| The internal quality non-destructive detection using near infrared spectroscopy (NIR) is very important for promoting the standardization process and quality control for kiwifruit production. The combined effecting mechanism of the detecting conditions on the spectral response characteristics of kiwifruit, spectroscopy data processing and prediction modeling approaches are the key elements for kiwifruit near infrared non-destructive testing technology at present.A plenty of research were conducted on the effecting mechanism on the spectral response characteristics of kiwifruit from single detecting parameter while less from combined parameters. The accuracy and versatility of prediction models is also low due to backward modeling approaches. Therefore, these models are seldom implemented into practice at present.Using the MPA Fourier's NIR Spectrometer make from the BRUKER company in Germany, taking kiwifruits as testing samples picked from different locations and orchards in Yangling Shaanxi with varied storage time, species and maturity, this research systematically studied the influences from the near infrared spectrometer instrument parameters, detection parameters and sample parameters on the spectral response characteristics of kiwifruit as well as on the accuracy of prediction models. This research has also implemented nonlinear optimization approaches such as genetic algorithm, wavelet analysis, artificial neural networks and clustering analysis on wavelength optimization, signal compression, quantitative and qualitative models developing in kiwifruit near infrared spectroscopy detecting. The following conclusions were drawn.1) The parameters of the near-infrared spectrometer such as the resolution, scanning frequency and attachments to scanning the fruit will affect the internal quality assessment of kiwifruit accuracy directly. The optimized parameters combination is 8cm-1 of resolution, 32 times of scanning frequency and the solid fiber of spectrum obtaining attachment, after the overall considering of the scanning speed, signal to noise ratio, model prediction accuracy and the requirements of online detection.2) The prediction models of sugar content built by the spectral data collected at different positions on the equatorial part of one kiwifruit showed no significant difference on measurement accuracy. Therefore, the direction of the fruit is not necessary to be considered when the spectra data of fruit is collected. The accuracy of prediction model built by average spectrum from multiple collecting points is higher than models built by single collecting point and multiple collecting points. Therefore, it is advised that a number of fiber-optic probes should be used to collecting multi-points spectrum and their average spectra is used to build the prediction model.3) It is feasible to detect the soluble solids content of kiwifruit from different locations and orchards with varied storage time and maturity in Yangling using near-Infrared Spectroscopy of 12000 - 4000cm-1. The differences from the kiwifruit samples such as growing location, storage period, size, sugar content have great impacts on the accuracy of the prediction model of sugar content. The more of the modeling samples and the ingredient content gradient, the high of the model versatility while the low of the reliability of the model. In the range of 11991.6cm-1～5446.2 cm-1, the near-infrared diffuse reflectance spectrum and the soluble-solids content in kiwifruit showed remarkable linear correlation with 93.65 of the decided coefficient R2 and 0.656of the RMSEP of the prediction model.4) The genetic algorithm is efficient to build models of sugar content of the kiwifruit with high prediction accuracy. The model takes the number of sub-intervals in the full spectrum as the number of chromosomes and the formula of R/(1+δ) as the objective function in which R is the correlation coefficient between the cross-check prediction value and standard value whileδis of the prediction mean square deviation. The optimized parameters combination is 20 of the initial population, 0.8 of the crossover probability, 0.05 of mutation probability, and 600 of iterations. The results showed 10 sub-interval that have useful information on prediction model development of soluble solids content . The accuracy of the model built by the spectrum data from the above spectral regions is higher than model built by all data from full spectrum region.5) The wavelet analysis algorithm is efficient for noise removing and data compression for kiwifruit spectrum data processing. The optimized model with 3 of filter length and 4 of the decomposition scale takes Coiflet as the wavelet basis and select the threshold value based on the Sqtwolog rule. The results showed that the wavelet analysis algorithm model removes excessive noise spectral areas and reduces the wavelength points to shorten the time for modeling.6)The prediction models of sugar content of the kiwifruit built by the three-tier BP artificial neural network results higher prediction accuracy than the models built by PCR and PLS with 89.8273 of the prediction model determination coefficient R2, RMSEP=0.3256 . The model takes the increasing-momentum as the method to improve the BP network performance and the initial weights and threshold are set by the MATLAB program at a certain range randomly. The model also takes Tan Sigmoid Function as the activation function in the hidden layer and the Pureline Function as the linear activation function in the output layer of the BP network. The other parameters combination of the BP network model is 10000 of the maximum training times, 5 of nodes in the hidden layer, 0.7 of the learning rate, 0.08 of the goal error, and 0.4 of the momentum coefficient. The results also showed that the artificial neural network used together with the genetic algorithm and PLS, is an efficient method to build NIR prediction models for kiwifruits.7) It is very important to choose useful wavelength range and to adopt the method of the data pre-processing to extract characteristic information for kiwifruit species and injured identification using near infrared spectroscopy. At the wavelength range of 4500-8000cm-1, the cluster analysis method, with the standards-based, the first range of calibration and the level reproducibility normalization, can be used for species identification by near infrared spectroscopy. The classification accuracy is up to 100% for the three species of the Hayward, Guyatt, and Qinmei kiwifruits. However, the accuracy of the Qualitative Identification Analysis Method is only 56% for these three species classification.8) The soluble solids content at the injured kiwifruits will increase initially, and the spectrum absorbance of injured kiwifruits is greater than normal kiwifruits. However, as time passed, the spectrum absorbance of injured kiwifruits is less than the normal position. This feature can be used to remove kiwifruits injured for a long time while not efficient for kiwifruits injured for a short time such as two days or so. Absorbance at a short period of time changes of species differences. The use of cluster analysis to identify damage to the Qiu-xiang Kiwi is the best time of injury 10 days. The best time to njury identified to Qinmei and Yate is 3 days after injury. Damage Identification of Hayward kiwifruit are the best time after injury for 14 days. NIRDRS database law does not apply to the damage identification of kiwifruit.The results of this research promoted theoretical and technical guidelines for the standardization of testing conditions, improving the accuracy of prediction models and development of automatic sorting systems for kiwifruits using near infrared spectroscopy technology.
|
PDF Full Text Request