Research On Modeling Based On Improved Slow Feature Analysis For Cement Rotary Kiln

Dry-process precalciner kiln technology,including the raw material and coal preparation,the raw materials decomposition,the clinker burning and the cement preparation,significantly increases production efficiency and has become the mainstream of the cement industry.Due to the complexity of the cement production process and the technical problems in the industry,it is difficult to detect the raw material decomposition rate and clinker product quality index online,which cause that the current cement production can't meet the needs of modern industrial process automation.The cement raw material decomposition rate is usually obtained by manual sampling and offline testing,and the rotary kiln sintering process still relies on "operators observing fire",which is restricted by some personal elements.Hence,it is necessary to apply machine learning technique to the engineering practice and develop new methods for the cement raw material decomposition rate soft-sensing and the burning state recognition based on image processing and machine learning theory,which can solve the control and optimization problem in rotary kiln.The process data of the complex industrial processes such as cement production are continuous change with time,so the modeling problem can be regarded as multivariate time series learning problem.The existing cement rotary kiln modeling methods are based on static data.However,the excessive resampling interval of static data can lead to the loss of dynamic information so that it is difficult to extract effective and robust features to establish accurate model.Compared to static data,time series often is ordered in time and space and have potential useful knowledge.So it is of great theoretical and practical significance to study how to effectively extract useful features from the time series.The invariant feature is the essential feature of time series data.Slow feature analysis is an important unsupervised learning algorithm,with the goal of extracting the slowly changing features from time-series.It is an urgent problem to establish the cement rotary kiln model based on dynamic multivariate time series and slow feature analysis through in-depth analysis of the characteristics of cement production process,which can raise the accuracy of the cement raw material decomposition rate soft-sensing and the burning state recognition.The dissertation has developed the cement raw material decomposition rate soft-sensing method and the burning state recognition method of the cement rotary kiln based on based on the machine learning and image processing technology,with the goal of raising the accuracy of the cement raw material decomposition rate soft-sensing and the burning state recognition.The major contributions of this study are summarized as follows:(1)The dynamic data of cement raw material decomposing process is of correlation structure in both space and time,that is to say,the data is tensor.The traditional SFA algorithm is based on the vector form,which can lead to the loss of useful spatial-temporal structure information.And the tensor-to-vector transform can result in ill-conditioned matrix which leads to the failure of the eigenvalue decomposition.To solve the problem,this dissertation proposes tensor slow feature analysis(TSFA)algorithm by combining the slow feature analysis with the tensor analysis,which can capture tensor correlation structure,control ill-conditioned matrix and simplify calculation.Then,we extend the TSFA to nonlinear TSFA,so as to better solve the nonlinear problem of data.In this dissertation,applying data mining and machine learning technique,we proposed a novel soft sensor scheme based on the TSFA and ?-SVR.The experimental results of soft sensors based on different feature extraction methods show the effectiveness of TSFA;(2)Some of the video data of sintering process in rotary kiln is labeled,but SFA algorithm is an unsupervised algorithm,only extracts the feature of unlabeled samples,and lack the guidance of supervision information,which leads to poor extraction results.To solve the problem,we propose a novel semi-supervised dimensionality reduction method called spatial-temporal semi-supervised slow feature analysis(ST-SSFA)based on the SFA algorithm and the marginal Fisher analysis(MFA)algorithm.ST-SSFA algorithm can use time-series information,reveal the manifold structure of the unlabeled samples and reflect the label information of the labeled samples,so it can extract more effective and easily identifiable feature.In this dissertation,applying image processing and machine learning technique,we proposed a new method for the burning state recognition based on ST-SSFA.The experimental results based on different feature extraction methods show the effectiveness of ST-SSFA;(3)In addition to the research on algorithm,a rotary kiln clinker sintering state recognition experimental system is designed and developed based on Apache + PHP +MySQL and MATLAB.The system contains three parts:industrial background introduction,model off-line training and model online simulation.The simulation experiment is done based on the simulation software system for the burning state recognition.And the simulation software system provides a necessary platform for cement rotary kiln modeling research.