Exploring prior-structured low-rank adaptation for predictive maintenance based on time series foundation model

Predictive maintenance for industrial equipment is hampered by nonstationary dynamics and the scarcity of labeled data. In contrast, large foundation models offer the potential for universal sequence representations and parameter-efficient adaptation. Existing practices adapt language or vision-pretrained backbones and inject external semantic knowledge, leaving the patterns of time-domain monitoring signals underrepresented. Moreover, the characteristics of monitoring data with band-limited impulsive transients and degradation trajectories call for domain-specific priors to guide parameter-efficient adaptation. To address these gaps, an exploratory instantiation of a prior-structured adapter (PSA) on the frozen Timer–XL time-series foundation model is presented. PSA modules are attached to selected projections and shaped by two complementary priors: a frequency-enhanced prior that emphasizes band-limited and impulsive signatures of rotating machinery, and a degradation-trajectory prior that encourages monotonic evolution for prognosis. Based on these priors, a condition-aware gate modulates adapter strength across operating regimes while the backbone remains frozen. Evaluation on bearing diagnosis and remaining useful life prediction datasets shows that the proposed PSA achieves competitive or improved results relative to classical baselines and other adaptation methods, with ablation confirming individual component contributions and few-shot experiments demonstrating enhanced data efficiency.