The application of parallel processing in the selection of spectral variables in beer quality control

Abstract

Parallel data analysis was investigated to improve performance in variable selection and to develop predictive models for beer quality control. A set of spectral near infrared (NIR) data from 60 beer samples and its primitive extracts as the original concentration was used. The dataset was distributed to Raspberry Pi 3 Model B devices connected to a network that was running a Machine Learning service. With more than 4 devices acting in parallel, it was possible to reduce time in 57% to find the best linear regression coefficient (0.999) with the lower RMSECV (0.216) if compared to a singular desktop computer. Thus, parallel processing can significantly reduce the time to indicate the best model fitted during the variable's selection.

Keywords: Methods of Parallelism; Partial Least Squares; Variable selection; WebService.