Prediction of Wind Turbine Lubricating Oil's Acid Value by Ordinary Least Square Method Based on Attenuated Total Reflectance-Fourier Transform Infrared Spectroscopy Through Higher-Order Derivative Combined with Angular Metric
10.19756/j.issn.0253-3820.241274
- VernacularTitle:衰减全反射-傅里叶红外光谱高阶导数结合角度量的最小二乘预测风电机组润滑油酸值
- Author:
Chun-Hui GE
1
;
Yan-Jun LIU
;
Meng-Shi CHEN
;
Ce YANG
;
Pei-Pei LIANG
;
Zhi-Xiang YAO
;
Kai ZHANG
Author Information
1. 华北电力大学,热电生产过程污染物监测与控制北京市重点实验室,北京 102206
- Keywords:
Lubricating oil;
Acid value;
Attenuated total reflectance-Fourier transform infrared spectroscopy;
Higher-order derivative;
Vector space angle;
Collinearity;
Ordinary least squares regression
- From:
Chinese Journal of Analytical Chemistry
2024;52(9):1254-1265,中插1-中插4
- CountryChina
- Language:Chinese
-
Abstract:
To address the key challenges in multivariate statistical modeling,a higher-order derivative approach combined with vector space angle multiplicative error correction was proposed for establishing an acid value prediction ordinary least squares(OLS)regression model based on attenuated total reflectance-Fourier transform infrared(ATR-FTIR)spectroscopy.By using acid values measured by potentiometric titration as reference,ATR-FTIR spectroscopy was utilized for direct calibration and prediction of acid values on 96 kinds of lubricating oil samples from a wind turbine.Firstly,the simulated hyperbolic(SH)method was employed to obtain accurate fourth derivative spectrum,resolving overlapping bands and enhancing spectral selectivity.Then,from the calibration set(48 samples),informative spectral regions were identified based on correlation coefficients.Next,the sample with the highest acid value was selected as the reference and1/(1+tan(θ/2))was used as the metric relation of the spectrum to suppress the multiplicity error caused by factors such as the change of effective optical path in ATR-FTIR spectroscopy.After pretreatment of the spectrum by the method of fourth-order derivative combined with angular quantity,the number of variables decreased from 1737 to 8,and the matrix condition number decreased from 1.85×1015 to 56.34,which effectively eliminated the collinearity issue for OLS regression.Direct OLS modeling on spectral preprocessed data achieved a determination coefficient of 0.981 for 47 validation samples,with a relative error range of-8.38%-8.22%,outperforming the commonly used partial least squares(PLS)method(Determination coefficient of 0.865,relative error of-27.82%-22.38%).It was proved that effective data preprocessing significantly improved the prediction accuracy of the model.Furthermore,when the number of calibration set was compressed to 25 and the number of validation set was expanded to 70,the model retained 8 variables with a condition number of 42.60,the determination coefficient of validation set was 0.972,and the relative error ranged from-10.80%to 12.31%.Comparing with the PLS method(Determination coefficient of 0.724,relative error of-34.26%-53.84%),the improvement was more obvious,which showed that the method could still have high prediction accuracy even with fewer modeling samples as well as robustness against multiplicative error interference.
