An Evolving Multivariate Time Series Compression Algorithm for IoT Applications

Area of Science:

• Computer Science
• Embedded Systems
• Data Compression

Background:

• Internet of Things (IoT) systems face challenges with high latency and energy consumption due to large real-time data transmission.
• Tiny Machine Learning (TinyML) offers a solution by enabling machine learning on resource-constrained embedded devices.
• Vehicle monitoring applications within IoT generate substantial data, exacerbating transmission issues.

Purpose of the Study:

• To develop and evaluate two novel online multivariate data compression approaches for TinyML applications.
• To leverage the Typicality and Eccentricity Data Analytics (TEDA) framework for data compression.
• To optimize compression performance without relying on predefined mathematical models or data distribution assumptions.

Main Methods:

• Developed two online multivariate compression techniques based on data eccentricity within the TEDA framework.
• Applied the compression approaches to the OBD-II Freematics ONE+ dataset for vehicle monitoring.
• Evaluated both parallel and sequential compression strategies.

Main Results:

• Both proposed compression approaches demonstrated significant improvements in execution time.
• The methods achieved notable reductions in compression errors.
• The study confirmed the effectiveness of eccentricity-based compression for TinyML.

Conclusions:

• The developed compression approaches enhance the performance of embedded IoT systems, particularly in vehicular applications.
• These findings contribute to improving the efficiency and sustainability of real-time data processing in resource-constrained environments.
• The TEDA framework provides a robust foundation for developing advanced data compression techniques for TinyML.