Research On Robust Recognition Method For Sintering State Of Rotary Kiln Based On The Coal-Fired Flame Images

Rotary kilns are widely used in nonferrous metallurgy, cement and other industry fields. The sintering process is a complex physical and chemical reaction of coal powder, air and other materials through mixture. It is a complicated combustion process and hard to model and control. The sintering temperature of clinker in a rotary kiln is essential in the production control process, and can directly influence the quality of products, energy consumption and pollutant emission levels. Therefore, the measurement and control of the sintering temparature in ratary kilns is one key element in the optimization of a sintering process, and is important for energy saving and emission reduction of the traditional coal industry.Applying image processing in the working condition analysis of sintering is a fast and effective way. Copmared to other meothds such as thermocouple, high temperature colorimetric pyrometer, and the use of thermaldata, it is easy for implementations, and gradually attracted widespread attentions in recent years. Previous research works have been mainly focused on clear flame images for feature extraction. In practice, the quality of flame images are seriously affected by the harsh field environment, and the acquied images are often blurry. This thesis addresses on the problem of rotary kiln working condition recognition using blurry and noisy images collected from industrial fields. A series of methods for image enhancement, illumination compensation and feture extraction are proposed. Also, a class of new robust learning machines are proposed for the robust recogntion of sintering conditions.The main contributions of this thesis are described as follows:Firstly, a segemention method for the jet regions of mixed pulverized coal and air ("black handle" or "pulverized coal region") in blurry flame images of a rotary kiln is proposed, and the The morphological features of region is extracted to recognize the working condition of rotary kilns.Blurry flame images of a rotary kiln have the defects of illumination variations and dust disturbances. To address this problem, an adaptive illumination compensation and fuzzy enhancement method is developed. Pulverized coal regions are segmented by the fast fuzzy C-means clustering algorithm. The morphological features such as rectangular degree and area proportion are extracted from the segmented regions and used for the working condition recognition of rotary kilns.Secondly, a set of dynamic characteristics on an image series are constructed to recognize special working conditions. In such special conditons, recognition errors can occur using sigle pulverized coal frame based statistical analysis. Instead, we compute the short-term energy and sample entropy features of illumination statistical features in a series of images, and use them as the working condition criteria of rotary kilns.Thirdly, a computer vision-based measurement approach is developed for calculating the positions, heights, and swing angle of the centroids of material regions. These dynamic morphological characters of material regions can be used to detect the sintering state of clinkers. Pre-processings such as illumination compesation and shadow correction are used to enhance the image edges of matetial regions. Then the material regions are segmented by the region-growing method. The characteristics of each image coordinate, height and swing angle of material region centroids are extracted to form the features in time series. Last, the short-term energy and sample entropy of feature series are computed as the criteria of clinker sintering states.Fourthly, a robust extreme learning machines based on Parzen windows is applied. There are outliers in feature data extracted from blurry image sequences in ratary kilns. To alleviate these outliners interference, the Extreme learning machine (ELM) is combined with robust estimation theory to form a new Parzen-ELM, which is contructed for robust output weight estimation using the weight matrix formed by Parzen window. Parzen-ELM can estimate working conditions more accurately, quickly and robustly.Experiment results from real-time video captured from industrial rotary kilns validate the robustness and applicability of the proposed methods in this thesis. An implementation of robust working condition recognition in sintering area of kilns is provided based on coal-fired image sequences. The evaluation results in a domestic aluminum plant demonstrate that the methods developed in this thesis can recognize the sintering working condition robustly and quickly, and suggets that the thesis work has important scientific significance and broad application prospects.