贝类重金属污染的多模态融合光谱开集检测及不确定度研究

Shellfish is a kind of seafood that is easily polluted by heavy metals. It is very important to guarantee the safety of shellfish products by detecting the heave metal pollution of shellfish in a rapid way. Current spectroscopic techniques cannot make the correct detection on unknown or partial unknown samples in the practice, because the calibration of spectroscopic models cannot includes samples with all pollution situations and random interference factors in the detection process. On the basis of the previous funding, this project will apply three main spectroscopic techniques for heavy metal detection in foods, namely infrared, Raman and laser-induced breakdown spectroscopies to study and develop open set detection algorithms for spectroscopic technique based heavy metal detection of shellfish, establish the data description borders, deny the unknown samples in models. Besides, this project will focus on obtaining the joint posterior probability distribution of prediction samples and training samples to give the evaluation of degree of uncertainty for partial unknown samples under test. In addition, this project will study unified construction algorithms for multi-mode hierarchical data structure to effectively fuse the multiple kinds of spectroscopic information, resulting in further improving the accuracy of the open set detection and the evaluation of degree of uncertainty. This project aims at improving the detection capability of spectroscopic techniques for rapid detection of heavy metal pollution in shellfish, expanding their applying areas, making them more to meet the demand of shellfish industry and guarantee the safety of shellfish products more efficiently.

贝类是一种易受重金属污染的水产品。贝类重金属污染的快速检测对于保障贝类食品安全具有重要意义。现有的光谱检测技术由于无法对实际过程中所有情况的样本信息以及检测过程中的随机干扰因素进行训练建模，导致难以对实际过程中未知或部分未知的样本信息做出正确的检测。本项目拟在前期基金的研究基础上，采用红外、拉曼以及激光诱导击穿光谱等三种主要的食品重金属光谱检测技术，研究和设计针对贝类重金属光谱检测的开集算法，构建数据描述边界，拒绝识别模型未知的样本。同时，研究待测样本和训练样本的联合后验概率分布，对模型部分未知的待测样本给出不确定度评估。此外，研究多模态层次化数据结构的统一搭建算法，实现异源光谱信息的有效融合，进一步提高开集检测和不确定度分析的精准度。此项目的开展旨在提高光谱技术进行贝类重金属污染快速检测的能力，扩大其应用范围，使其更加满足于实际贝类产业的需要，能够更有效地保障贝类食品安全。

近些年，随着城市工农业生产的快速发展，大量有害的重金属被不断地排放，导致土壤，河水等受到严重的生态威胁。贝类水产品作为一种栖息在滩涂上生长的虑食性水产品，极易受到重金属元素的污染，其富集的有害重金属具有隐蔽性、长期性等特性，对人类健康形成了潜在威胁。针对这个问题，本研究提出应用光谱技术对贝类重金属污染进行快速检测，扩大对贝类产品的抽检力度，保证广大群众的贝类食用安全。本研究内容主要包括，利用小波变换、模拟退火算法、连续投影算法、间隔偏最小二乘算法、改进无信息变量去除算法等，筛选能够反映重金属污染信息变化的光谱响应信号的指纹波段，研究结果发现3247-3251cm-1、1614 cm-1和1579-1581cm-1红外波段跟重金属富集信息相关性强,这些区域主要跟蛋白质成分变化相关,而1228 cm-1、1394 cm-1、1452 cm-1和2960 cm-1这些波段区域与重金属富集信息也有一定的相关性,而这些波段可能与磷脂质、脂肪酸等成分的变化相关；采用差异排序总和算法实现对提取的特征指信息进行融合研究，其定性和定量结果均优于单特征波段信息所构建模型的精度；基于高斯过程、单类支持向量机、融合偏最小二乘等算法对泥蚶重金属污染信息的定性和定量的检测分析,结果表明多种重金属污染的泥蚶样本多分类结果识别率达到了93.3%，单类模型识别率达到了99.17%，而基于光谱技术对铜元素构建的定量分析模型的预测集相关系数（Rp）和剩余标准偏差（RPD）分别达到了0.97 和2.77。此外，课题组还针对光谱数据问题提出了样本变量共识模型，变量选择算法，单类模型算法和线性回归的稀疏分类算法，解决了光谱信息变量冗余和光谱模型准确性的问题，提高了模型检测精度。本项目的研究成果不仅为贝类重金属的快速检测提供了一种新的研究思路，而且还对其他水产品的品质和安全检测提供理论指导和技术支持。