Bachelor Thesis: Diffusion Architectures for Synthetic Wearable Sensor Data Generation

If you are interested in this topic, please contact Max Burzer (burzer@teco.edu)

Background

Human Activity Recognition (HAR) using data from wearable sensors (e.g., accelerometers, gyroscopes) is an increasingly active research area with broad applications in healthcare, sports, mobile computing, and human-computer interaction. A key challenge in HAR is the scarcity of large, high-quality labeled datasets, as data collection and annotation are both time-consuming and costly. Additionally, class imbalance further hampers model performance.

To address these challenges, the generation of synthetic wearable sensor data has been actively explored. While Generative Adversarial Networks (GANs) have shown promise [1][2], they often suffer from training instabilities, such as mode collapse. More recently, diffusion models have emerged as a robust and promising alternative. Existing diffusion-based approaches [3] predominantly focus on generating time–frequency representations (e.g., spectrograms) of wearable sensor data, while only a few studies have directly targeted raw time-series data [4]. This focus on two-dimensional representations is largely due to the ability to directly adapt diffusion architectures originally developed for the image domain.

This thesis aims to develop multiple diffusion backbone architectures specifically designed to operate on raw, time-domain wearable sensor data windows. In particular, both U-Net- and Transformer-based (DiT) backbone designs will be implemented and compared to architectures that operate on time–frequency representations. The objective is to evaluate whether direct modeling in the time domain can achieve synthetic data quality comparable to or better than that of time–frequency approaches, while also providing potential efficiency benefits for downstream HAR tasks.

Tasks

  • Conduct a literature review on existing diffusion backbone architectures for time-series data.

  • Design, implement, and train different diffusion backbone architectures for synthetic wearable sensor data generation in the time domain.

  • Evaluate the quality and utility of the generated data by training HAR classifiers on synthetic datasets.

  • Benchmark and compare the performance and computational efficiency of time- against time-frequency–based approaches.

Requirements

  • Solid understanding of deep learning concepts.

  • Proficiency in Python and PyTorch.

  • Knowledge of diffusion models is advantageous but can be acquired during the thesis.

References

  1. Li, X., Luo, J., & Younes, R. (2020). ActivityGAN: Generative Adversarial Networks for Data Augmentation in Sensor-Based Human Activity Recognition. In Adjunct Proceedings of the 2020 ACM International Joint Conference on Pervasive and Ubiquitous Computing and Proceedings of the 2020 ACM International Symposium on Wearable Computers (pp. 249–254). ACM.
  2. Mohammadzadeh, M., Ghadami, A., Taheri, A., & Behzadipour, S. (2025). cGAN-based High Dimensional IMU Sensor Data Generation for Enhanced Human Activity Recognition in Therapeutic Activities. Biomedical Signal Processing and Control, 103, 107476.
  3. Oppel, H., & Munz, M. (2025). A Diffusion Model for Inertial Based Time Series Generation on Scarce Data Availability to Improve Human Activity Recognition. Scientific Reports, 15(1), 16841.
  4. Li, X., Sakevych, M., Atkinson, G., & Metsis, V. (2024). BioDiffusion: A Versatile Diffusion Model for Biomedical Signal Synthesis. Bioengineering, 11(4), 299.