Data-driven modeling of noise time series with convolutional generative adversarial networks.

Area of Science:

• Signal Processing
• Machine Learning
• Data Analysis

Background:

• Random noise is inherent in physical measurements and limits signal processing.
• Generative adversarial networks (GANs) show promise for data-driven modeling.
• Assessing GANs' ability to reproduce noise in time series is crucial.

Purpose of the Study:

• Empirically investigate the capability of GANs to faithfully reproduce various noise types in time series data.
• Evaluate two deep convolutional GAN architectures for time series noise generation.
• Provide insights into GAN limitations for noise modeling and establish a benchmark for future research.

Main Methods:

• Trained and evaluated two general-purpose time series GANs (direct and image-based) on simulated noise.
• Utilized short-time Fourier transform for the image-based GAN data representation.
• Tested GANs on diverse noise distributions: band-limited thermal, power law, shot, and impulsive noise.

Main Results:

• GANs successfully learned several noise types, demonstrating capability in noise modeling.
• GAN performance degraded with noise types ill-suited to the architecture, such as impulsive noise with extreme outliers.
• The study identified specific limitations of current time series GANs.

Conclusions:

• GANs show potential for modeling various time series noise characteristics.
• Architectural suitability is key for GANs to accurately reproduce complex noise patterns.
• This work offers a benchmark for developing advanced deep generative models for time series noise.