With the widespread installation of axle-box vibration sensors on domestic urban rail trains, vibration signal-based wheel degradation state evaluation has become a critical demand for metro maintenance. But vibration signals are easily interfered by irregular noise due to roadbed diseases, variable loads, and measurement errors, leading to low accuracy degradation state evaluation. Also, insufficient data from new vehicles makes it hard to characterize degradation trends, resulting in poor robustness of detection models. To solve these, this paper proposes a robust unsupervised tensor domain adaptation network, using cross-vehicle wheel degradation information to enhance state assessment. First, deep features are extracted using a deep convolutional autoencoder network, and core tensors are obtained via Tucker decomposition. Domain adaptation is achieved by minimizing the maximum mean discrepancy loss and tensorized trend loss between the core tensors of the two domains. Second, an alternating optimization algorithm is established to iteratively optimize tensor decomposition and domain adaptation, obtaining optimal core tensor representations and shared degradation features. Finally, a health indicator (HI) is constructed from the first principal component of the degradation features, and degradation inflection point are identified by analyzing the concavity and convexity of the HI curve. Tests on metro Line 6 data confirm the method's effectiveness in transferring wheel degradation information, offering a robust and accurate solution for cross-line and cross-vehicle maintenance.