| Infrared spectroscopy can reflect the chemical composition and properties of substances at the molecular level.Its combination with multivariate analysis methods forms infrared spectroscopy analysis technology,which has been broadly used in petrochemistry,agriculture,pharmacy,food,medicine and other fields.Due to the characteristics of instantaneous determination of types and various physicochemical properties of substances,infrared spectroscopy has great potential to develop as a tool for the perception of intrinsic information.However,there are still some issues that restrict its practical application in the field of artificial intelligence(Al).(1)The modeling and maintenance process of infrared spectral analysis is too complicated with heavy workload,resulting in high cost and long period;(2)At present,the objects that can be analyzed by infrared spectroscopy are still limited to samples with good light transmittance and uniformity,and the detection limit is not less than 5 wt‰.For different types of samples with similar compositions,complex shapes,uneven distribution,and susceptibility to moisture changes,effective identification has not yet been carried out;(3)For some samples rich in water,the spectral profiles of the tested components and the water components are highly overlapped,and the spectra are sensitive to temperature changes,which seriously interfere with the extraction of useful information.The commonly used multivariate analysis methods can no longer tackle these issues.This paper aims to carry out research on information mining methods of spectral databases to overcome the defects of modeling and maintenance;The‘dynamic'spectrum combined with image recognition technology is suggested to solve the problem of highly similar sample recognition;A method of Gaussian curve fitting combined with genetic algorithm is proposed to tackle the technical bottleneck of signal resolution in spectral analysis of complex systems and processes that are susceptible to changes in aqueous composition and temperature.The main research contents,results and innovations of this paper are as follows.Chapter 2.Research on a spectral databases mining based real-time quantitative analysis method.The propose of this chapter is to develop a spectral databases mining based real-time quantitative analysis method to overcome the drawbacks of traditional modeling methods,such as excessive workload,long cycles,and high cost,making it easier to realize the real-time determination of multi-properties of materials by infrared spectroscopy.In the experiment,the infrared spectral analysis of asphalt was taken as the research object.431 asphalt samples were collected from refineries,their wax content,penetration and softening point were measured by standard test methods and their infrared spectra were collected by using the attenuated total reflection accessory.All the samples are divided into library-building sets and validation sets.The asphalt spectral database is constructed by using the spectra and properties of the library building sets,while the validation sets are employed to verify the performance of the new method.The obtained root mean standard errors of prediction(RMSEP)for wax content,softening point and penetration respectively are 0.14%,0.55?and 4.71(0.1mm),which are less than the reproducibility error of the standard test method,indicating that the results of the new method and the standard method were consistent.Compared with the prediction results of two commonly used multivariate analysis methods(including partial least squares regression(PLS)and local densification modeling(LMD)),the results show that the prediction performance of the new method and PLS is at the same level.The new method avoids the complicated process of PLS modeling and maintenance,and effectively solves the technical problems that hinder the practical application of infrared spectrum analysis technology.Compared with LMD method,the new method has obvious improvement in repeatability,calculation speed and prediction robustness,and its prediction results are more accurate for samples located in areas with low density and unreasonable distribution in spectral database.Chapter 3.Research on pattern recognition method of‘dynamic'infrared spectroscopy combined with deep learning.Based on the difference between infrared spectra,rapid classification and identification of substances can be realized by means of pattern recognition method.However,for different types of samples with large morphological changes,uneven distribution,and highly similar chemical composition,the spectral difference on which classification is based is very weak,thus it is difficult to classify and identify them using common spectral pattern recognition methods.This is an unsolved classification and identification problem in the field of infrared spectroscopy.For this reason,this chapter proposes‘dynamic'spectra that can expand the discriminative information,and chemical images are constructed by two-dimensional correlation analysis,followed by spectral classification and recognition method based on GoogLeNet deep learning architecture and transfer learning.Cashmere textiles and cashmere/wool blended textiles,as well as pure cotton and mercerized cotton textiles are selected as research objects.Water perturbation was applied to the dried samples,samples with different moisture contents were prepared,and their moisture-dependent near-infrared spectra were measured.For dried samples and samples with different moisture contents,a total of 16 models,including soft independent modeling of class analogy(SIMCA)and support vector machine(SVM),have been established by respectively using their original,first-order derivative,second-order derivative,and multivariate scattering correction spectra.The classification model was established for comparison using the new method and moisture-dependent spectra.The experimental results indicate that the prediction accuracy of traditional spectral pattern recognition methods are less than 80%,which cannot meet the need in practice.Using the proposed method,the overall prediction accuracy of cashmere and cashmere-wool blend is 92.59%,and that of cotton and mercerized cotton is 94.62%,which meets the requirements in actual application.In this study,the image(two-dimensional matrix)classification method is applied to spectral(one-dimensional vector)classification and recognition for the first time,opening up a new way for the research on spectral analysis.The introduction of transfer learning effectively solves the issue that a small amount of samples available in spectral analysis cannot satisfy the need of training deep learning-based architecture,and provides a successful demonstration for applying advanced AI technology to solve chemical-related problems.Chapter 4.Research on an adaptive-weighted-fitting-based spectral classification and recognition method.For different kinds of natural samples that are highly similar in composition and susceptible to moisture,changes in ambient humidity have a great influence on their infrared spectra.Although the accuracy of prediction can be improved by drying or balancing moisture methods,it can deprive spectral analysis of the advantages of instant detection.Common spectral pattern recognition methods cannot effectively classify and identify them.To this end,an adaptive-weighted-fitting-based spectral classification and instant recognition method was developed in this chapter.The classification and identification of cashmere textiles and cashmere/wool blended textiles were selected as model system.A total of 120 samples of cashmere,wool and cashmere/wool blended fabrics with diverse colors and textures were collected from the market.Their types were determined using standard methods.Dried and moisture-containing samples were prepared,and their near-infrared spectra were collected by portable spectrometer.For dried samples and moisture-absorbing samples,SIMCA,SVM and new methods are respectively used to establish classification and identification models,and the effects of common spectral pretreatment methods and moisture changes on the model performance are studied in detail.The results show that the prediction performance of the three methods is at the same level for dried samples.For moisture-absorbing samples,the performance of the new method is far superior than that of other methods.The prediction accuracy rate of cashmere textiles is 93.33%,and that of cashmere-wool blended textiles is 96.60%,suggesting the proposed drying-free method meets the requirements in practical application.Chapter 5.Research on a multi-component overlapping spectral resolution method combining Gaussian peaks with genetic optimization.Gaussian curve fitting is a typical algorithm for fitting and separating overlapping spectral band and still not suitable for separating overlapping spectral bands of more than 3 components.This has been a hot spot in spectral component research in recent years.A multi-component overlapping spectral resolution method combining Gaussian curve fitting with genetic optimization was proposed in this chapter to solve the problem of multi-component(greater than 3)overlapping spectral resolution.The Amphiphilic thermosensitive hydrogel has great prospects in the field of biology,and its phase transition mechanism has also become a hot spot in this field.Both hydrophilic and hydrophobic segments involved in the hydrogel result in the coexistance of intramolecular and intermolecular hydrogen bonds.In addition,according to the theory of aquaphotomics,water has multiple(more than 6)"components" and is sensitive to changes in hydrogen bonding.Therefore,the thermosensitive mechanism is very complicated.Near-infrared spectrum is sensitive to hydrogen-containing groups,therefore,is suitable for studying temperature-sensitive hydrogel.However,the near-infrared band of water is wide,and the bands of different water components overlap highly.The common resolution method can only discriminate 2-3 water components rather than more,which hinders the in-depth analysis of the phase transition mechanism.In this chapter,ABA triblock hydrogel solution is taken as the research object,and the in-situ near-infrared spectra during its sol-gel phase transition process are collected online.Using the new method,the spectral bands of 6 different water components were successfully resolved.The change of the content of each water component during the phase transition was quantitatively studied,and the mechanism that S1 and S2 water components provide driving force for phase transition was revealed. |
PDF Full Text Request