Deep belief network (DBN) has recently emerged as a powerful tool in building nonlinear data driven models. However, a single DBN model can still lack reliability especially when the amount of data available for modelling is limited. This paper proposes a bootstrap aggregated deep belief network (BAGDBN) to improve model reliability and robustness. In the proposed method, bootstrap re-sampling with replacement is applied to the original modelling data to generate multiple replications. A DBN model is developed on each replication of the original modelling data. These individual DBN models are then combined to form a BAGDBN model. The proposed method is demonstrated on two application examples, modelling of a conic water tank and inferential estimation of polymer melt index in an industrial polypropylene polymerization process. The application results demonstrate that the proposed BAGDBN models can give more reliable estimation and prediction than single DBN models.