Binary Neural Networks for Real-Time Structural Health Monitoring on Edge Devices

Structural Health Monitoring (SHM) is critical to ensuring the safety of civil structures. SHM has been mostly addressed through cloud/workstation solutions, but new approaches are also considering edge devices, whose proximity to data sources allows minimizing latency and bandwidth. Typical SHM methods demand extensive data processing and high computational resources, that cannot be sustained by edge devices. In this paper, we investigate the application of binary neural networks (BNNs) to the SHM problem, exploiting their efficiency and reduced computational cost, without too much degrading classification performance, which makes them a very promising solution for edge computing. Using the Z24 Bridge dataset, our results show that BNNs can achieve state-of-the-art accuracy while reducing model size by approximately 90×, energy consumption by a factor 14, with a 12× latency cut, allowing real-time operation on a Raspberry Pi 4 device.