A synthetic dataset for time series super-resolution with deep learning.

Area of Science:

• Signal Processing
• Machine Learning
• Data Science

Background:

• Time-series analysis is crucial in fields like biomedical engineering and telecommunications.
• High-quality data is essential for training advanced machine learning models.
• Data acquisition at suitable resolutions faces ethical, economic, and practical limitations.

Purpose of the Study:

• Introduce CoSiBD (Complex Signal Benchmark Dataset for Super-Resolution).
• Provide a synthetic dataset for reproducible time-series super-resolution research.
• Address the need for accessible, high-resolution temporal data.

Main Methods:

• Generated 2,500 high-resolution synthetic signals (5,000 samples each).
• Created aligned low-resolution versions via uniform decimation (150-1,000 samples).
• Incorporated diverse non-stationary behaviors using frequency modulation and spline-based amplitude envelopes, including clean and noisy variants.

Main Results:

• Dataset distributed in NumPy, plain text, and JSON formats.
• Comprehensive metadata ensures full reproducibility (segment structure, generation parameters, seeds).
• Technical validation includes spectral analysis and baseline super-resolution benchmarking.

Conclusions:

• CoSiBD facilitates reproducible research in time-series super-resolution.
• The dataset supports benchmarking and transfer learning experiments.
• Enables advancement in machine learning model development for time-series data.