Enhancing Earthquake Preparedness in the Himalayan Region: A Machine Learning Approach using EEW System Parameters

Earthquakes pose serious risks to global populations, making early warning systems essential. Conventional methods relying on peak ground displacement and velocity often struggle with the non-stationary nature of seismic signals. This study leverages first-arriving P-wave signals to extract $$\tau _c$$ , $$P_d$$ parameters for earthquake event classification and proposes an ensemble-based machine learning architecture that outperforms existing classifiers. While the current state-of-the-art single model achieves 91% accuracy, our approach employs three different ensemble models, and finally proposed model (Architecture-3) delivers 96.56% accuracy, representing a notable improvement with a reduced false alarm rate, accomplishing a false positive rate of 1.96%, enhancing reliability for real-time earthquake early warning (EEW) in the Himalayan region. All models are deployed on the PYNQ-Z2 FPGA platform using IIT Roorkee’s PESMOS data, again our proposed Architecture-3 outperform all, achieving 8.1 ms inference latency, 3.0 W power consumption, and moderate PS utilization (48% CPU, 58% memory), confirming feasibility for real-world implementation. These results highlight the potential of ensemble ML for robust and efficient EEW systems.