Research Of Dye Liquor Spectral Classification Algorithm Based On Machine Learning

As China being the center of the global printing & dyeing industry,the traditional extensive artificial development mode has been unable to meet the increasing demand of the people,while industrial waste water of printing & dyeing industry also brought great pressure to the environment,thus achieving automatic control and real-time monitoring the various parameters of dyeing process are the inevitable trend of the development of printing & dyeing,where dye liquor concentration is the most important parameter.At present,the most commonly used dye liquor concentration detection method is spectrophotometry,which is based on the Lambert-Beer law,using a linear relationship between absorbance and concentration to realize the measurement of dye liquor concentration.But most of the mixed dye liquor components present a problem of absorption spectrum overlap or interference in the actual production,so the commonly used stoichiometry cannot meet the detection accuracy.In order to achieve the requirements of production and development,we must rely on smarter chemometrics or improved algorithm to improve the detection accuracy of mixed dye concentration.Based on the study of the existing theory of machine learning,in this thesis a machine learning algorithm is presented for the classification of dye liquor optical spectrum.Machine learning can find out the rule which cannot be obtained by theoretical analysis from the observation data by constructing a classification model with low structural risk and high generalization ability to realize the prediction of data.The support vector machine(SVM)is a kind of machine learning method based on the structural risk minimization(SRM)principle to build the largest interval classification surface.The neural network has strong ability of classification by adjusting the perceptron weights and thresholds of the perceptron to obtain the regularity of the data.SVM and neural network algorithm have more mature results in multi-classification.Deep learning,as a derivative of neural network,has the function of learning deeper features.However,SVM and deep learning are not perfect in multi-spectral classification.Based on the research of mixed spectral classification,in this thesis two multi-component mixed dyes liquor optical spectrum classification algorithms based on SVM and deep learning were proposed.The main contents of this thesis include the following four parts:(1)Design and build a spectral acquisition system with halogen as a light source;opticalfiber having a core diameter of 400 ?m as a light transmission channel to reduce energy loss;SMA Z-flow cell as a sample pool,storing the circulating mixed dye liquor;using USB2000 +optical fiber spectrometer to collect mixed dye liquor's absorbance data;(2)Put forward a data processing method combining normalization,Savitzky-Golay(SG)convolutional smoothing and SPXY method.The normalization method can reduce the noise and drift of the absorbance data;SG convolution smoothing preprocessing filter impurities on the influence of the absorbance data;SPXY method divides the sample set into 75 training sets and 5test sets by considering not only the relation between the absorbance data,but also the relationship between the absorbance and the concentration;(3)Put forward a spectral classification model of three components weak acid dye liquor solution based on successive projections algorithm(SPA)algorithm and SVM.The SPA algorithm is used to extract the 22 optical spectral features of the absorbance data;The nu-SVR model is improved by adding L2 regular terms to the original objective function.Using the RBF kernel function,its optimal value of regular parameter C and kernel parameter ? are obtained by cross-validation.The experimental results show that this algorithm can ensure the accuracy of optical spectrum classification and shorten the classification time,and provide the basis for realizing the later online classification of multi-component dyes optical spectrum;(4)Put forward a spectral classification model of three components reactive dye liquor solution based on genetic algorithm(GA),deep learning and back propagation(BP)algorithm.The deep learning's best number of hidden layer chosen by GA algorithm is 2,and each layer number is 60.After using the optimal deep learning model to extract the depth features of the absorbance data,the BP algorithm with two hidden layers of 14 and 8 cells is used to achieve the classification.The experimental results show that the proposed algorithm can effectively improve the optical spectral classification accuracy of multi-component mixed dye liquor.