Accurately and rapidly identifying the switching states in continuous lower limb movements is crucial for natural humanrobot interaction ( HRI ) with exoskeletons. The switching state sEMG signals contain both pre-and post-switching movement information, as well as transient information related to the switching, making them difficult to directly use for recognition. In order to quickly and accurately identify the switching states, this paper proposes a real-time recognition method called FMICMD-LACNN. An adaptive multi-component instantaneous frequency estimation method is proposed to improve the computational efficiency of the multivariate intrinsic chirp mode decomposition ( MICMD) , and a component energy penalty factor is proposed to enhance the decomposition accuracy of MICMD, thus forming the fast multivariate intrinsic chirp mode decomposition (FMICMD) algorithm. For the sEMG signals decomposed by FMICMD, a LACNN recognition model was established to achieve fast and accurate switching states identification. This study collected sEMG signals from 10 subjects in 8 common lower limb continuous motion switching states for experimental verification. The results show that for these 8 switching states, the average recognition accuracy of this method is 98. 35% , and the average recognition time is only about 8 ms, which is better than the CNN-LSTM, E2CNN and CNN-BiLSTM methods. This method has high accuracy and real-time performance, and can meet the needs of fast and natural interaction between the exoskeleton and the human body.