Efficient edge intelligence for onboard data anomaly classification in wireless structural health monitoring using knowledge distillation on low-cost IoT nodes

Abstract

Internet of Things (IoTs) systems with wireless data transmission have been increasingly adopted by structural health monitoring. However, sensor malfunctions and data anomalies have been identified frequently in these systems for long-term deployment, especially in harsh environments. Advances in machine learning (ML) offer promising solutions for effectively detecting, classifying and recovering data anomalies. Most state-of-the-art ML-based solutions rely on powerful workstations or cloud servers, where large amounts of datasets are assumed to be transmitted to a centralized station, leading to the issues of heavy load of data transmission and long delay of subsequent data analytics. To address the issues, this article proposed a novel decentralized approach integrating machine learning with edge computing, aka, edge intelligence, for efficient onboard data anomaly classification. The key innovations lie in the development of an effective ML-based feature extractor adapted to low-cost IoT nodes and an efficient edge computing deployment for ML execution suitable for resource-constrained microcontroller unit-based devices. The proposed solution has been successfully deployed on a newly developed IoT node, LiftNode, which is by first time to realize low-cost onboard data anomaly classification in structural health monitoring (SHM) community. Using the dataset collected from a real long-span cable-stayed bridge in China, the proposed solution is found to achieve the accuracy of 96.37% in PC with the model size of 0.29 MB, and comparably, the accuracy of 96.36% in LiftNode. The framework and deployment strategies are elaborated, highlighting the potential of edge intelligence for improving SHM systems’ responsiveness and resource utilization.