| The Chinese liquor is a treasure of Chinese culture with a history of thousands years,which Chinese people take pride in.However,many people are suffering from misfortunes since there are lots of fake and adulterate wines in China’s liquor market nowadays.Moreover,it always takes an extremely long time to perform experiments and analysis to distinguish them from true branded wines because of the limitation of current identification methods.Therefore,we urgently need to come up with a new detecting method of higher efficiency for China’s liquor market.For this purpose,this paper proposes a detecting analysis model based on ensemble learning of fluorescence spectrum,which successfully combines fast responding speed of laser-induced fluorescence technology and ensemble learning algorithm.This paper initially analyses the current condition of China’s liquor market.In this stage of analysis,we chose four liquors as our samples,“56 degree Red Star Erguotou”,“56 degree Sanjiu”,“42 degree Little Qinghua” and “42 degree Red Star Erguotou”.We used these four wines in our experiment to verify feasibility of our new method when facing wines of different brands and degrees.Firstly,we collected 100 fluorescence spectra from each sample and preprocessed the collected spectra by using the preprocessing algorithm to correct the noise and deviation from the acquisition process.Then,we added the original spectrum to these four kinds of preprocessed spectra and put these new five kinds of spectra into analysis to get the best preprocessing method of highest learning efficiency by applying AdaBoost algorithm.Next,we applied the LightGBM algorithm in order to enhance our identification model of wine sample spectrum and used exactly the same hyperparameters to train our model into a more powerful model based on this LightGBM algorithm.Through this model building process,we have found out that if we apply SNV preprocessing method,we can acquire better outcomes when using the AdaBoost method to build models,and we can acquire the best outcome when we select learning rate as 0.1 after 300 times of iterations.In addition,when building the model according to the same SNV preprocessing spectrum,the five-fold cross-validation equals to 1,and the required model training time is even less than the time required when models are trained based on AdaBoost method.In conclusion,we can achieve 100% of accuracy and well universality when we combine the SNV preprocessing method and LightGBM spectrum identification model.In this paper,it is the first time for the LightGBM method to be applied with laser-induced fluorescence technology in wine fluorescence spectrum identification.Moreover,the simulation results show that the model based on the combination of the LightGBM method and the SNV preprocessing method is worth further research.Our paper introduces an identification model which can quickly and accurately distinguish Chinese liquors.Furthermore,we will also provide a crucial example for building a more perfect liquor laser-induced fluorescence spectroscopy database in the future.Figure [32] table [4] reference [72]... |
