Determination of Base Composition Based on Flow Injection Analysis and Local Linear Embedding-Support Vector Regression Modeling Method

Abstract

A flow injection analysis (FIA) method was used to determine the alkaline constituents in boil water. The FIA method was based on the injection of an alkaline sample into a carrier stream of a standard acid solution. The sample and reagent zones were partially mixed due to dispersion and thereby the sample was partially neutralized by the acid. A flat pH electrode was used to sense the output voltage of the mixture. The signal was recorded as a typical FIA peak. For the mixed base with the same composition, the FIA peak areas were experimentally found to be proportional to the logarithm of total basicity of mixed base. The composition of the mixed base could be determined by the pH and the FIA peak area. In case of Na 2 CO 3 and NaOH mixture, the concentration of NaOH could be directly obtained from the pH value of the solution, due to the fact that the pH of the mixture was mainly decided by the NaOH concentration. The total basicity could be determined by the corresponding calibration curve. The concentration of Na 2 CO 3 could be calculated by the concentration of NaOH and the total basicity of mixture. In order to get the better precision of the FIA method, a local linear embedding (LLE), coupled with support vector regression (SVR) modeling, is also proposed to determine the constituents of an alkaline mixture instead of the previous linear calibration curve. All the output voltage of the pH electrode, instead of the FIA peak areas, was used as the input of the LLE-SVR model. The logarithm of total basicity of mixed base was the output of the model. LLE is a nonlinear dimensionality reduction method, which is suitable for the data that lies on the nonlinear manifold and can reveal the global nonlinear structure by combining the local linear relationship. SVR is considered as a substitute for traditional learning regression approach and has the excellent generalization performance especially in small samples of the nonlinear case. Using LLE and SVR method (LLE-SVR) together to determine alkaline mixture can avoid disturbance from the unknown compositions. Relative standard deviation from the linear calibrating curve between the peak area and the logarithm of total basicity of mixed base was 2.34 percent. The proposed LLE-SVR modeling can decrease the calibrating error between the FIA peak areas and the total basicity. Prediction relative error of NaOH from LLE-SVR model and SVR model was 0.7224 and 1.1857 percent respectively. Prediction relative error of Na 2 CO 3 from LLE-SVR model and SVR model was 1.1864 and 1.6885 percent respectively. But the LLE-SVR and SVR had the more computation than the liner calibration curve.