Quality Determination Of Citrus Fruit Using Electronic Nose,Electronic Tongue And Fusion System

Citrus is an important fruit crop cultivated widely in different agro-climatic conditions for its characteristics of high nutritional value, mildly sweet taste, and easy-to-eat. In this study, electronic nose (E-nose), electronic tongue (E-tongue) and Gas Chromatography-Mass Spectrometer (GC-MS) were applied to detect the quality of citrus, including quality tracing at different harvest stages and at different storage conditions, sensory analysis of mixed juices, high hydrostatic pressure (HHP) processed citrus juices, and the food additive, meanwhile, four quality parameters (total acid, total soluble solids, vitamin C, and sugar/acid ratio) were detected by traditional chemical analysis, and the sensory characteristic of mixed juices were analyzed quantitative descriptive analysis (QDA). The main objects of this research is as follows:(1) to explore the relationships between E-nose’s and E-tongue’s sensors, (2) to compare the effectiveness of data reduction methods based on feature extraction; (3) to optimize models (extreme learning machine (ELM), random forest (RF), and support vector machine (SVM)). The conclusions of this study are as follows:(1) Mandarin fruits (Citrus unshiu Marc.) at different harvest stages were detected by E-nose, E-tongue and GC-MS. The results of ANOVA partial least squares regression (APLSR) showed that the relationships between sensors, and quality indices and volatile components were existed but not obvious. Extreme learning machine (ELM), as an innovative neural network, were applied to process the signals of E-nose and E-tongue for sample classification and indices prediction. For mandarin classification, SVM and LVQ had same results (100%) as ELM models did in three data sets (E-nose, E-tongue, and the fusion dataset), but the training time of ELM were much shorter than that of SVM and LVQ. For prediction of quality indices and volatile components, ELM showed higher R2 and lower RMSE than SVM and MLR did.(2) The quality changes of mandarins stored at three different temperatures (4℃,10℃, and 20℃) were traced by E-nose, E-tongue, GC-MS, and traditional chemical analysis for 12 days (detected at every 3 days). Two-way multivariate analysis of variance (Two-way MANOVA) were used to analysis the two main effects of temperature and time on the quality of stored mandarins, and their interaction. Result based on Two-way MANOVA showed that the contents of vitamin C, sugar/acid ratio, limonene, terpinene, and myrcene were influenced not only by their main effects but also their interaction, but effects on soluble solid and total linalool showed no interaction. ELM models and SVM models based on E-nose data set and the fusion data set discriminated samples successfully. For quantitative analysis of quality parameters and volatile components, ELM showed better performance than Partial Least Squares Regression (PLSR) did.(3) E-nose, E-tongue, and GC-MS were applied to trace the changes of mixed juices, and QDA were utilized to evaluate the sensory characteristic of mixed juices. The results showed that the changing trends of main volatile components, such as limonene, valencene and myrcene, were increased (or decreased) with the decline of mixing ratio, but some other components were not paralleled with the mixing ratio. Same result were also found in QDA analysis. The fusion system composed of E-nose and E-tongue predicted the sensory scores successfully based on random forest (RF). And also, RF showed a satisfying performance in predicting the contents of volatile components.(4) Two varieties of citrus juices were processed by high hydrostatic pressure (HPP) ranged from 100 MPa and 500 MPa, and detected by E-nose.Three feature extractions (stable value, max value and area value) and three data reductions (PCA, LDA and locality preserving projection (LPP) were applied to preprocess the E-nose signals. According to the 2D-plot, the sample distribution based on LDA were much better than PCA, and LPP showed results similar to LDA; according to the discrimination results, SVM and ELM based on LPP or LDA had better classifying abilities than that based on PCA. Considering LPP is a unsupervised method, LPP was better than LDA which is a supervised method.(5) E-nose and E-tongue were utilized as new alternative detection systems to traditional technologies for inspecting the concentrations of food additives (benzoic acid and chitosan) which exist in foods with very small amount. According the PCA, the fusion system could solved some defects in the single system. For prediction of food additive, the RF models had better performance as the number of decision trees increased, whereas the ELM models were not satisfying when the number of hidden nodes closed to the number of samples. In general, RF and ELM showed better prediction performance than principal component regression (PCR) and SVM did.