Real-time gypsum quality estimation in an industrial calciner: A neural network-based approach

Abstract

Total bound moisture (TBM) is a typical quality indicator of industrial-grade gypsum. This gypsum is comprised of three distinct phases, namely anhydrite, dihydrate, and hemihydrate, of which only the latter is of much industrial use. TBM analysis is a lengthy laboratory procedure, and an artificial neural network (ANN) TBM inference measurement is proposed as a fast and online alternative. An ANN inference model for gypsum TBM based on plant data was developed. The inputs to the network were primarily focused on the plant's calciner, and different network topologies, data divisions, and transfer functions were investigated. Furthermore, the applicability of the TBM value as a quality indicator was investigated based on a gypsum phase analysis. A strong correlation between TBM and the gypsum hemihydrate and anhydrite content was found, validating the plant target TBM of 5.8% as a quality indicator. A network topology consisting of one hidden layer with logarithmic-sigmoid (logsig) and pure linear (purelin) transfer functions showed the best performance (R > 90%).