【作者】 葛红；

【导师】 闫勇； 魏振华；

【作者基本信息】 华北电力大学(北京) ， 控制科学与工程， 2020， 博士

【摘要】 近年来,随着生物质发电持续稳定发展,生物质锅炉燃烧状态监测与控制问题日益受到关注。我国可用生物质燃料品种多样,燃料偏离锅炉设计种类容易造成结焦和严重积灰等问题,直接影响生产安全和经济效益。本文基于火焰光谱分析算法和图像处理技术,研究生物质燃料种类识别及燃烧状态的监测方法。在单生物质燃烧特性的基础上,重点研究混合生物质的燃料识别和燃烧稳定性分析问题,为生物质锅炉适应燃料的多样性和机组的稳定经济运行提供依据。主要内容如下:1)针对生物质混燃问题,构造了基于火焰光谱6维特征与集成学习的生物质燃料识别方法。通过花生壳、柳木、麦秸、玉米芯四种单生物质及其等质量均匀混合的六种混合生物质进行燃烧实验;同时基于生物质燃烧火焰光谱的时域特性和频域特性的深入分析,构造了基于火焰光谱6维特征(由OH*、CN*、CH*、C2*-1和C2*-2等5类自由基辐射强度和火焰辐射强度信号组成)与集成学习的生物质燃料识别方法,并将识别结果与SVM和决策树两种常用算法进行了对比,结果表明,基于6维火焰光谱强度信号与集成学习的生物质燃料识别方法具有较高的识别率,单生物质高于99.32%,混合生物质均达到100%。该方法简化了特征提取过程,集成了支持向量机(Support Vector Machines,SVM)与决策树两种单分类器的优势,具有较好的泛化能力。2)构造了基于火焰图像与卷积神经网络的燃料识别方法。为克服传统基于图像的燃料识别方法对特征提取的依赖性,在分析生物质燃烧火焰图像的几何特征、光学特征和热力学特征的基础上,构造了基于火焰图像与深度卷积神经网络ResNet50的燃料识别方法,并分别对单生物质和混合生物质进行了燃料识别实验,结果表明基于火焰图像与ResNet50的燃料识别准确率高达97.41%,而其它机器学习算法最高仅达85.83%,可见该方法识别效果优于四种传统机器学习方法。3)定义了基于火焰图像和数据融合的燃烧稳定性指数。在提取火焰图像高度、平均灰度、平均温度和最高温度等特征参数的基础上,定义了基于火焰图像和数据融合的稳定性指数,并用来定量表征生物质火焰燃烧稳定性状态,并结合火焰闪烁频率对单生物质和混合生物质燃料稳定性进行了分析,分析结果表明所提方法能够有效表征生物质火焰燃烧稳定性状态。4)构建了一种生物质燃烧监测及稳定性判别的优化方案。基于混淆矩阵,对光谱分析与图像处理在生物质燃料识别与燃烧稳定性分析效果进行了对比分析,提出了以火焰光谱6维特征与集成学习进行生物质燃料识别、基于火焰图像和稳定性指数进行生物质燃烧稳定性分析的生物质燃烧监测优化方案。借鉴多元统计指标T2和SPE实现了不同生物质燃烧稳定与不稳定的分类,并根据分类结果确定了稳定性指数阈值,最终实现了燃料和稳定性的实时判定。更多还原

【Abstract】 In recent years,with the sustained and stable development of biomass power generation,the monitoring and control of combustion state of biomass boilers has attracted increasing attention.There are many kinds of biomass fuels available in China,and the deviation of fuel from boiler design is easy to cause coking and serious ash deposition,which directly affects production safety and economic benefits.In this thesis,research on the biomass combustion monitoring using flame spectrum analysis and image processing is conducted based on the in-depth analysis of the state-of-the-art in biomass combustion processes.The study focuses on the fuel identification and combustion stability analysis of blended biomass based on the combustion characteristics of a single type of biomass.The outcome of the results provides the basis for maintaining the continuous supply of various types of fuel and the stable economic operation of combustion in biomass power plants.The main achievement and contributions of the research are as follows:1)Aiming at the problem of biomass co-combustion,a biomass fuel identification method based on 6-dimensional characteristics of flame spectrum and ensemble learning is constructed.The method composed of the radiation intensities of five types of free radicals,such as OH*,CN*,CH*,C2*-1 and C2*-2,and the intensity of flame radiation signals,are selected based on the analysis of the time-domain and frequency-domain characteristics of the flame spectrum of biomass combustion.Combustion tests using our types of biomass(peanut shell,willow,wheat straw,corncob)and six different biomass blends with equal mass were conducted.Comparison analysis between the proposed 6-D method,SVM and decision trees were performed.Test results demonstrate that the proposed method with ensemble learning has a remarkable recognition rate of 99.32%,and 100%for pure biomasses and blended biomass,respectively.The proposed method not only simplifies the feature extraction process,but also integrates the advantages of support vector machines(SVM)and decision trees,therefore,has better generalization ability.2)A fuel identification method based on flame image and convolution neural network is constructed.To overcome the drawback of strong dependence on image feature extraction,which is crucial to traditional image-based fuel recognition methods,a biomass fuel recognition method using flame images and depth convolution neural network ResNet50 is proposed based on the analysis of the geometric,optical and thermodynamic characteristics of flame images of biomass combustion.The fuel recognition tests of pure biomass and blended biomass were carried out and the recognition results were compared with four traditional machine learning methods.The comparison results show that the accuracy of fuel recognition based on the proposed method excelled the traditional methods by 97.41%versus 85.83%at best.3)The combustion stability index based on flame image and data fusion is defined。A flame stability index is proposed to quantitatively characterize the combustion stability of biomass flame by fusing the height,average gray level,average temperature,maximum temperature and other characteristic parameters of biomass flames.Based on the stability index and flame flicker frequency,the stability of pure biomass and blended biomass fuel is analyzed.The analysis demonstrated that the proposed method effectively characterizes the combustion stability of a biomass flame.The stability index has the advantages of easy calculation,independent of external conditions such as burner structure,is suitable to use for online monitoring,and is guaranteed ranged within[0,1].4)An optimization scheme of biomass combustion monitoring and stability discrimination is constructed.An optimization scheme of biomass combustion monitoring is proposed based on biomass fuel identification and flame stability obtained by the 6-D method and the flame stability index.Comparative study on the effects of spectral analysis and image processing on biomass fuel identification and combustion stability was conducted based on the confusion matrix method.The classification of combustion stability and instability of different biomass is realized based on the multivariate statistical indexes T2 and SPE.Then,the real-time identification of fuel type and determination of combustion stability are realized based on the obtained threshold of stability index(according to the classification results).更多还原