稳健偏最小二乘类模型及其在清真香肠红外质量控制中的应用

Robust One-Class Partial Least Squares for Quality Control of Halal Sausage by Infrared Spectroscopy

针对在实际多元质量控制中经常遇到的奇异样品问题,本研究提出了一种稳健偏最小二乘类模型。本方法基于Stahel-Donoho奇异度和样品重加权策略,用稳健的类中心和模型误差构造稳健的决策区间。将本方法用于清真香肠的红外分析,建立了稳健的质量控制方法。在香肠样品的不同部位进行取样,充分研磨后制备溴化钾压片,以空气为背景,测量4000～400cm"1范围的红外透射光谱。基于73个清真香肠样品和78个非清真样品的光谱数据,研究了新提出的稳健类模型的统计效率和稳健性。在有奇异样品存在的情况下,本方法能有效检出奇异样品,为新样品的预测提供稳健的决策区间。排除奇异样品后,基于原始光谱的模型灵敏性为0.846,特异性为0.936;基于标准正态变量法的模型灵敏性为0.923,特异性为0.974。

To tackle the problem of outliers in practical multivariate quality control, a robust one-class partial least squares （ROCPLS） model was proposed. Based on Stahel - Donoho multivariate outly- ingness and re-weighting of samples, a robust decision region was constructed using robust estimators of sample location and model errors. ROCPLS was applied to infrared spectroscopic analysis of Halal sausages and the robust quality control methods were developed. Different parts of a sample were cut off, finely milled and made into thin KBr disks. Transmittance FTIR spectra ranging from 4000 to 400 cm-1 of 73 Halal and 78 non-Halal sausage samples were measured and the robustness and statisti- cal efficiency of ROCPLS model were investigated. When the data set was contaminated with outliers, ROCPLS could detect them and provide a robust decision region for predictions of unknown samples. With outliers removed, the sensitivity and specificity of ROCPLS model were 0. 846 and 0. 936 for raw data, 0. 923 and 0. 974 for standard normal variant （SNV） spectra, respectively.