LCANet: Lightweight Context-Aware Attention Networks for Earthquake Detection and Phase-Picking on IoT Edge Devices

Aftershocks monitoring system is a large-scale Internet of Things (IoT) system. Earthquake signal detection and phase-picking, as the core of aftershocks monitoring system, is time-sensitive. Generally, the data generated by seismic stations are transmitted to the cloud for integration, storage, and processing through the network, contributing to network congestion and affecting the real time of earthquake warning as well as the supported applications. Edge computing, a new paradigm for real-time IoT tasks, has emerged as a trend to address concerns of response time, requirement, saving bandwidth costs, as well as data safety and privacy. Therefore, a neglected problem was tackled in this article. First, to meet the real-time requirement, we proposed a novel lightweight deep learning model called lightweight context-aware attention networks for earthquake signal detection and phase-picking. We optimized the deep learning model and reduced the computation requirements to deal with edge devices that have lower computation power than cloud servers. Second, we deployed the model to the Jetson Nano, a small edge device, to offload computation tasks from cloud servers to edge devices. In this way, seismic stations do not need to send raw data to a centralized server. Third, experiments show that our method is robust and easily generalized to other databases. The entire model is only 3.7 MB and achieves the accuracy-latency trade-off.