In multivariate model predictive control (MPC) systems, the quality of multi-input multi-output (MIMO) plant models has significant impact on the controller performance in different aspects. Though re-identification of plant models can improve model quality and prediction accuracy, it is very time consuming and economically expensive in industrial practice. Therefore, the automatic detection and isolation of the model-plant mismatch is highly desirable to monitor and improve MPC performance. In this paper, a new closed-loop MPC performance monitoring approach is proposed to detect model-plant mismatch within state-space formulations through subspace projections and statistical hypothesis testing. A monitoring framework consisting of three quadratic indices is developed to capture model-plant mismatches precisely. The validity and effectiveness of the proposed method is demonstrated through a paper machine headbox control example.