| dc.description.abstract | Structural Health Monitoring (SHM) is critical for the safety, durability, and longevity
of critical infrastructures ranging from buildings to very big structures such as wind
turbines, and bridges. In traditional cloud-based SHM systems, high latency, energy
consumption, and low scalability are the challenges. By integrating Machine Learning
(ML) with edge computing via Wireless Sensor Networks (WSNs) leveraging device
learning, we propose a new approach to address these issues. Deep Neural Networks
(DNNs) are directly deployed on edge devices for real-time data analysis and anomaly
detection at sensor nodes using the framework. Thus, it reduces the need for continuous
data transmission to the centralized servers, reduces energy consumption, and improves
system efficiency. Real-time data are collected from key sensors, such as accelerometers
and strain gauges, and processed locally by DNNs. Adaptive retraining is enabled by
drift detection algorithms, which allow response to changing structural conditions. The
findings show that DNNs on the device provide both latency and scalability benefits and
are unable to accurately classify clean as well as noisy sensor data. On-device learning
in combination with adaptive retraining to keep the system accurate and reactive to
changing structural conditions. This proposed system also finds a quantized model
using TensorflowLite, for optimizing DNN deployment on resource-constrained devices,
to reduce computational overhead and memory footprint, while maintaining acceptable
inference accuracy for real-time processing and data transmission. This research also
provides a scalable, adaptive solution for real-time infrastructure monitoring, as well as
new avenues for adaptive re-training, predictive maintenance, and energy harvesting
for Structural Health Monitoring. | en_US |
