Vibration-based structural damage detection encounters challenge due to the inadequate sensitivity of features representing structural conditions. Recently, there has been growing interest in deep learning (DL) methods which avoid the process of manual feature extraction. However, the effectiveness of DL-based damage detection is often limited by class imbalance in the dataset, where instances of damage are much fewer compared to undamaged cases. This study proposes a damage detection method through data augmentation to address the issue of class imbalance in datasets for structural health assessment. First, a finite element model (FEM) of a three-span continuous beam bridge is built, from which a large volume of vibration data for the healthy condition and a smaller volume for the damaged condition are simulated under white noise excitation. These raw data are then normalized and sliced to create original datasets for both healthy and damaged conditions. Subsequently, a deep generative model, named Denoising Diffusion Probabilistic Model (DDPM), is employed to generate additional samples based on the limited number of original datasets for damaged condition. DDPM operates by gradually adding Gaussian noise to the original data with timesteps according to a pre-defined schedule in a forward process, followed by a reverse process to denoise and reconstruct the data by a U-Net network. The performance of this model is then compared with a mainstream deep generative model, Generative Adversarial Network (GAN), through the Fréchet Inception Distance (FID) between feature vectors calculated for original and generated data. Finally, the augmented dataset, combined with the original for healthy condition, are utilized for damage identification using a trained Convolutional Neural Network (CNN). The impact of the timesteps and the configuration of the noise schedule in DDPM as well as the measurement noise on the classification performance is evaluated in detail.