Multi-modal adaptive weighted distillation for industrial equipment health monitoring

Conventional knowledge distillation in industrial equipment health monitoring suffers from three critical bottlenecks: geometric misalignment of heterogeneous modalities, prohibitive offline storage overhead, and static parameters mismatching non-stationary data streams. This paper presents the first multi-modal adaptive weighted distillation (MAWD) framework that systematically addresses these limitations via three original contributions. First, heterogeneous vibration and acoustic emission features are mapped onto symmetric positive definite manifolds, where Riemannian distance and Karcher mean enable geometry-preserving cross-modal alignment beyond Euclidean space. Second, a dual-layer multi-teacher collaborative mechanism is proposed, assigning modality-exclusive teachers for feature extraction and a cross-modal teacher for relational knowledge fusion, achieving two-level knowledge transfer of both predictive distributions and intermediate representations. Third, a sliding-window-based online adaptation module dynamically calibrates distillation temperature and modal weights in real time, eliminating offline storage costs and mitigating knowledge drift caused by evolving equipment conditions. Experiments on three case studies show that MAWD outperforms state-of-the-art multimodal distillation methods by 7.4% in fault diagnosis accuracy on industrial datasets, while achieving a 68% model size reduction, 92% storage overhead decrease, over 50% energy consumption reduction on edge devices, and up to 748% inference throughput improvement. Ablation studies validate the independent and synergistic contributions of each core module, with SPD manifold alignment providing over 13% accuracy gain.