【作者】 张欣；

【导师】 韩璞；

【作者基本信息】 华北电力大学（河北） ， 热能工程， 2009， 博士

【摘要】 由于现代热力发电设备向着大容量和高参数的方向发展,机组设备越来越复杂,对生产过程的控制品质要求越来越高。火电厂煤粉锅炉燃烧状态的监测与评判,对于保障生产的安全性、经济性和环保性是非常有意义的。将人工智能算法引入锅炉燃烧状态监测与评判,可以提高监测与判别的准确性、实时性,能为机组运行的自动控制提供可靠依据。目前火电厂煤粉锅炉燃烧状态监测以火焰图像为主要对象,从火焰图像处理、火焰图像状态评判和火焰燃烧过程状态建模的研究状况来看,人工智能的研究与应用具有积极的意义,为锅炉火焰燃烧状态监测和评判开辟了智能化的新思路。本文在研究一些热门人工智能算法和分析煤粉锅炉的工作原理及影响锅炉燃烧状态的因素的基础上,开展了以下三个方面的研究工作:①应用智能算法实现火焰图像的预处理和特征提取;②利用火焰图像和智能算法评判锅炉燃烧状态;③利用智能算法建立基于火焰图像序列的锅炉燃烧状态识别模型。本文提出了自适应投票法快速中值滤波(AVMF)算法用于图像滤波。本文详细介绍了煤粉锅炉火焰燃烧状态监测系统的构成和火焰图像采集的原理。分析了火焰图像的噪声产生机理及常用噪声去除方法。AVMF算法结合了自适应判断噪声点技术和投票法滤波算法,与现有其它中值滤波改进方法比较,既具有极高的处理速度,又能较好地保留图像细节。本文提出将遗传算法和混合高斯马尔可夫随机模型用于火焰图像的分割。由于煤粉火焰图像的大噪声和边缘模糊的特点,采用传统的基于灰度梯度的图像分割方法不能取得满意的效果。将遗传算法应用于火焰图像区域的分割和合并。实验结果表明该方法较之传统方法具有更理想的分割效果。基于图像数据的局部相关性,马尔可夫随机过程理论用一个二维随机场模型来描述图像。它用条件概率描述图像的数据分布,用高斯分布特性描述单个像素及其邻域关系。实验结果表明利用混合高斯马尔可夫随机模型能够分割噪声图像。本文提出将隐马尔可夫随机模型(HMM)用于基于火焰图像的燃烧状态识别。在研究了煤粉燃烧机理的基础上,提出了火焰图像中反映燃烧特性的特征参数及其计算方法。根据火焰燃烧的随机性特点,率先建立了基于火焰图像的锅炉燃烧状态识别HMM。实验结果表明该方法能够获得较满意的识别效果。本文提出了将交互式学习神经网络用于基于火焰图像的燃烧状态识别。利用球面邻域理论构建神经网络,使用交互式学习技术降低网络构建的复杂度,提高模型识别精度。实验结果表明该技术的能够获得更好的识别效果。本文提出了将光滑支持向量回归(SSVR)算法应用于火焰图像序列预测。仿真结果表明该方法具有一定的预测效果。根据火焰图像具有异常噪声的特点,本文率先提出了变ε光滑支持向量回归(AεSSVR)算法,并用于火焰图像序列分析。仿真结果表明AεSSVR能有效避免异常噪声带来的虚报警。本文提出了将隐马尔可夫模型应用于基于火焰图像序列的燃烧状态建模。利用隐马尔可夫模型对于随机过程的强大模式分类能力及其与传统方法相比具备的独特的自适应特性,本文率先建立了基于火焰图像序列的锅炉燃烧状态的一维隐马尔可夫模型和伪二维隐马尔可夫模型。通过与神经网络的对比,仿真结果表明伪二维隐马尔可夫模型具有更高的识别精度。本文的主要理论创新点:(1)提出了自适应投票法快速中值滤波算法并用于火焰图像滤波;(2)提出了变ε光滑支持向量回归算法并用于火焰图像序列分析。本文的主要应用创新点:(1)将遗传算法和混合高斯马尔可夫随机模型用于火焰图像的分割;(2)将交互式学习神经网络用于基于火焰图像的燃烧状态识别;(3)将隐马尔可夫模型用于基于火焰图像的燃烧状态识别;(4)将光滑支持向量回归算法用于火焰图像序列预测;(5)将隐马尔可夫模型用于基于火焰图像序列的燃烧状态建模。更多还原

【Abstract】 As the modern thermal energy generating set develops toward to large capacity and high parameter, the equipment is getting more and more complex, and it requires higher controlling quality of production process. It is very meaningful to monitor and judge the combustion status of power plant pulverized-coal boiler for security, economic and environment protection, the introduction of artificial intelligence algorithm in boiler combustion monitoring and judgment improves the accuracy, timeliness of the monitoring and judgment, and also provides the reliable basis for automatic controlling of the boiler unit. At present, the combustion monitoring of power plant pulverized-coal boiler takes the flame image as the main object . Considering the study on flame image processing, the judgment of the flame image status and the modeling of flame burning process status, the research and application of artificial intelligence have positive significance and open up an intelligent new road for monitoring and judging of the boiler flame combustion status. On the foundation of the study on some popular artificial intelligence algorithm and the analysis of the working principle of pulverized-coal boiler and the factors that impact on flame combustion status , three aspects of the research following have been done: (1)Flame image pretreatment and its features extraction are achieved using intelligence algorithm;(2)Judging the boilercombustion state with flame image and intelligence algorithm;(3)Recognition model of boiler combustion state based on the flame image-serial is build by using intelligence theory.In this dissertation, Auto-adapted voting fast median filtering algorithm (AVMF) is proposed for image filtering. The component of the pulverized-coal boiler flame combustion monitoring system and the principle of flame image acquisition are introduced in details. The mechanism of flame image noise production and the methods commonly used to eliminate noise are analyzed. AVMF algorithm combines the determining noise technology of auto-adapted with the votting filtering algorithm. Comparing with other improved median filtering methods, AVMF not only has the extremely high processing speed, but also better able to maintain the image detail.In this dissertation, it is proposed that genetic algorithm and compound Gauss-Markov random model are used in flame image segmentation.Because pulverized-coal flame image has the features of big noise and fuzzy edge, therefore, the traditional gray gradient method of image segmentation can not obtain satisfactory result. Genetic algorithm is applied to dividing and uniting the region of flame image. Experiment shows this method makes a better performance in image segmentation than traditional methods.Based on the local relevance of image data, according to theory of Markov random process,the image is described by a model of two dimension random field, the data distribution of the image is described by conditional probability, A single pixel and its neighborhood relationship are described by Gaussian distribution. Experiment indicates this method achieves satisfactory image segmentation on the condition of including noises.In this dissertation, it is proposed that hidden Markov random model(HMM) is used in recognition of combustion state based on flame image. After studying the mechanism of pulverized-coal boiler combustion, characteristic parameters that reflect combustion and its calculating method are introduced. According to the random of combustion, HMM of recognition combustion state based on flame image is established first.Experiment result shows that this method can achieve satisfactory recognition effect. In this dissertation, it is proposed that interactive learning neural network technology is used in recognition of combustion state based on flame image. The construction of neural network is based on the sphere neighborhood principle, the complexity of neural network modeling is declined and the accuracy of the mode recognition is improved by using interactive learning technology. Experiment result indicates that this method can achieve a better recognition effect.In this dissertation, it is proposed that smooth support vector regression (SSVR)algorithm is used in the prediction of the flame image-serial. The simulation result shows that this method has a certain predictive effect. According to the feature of abnormal noise in flame image, alterableεsmooth support vector regression (AεSSVR)algorithm is introduced novelly and it is applied to flame image-serial analysis. The simulation result indicates that this algorithm can avoid the false alarm which is brought about by abnormal noise effectively.In this dissertation, it is proposed that HMM is used in modeling of combustion state based on flame image-serial. HMM has a strong ability of classification to random process and unique adaptability comparing with traditional modes. Making use of the features above, one dimension HMM and pseudo two dimension HMM of boiler combustion state based on flame image-serial are firstly established in this dissertation. Comparing to the neural network, the simulation result shows that using pseudo two dimension HMM can obtain higher accuracy of recognition.The main theory innovations of this dissertation are as fellows: (1) Auto-adapted voting fast median filtering algorithm is proposed and it is applied to flame image filtering.(2) Alterableεsmooth support vector regression algorithm is proposed and it is applied to flame image-serial analysis.The main application innovations of this dissertation are as fellows:(1) Genetic algorithm and compound Gauss-Markov random model is used in flame image segmentation.(2) Interactive learning neural network is used in combustion state recognition based on flame image.(3) Hidden Markov Model is used in combustion state recognition based on flame image.(4) Smooth support vector regression algorithm is used in the prediction of the flame image-serial.(5) Hidden Markov model is used in modeling of combustion state based on flame image-serial.更多还原