APLASE: Compression using Adaptive Piecewise Linear Approximation and Sparse Encoding

Abstract

This work focuses on compressing vast amounts of time series data from IoT sensors while achieving low reconstruction error, at a low compression ratio (ratio of output data size to input data size), for efficient storage and transmission. We investigate two lossy compression techniques: Adaptive Piecewise Linear Approximation (APLA) and Sparse Encoding (SE). APLA reduces data volume by leveraging the temporal correlation in time series data, while SE captures small, complex variations that APLA might miss. We identify a signal characteristic parameter, denoted by σ<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">d</inf>, which is used to derive error bounds for SE using σ<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">d</inf>, the dictionary used, and the input compression parameters. We also introduce APLASE, a novel compression technique that combines the two techniques by applying SE to the reconstruction error generated by APLA. We evaluate APLA, SE and APLASE using data from the NASA flight data recorder database consisting of recordings from 116 sensors. Our results show that APLASE consistently outperforms both APLA and SE individually, as well as the recent QoZ algorithm, particularly at lower compression ratios.