MTS-GAN: Synthetic multivariate time series generation with the Mamba architecture

The availability of high-quality data is fundamental to the development of robust machine learning workflows. Nevertheless, in complex real-world settings, acquiring qualitative datasets is often both costly and time-consuming. This limitation frequently necessitates the generation of synthetic data, not only to address data scarcity issues but also to provide a mechanism for information sharing while preserving privacy. This paper presents a novel method for multivariate synthetic time-series generation based on a Mamba architecture, namely MTS-GAN, that generates realistic and long-span multivariate time series. Our approach outperforms existing GAN-based methods in generative quality while overcoming the significant computational and inference-time limitations inherent in state-of-the-art diffusion models such as Diffusion-TS. Through a comparative analysis, we demonstrate that MTS-GAN achieves slightly superior results to Diffusion-TS across the majority of tested scenarios, but with a much more efficient architecture that scales linearly with sequence length. We validate our approach using both open-source datasets as a foundational benchmark and a real-world case study centered on predictive maintenance for a milling machine, confirming its practical efficacy and superior efficiency in handling long-sequence data. • Mamba-based Generative Adversarial Network (GAN): We introduce a novel Mamba-powered GAN architecture that achieves state-of-the-art performance in multivariate time series generation, with fully reproducible and open-source code. • Enhanced Evaluation Framework: We propose improved statistical metrics to more accurately and robustly assess the quality of generated sequences, overcoming the limitations of conventional evaluation methods.