A New Intelligent Modeling Algorithm Research And Application For Molecular Vibration Spectrum

Entering the new century,the industrial production process is developing towards large-scale,continuous and integrated direction,as a result the requirements for real-time and accurate detection of industrial production process is becoming more and more stringent.With the development of chemometrics,molecular vibration spectroscopy is widely used for real-time on-line detection in numerous different fields.In this paper,according to the characteristics of molecular vibration spectrum,a variety of wavelength selection algorithms which combine Importance Coefficient of Variable Projection(VIP)and Particle Swarm Optimization algorithm(PSO),and CNN model,local model based on different similarity,the combination of local model and wavelength selection are proposed,applied to the molecular vibration spectrum.The results show that the new wavelength selection algorithm and quantitative analysis model proposed in this paper can effectively,not only quickly but-also accurately analyze the attribute values of samples,and also achieve excellent performance in reducing computational complexity.The main contents of this paper can be summarized as follows:1.This paper proposes a new adaptive particle swarm optimization algorithm based on VIP coefficient,called VAPSO,to solve difficult control,fast convergence,easy to fall into the local advantages but unable to jump out,or other problems in BPSO.Result shows that VAPSO can find a subset of vwavelengths better and faster.2.The iVAPSO algorithm is proposed,which combines interval partial least squares.Firstly,we explore the best wavelength interval combination.And in second round best wavelength points in these interval are selected.IVAPSO algorithm reduces the particle length in iteration,and achieve higher and more stable prediction performance in less time.3.Due to the characteristics of the molecular vibration spectrum,the CNN model,which is always used in two dimension image domain,can also be applied in one dimension quantitative analysis.From the results of NIR spectrum of biodiesel,it is found that CNN shows great performance when it it applied to suitable dataset,such as faster convergence speed,higher prediction accuracy,etc.Compared with PLS,the RMSEC is reduced by 50%,and the RMSEP is reduced by 19%,superior than other models.With the collection of data sets,the performance of CNN will be more prominent.4.Three difference similarities,Euclidean distance,Euclidean distance between the net signal,and the spectral information divergence,are explored in local PLS model.Then put forward the local model combine with VAPSO algorithm to select wavelength subset and sample subset(Loc-VAPSO).This method costs shorter time but has better performance than the local model and VAPSO algorithm.Good prospect in the field of quantitative analysis of molecular vibration spectrum is foreseeable.5.The new wavelength selection algorithms and quantitative analysis models proposed before,are applied to detect the blending ratio of Biodiesel Blended Oil.The near infrared and Raman spectra were detected for the mixed oil samples,and the preprocessing and devision of sample were carried out respectively.The proposed new spectral characteristic wavelength screening method(VAPSO,iVAPSO)and the new quantitative analysis model(CNN model,Loc-PLS and Loc-VAPSO)were applied to biodiesel spectroscopy.The experimental results show that the new algorithm proposed in this paper have a great improvement in the detection accuracy and reduction of computational complexity of biodiesel blending ratio.