The loss of excitation fault of large condenser seriously affects the safety and stability of equipment and system. The reliability and selectivity of existing low-voltage and reactive power reverse criteria based on local static threshold are insufficient. In this paper, a loss of excitation protection principle based on intelligent identification of the global dynamic trajectory of the measured impedance is proposed, which can reflect the operating state of the condenser. From the point of view of kinematics, characteristic quantity time series that can accurately restore the measured impedance trajectory under loss of excitation and other conditions is formed, and statistics is further introduced to extract the highly explanatory features. Multiple kernel learning support vector machine (MKL-SVM) is trained by using the combination of global and local kernel functions of adaptive weights to ensure the learning ability of the classification model while enhancing its generalization ability. A two-stage recognition strategy based on the space distance of the classification core is proposed, which can improve the protection reliability while ensuring the system security. The model of condenser connected to power grid is built based on PSCAD simulation platform for verification, and simulation results show that the proposed method does not need to collect the electrical quantities at the rotor side with high identification accuracy, and it still has excellent applicability in the face of new energy access and unknown disturbances.