Master Thesis Defense: Bo Li
Speaker: Bo Li
Supervisor: Dr. T. Glatard
Examining Committee: Drs. B. Jaumard, E. Shihab, O. Ormandjieva (Chair)
Title: Multi-dimensional Data Stream Compression for Embedded Systems
Date: Tuesday, July 30, 2019
Place: EV 2.260
The rise of embedded systems and wireless technologies led to the emergence of the Internet of Things (IoT). Connected objects in IoT communicate with each other by transferring data streams over the network. For instance, in Wireless Sensor Networks (WSNs), sensor-equipped devices use sensors to capture properties, such as temperature or accelerometer, and send 1D or nD data streams to a host system. Power consumption is a critical problem for connected objects that have to work for a long time without being recharged, as it greatly affects their lifetime and usability. Data summarization is key for energy-constrained connected devices, as transmitting fewer data can reduce energy usage during transmission. Data compression, in particular, can compress the data stream while preserving information to a great extent. Many compression methods have been proposed in previous research. However, most of them are either not applicable to connected objects, due to resource limitation, or only handle one-dimensional streams while data acquired in connected objects are often multi-dimensional. Lightweight Temporal Compression (LTC) is among the lossy stream compression methods that provide the highest compression rate for the lowest CPU and memory consumption. In this thesis, we investigate the extension of LTC to multi-dimensional streams. First, we provide a formulation of the algorithm in an arbitrary vectorial space of dimension n. Then, we implement the algorithm for the infinity and Euclidean norms, in spaces of dimension 2D+t and 3D+t. We evaluate our implementation on 3D acceleration streams of human activities, on Neblina, a module integrating multiple sensors developed by our partner Motsai. Results show that the 3D implementation of LTC can save up to 20% in energy consumption for slow-paced activities, with a memory usage of about 100 B. Finally, we compare our method with polynomial regression compression methods in different dimensions. Our results show that our extension of LTC gives a higher compression ratio than the polynomial regression method, while using less memory and CPU.