In recent years, neural networks have been extensively used to model complex nonlinear chemical processes. This article studies their application in the development of virtual (software) sensors for product quality prediction. These sensors use, for the training and validation, information extracted from historical process operational databases. A method is proposed to select, out of the whole set of available measured process variables, an appropriate subset used as input to the software sensor. This method, based on principal component analysts and information theoric tests helps to discard redundant, superfluous and collinear variables, while keeping those that explain most of the variations of the sensors´s output. To show an application of these procedures, product quality sensors for a refinery distillation unit are developed using operational process data.