A Stacked Extreme Learning Machine Algorithm Based On NIR Spectroscopy And Its Application

Extreme Learning Machine(ELM)algorithm have fast modeling speed and strong generalization ability,but the error of forecast is large and unstable in processing the near infrared spectrum data of high dimension and small sample.In this paper,the elm was applicated stacked strategy and we proposed Stacked Extreme Learning Machine(SELM)for the characteristics of near infrared spectroscopy,and developed a prototype system based on traditional Chinese medicine process analysis.The main work of this paper is as follows:A stacked ELM algorithm is proposed,which is used to process the spectral data with high dimension and few samples.Ordinary ELM will encounter problems of flat matrix in the process of modeling,resulting in a low accuracy and instability for the model prediction.In the SELM algorithm,the column attributes of spectral data are stacked into several sub segments.Each sub segment was trained independently of the ELM model.Then weighted integration of all the sub segment models is then obtained,and the final SELM model is obtained.The weight of the integrated process is related to the Root Mean Square Error(RMSE)of each sub model,and the bigger the error is,the smaller the weight value is.The results of the sample components of SELM and ELM,SPLS,PLS for the open NIR spectral data sets are given.It is verified that the SELM algorithm is accurate by using near-infrared spectral data set.Through the method of cross validation to determine the near infrared spectral data number of segments and hidden node number of the elm model,which improved the stability of the model.The test uses three datasets:tablet,beer,and biscuit.In biscuit data set,there are four physicochemical indicators:fat,sugar,flour,water.Through the three datasets,the accuracy of the SELM algorithm is proved to be significantly improved compared with ELM,PLS and SPLS.The prototype of a SELM based software is designed and implemented,which is for analysis of the extraction process of Chinese medicine.The process and function of the SELM software includes data preprocessing,parameter setting,data saving,and the results of the software prototype of the extraction process of Chinese medicine were demonstrated.TCM data is extracted through multiple batches in a mixture of four kinds of medicinal herbs.Using the data of the former six batches to model SELM,then predicting the the content of the Chinese herbal medicine in seventh batches of Chinese herbal medicine.SELM get the higher accuracy of ELM,PLS,SPLS.