不饱和脂肪酸是鲜肉脂肪的基本组成成分, 其种类和含量直接影响鲜肉的风味和品质。 不同于用时长, 破坏样品的气相色谱法, 近红外分析可快速, 无损地对鲜肉中不饱和脂肪酸进行定量检测。 选取了驴肉, 牛肉, 羊肉和猪肉鲜肉样品共63个, 在4 000～12 500 cm-1波段上, 分别采集了肉块样品和粉碎口径3 mm的肉糜样品在5, 10, 15, 20, 25, 30, 35 ℃下的近红外漫反射光谱, 并使用气相色谱法作为参考检测了样品的不饱和脂肪酸含量。 为优化模型性能, 比较了不同温度下肉块和肉糜样品全波段光谱的棕榈油酸、 亚油酸、 油酸、 二十四碳一烯酸和总不饱和脂肪酸的偏最小二乘模型参数, 发现棕榈油酸和总不饱和脂肪酸的5 ℃肉糜光谱模型, 亚油酸的35 ℃和油酸的25 ℃肉糜光谱模型, 以及二十四碳一烯酸的15 ℃肉块光谱模型较佳, 且温度对模型的影响不具有明显的规律性。 在较佳模型的基础上, 分别以220, 440, 881和1 762个变量为窗口区间, 使用前向和反向区间偏最小二乘法对波长进行了优选, 发现棕榈油酸、 亚油酸、 油酸和总不饱和脂肪酸以220个变量作为窗口的反向区间偏最小二乘模型, 以及二十四碳一烯酸以440个变量为窗口的前向区间偏最小二乘模型较佳, 且预测性能均优于全波段PLS模型。 其中棕榈油酸的建模波段为: 4 425～4 636, 4 849～5 272, 5 486～5 696.7, 7 398.6～7 818, 8 031.1～8 666.5, 9 947～10 363.6, 12 495.5～12 498.4 cm-1; 亚油酸的建模波段为: 4 000.6～4 423.9, 5 273.4～5 698.6, 7 398.6～9 090.8, 10 576.7～10 787.8, 12 495.5～12 498.4 cm-1; 油酸的建模波段为: 4 000.6～4 423.9, 4 637～4 848.2, 7 398.6～8 242.3, 8 455.4～9 090.8, 9 947～10 787.8, 12 495.5～12 498.4 cm-1; 二十四碳一烯酸的建模波段为: 4 849.1～5 272.4 cm-1; 总不饱和脂肪酸的建模波段为: 4 000.6～4 423.9, 4 637～5 698.6, 9 097.5～9 515.1, 9 940.3～10 575.7, 11 646～12 060.6, 12 273.7～12 498.4 cm-1。 使用偏最小二乘法对优选波长光谱数据进行降维, 利用得到的潜在变量作为各指标最小二乘-支持向量机模型的输入, 并和各指标的区间偏最小二乘模型进行了性能的比较, 发现最小二乘-支持向量机模型的各指标定量结果最优。 棕榈油酸、 亚油酸、 油酸、 二十四碳一烯酸和总不饱和脂肪酸最优模型的校正集相关系数和均方根误差, 以及留一交叉验证相关系数, 均方根误差及相对预测误差分别为: 0.974, 1.403 mg·(100 g)-1, 0.973, 1.428 mg·(100 g)-1, 4.31; 0.99, 2.233 mg·(100 g)-1, 0.99, 2.263 mg·(100 g)-1, 7.21; 0.982, 8.194 mg·(100 g)-1, 0.982, 8.223 mg·(100 g)-1, 5.19; 0.921, 0.224 mg·(100 g)-1, 0.92, 0.225 mg·(100 g)-1, 2.52; 0.996, 24.21 mg·(100 g)-1, 0.995, 26.045 mg·(100 g)-1, 10.01。 其中, 亚油酸、 油酸和总不饱和脂肪酸最优模型的交叉验证相对预测误差超过5, 棕榈油酸最优模型的交叉验证相对预测误差接近5, 二十四碳一烯酸的交叉验证相对预测误差接近3, 各指标最优模型的预测性能较为令人满意。 研究结果表明, 波段优选, 偏最小二乘降维以及最小二乘-支持向量机算法的结合可对鲜肉中单个不饱和脂肪酸及总不饱和脂肪酸指标的近红外光谱定量结果进行有效的优化。

Unsaturated fatty acids (UFA) are basic composition of fresh meat fat. The composition and content of UFA in fresh meat directly determine its flavor and quality. Differing from being time consuming and causing sample destruction of Gas chromatography, Near-infrared spectroscopy can be used to determine UFA in meat rapidly and non-destructively. NIR diffuse reflectance spectra of sixty-three fresh meat samples including donkey meat, beef, mutton and pork were acquired in the band of 4 000～12 500 cm-1 at temperatures of 5, 10, 15, 20, 25, 30 and 35 ℃. Gas chromatography was used as the reference method to determine the composition and content of UFA in samples. Partial least square (PLS) Calibration models for individual UFA of palmitoleate, linoleic, oleic and tetracosenic acid, and total UFA (TUFA) were developed with all band spectra data of intact and minced samples (diameter of 3 mm) at different temperatures, respectively. The better performances of PLS models for palmitoleate and TUFA were attained with spectra of minced samples at 5 ℃; for linoleic and oleic acid with spectra of minced samples at 35 and 25 ℃ respectively; and for tetracosenic acid with spectra of intact samples at 15 ℃. The influence of sample temperatures on the performances of models for the five indexes was irregularly. Then forward and reverse interval PLS (FiPLS and RiPLS) with interval size of 1 762, 881, 440 and 220 variables were performed to select optimal bands based on all band PLS models. For palmitoleate, linoleic, oleic acid and TUFA, the method of RiPLS with interval size of 220 variables gained better prediction, while for tetracosenic acid the performance of FiPLS model with interval of 440 variables was better than the else iPLS models and all band PLS models. The optimal bands were 4 425～4 636, 4 849～5 272, 5 486～5 696.7, 7 398.6～7 818, 8 031.1～8 666.5, 9 947～10 363.6 and 12 495.5～12 498.4 cm-1 for palmitoleate; 4 000.6～4 423.9, 5 273.4～5 698.6, 7 398.6～9 090.8, 10 576.7～10 787.8 and 12 495.5～12 498.4 cm-1 for linoleic; 4 000.6～4 423.9, 4 637～4 848.2, 7 398.6～8 242.3, 8 455.4～9 090.8, 9 947～10 787.8 and 12 495.5～12 498.4 cm-1 for oleic acid; 4 849.1～5 272.4 cm-1 for tetracosenic acid; and 4 000.6～4 423.9, 4 637～5 698.6, 9 097.5～9 515.1, 9 940.3～10 575.7, 11 646～12 060.6 and 12 273.7～12 498.4 cm-1 for TUFA. The spectra data of optimal bands were compressed by PLS. The latent variables obtained from compression were used as input to Least squares-support vector machine (LS-SVM) models for the five indexes. The performances of LS-SVM models were optimal in comparisons with iPLS models. The correlation coefficients and root mean square error of calibration and leave-one-out cross validation, and ratio of prediction to deviation of cross validation (RPDcv) of the optimal models were 0.974, 1.403 mg·(100 g)-1, 0.973, 1.428 mg·(100 g)-1 and 4.31 for palmitoleate; 0.99, 2.233 mg·(100 g)-1, 0.99, 2.263 mg·(100 g)-1 and 7.21 for linoleic; 0.982, 8.194 mg·(100 g)-1, 0.982, 8.223 mg·(100 g)-1 and 5.19 for oleic; 0.921, 0.224 mg·(100 g)-1, 0.92, 0.225 mg·(100 g)-1 and 2.52 for tetracosenic acid; and 0.996, 24.21 mg·(100 g)-1, 0.995, 26.045 mg·(100 g)-1, 10.01 for TUFA. The RPDcv of linoleic, oleic acid and TUFA models were all more than 5, and the one of palmitoleate was near 5, and the one of tetracosenic acid near 3. The prediction performances of NIR models for the five indexes were satisfied. The results show that the method of combination band selection and PLS compression with LS-SVM can optimize the prediction performance of NIR quantitative models for individual UFA and TUFA in fresh meat.