A Time-Series Data Augmentation Model through Diffusion and Transformer Integration
Yuren Zhang, Zhongnan Pu, Lei Jing
TL;DR
This work tackles the scarcity of labeled time-series data by proposing a diffusion-to-Transformer augmentation framework, where a diffusion model generates the initial time-step data and a Transformer autoregressively predicts subsequent steps. The diffusion component adapts a Stable Diffusion backbone for time-series with a forward noising step $I^i_t = \sqrt{\hat{\alpha}_t} I^i_{t-1} + \sqrt{1 - \hat{\alpha}_t} \cdot \epsilon$ and a reverse denoising objective $\mathcal{L}(\theta) = \mathbb{E}_{t,I^i_0}[\| I^i_0 - \hat{I}^i_0(I^i_t,t) \|^2]$, while the Transformer incorporates an encoder–decoder structure with a view-mask and a sigmoid output to bound predictions. A tailored weighted loss, defined via $\Delta x_t = x_t - x_{t-1}$ and interval-based weights $w_i$ in $L = \sum_{i=1}^4 w_i \sum_{t\in I_i} \frac{1}{d} \sum_{j=1}^d (y_{t,j} - \hat{y}_{t,j})^2$, is alternated with standard MSE to address early-time emphasis and no-change phases. Evaluations on a custom sign-language dataset collected with a 34-feature data glove show that augmented data substantially improves downstream classification, with peak gains when using larger window sizes in the Transformer component; the results validate the potential of diffusion–Transformer augmentation for time-series tasks. The work highlights a practical, modular pipeline with significant impact on data efficiency, while recognizing the need for future end-to-end integration and more robust authenticity metrics.
Abstract
With the development of Artificial Intelligence, numerous real-world tasks have been accomplished using technology integrated with deep learning. To achieve optimal performance, deep neural networks typically require large volumes of data for training. Although advances in data augmentation have facilitated the acquisition of vast datasets, most of this data is concentrated in domains like images and speech. However, there has been relatively less focus on augmenting time-series data. To address this gap and generate a substantial amount of time-series data, we propose a simple and effective method that combines the Diffusion and Transformer models. By utilizing an adjusted diffusion denoising model to generate a large volume of initial time-step action data, followed by employing a Transformer model to predict subsequent actions, and incorporating a weighted loss function to achieve convergence, the method demonstrates its effectiveness. Using the performance improvement of the model after applying augmented data as a benchmark, and comparing the results with those obtained without data augmentation or using traditional data augmentation methods, this approach shows its capability to produce high-quality augmented data.