A Physics-Informed Deep Learning and Probabilistic Inference Framework for Real-Time Single-Station Earthquake Detection and Magnitude Estimation

Earthquake Early Warning (EEW) systems are important for mitigating casualties and damage to infrastructure by providing seconds of advance notice of damaging seismic waves. However, established EEW systems depend on dense seismic networks, handcrafted features, or multi-station triangulation to find an earthquake’s P-wave arrival time requiring substantial financial investment, time, and infrastructure unsuitable for resource-constrained areas. The contribution of this paper is to propose a hybrid single-station framework, which applies physics-informed preprocessing, deep learning, and probabilistic modelling. The tri-axial accelerometer signals are pre-processed via double integration and bandpass filtering and analysed via a U-Net + + encoder–decoder with dilated convolutions and Multi-Head Self-Attention (MHSA). This allows the U-Net + + architecture to simultaneously fine-tune recognition of spatial-temporal features and to model the global dependency context necessary for accurate P-wave detection. Gaussian label smoothing has been incorporated to adapt and enhance robustness to potential uncertainty in the annotation labels, while a Bayesian Markov Chain Monte Carlo (MCMC) method provides an easy procedure for obtaining probabilistic (and noise-invariant) magnitude estimation. The evaluation across two datasets: the Stanford Earthquake Dataset (STEAD), and IoT-based simulation results illustrate significant margins of improvement, e.g., approximately 10–12% greater F1-scores for P-wave detection, and an approximately 35% lower rate of magnitude-estimation errors. By demonstrating noteworthy improvements in earthquake detection and magnitude estimation, the hybrid single-station framework is a cost-effective and real-time EEW with multiple pathways for deployment at a global scale.