The Impact of Varying Speed on Remaining Useful Life Estimation for Rolling Element Bearing
Most data-driven bearing remaining useful life (RUL) estimation studies assume constant rotational speed, which does not reflect real industrial operating conditions. In applications such as electric vehicle drivetrains, bearings operate under continuously varying speed profiles that modulate vibration signal characteristics and obscure degradation signatures, making RUL estimation significantly more challenging. This study presents an experimental investigation of the effect of speed non-stationarity on deep learning-based bearing RUL estimation. Two run-to-failure accelerated life tests were conducted on SKF 6203-2RSH/C3 deep groove ball bearings: one under a constant rotational speed and one under a time-varying speed profile derived from the Worldwide Harmonised Light Vehicles Test Procedure (WLTP). Average speed and radial loading were kept identical across both experiments to isolate the effect of speed variability. A 1D residual convolutional neural network (CNN), a gated recurrent unit (GRU) network and an attention-augmented bidirectional gated recurrent unit network (Att-BiGRU) were evaluated under four training-test scenarios covering both within-condition and cross-condition configurations. Within-condition results show that all models can learn meaningful degradation patterns even under non-stationary conditions when training and test data are matched, yielding RMSE values of 12.93, 16.58 and 12.37 cycles for the CNN, GRU and Att-BiGRU respectively under the WLTP regime. These results demonstrate that variable speed operation represents the general case for industrial bearing prognostics and that dedicated strategies at the feature extraction, model architecture, and training levels are required to develop reliable RUL estimators for real-world applications.