Edge AI for Structural Health Monitoring: an FPGA-Based Approach on IoT Sensor Nodes

초록

This paper presents an FPGA-accelerated on-device AI framework for Structural Health Monitoring (SHM), enabling real-time vibration analysis using IoT sensor nodes. By integrating hardware-optimized deep learning inference, the proposed system addresses key limitations of server-based SHM approaches, including latency, power consumption, and network dependency. A deep neural network (DNN) was trained on structural vibration data and deployed to an FPGA-based inference engine using quantization and parallel processing techniques to ensure efficient execution at the edge. Experimental validation on a pedestrian bridge testbed demonstrated that the FPGAbased approach achieves classification accuracy comparable to software-based on-device AI (e.g., TensorFlow Lite), while significantly reducing inference time and improving energy efficiency.