Inertial Signal Forecasting with Foundation Model Techniques

This paper introduces Dual-View FM, a generative pre-trained Foundation Model (FM) specifically designed for forecasting human inertial signals from body-worn sensors. Unlike traditional Transformer-based approaches, Dual-View FM employs a novel architecture that mixes global and local signal properties in the time and feature domains to build a comprehensive understanding of inertial activity data. The global view identifies general trends in human inertial signals, while the local view addresses their spontaneous nature by individually analyzing short, sensor-specific signal frames. Our evaluation on eight established datasets shows that Dual-View FM accurately forecasts human inertial signals from previously unseen test datasets that involve various sensors, activities, and users. The pre-trained models even outperform a production-ready FM for time series forecasting by up to 36 % in mean absolute error (MAE), demonstrating the advantages of domain-specific FMs. Fine-tuning further enhances Dual-View FM’s forecasting capabilities so that it surpasses other state-of-the-art architectures by more than 7 % in MAE.