Enhanced prognostic reliability for rotating machinery using neural networks with multi-scale vibration feature learning and uncertainty quantification

Abstract Reliable remaining useful life (RUL) prediction contributes to fault analysis and preventive maintenance of rotating machinery. Existing artificial intelligence methodologies, however, are challenged by inaccurate feature extraction and uncertainty involved in the RUL prediction process. To this end, this paper proposes a reliable fault prognosis method for rotating machinery using neural networks with multi-scale vibration feature learning and uncertainty quantification. Specifically, the proposed fault prognosis framework starts with constructing a multi-scale semantic embedding module to identify the semantic information in mechanical vibrations. A neural network with local and global feature extraction capabilities is then created to capture information from each scale for RUL prediction. By quantifying the uncertainty of predictions, the framework provides a confidence level for each prediction, and therefore a confidence-based RUL decision fusion method is proposed to achieve the reliable RUL estimation. The feasibility, reliability, and superiority of the framework over state-of-the-art methods are validated by datasets from machinery. Overall, the proposed framework contributes to the safe operation and maintenance of rotating machinery systems.