Lightweight deep transfer learning for earthquake detection in resource-constrained IoT devices

Abstract

In a natural disaster, intelligent Internet of Things (IoT) systems can be utilized to respond appropriately. Recently, the application of IoT technology in seismology, particularly in earthquake detection, has garnered much attention. This approach’s attractiveness lies in its simplicity of installation, minimal processing power requirements, cost-effectiveness, and expansive coverage, even in areas lacking Internet connectivity. However, the locality of installed sensors brings variations in seismic and noise data, making the earthquake detection task very challenging because of the false alarms. Network-based systems connecting multiple IoTs can resolve the issue by running highly computation-intensive algorithms on a powerful server or cloud and aggregating the data sent from those sensors. On the other hand, Standalone IoT devices operate independently, making decisions locally using both traditional and machine learning methods to manage false alarms. However, these techniques struggle to handle diverse noise patterns and often fail to detect low-magnitude earthquakes in noisy environments. While deep learning models can enhance earthquake detection in such conditions, their high computational cost makes them impractical for resource-constrained devices. To address these challenges, this article introduces a lightweight deep learning model incorporating a transfer learning approach for standalone devices. The proposed model outperforms traditional machine learning methods in earthquake detection using IoT sensors while significantly reducing computational demands. Designed to operate without internet connectivity, the Multi-headed Convolutional Neural Network (MCNN) model achieves 99% accuracy without incurring additional processing costs. Furthermore, it demonstrates high adaptability and the ability to update rapidly with minimal configuration changes.