TinyHeal: TinyML-Assisted Predictive Link Degradation Recovery for Self-Healing IoT Mesh Networks

Tiny-scale Internet of Things (IoT) deployments increasingly rely on multihop mesh networking, yet their performance remains highly sensitive to wireless link variability caused by interference, mobility, and environmental fluctuations. Existing linkquality estimation techniques offer limited early-warning capability, and reactive routing protocols often incur costly route repairs, leading to packet loss, unstable forwarding paths, and excessive control overhead. To address these constraints, this work introduces TinyHeal, a fully on-device TinyML-assisted self-healing communication framework that predicts link degradation before failures occur and proactively stabilizes routing decisions. The proposed design integrates lightweight temporal feature extraction, micro-model inference, and a prediction-driven routing state machine engineered for microcontrollerclass IoT nodes. Extensive experiments on four real-world and large-scale datasets—Intel Lab, GreenOrbs, FIT IoT-LAB, and IoT-RPL—demonstrate that TinyHeal achieves higher prediction accuracy, reduces parent-switch frequency, improves packet delivery ratio by up to 18%, and lowers control overhead by more than 50% compared with state-of-the-art LQE and ML-based baselines. A robustness and sensitivity analysis further confirms that TinyHeal maintains strong reliability-energy tradeoffs across a wide range of operating parameters, validating its suitability for resource-constrained and dynamically varying IoT mesh environments.