鱼粉是以一种或多种鱼类为组成原料的高蛋白饲料, 其在养殖产业中占据着非常重要的地位。 鱼粉产地众多, 品质参差不齐, 有不法商家以次充好, 为维护市场秩序, 应当建立一种鱼粉的产地溯源方法, 以便更加精准地判别和分析鱼粉的品质。 采用近红外光谱技术并结合化学计量学方法实现对不同产地鱼粉进行快速准确的产地溯源。 采用径向基为核函数的支持向量机进行模式识别, 并采用灰狼算法对以径向基为核函数的支持向量机进行关键参数的选择, 模拟狼群狩猎行为, 由适应度高低来设置等级制度, 有秩序地对目标参数进行逐渐逼近围捕的移动, 每次移动后重新进行适应性评估, 经过狼群等级迭代最终捕获猎物, 搜索到最佳惩罚因子和核函数半径; 最后, 利用最佳参数建立支持向量机模型进行不同产地鱼粉的产地溯源。 灰狼算法可以提高支持向量机算法中关键参数的选择速度和精度, 并提高支持向量机分类正确率。 对来自浙江温岭、 山东荣成、 山东威海、 辽宁大连四个产地的鱼粉样品采样, 共获得144条光谱, 光谱范围为3 700～12 500 cm-1, 用获得的光谱进行产地溯源。 随机选取每个产地样品的70%作为建模训练样本集, 30%作为测试样品集。 首先对原始近红外光谱进行预处理, 采用多元散射校正算法计算所有光谱的平均光谱当作“理想光谱”, 其他光谱对平均光谱进行一元线性回归, 对光谱平移、 偏移进行基线校正。 采用小波变换对原信号分解, 对高频信号进行阈值化处理, 消除高频噪声达到光谱曲线平滑去噪效果; 利用灰狼算法优化的支持向量机进行十次平行实验, 降低误差干扰, 得到产地分类结果: 浙江温岭、 山东荣成、 山东威海、 辽宁大连识别正确率分别为100%, 98.89%, 96.43%和97.78%。 与网格搜索法相比, 改进后的灰狼算法搜索支持向量机的惩罚因子和核函数半径速度更快更精确, 分类准确率更高, 可见灰狼算法优化的支持向量机（GWO-SVM）对鱼粉光谱进行产地溯源是可行的。

Fish meal is a kind of high-protein feed made up of one or more kinds of fish, which occupies a very important position in the aquaculture industry. In order to maintain market order, a method of tracing the origin of the fish meal should be established to identify and analyze the quality of the fish meal more accurately. In this paper, near-infrared spectroscopy (NIRS) and chemometrics are used to trace the origin of fish meal from different habitats quickly and accurately. The support vector machine with radial basis function (RBF-SVM) as the kernel function is used for pattern recognition, and the gray wolf algorithm is used to select the key parameters of RBF-SVM. By simulating the hunting behavior of wolves, a hierarchical system is set up according to the fitness level. The target parameters gradually approximate the movement of encirclement. After each movement, the adaptability is re-evaluated. The prey is finally captured through the iteration of wolf pack rank, and the optimal penalty factor and the radius of the kernel function are searched. Finally, the optimal parameters are used to establish the support vector machine model to trace the origin of fish meal from different origins. Grey Wolf algorithm can improve the speed and accuracy of selecting key parameters in the support vector machine algorithm, and improve the classification accuracy of support vector machine. In this paper, 144 spectra of fish meal samples from four fishmeal producing areas in ZhejiangWenling, Shandong Rongcheng, Shandong Weihai and Liaoning Dalian were obtained. The spectrum ranges from 3 700 to 12 500 cm-1. The origin of fish meal was traced by the obtained spectra. Seventy percent of the samples from each producing area was randomly selected as the training sample set for modeling and 30 percent as the test sample set. First, the original near infrared spectra are pretreated, and the average spectra of all the collected spectra are calculated by multivariate scattering correction as “ideal spectra”. The other spectra are linearly regressed, and the baseline correction of spectral translation and migration is carried out. The original signal is decomposed by wavelet transform, and the high-frequency signal is thresholded to eliminate the high-frequency noise so as to achieve the smooth denoising effect of the spectral curve. Ten parallel experiments were carried out by support vector machine to reduce error interference, and the classification results were obtained as follows: Zhejiang Wenling, Shandong Rongcheng, Shandong Weihai and Liaoning Dalian were 100%, 98.89%, 96.43% and 97.78%, respectively. Compared with the grid search method, the Improved Grey Wolf algorithm searches for the penalty factor and the radius of the kernel function faster and more accurately, and the classification accuracy is high. It can be seen that the improved grey wolf algorithm’s support vector machine (GWO-SVM) is feasible for tracing the origin of fish meal.