Scholarlink

<p>The accurate prediction of °Brix values, which measure soluble solids concentration in agricultural products, is crucial for quality control and product development in the food industry. Traditional methods for °Brix measurement are often time-consuming and destructive, highlighting the need for rapid, nondestructive alternatives. This study addresses this challenge by applying near-infrared (NIR) spectroscopy combined with machine learning techniques to predict °Brix values. Specifically, we explore the efficacy of convolutional neural networks (CNNs), particularly 1D CNNs, in analyzing NIR spectral data. The research involves optimizing CNN models through Bayesian optimization to enhance predictive performance. We compare these models with partial least squares regression (PLSR) using various preprocessing techniques, including multiplicative scatter correction (MSC) and spectral derivatives. Our results demonstrate that 1D CNNs, especially when combined with MSC and spectral derivatives, outperform PLSR in terms of predictive accuracy and generalization to new data. This study highlights the potential of advanced machine learning models for improving nondestructive chemical property prediction in the food industry. The findings suggest that CNNs offer a robust alternative to traditional methods, with implications for both practical applications and theoretical advancements in spectroscopy and predictive modeling. Future research should focus on expanding the dataset and exploring multitask learning approaches to further enhance model capabilities.</p>

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

Optimizing 1D CNN models for accurate °Brix prediction

  • 期刊论文

  • 发表日期：

    2025年7月5日

查看全文
  • Ernesto Alonso Paiva-Peredo,
  • María Huánuco-Chipana,
  • Fenixe Ausejo-Gonzales,
  • Kevin Acuna-Condori,
  • Wiliam Trujillo

摘要

The accurate prediction of °Brix values, which measure soluble solids concentration in agricultural products, is crucial for quality control and product development in the food industry. Traditional methods for °Brix measurement are often time-consuming and destructive, highlighting the need for rapid, nondestructive alternatives. This study addresses this challenge by applying near-infrared (NIR) spectroscopy combined with machine learning techniques to predict °Brix values. Specifically, we explore the efficacy of convolutional neural networks (CNNs), particularly 1D CNNs, in analyzing NIR spectral data. The research involves optimizing CNN models through Bayesian optimization to enhance predictive performance. We compare these models with partial least squares regression (PLSR) using various preprocessing techniques, including multiplicative scatter correction (MSC) and spectral derivatives. Our results demonstrate that 1D CNNs, especially when combined with MSC and spectral derivatives, outperform PLSR in terms of predictive accuracy and generalization to new data. This study highlights the potential of advanced machine learning models for improving nondestructive chemical property prediction in the food industry. The findings suggest that CNNs offer a robust alternative to traditional methods, with implications for both practical applications and theoretical advancements in spectroscopy and predictive modeling. Future research should focus on expanding the dataset and exploring multitask learning approaches to further enhance model capabilities.