Burning State Recognition For Rotary Kiln Sintering Process Based On Burning Image Sequence Analysis

Rotary kiln is a large core thermal equipment and used in cement,alumina and other industrial raw materials production process,it is used for physical and chemical treatment of materials to produce clinker.But the key process parameters can not be detected online and thus it is difficult to realize the automatic control of the kiln production process because the continuous rotation of the kiln in the production process,coupled with the complicated combustion in the kiln,the convection heat transfer between the flue gas and the material and other factors.At present,rotary kiln clinker sintering process still relies on the industrial television mode.However,this operation pattern is restricted by some personal elements,easily leading to low quality of product,low efficiency of kiln,low productivity and high energy consumption.The clinker sintering state of burning zone are related to the clinker quality tightly,images of burning zone are important to determine the clinker burning state,this is a solid foundation to research rotary kiln clinker burning state recognition technology based on the video image.However,due to the continuous rotation of the kiln and the interference of kiln dust,the noise of burning state video is large,the significant areas are strong coupled and has blur boundary.At present,the recognition method based on static images is easily affected by various noise information,resulting in a lower accuracy of identification.The burning zone video sequence contains more complete and more robust information about clinker sintering states,which can reflect the dynamic changes of clinker sintering states more accurately.Hence,using the research achievements in video processing,machine learning and deep learning has important theoretical significance and application value to study on burning state recognition of rotary kiln clinker based on burning zone video.For above problem,the dissertation is supported by the national natural science fund project "product quality parameter prediction modeling based on the fusion of video and process data about rotary kiln".Taking cement rotary kiln as the research object,the video sequences of burning zone were collected.Considering the characteristics of time series data,combined with the semi-supervised learning method,the research on the clinker sintering state identification method based on video images of burning zone was carried out with machine learning and depth learning technology.The major contributions of this paper are summarized as follows:(1)Aiming at the shortcomings of current clinker sintering state identification methods based on static images and supervised learning methods which require a large number of manually calibrated samples,the rotary kiln sintering state identification method of semi-supervised learning and hidden Markov models(HMM)based on the dynamic information of the video image sequences are studied.Firstly,the image sequence information should be preprocessed and extracted features using PCA,and then processed into two sample sequences of fast and slow scales.Fast and slow scale samples refer to video samples collected by the image acquisition card at a certain frame rate(milliseconds),extracting a single frame or a plurality of frames within the time interval according to a short time interval(such as every frame or several frame intervals,millisecond intervals)or a long time interval(for example,one identification period,second level)information,and then processed into a fast or slow scale sample(such as can take the mean or expansion of the eigenvector),and then generate fast or slow scale sample sequence.Comparing with the static image,the fast-scale sample sequence reflects the characteristic change in a short time during the sintering process;the slowscale sample sequence reflects the switching and transition characteristics of the sintering process in a long time.Then,a fast-scale dynamic information classifier such as Gaussian Mixture Model-Hidden Markov Model(GMM-HMM)or Support Vector Machine Model(SVM)is developed by using supervised learning method with fast scale sample sequence with class calibration.Slow-scale dynamic classifier(GMM-HMM)is developed by using unsupervised learning method with slow-scale sample sequence.Then,based on HMM filter framework,a semi-supervised time series classifier method based on fast and slow scale dynamic information is proposed.Furthermore,combing measure propagation(MP)semi-supervised learning method with fast scale dynamic information classifier based on supervised learning method,a semi-supervised fast-scale dynamic information classifier based on MP is proposed.The above two methods can improve the accuracy of burning state identification due to the unlabeled samples and fast-slow scale information of time series data.(2)Aiming at the complexity of feature expression and classification of time series data,deep neural network has a larger capacity than GMM-HMM,and it can discover hidden useful information from massive data to learn feature representation.deep neural network can also solve dynamic timing problems(Such as RNN),so further study is conducted by using deep learning methods.Two classifier methods based on the dynamic information of the video sequence in the burning zone are proposed:The first one is to use the recurrent neural network(RNN)as a classifier to classify the fast-scale sample sequence after the preprocessing of the rotary kiln image sequence data and the preliminary feature extraction of the PCA.The second is to use CNN-RNN deep learning classifier to extract and classify further features of fast scale sample sequences.During the construction of the model,the hyper parameters,such as the learning rate of the optimization algorithm,the L2 regularization coefficient,the dropout ratio of the convolutional neural network and the number of layers of the network are chose reasonably,thus the model structure is determined.(3)The experimental research on the above method was carried out on a sample set of 10%,30%and 50%labeled samples using the rotary clinker sintered video sequences of a cement plant.The experimental results show that compared with the traditional static classifier,the classification accuracy of the two kinds of semi-supervised learning classifiers based on dynamic information presented in this paper has been greatly improved.In addition,the validity and universality of the two kinds of structural frameworks are verified by combining SVM based on dynamic information samples.Finally,two deep learning models are constructed by TensorFlow.The model structure is determined by optimizing the hyper parameters.Finally,the fast scale GMM-HMM,RNN and CNN-RNN time series classifiers are compared and analyzed.