A neuro-fuzzy system for prediction of pulp digester K-number

A neuro-fuzzy system (NFS) has been developed for the prediction of K-number (K ) in a continuous wood pulp digester. The NFS architecture uses a K-means fuzzifier and a modified center-average defuzzifier for fuzzy/crisp conversion. The fuzzy rule base is determined from observed input/output data via an iterative rule-confidence matrix training algorithm, and a max-min fuzzy inference engine is used for rule interpretation. A hybrid backpropagation/genetic-algorithm training routine was developed for tuning of all membership functions. K modeling experiments were conducted on six months of observed industrial digester data at a fifteen-minute sample rate. A variable transform was developed and successfully applied to process variable representation, effectively describing the history of an input variable over a specified window of time