Scholarlink

In this research, the feasibility of employing multiple machine learning algorithms with FTIR-ATR spectroscopy for the simultaneous quantification of ethanol and legal limits of methanol in distilled and artisanal beverages characteristic of Ecuador is investigated, based on spectral matrix similarity. Initially, spectra acquired in the range of 4000 to 400  \({\text{cm}}^{-1}\) underwent spectral preprocessing including baseline correction, smoothing, normalization, first and second derivative, and their combinations. Forty-eight distinct treatments were used to construct models employing Principal Component Regression (PCR) and Partial Least Squares 2 (PLS2). The treatment yielding superior metrics was employed for constructing an Artificial Neural Network combined with Principal Component Analysis (PCA-ANN) and Recursive Feature Elimination (RFE-ANN) utilizing a Decision Tree Regressor as a variable selector. Based on confidence intervals and hypothesis testing of statistics such as root mean squared error of prediction (RMSEP) the PCR and PLS2 models exhibited superior performance. PCR achieved detection and quantification limits of 0.25% and 0.7%, respectively. In commercial beverages, predictions were compared with results obtained via gas chromatography, where again PCR and PLS2 demonstrated the finest metrics, showcasing speed and high cost-effectiveness, rendering them viables alternatives for preliminary quality control analysis.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multi-Method Spectral Predictive Models with FTIR-ATR in the Simultaneous Quantification of Ethanol and Legal Methanol Limits in Ecuadorian Clear Spirits

  • 会议论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年

查看全文
  • Wilson Cajilima,
  • Pablo Arévalo

摘要

In this research, the feasibility of employing multiple machine learning algorithms with FTIR-ATR spectroscopy for the simultaneous quantification of ethanol and legal limits of methanol in distilled and artisanal beverages characteristic of Ecuador is investigated, based on spectral matrix similarity. Initially, spectra acquired in the range of 4000 to 400  \({\text{cm}}^{-1}\) underwent spectral preprocessing including baseline correction, smoothing, normalization, first and second derivative, and their combinations. Forty-eight distinct treatments were used to construct models employing Principal Component Regression (PCR) and Partial Least Squares 2 (PLS2). The treatment yielding superior metrics was employed for constructing an Artificial Neural Network combined with Principal Component Analysis (PCA-ANN) and Recursive Feature Elimination (RFE-ANN) utilizing a Decision Tree Regressor as a variable selector. Based on confidence intervals and hypothesis testing of statistics such as root mean squared error of prediction (RMSEP) the PCR and PLS2 models exhibited superior performance. PCR achieved detection and quantification limits of 0.25% and 0.7%, respectively. In commercial beverages, predictions were compared with results obtained via gas chromatography, where again PCR and PLS2 demonstrated the finest metrics, showcasing speed and high cost-effectiveness, rendering them viables alternatives for preliminary quality control analysis.