Temporal latent diffusion model for machine degradation trend forecasting

Predictive maintenance technology is critical for maximizing machine utilization and saving manufacturing costs. Conventional remaining useful life prediction techniques often suffer dilemmas in obtaining accurate life labels in scenarios with extremely scarce degradation data. Meanwhile, traditional deep time-series prediction models are prone to overfitting or lack generalizability under data-scarce conditions. Moreover, diffusion models (DMs) designed for data scarcity issues often struggle to capture temporal dependency information. In light of the above challenges, we propose a novel temporal latent diffusion model (TLDM) for machine degradation trend forecasting (DTF), which relies solely on limited historical monitoring data to forecast the future degradation trends of machines. Specifically, we first develop a temporal pre-enhancement module based on bidirectional gated recurrent units (BiGRU), replacing the commonly used temporal inputs with hidden states of BiGRU to guide the latent diffusion model (LDM) training in a low-dimensional space. Furthermore, we introduce a soft conditional guidance-based temporal sampling strategy, providing more comprehensive temporal dependency information for DTF from global and local perspectives. Extensive experiments on two mechanical platforms indicate the effectiveness and superiority of our method for machine DTF.