The integration of imitation learning and reinforcement learning has demonstrated significant success in whole-body control for humanoid robots, enabling them to execute a diverse range of locomotion behaviors. However, due to the limitations of imitation targets in imitation learning and the limited exploration capacity of reinforcement learning, single-skill policies often outperform multi-skill ones on specific tasks. Furthermore, real-world tasks typically require a combination of multiple skills, which restricts the applicability of imitation learning. To address this issue, we propose a training framework based on the Fuzzy C-means clustering for imitation learning state initialization. This framework analyzes the state information of robots performing various skills within the imitation learning dataset, identifies similar motion states across different tasks, and utilizes these states to initialize training for robots learning similar skills. This framework extends the adaptability of humanoid robot motion control by leveraging the Fuzzy C-means clustering to explore the optimal training initialization for stable and flexible switching between multiple skills. It ultimately enables smooth transitions between diverse skills and control policies. In our experiments, we demonstrated the effectiveness of this approach on the H1 humanoid robot, achieving a 90.83% success rate when switching among four distinct tasks.