Estimating the remaining useful life (RUL) of a bearing is required for maintenance
scheduling. While the degradation behavior of a bearing changes during its lifetime, it is
usually assumed to follow a single model. In this letter, bearing degradation is modeled by a
monotonically increasing function that is globally non-linear and locally linearized. The
model is generated using historical data that is smoothed with a local linear estimator. A
neural network learns this model and then predicts future levels of vibration acceleration to …