Filtrace naměřených veličin v rámci smart home care

Abstract

The goal of this work was to optimize Wavelet Transform settings for filtration of unwanted noise from predicted signal of CO_2 concentration produced by Neural network, so that the resulting signal matches the reference one to the most possible extent. Six wavelets were chosen to be experimentaly tested with various settings to achieve appropriate filtration. Testing was done on 12 real world datasets which have different lengths of measurement (day, week and month). For evaluation of appropriate settings were used three metods: correlation coefficient, mean square error and Euklidean distance. The results were divided by the lengths of predicition, which means for every length of prediction was choosen one wavelet with appropriate settings. Wavelets with the highest correlation coeficient against the reference CO_2 concetration signal were chosen as the best performing ones.