The design of a soft-sensor to improve the monitoring of the sugar crystallization process is examined. Information about the mother liquor purity is relevant to improve the manufacturing process, especially during the last stage of three massecuites, called C crystallization. However, this piece of information is not available on-line and requires the development of a soft-sensor. In industrial context the measurements are often incomplete and/or noisy, therefore an input-output model is chosen instead of a knowledge one. An artificial neural network model is used to predict on-line the evolution of the purity of the solution during the crystallization process. The validation step is performed using industrial databases and experimental results show the efficiency of the proposed soft-sensor.