Optimized Deep Siamese Network for Structural Health Monitoring Using Vibration-Based Measurements

Abstract

Structural health monitoring (SHM) is essential for evaluating infrastructure conditions and ensuring its safety. Over time, the structural integrity of buildings and civil infrastructure can deteriorate, leading to a decline in performance and reliability. SHM has increasingly used deep learning (DL) models for effective damage detection. However, for onsite damage detection of smart structures, there is a need to deploy these trained DL models on resource-constrained devices such as Raspberry Pi, Arduino, mobile devices, or microcontrollers. To accomplish this task, the trained DL models need to be optimized for deployment that enables efficient and timely damage detection in realworld applications by testing the vibration (acceleration) measurements coming from the sensors placed over the structures. As per the knowledge of the authors, this is a crucial aspect that has yet to be fully explored in SHM. In the present work, vibration signals are initially transformed into a 2D matrix format. These transformed 2D matrices are used as input to a DL model, specifically the deep-siamese network (DSN). The DSN operates based on similarity measures between pairs of training samples. This enables the proposed SHM approach to detect and report previously unseen damage scenarios during testing. To enable real-time inference, the DSN model is optimized using pruning, quantization, and knowledge distillation. These techniques reduce the size of the model and the inference time, making it suitable for deployment on resource-constrained devices. To validate this approach, the IASC-ASCE benchmark structure with varying damage conditions is used. The results show that the optimized model maintains high detection accuracy while enabling fast and efficient decision making. Each optimization technique is used to convert the DSN model into a lightweight (.tflite) format. The model is then deployed on a virtual machine running Raspberry Pi OS with Debian Linux. These optimizations support real-time structural health monitoring with high detection accuracy. The model also handles unseen damage during testing, making it suitable for real-world deployment. This makes it a promising solution for real-world SHM, bridging the gap between DL advancements and practical online monitoring.