Cement Clinker F-CaO Prediction Method Based On Extreme Learning Machine

f-CaO is the free lime existing in the cement clinker. The concentration of f-CaO is an important index for the quality judgment of cement clinker. At present, f-CaO concentration is mostly measured offline in laboratory every one hour, making it difficult for the workers to make judgement about the process condition and to adjust the process parameters in time. It is hence necessary to develop a method to predict the f-Cao concentration automatically.However, the process mechanism in cement production is rather complex and is characterized by large delay, nonlinearity, strong coupling and time-varying, making it difficult to develop mechanism-based prediction models. Extreme learning machine(ELM) is a new single-hidden layer feed-forward neural network, which belongs to data-driven model and has the advantages of simplicity in parameter setting, fast learning speed, global optimization and good generalization, and find many applications in the modelling of complex process.The present work, in the framework of a NSFC project and a NSFHN project, aims to study an ELM-based prediction method for the concentration of f-CaO in cement clinker for a cement plant located in Jiang Xi province. The research work accomplished and the main conclusions are as following:(1) The calcining process of cement clinker was analyzed, including the movement of material, airflow and combustion. The important process equipments involved and the traditional manual measuring methods of f-CaO are also described.(2) Based on the cement production process and practical experience, three easily measurable process variables were chosen as model inputs(the current of the drive motor, the temperature of the decomposing furnace and the pressure of the cooler), and the structure of ELM-based f-Cao prediction model was designed. The time matching problem of inputs and output s was also studied due to the time-delay features of the cement process.(3) Numerous data had been collected in a cement plant in Jiang Xi province for 3 months, including all measurable process parameters covering the pre-heater and the cooler, as well as the data of f-CaO recorded by the lab. The original data was preprocessed, and 720 reliable training and testing sets were obtained.(4) The ELM-based f-CaO prediction model was realized by MATLAB programming, and verified by the training set and testing set with a mean square error of 0.247 and 0.196, respectively, which is less than the error of other models(about 0.5) reported in literature, such as BP neural network and support vector machine.(5) The effect of hidden nodes number, data filtering algorithm and time matching method on model prediction results was also analyzed. An optimum was obtained with 310 hidden nodes, 5-minute filtering and time matching method C.(6) The ELM prediction result were compared with SVM prediction result, and was found to have better performance in the maximum absolute error, mean absolute error and the mean square error.