A multimodal dynamic parameterized bilinear factorized framework for remaining useful life prediction under variational data

Abstract

Remaining useful life (RUL) prediction, as the cornerstone of prognostics and health management, transforms machine maintenance from a passive to a predictive workflow. With the growing availability of sensor types and installations, monitoring data are becoming progressively multimodal, providing comprehensive perspectives on machine health status. In this paper, we propose a dynamic parameterized bilinear factorized framework to tackle some challenges encountered in exploiting multimodal data. To overcome blending and interference among different modalities, a parallel strategy equipped with distillation blocks is adopted to independently process each modality, extracting specialized sensitive features. Subsequently, high-dimensional features are explicitly fused by a bilinear factorized attention module based on their degree of contained degradation-related information. Notably, we design and validate a new loss function called RUL loss, tailored for industrial requirements, considering the asymmetric risk of different prediction errors. Additionally, a dynamic structure distinct from the regular network paradigm is developed, allowing weight parameters to adapt quickly to data variations rather than being fixed during the reasoning process. The superiority of the proposed method is demonstrated using gear life-cycle experimental data. Experimental results show the effectiveness of each component and the overall framework for RUL prediction, showcasing improvements in accuracy and robustness.