A physics-informed integrated prognostic framework for high-speed components in filter rod making machines using multi-feature fusion and DL-BiGRU-Attention

Abstract

The main transmission box is the critical power core of filter rod making machines operating under ultra-high-speed conditions of 600 m min−1, where frequent sticking faults severely disrupt production continuity. To bypass the prognostic bottlenecks of insufficient multi-physics feature coupling and fragile early degradation identification, this paper proposes a physics-informed integrated prognostic framework. First, a Dirichlet random weighting mechanism is developed to comprehensively evaluate the monotonicity, correlation, and robustness of features, guaranteeing the non-biased identification of robust degradation indicators. Second, multi-dimensional thermodynamic and kinematic variables are fused via structural pooling to synthesize a high-fidelity unified health indicator. To overcome industrial sample scarcity, a physics-informed data augmentation strategy expands the training repository by generating evolutionary components compliant with kinematic degradation laws. Finally, a deep hierarchical sequence network combining dual-layer bidirectional gated recurrent units with a self-attention mechanism (DL-BiGRU-Attention) captures multi-scale temporal variations. Empirical evaluations on actual ZL29 trajectory data demonstrate that the proposed framework consistently delivers a coefficient of determination (R2) of 0.9914, with suppressing the mean squared error a minimal magnitude of 1.18 × 10−4. Architectural ablation benchmarks further confirm that the full framework outpaces traditional recurrent networks by precisely capturing critical transient mutation phases, providing a dependable, interpretable mathematical vehicle for condition-based maintenance of high-speed industrial transmission systems.