Model predictions often exhibit discrepancy with respect to experimental observations, due to assumptions and approximations in the model. Bayesian approaches for estimating discrepancy in single models have been studied in the past. In this paper, we approach the problem of discrepancy estimation in coupled models (especially multi-disciplinary models) using Bayesian state estimation methods. The proposed state estimation-based approach is found to have significant advantages over the previously studied Kennedy-O'Hagan method, in the estimation of discrepancies of hidden states, and in the identification of the sources (namely, model form errors) of the discrepancies. We adopt a partitioned approach to estimate discrepancies in weakly coupled systems and perform state estimation separately for the individual disciplines. The proposed approach is illustrated for a four-discipline problem related to aero-thermo-elastic response prediction of a hypersonic aircraft panel.