TinyML: Accelerometer-Based Processing for Seismic Detection in Buildings

During emergencies, fast and accurate situational awareness of building collapses supports critical decisions that can save human lives. To tackle this problem, we develop a compact on-device learning model (TinyML) that infers seismic activity from accelerometer readings collected in buildings. Indeed, we explore classification approaches, including Support Vector Machine (SVM), Random Forest (RF), k-Nearest Neighbors (kNN), and Decision Tree (DT) to detect anomalies in buildings. We also provide a performance comparison using the Designsafe-CI dataset. Our simulation results revealed that both kNN and DT achieved the highest scores among the models for predicting building anomalies, with accuracies of up to 98%. Given these outcomes and the models’ computational footprint, we selected the DT model for the final deployment on the STM32 microcontroller.