Edge-VisuAlign: A Lightweight Virtual Sensor for RUL Estimation via Cross-Modal Knowledge Distillation

Accurate estimation of Remaining Useful Life (RUL) is critical for Predictive Maintenance (PdM) in Industry 5.0. While multimodal deep learning models achieve high accuracy by integrating one-dimensional (1D) telemetry signals and two-dimensional (2D) visual representations, their substantial computational demands hinder deployment on resource-constrained Internet of Things (IoT) devices. To overcome this challenge, Edge-VisuAlign is proposed as a lightweight virtual sensor framework tailored for edge computing. This framework employs a cross-modal knowledge distillation strategy to transfer physics-aware feature representations from a complex multimodal teacher network to a highly efficient, univariate student network. Evaluation on the NASA C-MAPSS dataset demonstrates that the edge-student model sustains robust predictive performance (RMSE of 16.22) while achieving a 99.58% reduction in trainable parameters compared to its teacher (from 8.7M to 37k). This significant compression enables reliable, real-time prognostics on embedded microcontrollers, thereby advancing sustainable and resilient industrial monitoring.