SensorGPT: A New Paradigm for Synthetic Sensory Data Generation

Human sensing applications are vital across diverse domains, including healthcare, fitness tracking, and security. These applications rely on the availability of high-quality annotated data, which is essential for training robust and accurate models. Sensory data synthesis methods have emerged as a promising solution for obtaining annotated data, avoiding practical and ethical constraints of real-world data collection. However, existing methods suffer from several limitations. A key issue is the limited diversity of the synthesised datasets, often stemming from deterministic mapping processes that restrict the richness of the generated data. Furthermore, integrating multiple target synthetic modalities, such as simultaneously generating accelerometer and gyroscope data, remains a challenge.This paper introduces SensorGPT, a novel paradigm that addresses these limitations, by breaking the synthesis process into three key sub-processes: 1) generating the physical phenomenon to be perceived by the target sensors (e.g., human motion), 2) analytically mapping the physical phenomenon to the target sensor measurements (e.g., accelerometer), and 3) applying unsupervised domain adaptation to align the synthetic data with real-world distributions, eliminating the need for labelled real-world data. We evaluate the effectiveness of SensorGPT in a human activity recognition application, where it generates accelerometer and gyroscope data from simple text descriptions of actions. Compared to the current state-of-the-art, SensorGPT achieves a threefold improvement in accuracy on benchmark datasets such as UCI, HHAR, and MotionSense. By combining the capabilities of deep generative models with the reliability of classical theoretical models, SensorGPT sets a new standard for the next generation of synthetic data generation techniques.