic tasks using these methods. This study proposes a methodology for simulating the vibration signals during the degradation process of bearings in order to mitigate the need for historical data for training the models. Simulations are realized using a phenomenological model whose free parameters are adapted based on real measurements so that the simulated run-to-failure datasets are under the same influence of speed as the real dataset with almost the same degradation rate. The simulated dataset is used for model training. Moreover, the proposed methodology is able to react to the shaft speed and be flexible at the predictions when the speed of the bearing varies. The proposed model can take extra information regarding the operating speed and the sequential ordering of the measurements to be aware of the working conditions and the dynamics of the damage progression. The positive effect of the extra information is shown in the results. Model training is based on an unsupervised domain adaptation approach to reduce domain discrepancy between the simulated and real feature space. The effectiveness of the proposed method is examined according to bearing run-to-failure tests under varying operating conditions.