Bio-VQVAE: multivariate and multi-scale time-series generation for bioprocess data augmentation

Biomanufacturing digitalization relies heavily on data-driven modeling, yet the scarcity of high-quality labeled data remains a bottleneck. Multivariate fermentation data exhibit temporal patterns at multiple scales and complex statistical dependencies among process-state and operational variables. Current time-series generation methods often have difficulty preserving multi-scale temporal patterns and statistical dependencies among process variables. This paper proposes Bio-VQVAE, a generative framework designed to improve the representation of these characteristics in a vector-quantized latent space. We introduce a Large Selective Kernel mechanism to adaptively capture temporal patterns at different scales by dynamically adjusting the effective receptive field. Furthermore, an Efficient Multi-Scale Attention module is integrated to enhance cross-dimensional feature interaction and facilitate the empirical learning of dependencies among process variables. Evaluation on the IndPenSim benchmark shows that Bio-VQVAE achieves the lowest mean values across all four quantitative metrics among the evaluated baseline models. Notably, the method achieves superior fidelity in preserving multivariate correlation structures and temporal autocorrelation consistency, even under constrained data regimes. These results demonstrate the potential of Bio-VQVAE for in-distribution augmentation of multivariate bioprocess time series and for supporting downstream predictive modeling.