Enhanced predictive modeling of rotating machinery remaining useful life by using separable convolution backbone networks

Accurate prediction of Remaining Useful Life (RUL) is a critical aspect in the field of prognostics health management (PHM). Striking a balance between prediction precision and model complexity is a substantial challenge when deploying deep learning (DL) methods in PHM. In response to this challenge, the present study introduces a novel approach called the Separable Convolution Backbone Network (SCBNet), which is designed to address the intricate mapping between degradation patterns and RUL by leveraging a structure based on separable convolutions. This innovative architecture aims to enhance prediction accuracy without unduly increasing the model intricacy. Furthermore, a backbone network is introduced to amplify the mapping features obtained from separable convolutions. To further strengthen the model, a novel strategy is devised to seamlessly integrate adjacent backbones. Through empirical experiments conducted on two bearing lifecycle degradation datasets, the proposed SCBNet demonstrates remarkable superiority over existing mainstream methods in terms of both prediction accuracy and model complexity. This study contributes valuable insights and a practical solution to enhance the effectiveness of DL-based methods in PHM applications. •Proposing SCBNet for balancing accuracy and model complexity in RUL prediction.•Introducing a separable convolution-based structure to capture more information.•Designing a backbone network to streamline the number of model parameters•Designing strategy to assemble adjacent backbones for a more robust model.•Evaluating proposed method via experiments on bearing degradation datasets.