Identity-preserved 3D human skeleton sequence generation is crucial for human-robot interaction, motion synthesis, and animation. These sequences contain both common kinematic patterns and unique identity-specific characteristics. We introduce a novel approach based on a disentangled information bottleneck framework, utilizing Semantics-Guided Neural Networks (SGNs) as the encoder and reverse SGNs as the decoder. Our method explicitly decomposes skeleton sequence inputs into identity-relevant and identity-irrelevant features, enhancing the robustness of identity-aware generation. Additionally, we incorporate a density estimator for precise probability estimation, further enhancing reliability. Experiments on multiple datasets demonstrate that our approach achieves competitive performance in identity recognition accuracy, diversity, and multi-modality compared to state-of-the-art techniques. This work provides practical architectural advancements for identity-preserved 3D human skeleton sequence generation, with our representation disentanglement demonstrating effective separation of identity and identity-irrelevant motion characteristics.