Ultrasonic sensor for predicting sugar concentration using multivariate calibration

This paper presents a multivariate regression method for the prediction of maltose concentration in aqueous solutions. For this purpose, time and frequency domain of ultrasonic signals are analyzed. It is shown, that the prediction of concentration at different temperatures is possible by using several multivariate regression models for individual temperature points. Combining these models by a linear approximation of each coefficient over temperature results in a unified solution, which takes temperature effects into account. The benefit of the proposed method is the low processing time required for analyzing online signals as well as the non-invasive sensor setup which can be used in pipelines. Also the ultrasonic signal sections used in the presented investigation were extracted out of buffer reflections which remain primarily unaffected by bubble and particle interferences. Model calibration was performed in order to investigate the feasibility of online monitoring in fermentation processes. The temperature range investigated was from 10 °C to 21 °C. This range fits to fermentation processes used in the brewing industry. This paper describes the processing of ultrasonic signals for regression, the model evaluation as well as the input variable selection. The statistical approach used for creating the final prediction solution was partial least squares (PLS) regression validated by cross validation. The overall minimum root mean squared error achieved was 0.64 g/100 g.