基于二次迭代Monte Carlo的羊肉硬度定量检测研究

随着居民生活水平的提高和对健康饮食结构的重视，羊肉作为一种高蛋白且低脂肪和胆固醇的畜肉，需求量逐年上涨。根据国家统计局统计，2012年-2019年我国畜肉产业中羊肉产量占比从6.27%上升到9.02%。研究提出了一种基于二次迭代Monte Carlo（MC）算法剔除异常样本的羊肉硬度定量检测PLSR模型。采用GaiaSorter高光谱分选仪的Image-λ-V10E-H相机采集羊肉样品400~950 nm的高光谱数据，Image-λ-N17E相机采集羊肉样品900~1 650 nm的高光谱数据。首先，对比分析了S-G平滑、二阶求导、多元散射校正（MSC）、标准正态变换（SNV）等光谱预处理方法在消除噪声影响，提高光谱分别率等方面的能力，选取最佳光谱预处理方法。然后，在第一次MC抽样中，计算所有样本预测误差均值和标准差的平均值，以该平均值的2.5~3倍作为可疑样本阈值，3倍作为异常样本阈值；剔除异常样本，保留并标注可疑样本，进行第二次MC抽样，以样本预测误差均值和标准差的3倍值为阈值进行异常样本二次剔除；对第一次MC抽样中标注可疑样本进行二次检测。最后，对比分析了基于全波长建立的偏最小二乘回归（PLSR）模型和基于回归系数法（RC）提取的特征波长建立的PLSR模型。研究结果表明，所提出的二次迭MC算法可以准确判别可疑样本是否为异常样本，有效优化样本集，为建模提供良好的数据基础。以MSC作为光谱预处理算法基于400~950和900~1 650 nm两段高光谱数据建立PLSR模型的R2_P分别为0.947 2和0.978 3，RMSE_P分别为47.789 9和30.590 1 g，优于其他三种光谱预处理算法。另外，基于900~1 650 nm建立的PLSR模型明显优于基于400~950 nm波长样本集建立的模型。通过RC算法选取出羊肉硬度在400~950和900~1 650 nm波长范围的特征波长分别为14个（410，438，450，464，539，558，612，684，701，734，778，866，884和935 nm）和10个（915，949，1 085，1 156，1 206，1 262，1 318，1 384，1 542和1 580 nm）。其中，基于900~1 650 nm波长建立的PLSR模型的R2_P为0.985 0，RMSE_P为24.397 0 g，为羊肉硬度预测的最佳模型。结果表明，所提出的融合二次迭代MC算法的PLSR模型可以有效预测羊肉冷藏过程中硬度特性变化趋势，为羊肉品质无损检测相关研究提供参考。

Quantitative Detection of Mutton Hardness Based on Twice Iterative Monte Carlo Method

Mutton, as a kind of meat with high protein content and low fat and cholesterol content, is becoming more and more popular with consumers. The demand for mutton is on the rise. According to the National Bureau of Statistics, China’s mutton production rose from 6.27% to 9.02% from 2012 to 2019. This study proposed a quantitative detection PLSR model of mutton hardness based on the twice iterative Monte Carlo (MC) method. In this study, the Image-λ-V10E-H camera of the GaiaSorter hyperspectral sorter was used to collect the hyperspectral data of mutton samples at 400~950 nm, and the Image-λ-N17E camera was used to collect the hyperspectral data of mutton samples at 900~1 650 nm. Firstly, the study compared and analyzed four spectral pretreatment methods (S-G smoothing, 2 derivations, MSC and SNV) in eliminating interference factors, such as noise and baseline drift. Then, in the first MC sampling, the samples were divided into normal samples, suspicious samples and abnormal samples according to the 2.5 and 3 times the means of the prediction error means and standard deviations of each sample. The second MC sampling was performed based on rejecting abnormal samples, retaining and labeling suspicious samples. The new abnormal samples were eliminated by 3 times of the means of the prediction error means and standard deviations of each sample. Finally, the PLSR model based on the full wavelengths and the characteristic wavelengths extracted by the regression coefficient method (RC) were established and analyzed. The experiment results show that the twice iterative Monte Carlo method proposed in the study could abnormal samples, optimize the sample set, and provide a good foundation for modeling. With MSC as the spectral preprocessing algorithm, the PLSR model based on 400~950 and 900~1 650 nm hyperspectral data was superior to the other three spectral preprocessing algorithms R2_P=0.947 2 and 0.978 3，RMSE_P=47.789 9 g and 30.590 1 g. And, the accuracy and stability of the PLSR model based on 900~1 650 nm were significantly better than that based on 400~950 nm. 14 characteristic wavelengths (410, 438, 450, 464, 539, 558, 612, 684, 701, 734, 778, 866, 884, 935 nm) and 10 characteristic wavelengths (915, 949, 1 085, 1 156, 1 206, 1 262, 1 318, 1 384, 1 542 and 1 580 nm) of mutton hardness were selected by RC algorithm from 900~1 650 and 400~950 nm. The PLSR model based on 900~1 650 nm was the optimal model for predicting the hardness of mutton with R2_P=0.985 0 and RMSE_P=24.397 0 g. In conclusion, the PLSR model based on the twice iterative MC algorithm can effectively predict the changing trend of mutton hardness during cold storage and provide a reference for related research on non-destructive detection of mutton quality.