燃煤烧结工况的鲁棒视觉检测方法研究与应用

Coal-fired kilns are widely used in electricity, metallurgy, cement and other industrial production fields. Identifying the sintering condition by analyzing coal burning flame is the key process to realize the optimized control, energy saving and emission reduction. Using the industrial alumina rotary kiln as application object, focusing on problem that the flame images are susceptible to interference and light environment dust, this project analyze the robust identification of coal sintering condition by fuzzy flame image sequence. First of all, an image enhance preprocessing method for coal-fired image based on dark color prior and nonlinear illumination compensation is presented; Secondly, when the details of image are lacking, the edge of the coal region, material region and flame region are segmented by the theory of phase congruency combined with time domain processing method; Then, based on the flame image sequence, the steady visual feature extraction for working condition detection are presented, to eliminate the false features under complex conditions caused by the complex working condition; At last, a set of HMM recognition model for condition detection are established based on the dynamic features. The conditions of every region are parallel recognized by the HMM and robust fused by the minimum risk Bayes theory. A set of robust system for sintering conditions detection based on fuzzy flame image sequence is achieved. This method is of great significance for improving the robustness and real-time application of detection and control of coal combustion.

燃煤窑炉广泛应用于发电、冶金、水泥等工业生产领域，通过分析窑内燃煤火焰来识别烧结工况是实现此类复杂过程优化控制和节能减排的基础和关键。课题以工业氧化铝回转窑为应用对象，针对现场火焰图像模糊且易受环境粉尘和光照影响的特点，研究燃煤烧结工况的鲁棒视觉检测方法。首先，研究一种基于暗原色先验和光照补偿的火焰图像快速去雾增强方法，降低现场烟雾和光照对后续特征提取的影响；其次，借鉴视觉感知机理，利用信号奇异点处谐波相位的一致性，解决煤粉区、物料区和火焰区的稳健分割问题，并构造各区域动态工况特征序列，利用多帧信息互补降低图像特征的不稳定性；最后研究一种结合最小风险贝叶斯的多层HMM工况融合识别模型，利用并行HMM建立各区域的识别模型，并对其输出进行最小风险融合，最终实现一套燃煤烧结工况稳健视觉检测方法并现场应用。课题研究对提高燃煤过程检测及其控制的鲁棒性具有重要意义，也可为其他工业软测量提供借鉴。

课题以工业氧化铝回转窑为应用对象，针对现场火焰图像模糊且易受环境粉尘影响的特点，研究燃煤烧结工况的鲁棒视觉检测方法。在大粉尘条件下采集的模糊熟料烧结图像的去模糊增强、鲁棒图像特征提取、鲁棒工况分类器设计等关键问题上取得了一些理论研究成果，并在工业现场搭建了耐高温高速视频图像采集平台，为后续的算法实施奠定了基础。首先，结合最小颜色通道与传播滤波，提出了一种新的图像去雾算法，该算法可应用于工业现场粉尘干扰严重的燃煤图像中，解决粉尘干扰严重的工业图像增强问题；其次，基于稀疏表示的图像去噪，从字典的相干性边界条件出发，提出了一种新的非相干字典学习算法，可有效去除工业现场粉尘和电磁干扰导致的模糊图像噪声；第三，为了针对模糊图像各区域提取的稳健特征，对其进行当前窑内工况分类识别，课题组提出一种鲁棒的特征分类器：二重分布学习机，实验结果表明该分类器具有较好的泛化性能和鲁棒性；第四，课题组在理论研究中发现窑前火焰图像序列具有一定的混沌特征，针对超混沌系统展开了一系列相关的理论研究；最后，在实践方面，设计了一套耐高温高速回转窑烧结带图像采集系统，并将其融入现有回转窑专家控制系统中，在大唐内蒙古再生资源开发有限公司氧化铝厂烧成车间实地运行，为算法的后续实施奠定了良好的基础。