In bridge health monitoring, traditional supervised learning-based damage identification methods depend on large labeled damage datasets for training. However, acquiring sufficient damage data is technically infeasible in practice, as bridge damage events are extremely rare and difficult to replicate artificially. To overcome this challenge, this paper proposes an unsupervised structural damage identification approach based on deep Res-UNet and multi-source damage-sensitive feature fusion. First, a novel unsupervised model, Res-UNet-AE, is developed by integrating the strengths of ResNet, UNet architecture, and convolutional autoencoders. A perceptual feature encoder model from computer vision is also introduced. Second, leveraging the Res-UNet-AE and the perceptual encoder, three novel damage-sensitive features are derived: Reconstruction Error (measured by Mean Absolute Error, MAE), Signal-to-Noise Ratio (SNR), and Perceptual Loss (PL). These features are then utilized to construct a three-dimensional feature space. Finally, unsupervised clustering and identification of different damage conditions are performed using a Gaussian Mixture Model (GMM). Experimental validation on the Japanese ADA Bridge and a Benchmark steel frame structure demonstrates that the proposed framework achieves 100% accuracy in both detecting the presence of structural damage and precisely classifying various damage scenarios. This method offers an effective solution to the scarcity of damage samples in practical engineering and shows promising application prospects for bridge structural health monitoring.