Progressive Bounded Error Piecewise Linear Approximation with Resolution Reduction for Time Series Data Compression.

Area of Science:

• Data Science
• Artificial Intelligence of Things (AIoT)
• Data Compression

Background:

• AIoT devices generate vast amounts of time series data, leading to high energy costs for transmission, storage, and processing.
• Existing lossy compression methods offer better ratios by sacrificing accuracy, but different applications require varying data fidelity.
• Storing multiple compressed versions for different error bounds is inefficient.

Purpose of the Study:

• To develop a method that efficiently generates multiple time series approximations with different error bounds from a single compressed representation.
• To reduce the overall data size and energy consumption associated with storing and managing time series data.
• To dynamically provide data versions tailored to specific accuracy requirements.

Main Methods:

• Progressive decomposition of time series into piecewise linear functions, starting with the largest error bound.
• Utilizing Swing-Recursive Residual (Swing-RR) algorithm to generate Bounded Error Piecewise Linear Approximations (BEPLA) at each decomposition step.
• Aggregating multiple BEPLAs to create approximations for successively smaller error bounds.

Main Results:

• The proposed method, PBEPLA-RR, was evaluated on eight real-world datasets for error bounds of 5%, 1%, and 0.5%.
• The combined data size of multiple BEPLAs was comparable to storing a single version at the smallest error bound.
• This significantly reduced storage requirements compared to maintaining independent compressed versions for each error bound.

Conclusions:

• PBEPLA-RR effectively achieves high compression ratios for time series data.
• The method provides a flexible way to obtain multiple approximations with different error bounds from a single compressed model.
• This approach offers substantial energy and storage savings for AIoT applications.