To achieve real-time gait recognition under multi-user, multi-cadence, and multi-gait-type conditions, a gait recognition method integrating a cycle-adaptive mechanism and knowledge distillation is proposed. This method comprehensively considers both recognition accuracy and real-time performance, aiming to enhance the adaptability of lower-limb exoskeleton systems in dynamic and complex environments. First, a cycle-adaptive sliding window mechanism based on spectrum analysis is designed to dynamically adjust the window length according to the signal’s periodicity, precisely adapting to variations in users and gait cadences. Then, a graph neural network (GNN) is used as the teacher model to fully extract the spatiotemporal relationships in multi-channel IMU data, and a multi-layer perceptron (MLP) model is used as the student model. Knowledge transfer is achieved through knowledge distillation. Experiments were conducted using five types of gait data for comparison and validation. Under the same data conditions, the adaptive sliding window scheme improved the overall classification accuracy from 91.6% to 94.2%, an increase of 2.6%. By combining hard labels with soft labels from the teacher model and optimizing the distillation loss function parameters, the student model learned the rich features and information from the teacher model, improving accuracy by 1.94%. Meanwhile, the average recognition time per window for the student model was reduced from 17.4 ms (teacher model) to 4.9 ms, significantly enhancing real-time responsiveness. Experimental results show that the method demonstrates good recognition stability and generalization across multiple users, cadences, and gait types, combining high accuracy and low latency, with strong practicality and deployment value. It is suitable for scenarios such as lower-limb exoskeletons that require high real-time performance and recognition accuracy.