种鸡蛋孵化过程耗时、 耗能, 但种蛋由于未受精导致胚胎发育率不足100%, 种鸡蛋孵化早期快速判别, 可有效提高种蛋孵化效率和降低经济损失。 近红外光谱分析技术可实现快速、 无损检测, 用于判别研究孵化早期种鸡蛋受精信息, 但现有光谱检测方法未能满足孵化早期受精判别位置要求, 亟需构建采集数据时的检测位置与内部信息之间的关系。 该研究利用可见/近红外光谱检测系统采集种蛋壳漫反射光谱强度, 试验选用壳色相近、 表面无裂纹的181个新鲜种鸡蛋, 从中随机选择61个样本完成交叉验证。 为消除暗电流的影响, 进行光谱校正, 得出种蛋壳漫反射率, 发现受精蛋与未受精蛋的光谱曲线趋势相同, 受精蛋和无精蛋赤道处的光谱曲线均高于两端。 剔除光谱曲线两端噪声大、 信噪比低的波段, 选择440.27~874.6 nm作为有效光谱波段, 并采用Savitzkg-Golay(S-G)平滑法、 二阶导数法Second Derivative、 标准正态变量校正SNV、 变量标准化Normalization和多元散射校正MSC预处理法分别构建PCA-SVM判别模型。 通过对入孵后24, 48, 72, 96和120 h不同位置所采光谱数据, 与受精信息综合分析, 结果发现二阶导数法Second Derivative和多元散射校正MSC两种预处理方式在同一时间、 不同位置的训练集和验证集的判别准确率均相同, 说明这两种预处理方法对数据来源位置不敏感; 变量标准化Normalization预处理和S-G平滑法预处理验证集的判别准确率则随着时间的推移在一定范围内波动, 在种鸡蛋胚胎发育120h时的判别准确率可达91.71%; SNV预处理赤道处验证集判别准确率随着时间的推移呈现上升趋势, 对数据来源时间及位置敏感, 种鸡蛋胚胎发育越长判别效果越佳, 120 h赤道侧的判别准确率为91.16%。 且平滑处理、 标准正态变量校正和变量标准化三种预处理方式均存在赤道侧判别准确率高于种蛋端部的现象, 这主要是由于赤道侧较种鸡蛋两端表面平缓, 采集到的信息更多, 该研究为早期判别受精蛋与无精蛋的数据采集位置提供新思路和新方法。

It takes more time and energy for eggs to hatch, but the circumstances of hatching egg embryo growth are less than 100%. The early discrimination of hatching eggs can reduce the economic loss and improve the efficiency. The near-infrared spectral analysis technology used in detecting the early fertilization information of hatching eggs because of speed and harmless. However, the existing detection method can not meet the requirement of detecting position. The necessitated problem is to build the relationship between the detecting position and the internal information. The visible/near-infrared spectroscopy detection system was used to collect the diffuse reflectance spectrum intensity of eggshell. 181 fresh eggs with similar shell color and no surface cracks were selected for analysis, and 61 samples were randomly selected for cross-validation. In order to eliminate the influence of dark current, the diffuse reflectance of eggshell was obtained by spectral correction. It was found that the trend of the spectral curve of fertile egg and the infertile egg was the same, and the spectral curves of position 3 and 4 were higher than position 1 and 2. The effective spectral bands of 440.27~874.6 nm were selected for the study. SGolay smoothing, second derivative, SNV, normalize and MSC pretreatment method were used to construct the PCA-SVM discrimination model. Then the data after 24, 48, 72, 96 and 120 h was collected at different positions. The results showed that the accuracy of derivative was as same as the accuracy of MSC, which indicated that the two pretreatment methods were not sensitive to the change of position through the analysis of data and ertilization information. The accuracy of the validation set was fluctuated in a certain range, and the accuracy rate after 120 h was 91.71% when the pretreatment methods of Normalize and SGolay were used to reduce noise. The accuracy rate of SNV pretreatment at the equator showed an upward trend with time, and it was sensitive to the time and position. The longer the embryo development, the better the discrimination effect. The best discrimination accuracy rate was 91.16% at the equator after 120 h. Moreover, smoothing, SNV and normalize have the highest discrimination accuracy at equator, which was mainly because the equator’s surface is flat and more information was collected. This study provides a new idea and method for the early identification of the data acquisition position.