A diffusion model for inertial based time series generation on scarce data availability to improve human activity recognition

• Published in: Scientific reports Vol. 15; no. 1; pp. 16841 - 15
• Main Authors: Oppel, Heiko, Munz, Michael
• Format: Journal Article
• Language: English
• Published: London Nature Publishing Group UK 15.05.2025; Nature Publishing Group; Nature Portfolio
• Subjects: 631/114/1305; 639/705/117; Accelerometers; Accelerometry - methods; Adult; Algorithms; Classification; Datasets; Diffusion Model; Diffusion models; Female; Fourier transforms; Human Activities; Human activity recognition; Humanities and Social Sciences; Humans; Male; Movement - physiology; multidisciplinary; Science; Science (multidisciplinary); Sensors; Synthetisation; Time series; Diffusion Model; Synthetisation; Human activity recognition; Time series
• ISSN: 2045-2322; 2045-2322
• DOI: 10.1038/s41598-025-01614-x

The domain of human activity recognition is able to differentiate between human movements based on sensory driven systems, e.g. in the form of an IMU. Though, in order to perform those differentiation tasks, a measurement setup has to be established and subjects have to be recorded. As this is a time and cost consuming process, research groups are focused to synthetically generate data resembling human movements to improve the underlying recognition task. So far, work groups are able to generate univariate and multivariate synthetic sequences on basis of an accelerometer or six axis IMU. Yet, they lack in generalizing on unseen subjects and are not able to expand further than a single six axis IMU. In this paper, we aim to fill this gap by using the backbone of a denoising diffusion probabilistic model from the vision domain to synthetically generate multiple IMUs which are able to generalize on unseen participants. The model was adapted to fulfill the criteria of generating meaningful human motion sequences. We then evaluated the quality of the data in two ways: (1) by a subjective visual analysis with the help of a clustering approach new to this domain and (2) by the classifier improvement when adding synthetic samples. The results show a significant improvement in the classification task when synthetic samples were added to the pool of training data. One of the key findings is the benefit of improvement, even in a scarce data set of only 2 samples per subject. This is a huge advantage in the domain of HAR as it reduces the time of a subject to perform a task.