An AIoT System for Earthquake Early Warning on Resource Constrained Devices

Seismic wave picking is an essential task the implementation of Earthquake Early Warning (EEW) systems. While artificial intelligence picking methods show excellent accuracy, most were designed for devices with high computational resources. At the same time, distributed approaches for Early Warning systems show promise for the implementation of viable, widespread alert systems. This paper introduces a complete AIoT system for earthquake picking on resource-constrained devices. An algorithm has been developed to enable AIoT devices to switch between a detection mode, in which inferences are run on the data measured by the device, and a transmission mode, in which the device transmits alarms or other event information. To reduce inference times and input window duration, a set of deep learning pickers, called Fast PNet, derived from the PhaseNet model, were developed, achieving 74.4% inference time reduction for the shortest input model, which also showed a significant decrease in computational and power consumption. Despite a slight reduction in picking precision compared to the baseline model (from 0.09 to 0.25 s), detection performance remains high (96% precision and 98% recall). The overall system has been tested on a real event from Central Italy, displaying a 0.16-second picking error for the selected event, besides also showing its ability to reduce the amount of data to be processed for transmission to about 26.5% of the total observed input data, a great benefit compared to fully centralized EEW systems.