Person re-identification (re-ID) in the unsupervised domain adaptation (UDA) settings faces the problem of domain shift caused by pseudo-label noise, cross-camera differences and background interference. Existing methods, such as mutual mean teaching, are often limited by error accumulation and insufficient network diversity, which makes it difficult to fully learn discriminative features. To this end, we propose an identity-preserved collaborative learning framework (IPCL) to address the above challenges by enhancing the robustness and feature consistency of pseudo-labels. We propose a collaborative contrast refinement (CCR) strategy, which combines soft pseudo-label supervision and contrast alignment in a dual network structure to achieve robust and structure-aware learning under noisy conditions; and an identity-preserved filtering (IPF) module, which is used to suppress style-related noise and highlight identity discriminative areas. Experiments on six UDA benchmark datasets show that IPCL outperforms existing advanced methods, verifying the effectiveness of its joint optimization mechanism and identity preservation mechanism.