Abstract: Advanced multi-function radars can generate various modulated and flexible parameterized work modes and can perform multiple different tasks simultaneously on the timeline, thus posing huge challenges to the electronic support system. This paper proposes a change point detection method for the “symbol-pulse” layer work mode in the multi-functional radar hierarchical model. This method is based on an autoencoder and is designed for the non-ideal observation situation. This method can detect work mode change points accurately not only in an unsupervised manner, but also free of parametric model assumptions. The simulation experiment verified the effectiveness and superiority of the algorithm proposed in this paper compared with the change point detection algorithm, it also has the best performance under non-ideal observations including 0~20% missing pulse ratios.