Detection of sugar content in food based on the electrochemical method with the assistance of partial least square method and deep learning
Detection of sugar content in food based on the electrochemical method with the assistance of partial least square method and deep learning
复制标题
DOI：
10.1007/s11694-023-01973-x
复制
发表时间：
2023-06-15
影响因子：
3.4
通讯作者：
Song, Dapeng
中科院分区：
文献类型：
--
作者：
Chang, Jin;Song, Dapeng
水果是纤维和维生素的宝贵来源，其内部成分的多样性和质量是消费者的基本因素。特别是，糖含量被认为是梨品质的最重要指标。在这项研究中，电化学指纹图谱，偏最小二乘（PLS）回归和深度学习技术被用来开发各种水果糖预测模型。具体而言，不同的PLS模型建立了电化学指纹和糖信息与各种电化学指纹预处理方法。采用竞争自适应加权算法，减少PLS模型的计算量，获得最准确的PLS模型预测。利用MobileNetV2网络构建了基于深度学习的水果糖度预测模型，并对模型参数进行了优化。实验结果表明，采用标准正态变量（SNV）对电化学指纹图谱进行预处理的PLS模型具有最高的预测精度。特征波长筛选可以减少计算负荷，但可能会稍微降低预测模型的准确性。结果表明，基于MobileNetV2网络的水果糖度预测模型具有一定的可行性，但对全光谱进行SNV预处理的PLS模型准确率最高。PLS建模仍然是一种简单而有效的方法，用于建立小批量样本的模型。
Fruits are a valuable source of fiber and vitamins, and the diversity and quality of their internal components are essential factors for consumers. In particular, sugar content is considered the most important indicator of pear quality. In this study, electrochemical fingerprinting, partial least squares (PLS) regression, and deep learning techniques were used to develop various fruit sugar prediction models. Specifically, different PLS models were built using electrochemical fingerprinting and sugar information with various electrochemical fingerprint pre-processing methods. The competitive adaptive reweighted algorithm was utilized to reduce the calculation load of PLS models and obtain the most accurate PLS model prediction. Furthermore, a fruit sugar prediction model based on deep learning using the MobileNetV2 network was constructed, and the model parameters were optimized. Experimental results showed that the PLS model with electrochemical fingerprint preprocessing using standard normal variate (SNV) achieved the highest prediction accuracy. Feature wavelength screening can reduce the computational load, but may slightly decrease the prediction model accuracy. The MobileNetV2 network-based fruit sugar prediction model showed some feasibility, but the PLS model with SNV preprocessing for the full spectrum had the highest accuracy. PLS modeling remains a simple and efficient method for building models with small batch samples.