Design of soft sensors for monitoring and control of cement manufacturing processes

Abstract

In most of the process industries the end product quality is not measurable online. Unavailability of reliable hardware sensors for continuous quality monitoring at times results in production of low quality products. Soft sensors or inferential sensors are process models which use the information of easily measurable process variables and produce as output, the estimated/predicted values of difficult to measure process variables. Most of the present day approaches for soft sensor development are data based because of the various difficulties associated with development of mathematical models for complex processes. Clinker composition and cement particle size are two of the most important quality parameters in cement manufacturing process. Unfortunately, there are no hardware sensors available for continuous monitoring of these quality parameters. Therefore, the focus of this research is to develop data-driven soft sensors for online monitoring of cement clinker quality and cement fineness. The two processes focused upon are, (1) a chemical process where the raw mix is converted to cement clinker at high temperature in a rotary cement kiln and (2) a physical process where the clinker is ground in a vertical roller mill for cement production. The required data for these two processes were collected from a cement plant. The grinding process data set comprised of data for three input variables and one output variable. The three input variables are measured continuously by installed hardware sensors and therefore contain some outlying observations. These outliers were detected and removed by the robust Hampel's method of outlier detection. The processed data set after outlier removal, consisted of 158 samples of input-output data. This data set was equally divided into a training set and a validation set consisting of 79 observations each. The data division was performed using Kennard-Stone maximal intradistance criterion. The training set was used for development of different kinds of datadriven soft sensors. The different soft sensors developed include, linear and support vector regression, artificial neural network, fuzzy inference and neuro-fuzzy models of Abstract the grinding process. The performances of the developed models were assessed with the validation data set by measuring six different statistical model performance indicators. The neuro-fuzzy and the back propagation neural network models showed better performances than the other models. The accuracy of both these models are quite acceptable as per the model acceptability criteria reported in the literature. The clinkerization process taking place in the rotary kiln, involves nine inputs and eight outputs. Out of the nine inputs, 4 are raw meal quality parameters (measured in the laboratory) and five are kiln operating parameters. The operating parameters are measured continuously by installed hardware sensors and therefore contain outliers. After performing a comparison of different multivariate outlier detection techniques, the outliers present in the kiln input data were removed by the technique of closest distance to center method. The processed data set consisted of 223 pairs of input-output data. Using Kennard-Stone algorithm, the total data set was divided into a training set consisting of 112 samples and a validation set containing 111 samples. The training set was used to develop three kinds of feed forward neural network models and two types of fuzzy inference models. The performances of the developed soft sensors were analyzed with the validation data set by evaluating six statistical model evaluation parameters. The analysis showed that the soft sensor based on Takagi-Sugeno fuzzy inference modeling technique, has the highest accuracy in estimating eight cement clinker quality parameters.