| The calcination of cement clinker is one of the most important procedures in the cement production process. Cement clinker directly determines the yield and quality of the cement. Rotary kiln is the thermal equipment used in the cement clinker sintering process and its main function is to sinter the raw material slurry into the qualified clinker. The behavior of the rotary kiln directly relates to the cement clinker quality.The construction and mechanism of the rotary kiln is complex, due to the length, size, and continuous rotation of the rotary kiln. The rotary kiln sintering process includes physical and chemical reaction of material, fuel combustion, heat transfer between gas, material, and lining, strong coupling among gas and material movement. Therefore, the clinker quality index (free calcium oxide, f-CaO) of the rotary kiln sintering process fails to be continuously detected online, and the burning state closely related to the clinker quality index is difficult to be recognized accurately. The rotary kiln sintering process still relies on "operators observing fire", i.e. operators observe the material sintering status, combined with process dataset to recognize the current burning state, and then regulate the manipulated variables to control the controlled variables that are closely related to the f-CaO content into the preset ranges, to realize sufficient burning of the raw material and obtain satisfied clinker. The burning state recognition result of the human operation pattern is restricted by some personal elements such as operator’s experience, responsibility, and attention etc. Hence, it will easily lead to low qualified rate of clinker, short lifetime of kiln lining, low running efficiency of kiln, low productivity, high energy consumption, and high operator’s workload, etc.Flame image containing rich burning zone temperature field and clinker sintering status information, combined with process dataset information, is the main foundation for operators recognizing the burning state. Appropriate burning state means satisfied clinker. Therefore, the clinker quality index is closely related to the burning state. However, owing to the influence of coal powder, smoke, and dust inside the kiln, strong coupling is between region of interests (ROIs) of the flame image, and ROIs has blur boundary. Additionally, the process dataset includes a lot of complex noise. The performance of the previous image segmentation technique based or process dataset based to recognize the burning state and construct the soft-sensing model of f-CaO content methods is weak. Hence, it is necessary to utilize the flame image and process dataset to develop new methods for the burning state recognition and soft-sensing of f-CaO content based on image processing and machine learning theory. Such^research can apply machine learning technique to the engineering practice to turely realize the "machine observing fire" replacing the "operators observing fire", and lay the foundation for the rotary kiln sintering process monitoring and closed-loop control of the clinker quality index.The dissertation has made detailed research on the processing and analysis of the flame image and process dataset of the cement rotary kiln, and developed the burning state recognition method and the f-CaO content soft-sensing method, with the goal of raising the accuracy of the burning state recognition and the clinker quality index soft-sensing. The major contributions of this study are summarized as follows:①A new method based on compact Gabor filter bank designing is proposed to filter flame image ROIs with discriminative texture characteristics to enhance the separability of ROIs to facilitate the burning state recognition, aiming at the problem of strong coupling between ROIs and blur ROIs boundary by the disturbance of smoke and dust. A new method without image segmentation technique is proposed to extract and select flame image features, aiming at the drawback of previous ROIs acquisition and feature extraction based on image segmentation technique. Firstly, the color feature of ROIs is extracted by the multivariate image analysis (MIA) technique, then the global configuration feature is extracted using improved principal component analysis (PCA), finally the local configuration feature is extracted based on scale invariant feature transform (SIFT) operator combined with "bag of visual words"(BoVW) and improved latent semantic analysis (LSA). Four kinds of pattern classifiers, i.e. probabilistic neural network (PNN), back propagation neural networks (BPNN), support vector machines (SVM), and extreme learning machines (ELM), are proposed to recognize the burning state for the extracted three flame image features, aiming at different pattern classifiers with different feature space partitioning power. A new method based on fuzzy integral is proposed to achieve the decision level fusion for burning state recognition results of three flame image features, to avoid the possible "curse of dimensionality" phenomenon by the direct feature level fusion. The experiment results show that our compact Gabor filter bank designing method not only enhances the separability of ROIs much more but also saves computational cost. For the designed pattern classifiers based on three individual flame image features, the best burning state recognition results are85.55%,88.57%, and92.12%. Compared with individual flame image feature based, image segmentation technique based, and other fusion methods,the best recognition result of the decision level fusion based on three flame image features is95.37%with PNN classifier.②Aiming at the process dataset as the important complement to the flame image during the burning state recognition, firstly, a new method based on the improved median number absolute deviation filter is proposed to filter process dataset to remove the noise disturbance. Secondly, a new method based on KPLS is proposed to extract the best compact process dataset feature vector subsets to avoid the possible collinearity and nonlinear characteristics existing in the process dataset. Thirdly, the extracted process dataset feature vector is fed into PNN, BPNN, SVM, and ELM pattern classifiers to recognize the burning state. Fourthly, a method based on fuzzy integral is proposed to achieve the decision level fusion for the flame image and process dataset burning state recognition results. The experiment results show that compared with traditional median number filter, our proposed improved median number filter not only removes fault data resulted by mutagenesis or pulse disturbance but also removes random noise. For the designed pattern classifiers, the best burning state recognition results based on the process dataset feature vector is92.56%. Compared with individual flame image recognition result and process dataset recognition result, the best burning state recognition result of the decision level fusion based on flame image and process dataset is96.67%with PNN classifier, which not only truely realizes the "machine observing fire" replacing "operators observing fire" but also provides technical support for monitoring the rotary kiln sintering process further.③A new method based on flame image and process dataset is proposed to construct the soft-sensing of clinker quality index f-CaO content, aiming at the problem of f-CaO content offline measurement. Firstly, flame image and process dataset are pre-processed, then the color feature, global configuration feature, and local configuration feature of flame image ROIs are extracted, combined with filtered process dataset to form the input vector of the soft-sensing model, with the f-CaO content laboratory values as the output vector of the soft-sensing model, using KPLS to extract latent variables as feature vectors, finally feeds the feature vectors into the ELM regressor and support vector regressor (S VR) to form the f-CaO content soft-sensing model respectively. The experiment results show that compared with support vector regressor based, PLS feature extraction based, and direct feature level soft-sensing model based method, with ELM regressor, RMES of our proposed f-CaO content soft-sensing model is0.0638, and goodness of fit is0.9987. Such results are to lay the foundation for the closed-loop control of the clinker quality index.This paper aiming at the problem of the complex industrial process modeling, has made deep research on the burning state recognition and the f-CaO content soft-sensing of the,: cement rotary kiln, based on image processing and machine learning technique. The experiment results show that better recognition and soft-sensing results can be obtained by pattern classifiers fusion technique of various features, and process dataset can improve the pattern classifier results of the image features. This study lays the foundation for the intelligent monitoring of the rotary kiln sintering process and raises the yield and quality of the cement effectively. |
PDF Full Text Request