Wavelet transform and k-NN method in sensor data analysis
Falkenberg, Tuukka · LUTPub (LUT University) · 2016
Industrial and commercial processes are packed with sensors and devices which produce signal data from the processes. These signals are used to monitor, control, and improve the processes or parts of them, usually in quite simple and limited ways. However, they can also provide much more information and value when analyzed more thoroughly with advanced signal analysis tools. This thesis studies the use of the wavelet transform, which is a type of time-frequency transformation, and a data classification and regression method known as the k-Nearest Neighbors method, as such signal analysis tools. The mathematical theory of the wavelet transform and k-Nearest Neighbors method is presented briefly, and their suitability in sensor signal analysis is discussed. The objective of this thesis is to develop a flexible, adjustable, and highly automated method for signal analysis. An algorithm which utilizes the wavelet transform and k-NN method as a machine-learning-based tool is presented, and the principle and development possibilities of the tool are explained. A prototype algorithm based on the presented tool is tested with sensor data from a highway traffic sensor, and its functionality is examined with an underlying emphasis on the possibility of developing the algorithm further into an automated signal analysis tool that requires minimum user involvement.