Near Infrared Spectral Modeling Based On Stacked Autoencoder

In the traditional detection technology,the content of the target substance is usually obtained by sampling offline analysis,which is time-consuming and labor-intensive,and cannot meet the real-time requirements of system control,optimization,and monitoring.The detection technology based on near infrared analysis has the advantages of rapidity,non-destructiveness and wide detection range,and is widely used in agriculture,petrochemical,pharmaceutical,food fermentation and other industries.In the near infrared analysis technology,an accurate calibration model is crucial to detection accuracy.Based on Stacked Autoencoder(SAE)and transfer learning,this dissertation investigates nonlinear,dynamic,and small-sample problems in near infrared spectroscopy modeling and its applications,respectively.The main research contents include:(1)Nonlinear modeling for near infrared spectral based on SAE.The complex industrial field environment has a great influence on near infrared spectroscopy measurements,and accurate results may not be achieved with a simple linear model.To this end,SAE,a nonlinear modelling approach,is employed to address this problem.Nevertheless,the pre-training in SAE is an unsupervised procedure,where features of the input data can be obtained,whilst physical and chemical values are not included.A near infrared spectral modeling method based on Stack Supervised Autoencoder,therefore,is proposed,where pre-training and fine-tuning are realized in a supervised way.Physical and chemical values can be added to the network pre-training stage to make the pre-training parameters closer to the optimal parameters as well as reduce calculation during fine-tuning.(2)Dynamic nonlinear modeling of near infrared spectral data.Considering dynamic correlation of physical and chemical values,a near infrared spectroscopy modeling method is proposed based on Stacked Dynamic Supervised Autoencoder(SDSAE),which addresses the dynamic problem by constructing a dynamic spectral matrix for modelling.The SDSAE algorithm is applied to the determine of total nitrogen of citric acid fermentation raw materials;Compared with other methods without considering the dynamic of data,it can deal with the dynamic of physical and chemical values effectively.(3)Deep transfer learning modeling of near infrared spectroscopy with small samples.In the near infrared spectrum model,few near infrared spectral data with physical and chemical values are included in the target domain.The novel approach combines deep learning and transfer learning,similar model parameters,from the source domain to the target domain,are used for model training with less training sample and high precision.The logic behind the proposed approach is: deep learning can extract deep feature from the data,while transfer learning can apply the data from source domain to target domain as well as reduce the data use from of target domain.The result of a melamine detection experiment proves that deep transfer learning can model the near infrared spectroscopy more accurately by using fewer samples.