Evaluating the Influence of Time Domain Feature Distributions on Estimating Rolling Bearing Flaking Size with Explainability

To enhance the maintainability of rotating machines, such as wind turbines, where the response to bearing damage is both costly and time-consuming, it is essential to predict the progression of flaking, which is a common rolling bearing fault. Conventional rule-based methods estimate the magnitude of flaking by analyzing the time interval of feature vibrations. However, this method requires trial-and-error adjustments by experts, limiting its applicability to a wide range of rotating machines. To overcome this limitation, we developed a deep learning-based estimation model and demonstrated that its performance depends on the distribution of time-domain features in the training data, which are associated with flaking damage. We then analyzed the manner in which these feature distributions impose limitations on the estimation accuracy of the model. Additionally, we incorporated explainability using Grad-CAM to verify that the extracted features were aligned with the physical phenomena of flaking damage, thereby confirming the link between the feature vibrations and estimation results. Our experiments under various training–test split conditions indicate that time-domain shifts of these features affect the model’s performance, providing insight into how feature distributions constrain the estimation of the flaking size.