The effect of flow rate in acoustic chemometrics on liquid flow: Transfer of calibration models

Abstract

Previous studies on acoustic chemometrics on liquid flow have demonstrated that flow rate, accelerometer location and temperature affect the passive acoustic spectra and prediction results. Changes in the flow rate result in spectral variations, causing the resultant local calibration model to perform poorly predicting new samples measured at other flow rates. Developing good and robust calibration models can be done using several approaches. Global calibration methods were discussed in a previous study on acoustic chemometrics on liquid flow. In the present paper preprocessing techniques such as multiplicative scatter correction (MSC) and piecewise multiplicative scatter correction (PMSC), and variable selection methods such as interval partial least squares (iPLS) regression and powered partial least squares (PPLS) regression are studied. The preprocessing and variable selection methods were validated on two experimental data sets from passive acoustic measurements of liquid flow through an orifice plate. Acoustic spectra were registered at four different accelerometer locations. The liquids were two-component mixtures of sucrose and water, and three-component mixtures of ethanol, sucrose and water. MSC resulted in the improvement of model performance predicting new (preprocessed) samples measured at other flow rates. Sucrose prediction in two-component mixtures and ethanol prediction in three-component mixtures were improved in terms of bias and correlation coefficients respectively. Absolute bias values for sucrose prediction were in the range of 0.84–2.57 wt.% for spectra preprocessed by MSC compared to 1.17–22.38 wt.% for the uncorrected spectra using an accelerometer located at the orifice plate and the highest of studied flow rates as calibration flow. Correlation coefficients for prediction of ethanol were in the range of 0.80–0.97 for MSC spectra compared to 0.76–0.97 for the uncorrected spectra using an accelerometer located at the orifice plate and the highest of studied flow rates as calibration flow. Limited systematic improvement was observed for the sucrose and water prediction in three-component mixtures. PMSC slightly improved sucrose and ethanol prediction in the three-component mixture compared to MSC. iPLS regression indicated some intervals in acoustic spectra which were less affected by flow rate fluctuations. Regression using these intervals instead of full acoustic spectra resulted in lower prediction errors for sucrose, ethanol and water prediction in three-component mixtures compared to full spectra models. PPLS regression on frequency (peak position) matrix derived from full acoustic spectra did not determine any peaks robust to flow rate fluctuations. Effect of the flow rate on positions of the peaks important for chemical composition was difficult to establish. However, a shift to lower frequency with increasing flow rate could be observed.