MicroAttn: Explicit Feature Reweighting for Noise-Robust Predictive Maintenance

Predictive Maintenance (PdM) models deployed in industrial environments often suffer from noisy sensor measurements, severe class imbalance, and nonstationary operating conditions. While attention mechanisms have been widely adopted to improve representation learning, they are typically embedded within complex architectures and offer limited insight into feature-level robustness. This paper introduces MicroAttn, a lightweight, model-agnostic feature-level attention mechanism that adaptively reweights sensor inputs under noisy conditions. MicroAttn can be integrated as a simple preprocessing layer without modifying the underlying model architecture. Experiments on a real-world industrial dataset using strict time-based evaluation demonstrate that MicroAttn improves fault-ranking performance and yields stable, interpretable feature weights across multiple training seeds. The results indicate that explicit feature-level attention provides a practical and efficient approach for enhancing robustness in predictive maintenance systems.