本文關(guān)鍵詞：基于機(jī)器學(xué)習(xí)的染液光譜分類算法研究　出處：《浙江理工大學(xué)》2017年碩士論文　論文類型：學(xué)位論文

  更多相關(guān)文章： 三組分 機(jī)器學(xué)習(xí) 支持向量機(jī) 連續(xù)投影算法 深度學(xué)習(xí)

【摘要】：我國作為全球印染行業(yè)的中心,傳統(tǒng)粗放的人工發(fā)展模式已經(jīng)不能滿足人們?nèi)找嬖鲩L的需求,同時(shí)印染行業(yè)產(chǎn)生的工業(yè)廢水也給環(huán)境帶來了巨大壓力,因此實(shí)現(xiàn)自動控制印染生產(chǎn)并實(shí)時(shí)監(jiān)測染液各個(gè)參數(shù)是印染發(fā)展的必然趨勢。染液濃度是參數(shù)監(jiān)測中最重要的一部分,目前國內(nèi)外最常用的濃度檢測方法是分光光度法,此方法以朗伯比爾定律為基礎(chǔ),利用吸光度與濃度之間的線性關(guān)系實(shí)現(xiàn)染液濃度的測量。但實(shí)際生產(chǎn)中大部分混合染液存在吸收光譜重疊或干擾等問題,現(xiàn)有常用的化學(xué)計(jì)量法并不能達(dá)到印染行業(yè)實(shí)際要求的檢測精度。為滿足企業(yè)生產(chǎn)和發(fā)展的要求,混合染液濃度檢測必須借助于更智能的化學(xué)計(jì)量法或其改進(jìn)算法來提高檢測精度。本文在學(xué)習(xí)現(xiàn)有機(jī)器學(xué)習(xí)理論的基礎(chǔ)上,提出了基于機(jī)器學(xué)習(xí)的染液光譜分類算法研究。機(jī)器學(xué)習(xí)能從觀測數(shù)據(jù)出發(fā)尋找出尚不能通過理論分析得到的規(guī)律,通過構(gòu)造具有低結(jié)構(gòu)風(fēng)險(xiǎn)和高泛化能力的分類模型實(shí)現(xiàn)對數(shù)據(jù)的預(yù)測。其中,支持向量機(jī)(SVM)是一種基于結(jié)構(gòu)風(fēng)險(xiǎn)最小化(SRM)原則構(gòu)建最大間隔分類面的機(jī)器學(xué)習(xí)方法,對于非線性回歸問題SVM可通過核函數(shù)轉(zhuǎn)化為線性可分問題;神經(jīng)網(wǎng)絡(luò)可通過調(diào)節(jié)感知器的權(quán)值和閾值得到數(shù)據(jù)的規(guī)律,具有很強(qiáng)的分類能力。SVM算法和神經(jīng)網(wǎng)絡(luò)在多分類領(lǐng)域都已有較成熟的成果,深度學(xué)習(xí)作為神經(jīng)網(wǎng)絡(luò)的衍生,具有學(xué)習(xí)更深層次特征的能力,但目前SVM和深度學(xué)習(xí)在混合光譜分類方面的研究還不夠完善。基于混合光譜分類的研究現(xiàn)狀,本文提出了基于支持向量機(jī)和深度學(xué)習(xí)的多組分混合染液光譜分類算法。本文的研究內(nèi)容主要包括以下四個(gè)部分:(1)設(shè)計(jì)搭建光譜采集系統(tǒng),鹵鎢燈作為光源;芯徑400μm的光纖作為光源傳輸通道,可減少能量損失;SMA Z-流通池作為樣品池,盛放循環(huán)的混合染液;采用USB2000+光纖光譜儀采集混合染液吸光度數(shù)據(jù);(2)提出一種將歸一化、Savitzky-Golay(SG)卷積平滑法和SPXY法進(jìn)行組合的數(shù)據(jù)處理方法,歸一化能減少噪聲和漂移對吸光度數(shù)據(jù)的影響;SG卷積平滑法可濾除雜質(zhì)對吸光度數(shù)據(jù)的影響;SPXY法通過綜合考慮吸光度數(shù)據(jù)之間和吸光度與濃度之間的關(guān)系將樣本集劃分為75個(gè)訓(xùn)練集和5個(gè)測試集;(3)提出一種基于連續(xù)投影算法(SPA)和支持向量機(jī)的三組分混合弱酸性染液光譜分類模型。利用SPA算法提取吸光度數(shù)據(jù)的22個(gè)光譜特征;通過在原目標(biāo)函數(shù)上增加L2正則項(xiàng)得到改進(jìn)的nu-SVR模型,nu-SVR分類模型采用RBF核函數(shù),其正則參數(shù)C和核參數(shù)σ通過交叉驗(yàn)證得到最優(yōu)值。實(shí)驗(yàn)證明,該算法在提高混合染液光譜分類精度的同時(shí),模型分類時(shí)間也得到了進(jìn)一步縮短,可為后期實(shí)現(xiàn)多組分染液光譜在線分類提供參考;(4)提出一種基于遺傳算法(GA)、深度學(xué)習(xí)和反向傳播(BP)算法的三組分混合活性染液光譜分類模型。利用GA算法得到深度網(wǎng)絡(luò)的最佳隱層數(shù)為2,每層單元數(shù)為60,通過最佳深度網(wǎng)絡(luò)模型提取吸光度數(shù)據(jù)的深度特征,BP算法兩個(gè)隱含層的單元數(shù)為14和8。實(shí)驗(yàn)表明,該算法能有效提高多組分混合染液的光譜分類精度。
[Abstract]:China as the world's printing and dyeing industry center, artificial development mode of the traditional extensive has been unable to meet the growing needs of people at the same time, the industrial wastewater of printing and dyeing industry has brought tremendous pressure to the environment, thus to realize the automatic control and real-time monitoring of the dye printing and dyeing production parameters is the inevitable trend of the development of printing and dyeing is a part of the dye concentration. The important parameters in the monitoring of the concentration detection method most commonly used at home and abroad are spectrophotometry, this method is based on Longbow Bill's law, the linear relationship between the absorbance and the concentration of the dye concentration measurement. But in the actual production of most existing mixed dye absorption spectra overlap or interference, the existing chemical measurement method the printing and dyeing industry and can not achieve the detection accuracy. In order to meet the actual requirements of the enterprise production and the requirements of the development of mixed dye concentration detection. Test must rely on the stoichiometry of more intelligent or its improved algorithm to improve the detection precision. Based on the study of existing machine learning theories, put forward the research of dye spectral classification algorithm based on machine learning. From the observation data of machine learning to find out is not obtained by law theory to forecast data the classification model has low risk structure and high generalization ability by constructing. Among them, the support vector machine (SVM) is a kind of based on structural risk minimization (SRM) principle to build the largest interval classification surface machine learning method for nonlinear regression problem by SVM kernel function into a linear separable problem; neural network can be obtained the rules by adjusting the perceptron weights and thresholds, with the ability of classification.SVM algorithm and neural network is very strong in many areas are already mature classification results Deep learning, the neural network is derived, with the ability to learn more profound features, but the current SVM and deep learning research in the classification of spectral mixture is still not perfect. The research status of mixed spectral classification based on support vector machine is proposed in this paper and deep learning of multicomponent mixed dye spectral classification algorithm based on research. The main contents of this paper include the following four parts: (1) design of spectrum acquisition system, tungsten halogen lamp as the light source; the fiber core diameter of 400 mu m as a light source transmission channel, can reduce energy loss; SMA Z- circulation pool as the sample pool, a mixed dyeing cycle; using USB2000+ fiber optic spectrometer to collect mixed absorbance data; (2) propose a normalized Savitzky-Golay (SG) method combined data smoothing method and SPXY method can reduce the noise and drift of normalized absorbance data The influence of impurities on the filter; can affect the absorbance data of SG convolution smoothing method; the relationship between the SPXY method by considering the absorbance data and the absorbance and the concentration of the sample set will be divided into 75 training sets and 5 test sets; (3) put forward an algorithm based on successive projection (SPA) and support vector machine the three component mixed weak acid dye spectral classification model. The extraction of 22 spectral absorbance data using the SPA algorithm; by increasing the L2 regularization in the original objective function to get the improved nu-SVR model, the nu-SVR classification model using RBF kernel function and the regularization parameter C and kernel parameter to get the optimal value through cross validation experiments., this algorithm can improve the classification accuracy of mixed dye spectra at the same time, the classification model of time has been further shortened, multicomponent dye spectrum online classification and provide reference for later implementation; (4) a proposed In the genetic algorithm (GA), deep learning and back propagation (BP) algorithm three component mixed spectral classification model. The best reactive dye hidden layer uses GA algorithm to get the depth of the network is 2, the number of units of each layer is 60, the best depth network model to extract depth characteristic absorbance data, BP algorithm two the hidden layer units showed that the 14 and 8. experiments, the algorithm can effectively improve the multicomponent dye spectral classification accuracy.

【學(xué)位授予單位】：浙江理工大學(xué)
【學(xué)位級別】：碩士
【學(xué)位授予年份】：2017
【分類號】：TS190;TP181

【相似文獻(xiàn)】

相關(guān)期刊論文 前2條

1 孟憲江,張鐵強(qiáng),白英奎,申鉉國;利用BP神經(jīng)網(wǎng)絡(luò)光譜分類法研究肉品新鮮度[J];光譜實(shí)驗(yàn)室;2004年05期

2 ;[J];;年期

相關(guān)博士學(xué)位論文 前1條

1 徐洋;稀疏與低秩先驗(yàn)下的高光譜分類與檢測方法[D];南京理工大學(xué);2016年

相關(guān)碩士學(xué)位論文 前3條

1 張銀露;基于機(jī)器學(xué)習(xí)的染液光譜分類算法研究[D];浙江理工大學(xué);2017年

2 葉虎平;黃東海海洋光譜分類方法研究[D];國家海洋技術(shù)中心;2009年

3 羅磊;基于模糊模式識別的光譜分類研究[D];鄭州大學(xué);2006年

，

本文編號：1378896