基于光谱技术的原料肉新鲜度指标在线检测系统开发及试验

为了实现原料肉新鲜度参数的无损在线实时评估，基于双波段可见/近红外反射光谱（350~1 100和1 000~2 500 nm）技术建立了原料肉新鲜度主要指标的在线检测系统。研究设计了装置的光源单元、光谱采集单元、控制单元和驱动单元，优化设计了光源固定支架和安放角度，编写了相应的控制程序，开发了实验室用和便于在不同生产线应用的两套在线检测系统。首先，对试验参数（传送带速度和样品到透镜入光口距离）进行了优化研究，通过光谱相似度比较和显著性分析，确定传送带速度是275 mm·s-1、距离是12 cm时能够获得更加稳定的光谱信号。然后，基于该试验参数，分别在静止和在线条件下采集了贮藏时间为1~13 d共50个猪肉样本的反射光谱，并利用抛物线拟合法对双波段光谱进行融合，以获取整条覆盖可见及近红外区域的完整光谱。为了使两个波段范围内的光谱数据点权重相同，在整个波段范围内均匀分布，借助三次样条插值法将所有光谱数据点以2 nm为间隔进行重新排布。采用窗口移动多项式最小二乘拟合法对光谱作平滑处理，采用标准正态变量变换对每条光谱进行标准化预处理，分别建立了静止和在线条件下新鲜度主要表征指标-颜色（L*，a*和b*）、pH和挥发性盐基氮的预测模型，以此验证所搭建系统的可靠性。经过对比分析，发现在线条件下的建模结果不如静止状态下的建模结果，这可能与在线采集时光谱存在漂移现象有关。进一步尝试利用一阶导数处理来消除基线漂移强化谱带特征，并对一阶导数和标准化处理顺序对建模结果的影响进行了探讨。结果发现先经过一阶导数再经过标准化处理，能更好地消除外部干扰造成的影响，建模结果更佳。在该处理方式下，基于第一波段光谱建立了颜色参数（L*，a*，b*）的预测模型，基于双波段光谱建立了pH和挥发性盐基氮的在线检测模型，预测相关系数分别为0.955 3，0.924 7，0.955 1，0.961 5和0.966 8。最后，为了验证模型的适用性，基于开发的便于在不同生产线应用的在线检测系统，利用独立的20个样本对在线模型进行外部验证，对颜色参数（L*，a*，b*），pH和挥发性盐基氮的预测相关系数分别为0.918 9，0.914 1，0.947 7，0.950 4和0.960 6。研究结果表明，该系统通过双波段光谱的实时采集和融合，可以获取更多反应样本内部信息的光学信号，具有更强的检测能力。结合设计的光路等其他硬件单元，可以同时获取样本表面更大区域的反射光谱信息，从而实现对原料肉新鲜度主要表征参数的无损、在线、实时评估。该系统便于组装和拆卸，可以适应不同企业生产线的实际需要，具有较强的实用价值和较好的市场前景。

Development and Test of On-Line Detection System for Meat Freshness Evaluation Based on Spectroscopy Technology

In order to realize real-time, on-line and non-destructive evaluation of main freshness attributes of raw meat, an on-line detection system based on dual-band visible/near-infrared reflectance spectroscopy(350~1100 and 1000~2500 nm) was established in this paper. The hardware which includes the light source unit, the spectrum acquisition unit, the control unit and the driving unit was designed. The light source fixing support and installation angle were optimized, and the corresponding control program was developed. Based on those, two sets of on-line detection systems were developed for laboratory use and to satisfy the demands of different production lines. Firstly, the experimental parameters including the conveyor speed and the distance between sample surface and the entrance of the lens were optimized. By comparing the spectral similarity and the significance analysis, the conveyor speed and the distance were determined as 275 mm·s-1 and 12 cm to obtain stable spectra. Then, based on the experimental parameters, the reflectance spectra of 50 pork samples stored for 1~13 days were collected under static and on-line conditions, respectively. The dual-band spectra were fused by parabolic fitting to obtain a complete spectrum which covered the whole visible and near-infrared region. Subsequently, all spectra were rearranged at 2 nm intervals by means of cubic spline interpolation to make the spectral data points distribute evenly over the two bands. Based on this, the spectrum was smoothed by the moving window polynomial fitting least square method and normalized by standard normal variable transformation. Then the prediction models for L*, a*, b*, pH and total volatile base nitrogen under static and on-line conditions were established and compared to verify the reliability of the constructed system. It was found that the modeling results for on-line detection performed worse than those under static conditions, and the reason may be attributed to the spectrum drift. Therefore, first derivative was further employed to eliminate the baseline drift and enhance the band characteristics. The influence of processing sequence of first derivative and standardization on the modeling results was also discussed. The results showed that the first derivative followed by standardization worked more successfully to eliminate the external interference. Then the prediction models for L*, a*, and b* were established based on the first band, and the models for pH and total volatile basic nitrogen were established based on the dual-band spectrum, with correlation coefficients of 0.955 3, 0.924 7, 0.955 1, 0.961 5 and 0.966 8. Finally, 20 independent samples were detected using the developed on-line inspection system to verify the model applicability, and the correlation coefficients for L*, a*, b*, pH and total volatile basic nitrogen were 0.918 9, 0.914 1, 0.947 7, 0.950 4 and 0.960 6, respectively. The results showed that by the real-time acquisition and fusion of dual-band spectroscopy, more optical signal were collected to reflect the internal information of tested samples. Combined with the designed optical path and other hardware units, the spectral information within a larger area of the sample surface were obtained. Thus, the non-destructive, online, and real-time assessment of the main attributes for raw meat freshness was achieved. The system was easy to assemble and disassemble, which made it possible to satisfy the actual needs of different production lines and had strong practical value and market prospects.