【摘要】 以腹泻性贝毒(diarrheal shellfish poison, DSP)污染和未污染良好贻贝为研究对象，利用近红外光谱仪采集950～1 700 nm波长内各120个样本的光谱数据；分析确定适合贻贝近红外光谱(near-infrared spectroscopy, NIS)的最佳预处理方法来消除环境光的影响；构建多层感知机(multi-layer perceptron, MLP)模型作为检测DSP污染贻贝的分类器。将240个样本构成的数据集按照7∶3随机划分为训练集和测试集，将运行50次模型得到的准确率的平均值作为最终评价指标，检测DSP污染贻贝的准确率达到99.94%。该研究所构建的基于NIS的MLP模型对DSP的检出限为35μg/kg。对于实际应用中存在的数据集中训练集所占比重不同、小样本和类别不均衡等问题，分析了MLP模型的检测性能。实验结果表明，基于一阶导数光谱预处理的MLP模型对以上3种问题不敏感，说明该分类器泛化能力及鲁棒性较强。因此，将NIS与MLP分类器结合是一种可行的贝毒无损鉴别的新方法。更多还原

【Abstract】 In this study, diarrheal shellfish poison(DSP)-contaminated and non-contaminated mussels were used as the research objects, the near-infrared spectrometer was used to collect the spectral data of 120 mussel samples of each class in the wavelength range of 950-1 700 nm. The best preprocessing method for near-infrared spectroscopy(NIS) of the mussels was determined to eliminate the influence of ambient light. Multi-layer perceptron(MLP) model was constructed as a classifier to detect DSP-contaminated mussels. The dataset composed of 240 samples was randomly divided into training and test datasets according to the ratio of 7∶3, the average accuracy of the model by running 50 times was the final evaluation index, and the accuracy of detecting DSP-contaminated mussels reached 99.94%. The detection limit of the MLP model based on NIS for DSP was 35 μg/kg. The detection performance of the MLP model was analyzed for the problems of different ratios of training sets, small sample datasets, and unbalanced classes in practical application. The experimental results showed that the MLP model based on first derivative spectral preprocessing was insensitive to these three problems, which indicated that the classifier had strong generalization ability and robustness. Therefore, the combination of NIS and the MLP classifier provided a feasible new method for the non-destructive identification of shellfish toxicity.更多还原