| Fixing is a key process in the green tea processing process.At present,the moisture content of green leaves is mostly judged by experienced tea masters through sensory methods to judge the moisture content of green leaves,and the sensory way is subjective,can not achieve quantification and accurate judgment of moisture content.The accurate measurement of the moisture content of fixing leaves requires the weighing and drying measurement calculation of the national standard oven method,which is time-consuming and labor-intensive,and cannot meet the needs of rapid mass production in tea processing.Therefore,there is an urgent need for near-infrared spectroscopy technology that can be fast,non-destructive and accurate to measure the moisture content of green leaves in real time.At the same time,the problem that the near-infrared spectrometer is susceptible to interference from the external environment,especially the temperature,is solved.This study confirms the feasibility of using temperature correction transfer algorithm and fully convolutional neural network to realize the online detection of moisture content of green tea green leaves.The main research results are as follows:(1)Based on the original information of the near-infrared spectral data collected from300 green leaves,Savitzky-Golay(S-G)convolutional smoothing,detrending,multiplicative scatter correction(MSC),and standard normal variety(SNV)preprocessing methods were used,where,The correlation coefficients of the model prediction set were0.9446,0.9403,0.9447 and 0.9398,and the rmsep of the model prediction set were 5.13,4.69,4.74 and 3.85,respectively,and SNV was selected.(2)Temperature interferes with the near-infrared spectral signal,and the model at a single temperature can be applied to this temperature,but the spectral signal acquired at other temperatures cannot be accurately detected.Therefore,the temperature correction algorithm needs to be used to correct the near-infrared spectrum data obtained at different temperatures,so that the model can accurately detect the near-infrared spectrum obtained at different temperatures.Based on the optimal preprocessing SNV results,the conventional correction algorithms used in this thesis include direct standardization(DS)algorithm,external parameter orthogonalization(EPO)algorithm,and piecewise direct standardization(PDS)algorithm,wherein The correlation coefficients of the model prediction set Rp were 0.9551,0.9483 and 0.9504,and the root mean square errors of the model prediction set were 1.93,2.12 and 2.03,respectively,and the optimal algorithm was selected as DS.(3)Based on the correction algorithm of deep learning,a fully convolutional neural network green tea water regression analysis model is established.The model parameters of CNN are established and the model construction is completed.The correlation coefficient of the model prediction set is 0.9875,and the rmsep error is 0.99.The results show that the model based on the fully convolutional neural network is feasible to directly predict the near-infrared spectral data collected at high temperature,and compared with the traditional temperature correction optimal algorithm,the correlation coefficient of the model prediction increases from 0.9551 to 0.9875,and the rmsep decreases from 1.93 to 0.99.The research results provide a new idea for solving the problem of temperature interference on near-infrared instruments. |
