【作者】 李鹏；

【导师】 钟南；

【作者基本信息】 华南农业大学 ， 农业电气化与自动化， 2023， 博士

【摘要】 三文鱼是世界范围内公认的营养价值和商业价值较高的海产品之一。目前,三文鱼市场出现了假冒、掺假和腐败变质等肉类食品质量安全问题,常规的理化分析方法有损样本、操作复杂、费时费力和成本高等,如何快速、无损和精准地检测三文鱼肉真伪和新鲜度是目前亟需解决的问题。本文采用拉曼光谱和高光谱成像检测技术,结合深度学习、宽度学习系统和光谱波长选择等相关算法,构建了高精度和强适应型的光谱检测模型,实现了三文鱼肉假冒、掺假和新鲜度的快速、无损和精准地检测。本文主要研究内容和结果如下:(1)研究拉曼光谱的三文鱼肉假冒鉴别方法。分析了真品和伪品三文鱼样本的光谱差异;引入了生成对抗网络(Generative Adversarial Network,GAN)来解决小样本拉曼光谱扩充问题并采用主成分分析方法评估了生成光谱的品质;设计了用于鉴别三文鱼肉假冒的一维卷积神经网络(Convolutional Neural Network,CNN)模型。结果表明:GAN方法生成的光谱具有和真实光谱高度相似的特征,将生成光谱添加到原始数据集后,更好的训练了CNN模型,从而提高了模型性能。与常用的机器学习模型相比,CNN取得了最佳性能,测试集上平均的准确率、精准率、召回率和F1-Score分别为96.67%,96.79%,96.67%和96.69%。(2)分别基于可见光-近红外(Visible Near-infrared,VNIR)和短波红外(Short Wave Infrared,SWIR)高光谱成像研究三文鱼肉假冒鉴别方法。提出基于宽度学习系统(Broad Learning System,BLS)的鉴别方法;将四种预处理方法与PLS-DA、KNN、RForest、SVM、BPNN、ELM和BLS模型相结合建立了全波长鉴别模型;采用单一的和混合的波长选择方法筛选特征波长,并结合BLS建立了基于所选特征波长的简化模型。结果表明:BLS模型的性能表现优于其他模型,VNIR和SWIR高光谱均能实现三文鱼肉的假冒鉴别,但VNIR高光谱的鉴别效果更好。其中,在全波长条件下,VNIR和SWIR范围内BLS模型的测试集识别率分别达到了100%和95.33%。在特征波长条件下,综合考虑识别率和波长数目,VNIR范围内优选出IRF-SPA-BLS和UVE-SPA-BLS两个模型,均选择了12个特征波长,测试集识别率均为98.67%;SWIR范围内优选出UVE-SPA-BLS模型,筛选了12个关键波长,测试集识别率为89.33%。(3)分别应用VNIR和SWIR高光谱成像预测三文鱼肉掺假比例。制备了包含11种掺假比例(0–100%,间隔10%)的样本并提取光谱数据;将五种光谱预处理方法和五种波长选择算法与CNN模型相结合建立了光谱反射率与不同掺假比例之间的定量关系。结果表明:SWIR范围内尽管通过波长选择方法提高了CNN模型预测精度,但依然差于VNIR范围内的建模结果。其中,在全波长条件下,VNIR和SWIR范围内的最佳预处理方法均是标准正态变量变换(SNV),其结合CNN模型得到两个波段范围内的平均预测结果分别为R_P~2=0.9885、RMSEP=3.3526、RPD=9.6882和R_P~2=0.9783、RMSEP=4.6437、RPD=6.8676。在特征波长条件下,VNIR和SWIR范围内的最佳波长选择策略均是VCPA-IRIV,其结合CNN模型的平均预测结果分别为R_P~2=0.9870、RMSEP=3.5778、RPD=9.0157和R_P~2=0.9839、RMSEP=3.9926、RPD=8.0251。最后使用最佳预测模型实现了制备样品中掺假物的分布可视化。(4)研究拉曼光谱的三文鱼肉贮藏时间鉴别方法。分析不同贮藏时间的三文鱼肉光谱差异;应用线性和非线性聚类方法探索所收集光谱样本的聚类趋势;比较深度信念网络(Deep Belief Net,DBN)、堆栈自编码(Stacked Auto Encoder,SAE)、长短期记忆网络(Long Short Term Memory,LSTM)和CNN共四种流行深度学习模型的鉴别准确率和抗噪性能。结果表明:非线性t-SNE方法的聚类效果最好,其轮廓系数为0.74。四种模型的测试集识别率均在96.00%以上,其中CNN模型的识别率最高(99.62%),并且其抗噪性能也优于DBN、SAE和LSTM模型。(5)采用VNIR高光谱成像检测表征三文鱼新鲜度的挥发性盐基氮(TVB-N)含量。使用PLSR结合不同的预处理方法和特征波长选择方法建立TVB-N定量预测模型。结果表明:原始光谱经去趋势方法预处理后建立的基于全波长的PLSR模型具有最优的预测结果,预测集的R_P~2=0.8712,RMSEP=2.9471,RPD=2.8134。比较特征波长的建模结果发现,CARS方法筛选出30个特征波长,CARS-PLSR模型的预测结果最好,其R_P~2=0.8608,RMSEP=3.0638,RPD=2.7062,稍差于全波长的建模结果。最后利用全波长模型反演了三文鱼肉冷藏过程中TVB-N含量的变化。此外,根据TVB-N含量将三文鱼肉新鲜度分为新鲜、次新鲜和变质三个等级,并使用SVM建立新鲜度分级定性鉴别模型以及提出基于麻雀搜索算法(Sparrow Search Algorithm,SSA)的SVM模型参数优化策略。结果表明:与传统的网格搜索(GS)、遗传算法(GA)和粒子群(PSO)方法相比,SSA方法更优,将SVM模型的鉴别准确率提高到了90.38%。更多还原

【Abstract】 Salmon is one of the world-wide recognized seafood with great nutritional and commercial value.Currently,meat quality and safety such as counterfeit,adulteration and spoilage often occur in the salmon market.Traditional physicochemical analysis methods are destructive,complicated,laborious and costly,thus how to rapidly,nondestructilvely and accurately detect the authenticity and freshness of salmon is a great need to solve.In this paper,Raman spectroscopy and hyperspectral imaging detection techniques coupled with deep learning,broad learning system(BLS),spectral wavelength selection and other related algorithms were employed,to construct the highly accurate and strongly adaptable spectral detection models,which can achieve rapid,nondestructive and accurate detection of counterfeit,adulteration and freshness of salmon.The main research content and results of this paper are as follows:(1)Raman spectroscopy was applied to identify counterfeits of salmon.The spectral differences between genuine and counterfeit salmon samples were analyzed.To extend small-sample Raman dataset,Generative Adversarial Network(GAN)was used to generate fake spectra,and the quality of the generated spectra was evaluated by principal component analysis.A cnvolution neural network(CNN)model was developed to identify salmon counterfeits.The results showed that the spectra generated by the GAN were highly similar to the real spectra.The generated spectra were added to the original dataset,which made the CNN model better trained,thus improving the accuracy of model.Compared with several popular machine learning models,CNN model achieved the best performance,with average accuracy,precision,recall and F1-Score of 96.67%,96.79%,96.67%and 96.69%,respectively.(2)Visible near-infrared(VNIR)and short wave infrared(SWIR)hyperspectral imaging were used to identify counterfeits in salmon fillets.A counterfeit identification method based on BLS was proposed.Several full wavelength models were established by combining four preprocessing methods with PLS-DA,KNN,RForest,SVM,BPNN,ELM,and BLS.Single and hybrid variable selection methods were employed to select feature wavelengths,and then which were fed into BLS model to establish simplified models.The results showed that the performance of BLS model was superior to popular machine learning models.Both VNIR and SWIR hyperspectral systems could be applied to identify salmon authenticity,but VNIR performed better than SWIR.In particular,under the full wavelengths,the accuracy of BLS models for the VNIR and SWIR were 100%and 95.33%,respectively;under the feature wavelengths selection,considering both model accuracy and wavelength number,two models were recommend for the VNIR,namely,IRF-SPA-BLS and UVE-SPA-BLS,both of which selected 12 characteristic wavelengths with an accuracy of 98.67%;The UVE-SPA-BLS model was recommended for the SWIR,it selected 12 key wavelengths and achieved the accuracy of 89.33%.(3)VNIR and SWIR hyperspectral imagings were used to predict the adulteration levels in minced salmon.Samples containing 11 adulteration levels(w/w,0–100%,10%interval)were prepared and spectral data were extracted.Five spectral preprocessing tools and five wavelength selection algorithms were combined with CNN model to establish the quantitative relationship between spectral reflectance and different adulteration levels.The results showed that although the prediction accuracy of the CNN model was improved by the wavelength selection method in the SWIR range,it was still worse than that in the VNIR range.In particular,under the full wavelength,the best preprocessing methods in the VNIR and SWIR ranges were standard normal nariate transformation(SNV),and which coupled with CNN model obtained the average results,with R_P~2=0.9885,RMSEP=3.3526,RPD=9.6882 and R_P~2=0.9783,RMSEP=4.6437,RPD=6.8676,respectively;under the characteristic wavelength selection,the optimal wavelength selection strategies in VNIR and SWIR ranges were VCPA-IRIV,and which coupled with CNN achieved the results of R_P~2=0.9870,RMSEP=3.5778,RPD=9.0157 and R_P~2=0.9839,RMSEP=3.9926,RPD=8.0251,respectively.Finally,the optimal prediction models were successfully used to visualize the distribution of adulterants in the prepared samples.(4)Raman spectroscopy was used to discriminate the salmon fillets storage time.The spectral differences of salmon fillets stored at different times were analyzed.Several linear and nonlinear clustering methods were applied to explore the clustering trend of collected spectral samples.Four popular deep learning models,including deep belief network(DBN),stack autoencoder(SAE),long short-term memory(LSTM)and CNN,were compared for modeling accuracy and noise resistance.The experimental results showed that non-linear t-distributedstochastic neighbor embedding(t-SNE)method performed the best clustering effect with a silhouette coefficient of 0.74.The test set recognition rate of the four models was above 96.00%,of which CNN model delivered the highest accuracy(99.62%)and better noise resistance than DBN,SAE and LSTM models.(5)VNIR hyperspectral imaging was applied to predict total volatile basic nitrogen(TVB-N)that characterizes the salmon fillets freshness.Several quantitative prediction models of TVB-N content were established by combining PLSR with different spectral pretreatment methods and characteristic wavelength selection methods.The results showed that the PLSR coupled with pretreatment methods de-trending(DT)achieved the best prediction performance,with R_P~2=0.8712,RMSEP=2.9471 and RPD=2.8134.Comparing the characteristic wavelength modeling results,it was found that wavelength selection CARS method selected 30 characteristic wavelengths,and CARS-PLSR model achieved the optimal results,with R_P~2=0.8608,RMSEP=3.0638,RPD=2.7062.Further,the salmon freshness was divided into three levels according to TVB-N content,including freshness,sub-freshness and spoilage.The qualitative discriminative model of freshness levels was established using SVM,and a novel optimization strategy of SVM parameter was proposed,namely,sparrow search algorithm(SSA).The results showed that compared with traditional optimization methods such as grid search(GS),genetic algorithm(GA)and particle swarm optimization(PSO),the SSA method was superior,improving the identification accuracy of the SVM model to 90.38%.更多还原