To address the data scarcity in structural health monitoring (SHM), this study proposes a health index (HI) prediction framework based on a feature-transfer-based conditional denoising diffusion probabilistic model (Conditional DDPM) for data augmentation. Specifically, guided wave signals from piezoelectric (PZT) sensors are first processed using the Hilbert–Huang Transform (HHT) to extract two-dimensional feature maps. Subsequently, a novel HI construction method based on the inverse Weibull distribution function is proposed as the condition for data generation in the Conditional DDPM. A two-stage training strategy is then implemented, where the model is pretrained on normalized feature maps to comprehensively learn data distribution patterns, followed by transfer learning on raw feature maps to leverage pretrained knowledge for generating high-quality augmented data. Experimental results demonstrate that the proposed method achieves significantly improved HI prediction performance compared to traditional augmentation techniques and other intelligent generative models such as conditional variational autoencoders (VAEs) and generative adversarial networks (GANs). Cross-validation experiments on a compression-compression fatigue dataset show a 25.5% reduction in root mean square error (RMSE). Furthermore, a generalization test on a tension–tension fatigue dataset indicates the method's potential effectiveness across different datasets. The work provides a novel and efficient data augmentation strategy for SHM of composite structures, showing promising potential to improve structural performance and reliability.