A virtual sensor that estimates product compositions in a middle-vessel batch distillation column has been developed. The sensor is based on a recurrent artificial neural network, and uses information available from secondary measurements (such as temperatures and flow rates). The criteria adopted for selecting the most suitable training data set and the benefits deriving from pre-processing these data by means of principal component analysis are demonstrated by simulation. The effects of sensor location, model initialization, and noisy temperature measurements on the performance of the soft sensor are also investigated. It is shown that the estimated compositions are in good agreement with the actual values.