Adaptive Modeling Strategy For Near Infrared Spectroscopy Of Complex Substances And Its Application

Coal,fuel oil and soil are the main resources on which human beings depend today,and they play an important role in the resource structure of various countries in the world.In this situation,to strengthen the supervision of resource efficiency and reduce environmental pollu-tion,the rapid detection technology of these substances has put forward higher requirements.The traditional detection principle of these substances generally adopts laboratory standard manual method,but this method is slow in analysis speed and high in analysis cost,which cannot meet people's needs in real industrial and agricultural production.Near infrared spec-troscopy is one of the effective and feasible methods to solve this problem.However,each wavelength of near infrared spectroscopy overlaps a large amount of information about other components.Its spectral characteristics show that the structure is complex,the peaks overlap seriously,the effective information band of the measured substances is wider,the spectral var-iation is larger,and the spectral data interferes with more noise.Therefore,the operation of practical analysis is slightly difficult,especially for some substances with more complex com-ponents.Therefore,how to use appropriate chemometrics methods to establish an effective near infrared spectroscopy calibration model for the spectral data of complex substances will play a huge role in promoting this field.In this dissertation,artificial intelligence algorithm is used as the core,and the near infrared spectroscopy data of coal,gasoline,diesel,soil and other complex substances are taken as the research object.Some intelligent near infrared spectroscopy model-ing strategies suitable for the analysis of complex substances are established,and the strategies are applied to the actual near infrared spectroscopy data measurement.To this end,this disser-tation has carried out the following innovative work:In order to overcome the shortcomings of traditional optimization algorithms in establish-ing near infrared spectroscopy calibration model for complex substances,such as the difficulty of interpreting the optimal solution with reasonable chemical mechanism,and the inadequate optimization of the parameters of pretreatment methods and the order in which they participate in calculation,a kind of linear synergy adaptive modeling strategy(LSA-XM-X)was proposed.LSA-XM-X can be divided into linear synergy adaptive interval modeling strategy(LSA-IM-X)and linear synergy adaptive moving window modeling strategy(LSA-MWM-X).LSA-IM-X combines partial least squares regression algorithm with genetic optimization algorithm to generate a synergy adaptive interval partial least squares method(SA-IPLS-GA)based on ge-netic algorithm.LSA-MWM-X generates a synergy adaptive moving window partial least squares(SA-MWPLS-ICA)method based on immune cloning algorithm by combining partial least squares regression algorithm with immune cloning optimization algorithm.These two kinds of algorithms realize precise location of spectral effective information area,efficient au-tomatic optimization of spectral data processing method parameters and processing sequence,and minimize the computational complexity and the difficult problem that analysts need to try to configure parameters artificially.In order to overcome the problem of non-linear model in practical work,a synergy adap-tive moving window support vector machine(SA-MWSVR-ICA)based on simplified immune cloning algorithm is proposed,which is a complement and extension of LSA-XM-X.The anti-body gene structure of the algorithm can optimize the parameters of spectral data preprocessing method and its processing sequence,calibrate the model parameters and spectral wavelength variables at the same time.By combining the special antibody gene with the decoding method,the algorithm achieves the co-optimization of changeable resolution grid search method and immune cloning algorithm,which has higher preprocessing accuracy than other algorithms.In this dissertation,the frequency domain characteristics of near infrared spectroscopy of coal are studied,and a frequency domain adaptive analysis method is proposed,which can au-tomatically establish the optimal quantitative analysis model according to the frequency domain characteristics of coal.The method is validated by taking the heating value of coal as the target to be measured,and the experimental results are relatively good.Compared with the traditional methods such as principal component regression,partial least squares regression and back prop-agation neural network,this method has higher prediction accuracy,and effectively avoids the drawbacks of potential over-fitting and false effective models of random search in frequency domain.In this dissertation,a genetic algorithm-based depth cooperative adaptive moving window partial least squares(DSA-MWPLS-GA)method is proposed.By using this method,the char-acteristics of near infrared spectral prediction models of moisture,ash,volatile matter and heat-ing value in coal are analyzed in detail.The results show that DSA-MWPLS-GA can automat-ically acquire prior knowledge through the combination of the "Search Basic Calibration Model" and "Search Evolution Model",thus ensuring the performance indicators of the model and improving the efficiency of model derivation.In this dissertation,a series of adaptive near infrared spectroscopy modeling strategies are proposed to solve the drawbacks of expert experience knowledge,high computational com-plexity and difficulty in optimizing the calibration model of near infrared spectroscopy for com-plex substances.The important performance indexes of the near infrared spectroscopy model are greatly improved,and the strategies are applied to the actual on-line coal quality measure-ment device.They have broad application prospects.