Due to physiological and environmental influences, PPG signal distributions dynamically change over time, potentially impairing model generalization. To address this, we propose a single-domain generalization disentangled representation learning framework (PPGDRL) for PPG biometrics. First, data augmentation via discrete Fourier transform (DFT) enriches distribution diversity. Then, a deep disentangled learning framework (D1DViT) based on a 1D vision Transformer is built, comprising local and disentangled encoders. The local encoder extracts PPG features into sequences, while the disentangled encoder separates them into domain-invariant and domain-specific features, capturing discriminative and environmental information, respectively. Finally, the recognition results of overall features and invariant features are predicted through a dual-branch classifier. Experimental results on public PPG databases demonstrate that PPGDRL achieves strong generalization performance under distribution shift.