Style Transfer for High-Fidelity Time Series Augmentation

Time series style transfer (TSST), the task of recombining the structural content of one time series with the stylistic characteristics of another, holds substantial promise for applications such as domain adaptation, scenario simulation, and data augmentation. Yet, despite its potential, TSST remains an underexplored area, with existing approaches primarily relying on rigid, heuristic-based techniques that lack generative flexibility. To bridge this gap, we introduce DiffTSST, the first diffusion-based generative framework for time series style transfer. Our method disentangles input sequences into content and style components, representing global structure and local variability, respectively. A conditional diffusion model then synthesizes new time series from noise, guided by content and style components to ensure structural coherence and faithful style integration. Unlike prior methods that merely overlay stylistic features, DiffTSST learns to capture and reproduce underlying style patterns in a data-driven, principled manner. Extensive experiments across varied domains demonstrate the model’s capacity to generate realistic, diverse, and high-quality time series. We further validate its practical utility in a data augmentation task, where it leads to significant improvements in downstream model performance.