The climate in modern livestock production buildings is controlled using a simple state controller. State controllers are typically not equipped to handle abnormal situations, e.g. sensors providing false or no readings, and they may therefore produce malformed climate control signals which may have severe consequences for the livestock. To make the system more robust, a sensor fusion system can be used to combine the different sensor readings and thereby pro- duce a more reliable estimate of the climate state. A dy- namic Bayesian network (DBN) model is constructed for this purpose. The model is tested in an online setup in a climate laboratory, where realtime behavior is archived by using the Boyen & Koller approximate inference algorithm. Preliminary experimental results show that the proposed model provides a promising framework for sensor fusion in livestock buildings.