Four techniques are studied to design a soft sensor for dry substance content estimation (% DS) in the sugar industry. Dry substance content sensors are in general expensive and inaccurate, so it is interesting to study and develop soft sensors for this variable. Concretely, the dry substance content of the juice leaving the evaporation station has been estimated. For that purpose, four methods have been proposed. The first one is based on indirect measurements, using physicochemical properties. The second one uses neural networks where the inputs to the net are selected manually, based on a correlation study of the variables of the evaporation station. The third one uses neural networks whose inputs are the scores calculated by means of Principal Component Analysis (PCA). The last method uses an estimation based on Partial Least Squares (PLS) regression. This paper explains, compares and analyses the results obtained using real data collected from the plant.
Asymptotic Prediction Error Variance for Feedforward Neural Networks
