A Study On Burning State Recognition Of Clinker Based On Burning Zone Image

Rotary kiln is a large, core thermal equipment for the production of alumina, cement and other industrial raw materials. Because of being restricted by the complex rotary kiln kiln body and the fuel combustion, material gas convection heat transfer factor, the existing measuring methods are difficult to realize the key process parameters of rotary kiln on-line detection with temperature and clinker sintering product quality, making it difficult to achieve automatic control of rotary kiln process.Rotary kiln sintering process rely on industrial television " operators observing fire " mode in long-term, i.e. operators observe the burning zone image, combined with process dataset and material burning state, and then regulate the manipulated variables to ensure the quality of the products. However, this operation pattern is restricted by some personal elements, easily leading to low quality of product, low running efficiency of kiln, low productivity, high energy consumption, etc.Burning zone image containing rich burning temperature field and clinker sintering status information, this provides a good basis for research and development of rotary kiln burning state-based image recognition. However, due to the rotation of the kiln and the interference of kiln dust, the noise of burning state image is large, strong coupling is between the significant area and has blur boundary. The existing image segmentation technique based and scale invariant feature transform method based difficult to extract effective feature, leading to the low recognition accuracy of burning state. Hence, it is an effective way to solve the problem of detection and control and optimization in rotary kiln, and has important theoretical significance and application value to study on burning state recognition of rotary kiln clinker based on burning zone image in using the latest research achievements in image processing, machine learning and other fields.For above problem of rotary kiln clinker burning state recognition, the dissertation is aimed at improving the accuracy of burning state recognition and supported by the national natural science fund project "product quality parameter prediction modeling based on the fusion of image and process data about rotary kiln", developed the burning state recognition method with the goal of raising the accuracy of the burning state recognition. The major contributions of this paper are summarized as follows:1. On the basis of review domestic and foreign existing image processing method and technology in process industrial and the analysis of the research difficulties in rotary kiln clinker burning state recognition, we applied the visual cortex cognitive theory and computational methods to the problem of rotary kiln clinker burning state recognition for the first time, and developed a novel technique about rotary kiln clinker burning state recognition, which has two new characteristics, the new image feature extraction methods based on deeping learning and independent subspace analysis, the unique design of integrated classifier based on extreme learning machine and fuzzy integral. The technique mainly consists of image preprocessing methods, feature extraction methods and classifier design methods.2. Image preprocessing includes significant area dividing, image patch extraction, image denoising and dimension reduction. According to the characteristic of image acquisition device fixed position and the significant areas of the burning zone image are invariant, and reference to the experience of field site operator when discriminate on kiln clinker burning state, we directly divide the burning zone image into three significant areas:black handle area (coal powder area), flame area and material area. We take the method of random sampling increasing size from three different signifincant areas in order to study feature expression, and reduce the correlation and redundancy between the image patches based on principal component analysis on the basis of visual cortical information processing and local receptive fields of neurons.3. A novel feature extraction method was proposed by combing deep learning, independent subspace analysis and bag of words model. Dataset of different significant areas local feature expression was constructed by image patches extracted random from different significant areas. Through deep learning and independent subspace analysis, local feature representation model based on increasing size image patches dataset was layer wise learned which possesses selectivity, invariance and low computational complexity. Significant image feature was extracted from different significant areas by local feature representation models through sliding-window method, and then composes feature set. Bag of words model was established to further reduce the feature dimensionality and learn effective feature representation.4. An integrated classifier design method based on extreme learning machine and fuzzy integral was proposed in this paper. The different significant area feature extraction model was applied to the all significant areas, all the significant areas image features are obtained respectively, constructing the classifier learning sample set of significant area, to avoid the possible "curse of dimensionality" phenomenon by the direct feature level fusion.Each sub-classifier of different significant areas was designed with extreme learning machine, while fusion mechanism of classifier decision level was designed with fuzzy integral.5. Experimental study which based on rotary kiln clinker burning image data was carried out to investigate above methods. The results show that preprocessing methods used in this paper can effectively reduce the correlation and redundancy between the image patches. Feature extraction methods based on deep learning, independent subspace analysis and bag of words model can extract effective features of different significant areas. The accuracies of burning state recognition sub-classifiers based on the three significant areas are 93.58%,92.35%,92.39% respectively, while the recognition accuracy of integrated recognition classifier is 94.57%.