Generating Synthetic Time Series Data for Cyber-Physical Systems

Deep learning is at the cutting edge of machine learning in fields as critical and diverse as predictive maintenance in industry and anomaly detection in healthcare. Data augmentation can facilitate deep learning applications in these time series domain tasks, which are often data poor by virtue of involving relatively rare systemic malfunction. A literature gap exists around the application of the transformer to data augmentation in time series. This work presents a transformer based architecture hybridizing several successful priors, and a quality testing metric using power spectrum Wasserstein distances. Though promising in theory, results suggest the challenge of this task and domain combination, and highlight several valuable directions for future work.