A Study On Image Processing And State Recognition Of Rotary Kiln Based On Slow Feature Analysis

The calcination of cement clinker is one of the most important procedures in the cement production process,and it directly affects the yield and quality of cement.The core thermal technology equipment in the cement clinker sintering process is called rotary kiln of which the main function is to sinter the raw material slurry into the qualified clinker,and its operation situation directly decides the quality of cement clinker production.But due to the rotation of the rotary kiln,fuel combustion,and heat transfer between gas and material,the key technical parameters of the rotary kiln sintering process fail to be detected accurately,which causes that the rotary kiln sintering process still relies on "operators observing fire" i.e.operators observe the material sintering status to recognize the current burning state.The burning state recognition result of the human operation pattern is restricted by some personal elements,which will cause that the quality of cement clinker production can’t meet the requirements.Flame image of rotary kiln contains rich clinker sintering status information.Hence,it is necessary to utilize the flame image to develop new methods for the burning state recognition.However,owing to the influence of coal powder,smoke,and dust inside the kiln,the flame image has blur boundary.The performance of the previous image technique based on static image scale invariant feature transform(SIFT)is weak and its identification precision can’t meet the requirements.Hence,considering the previous research of the cognitive model of visual cortex cells,it is necessary to utilize the flame video to develop new methods for the burning state recognition based on image processing and machine learning theory.Such research can lay the foundation for the rotary kiln sintering process monitoring and closed-loop control of the clinker quality index.For above problem of rotary kiln clinker burning state recognition,the dissertation,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 thesis are summarized as follows:(1)Aiming at the drawbacks of rotary kiln clinker sintering condition recognition which based on static images,after analyzing some reviews about image processing,we apply visual cortex cognitive theory and deep learning methods to rotary kiln clinker sintering condition recognition,proposing a new method based on video images in the sintering zone to solve our problem.This method is composed by video image pretreatment and the spatial-temporal features extraction based on hierarchical convolution slow feature analysis(HC-SFA)and bag of words model,as well as the random vector function link(RVFL)network classifier method;(2)Video image preprocessing methods mainly including spatial-temporal blocks extraction and whitening dimensionality reduction.Based on the stratification in visual cortex information processing system and the mechanism of neuron local receptive field gradually expand,our thesis randomly extract spatial-temporal block with increasing size step by step at different sintering conditions video areas,benefiting to hierarchical feature representation.The preprocessing method based on principal component analysis is used to reduce the correlation of image blocks and redundancy;(3)Feature extraction methods is composed by HC-SFA and bag of words model.In order to obtain effective and invariant feature representation,reducing the learning complexity and improve the learning efficiency of the model,the HC-SFA unsupervised learning method is used to build a slow component analysis model by hierarchical learning on the increasing size spatial-temporal blocks,adopt slowly varying elements to establish local slowly varying morphological characteristic;The local slowly varying spatial-temporal features arise from spatial-temporal blocks slowly varying composition mean square derivative of time(TMSD).In the end,we building the bag of words model of video images to get the global spatial-temporal morphological features.Moreover,the feature dimension is further reduced and benefit to the learning of classifier.(4)Considering the advantages of RVFL network,such as high learning speed and good generalization,we adopt RVFL network as the working condition recognition classifier,and we use grid search and bootstrap method to identify the number of RVFL network hidden layer node.In order to verify the feasibility,this thesis has done some experiments towards the proposed method.The experimental result indicates that the proposed preprocessing method effectively reduces the correlation of image block and redundancy;the feature extraction method based on hierarchical convolutional slow feature analysis and bag of word model can extract distinctive and robust morphological feature representation from rotary kiln video.The average classification accuracy based on RVFL network is 93.38%.In contrast,the average classification accuracy based on SIFT is 92.12%.The classification result indicate the proposed method have a better recognition and prediction performance towards rotary kiln sintering condition.