College of Engineering, China Agricultural University, Beijing 100083, China.
Environmental Microbial and Food Safety Laboratory, USDA-ARS, Beltsville, MD 20705, United States.
Food Res Int. 2024 Dec;197(Pt 1):115184. doi: 10.1016/j.foodres.2024.115184. Epub 2024 Oct 9.
Changes in the freshness of pork due to microbial action during complex transportation and storage indicate an urgent need for in-situ, real-time monitoring techniques for chemical spoilage of meat. This study reported the use of a portable detection device based on a miniaturized visible/near-infrared spectrometer, combined with data noise reduction and machine learning methods, to predict the total viable count (TVC) in pork samples. A rapid detection device for pork TVC was designed based on the miniaturized spectrometer; a spectral preprocessing method based on the resolution of the spectrometer was proposed; the effects of different preprocessing methods on the full-wavelength modeling effect were compared; and four different feature wavelength extraction algorithms were utilized for the feature wavelengths of pork TVC. The modeling effects of different simplified models were compared. The results showed that resolution interval correction combined with standard normal variation was the most effective in full-wavelength modeling, with correlation coefficients of prediction set (R), root mean square errors in prediction set (RMSEP), and relative percent deviation (RPD) of 0.918, 0.464 (lg CFU/g), and 2.508, respectively; interval random frog - partial least squares regression (iRF-PLSR) had the best predictive ability among all simplified models, the number of wavelengths used in the simplified model was reduced by 85.45% compared with the full-wavelength model. In contrast, the model performance was improved with R, RMSEP, and RPD of 0.948, 0.392 (lg CFU/g) and 2.970, respectively. The combination of a rational spectral acquisition setup and a data processing methodology, the miniaturized spectrometer showed competitive results with the complex detection system in predicting meat TVC.
由于微生物在复杂运输和储存过程中的作用，猪肉的新鲜度发生变化，这表明迫切需要用于肉类化学变质的原位、实时监测技术。本研究报告了一种使用基于微型可见/近红外光谱仪的便携式检测设备，结合数据降噪和机器学习方法，预测猪肉样品中总活菌数（TVC）的方法。基于微型光谱仪设计了用于猪肉 TVC 的快速检测设备；提出了一种基于光谱仪分辨率的光谱预处理方法；比较了不同预处理方法对全波长建模效果的影响；并利用四种不同的特征波长提取算法提取猪肉 TVC 的特征波长。比较了不同简化模型的建模效果。结果表明，分辨率间隔校正与标准正态变化相结合在全波长建模中效果最佳，预测集相关系数（R）、预测集均方根误差（RMSEP）和相对偏差（RPD）分别为 0.918、0.464（lg CFU/g）和 2.508；间隔随机蛙 - 偏最小二乘回归（iRF-PLSR）在所有简化模型中具有最佳预测能力，与全波长模型相比，简化模型中使用的波长数量减少了 85.45%。相比之下，模型性能得到了改善，R、RMSEP 和 RPD 分别为 0.948、0.392（lg CFU/g）和 2.970。合理的光谱采集设置和数据处理方法的结合，微型光谱仪在预测肉类 TVC 方面表现出与复杂检测系统相当的结果。
