为了对鸡种蛋胚胎进行雌雄识别, 探究利用紫外-可见-近红外透射光谱进行鸡胚雌雄识别的可行性, 搭建了鸡种蛋透射光谱检测系统, 采用横向和竖向大头朝上2种放置方式获取210枚鸡种蛋孵化0~15 d的光谱, 光谱范围为360~1 000 nm。 构建极限学习机（ELM）鸡胚雌雄识别模型, 通过比较不同放置方式和孵化天数下模型的识别准确率, 发现竖向放置且孵化第7 d的识别效果最好; 将竖向放置孵化第7 d的光谱初步分为紫外（360~380 nm）、 可见光（380~780 nm）、 近红外（780~1 000 nm）、 紫外-可见光（360~780 nm）和全波段（360~1 000 nm）5个不同的波段范围来分析, 预测集准确率分别为8286%, 7714%, 7571%, 8429%和8143%, 筛选出360~780 nm的紫外-可见光波段为有效波段; 在紫外-可见光（360~780 nm）波段, 采用多元散射校正（MSC）去噪, 并用竞争性自适应重加权采样算法（CARS）和连续投影算法（SPA）筛选特征波长降维, 建立不经筛选特征波长、 CARS筛选特征波长和SPA筛选特征波长的3种ELM模型。 其中不经筛选特征波长的ELM模型识别效果最好, 但输入变量最多, 隐含层神经元为680且激活函数为sig时, 预测集准确率为8429%。 SPA筛选特征波长的ELM模型识别效果次之, 输入变量有9个, 隐含层神经元为840且激活函数为hardlim时, 预测集准确率为8143%。 CARS筛选特征波长的ELM模型识别效果最差, 输入变量有27个, 隐含层神经元为100且激活函数为sig时, 预测集准确率为7857%; 用遗传算法（GA）优化ELM模型的权值变量和隐含层阈值, 不经筛选特征波长建立的GA-ELM模型, 预测集准确率为8714%, SPA筛选特征波长建立的GA-ELM模型, 预测集准确率为8714%, CARS筛选特征波长建立的GA-ELM模型, 预测集准确率为8143%。 紫外-可见光波段不经筛选特征波长的GA-ELM模型识别效果和经SPA筛选特征波长的GA-ELM模型相同, 表明SPA筛选的特征波长变量能够有效反映360~780 nm波段的信息, SPA使用的变量数仅占紫外-可见光波段的214%, 因此, 雌雄识别最佳模型为紫外-可见光波段经SPA筛选特征波长的GA-ELM模型, 预测集准确率为8714%, 其中, 雌性识别率为8857%, 雄性识别率为8571%, 单个样本平均判别时间0080 ms。 结果表明紫外-可见透射光谱技术和ELM模型为孵化早期鸡胚蛋雌雄识别提供了一种可行方法。

In order to identify male and female embryos of chicken eggs, the feasibility of using UV/Vis/NIR transmission spectrum to identify the male and female embryos is explored. The transmission spectrum detection system of chicken eggs is established with blunt end vertically placed upwards and horizontally placed separately to obtain the 0~15 d spectrum (ranging from 360 to 1 000 nm) of 210 hatched eggs. The identification model of the embryo learning male and female of the extreme learning machine (ELM) is constructed. By comparing the identification accuracy of different placement and the number of hatching days, it is found that the recognition effect of the vertical placement hatching on the 7th day is the best. The spectrum of the 7th day of vertical incubation is initially divided into ultraviolet (360～380 nm), visible light (380～780 nm), near-infrared (780～1 000 nm), ultraviolet/visible (360～780 nm) and full-band (360～1 000 nm). Five different band ranges are analyzed, and the prediction set accuracy rates are 8286%, 7714%, 7571%, 8429%, and 8143%, respectively. The ultraviolet/visible bands of 360～780 nm are selected as effective bands; In the ultraviolet/visible (360～780 nm) band, Multiplicative scatter correction (MSC) is used to denoise, and the characteristic wavelength reduction is selected by Competitive adaptive reweighted sampling (CARS) and Successive projection algorithm (SPA). Three kinds of wavelengths without screening, CARS screening characteristic wavelength and SPA screening characteristic wavelength are established ELM model. Among them, the ELM model without screening characteristic wavelengths has the best recognition effect, but the input variables are the most. When the hidden layer neuron is 680 and the activation function is sig, the prediction set accuracy is 8429%. The ELM model of the SPA screening characteristic wavelength has the second recognition effect, and there are 9 input variables. When the hidden layer neurons are 840 and the activation function is hardlim, the prediction set accuracy is 8143%. The ELM model with the CARS screening characteristic wavelength has the worst recognition effect, and there are 27 input variables. When the hidden layer neurons are 100 and the activation function is sig, the prediction set accuracy is 7857%. Using Genetic algorithm (GA) to optimize the weight variable and hidden layer threshold of ELM model, the prediction set accuracy rate is 8714%, 8714% and 8143% separately under the condition of the GA-ELM model established without screening the characteristic wavelength, the GA-ELM model established by SPA screening characteristic wavelength, and the GA-ELM model established by the CARS screening characteristic wavelength. The recognition effect of GA-ELM model in the ultraviolet/visible band without screening characteristic wavelength is the same as that in the GA-ELM model with SPA screening characteristic wavelength, which indicates that the characteristic wavelength variable of SPA screening can effectively reflect the information of 360～780 nm band. The number of variables used by the SPA is only 214% of the ultraviolet/visible range. Therefore, the best model for male and female identification is the GA-ELM model for screening characteristic wavelengths with SPA in the ultraviolet/visible range. The accuracy of the prediction set is 8714%, of which the female recognition rate is 8857%, the male recognition rate is 8571%, and the average discrimination time of a single sample is 0080 ms. The results show that UV/Vis transmission spectroscopy and ELM model provide a feasible method for the identification of chicken embryo eggs in early hatching.