Abstract Developing digital twin and condition monitoring models for Floating Offshore Wind Turbines (FOWTs) mooring systems requires massive data across various health, operational, and metocean conditions. The scarcity of real damage-associated data may represent a significant challenge. Deep generative models (DGMs) have recently been introduced as powerful tools for oversampling scarce data. However, most oversampling methods focus on minority intra-class information. The inter-class dynamics between minority and majority classes are often ignored, increasing the risk of overfitting, especially in scenarios with high imbalance ratios. This study proposes a novel hierarchical variational autoencoder (HVAE) utilizing the diffusion probabilistic architecture, healthy (majority) data distribution, and the relation between healthy and damage-associated data in mooring systems of FOWTs to learn the damaged state distribution. We first evaluate HVAE’s ability to augment minority data based on majority distribution, using the MNIST benchmark image dataset for validation. This experiment compares the performance of HVAE with conventional and recent oversampling techniques. The second use case is the OC4-DeepCWind FOWT benchmark. The fine-tuned HVAE can generate damage-associated platform records for various sea states. Experimental results on MNIST indicate that HVAE achieves significant improvements over alternative oversampling techniques in downstream classification tasks, particularly in case of extreme imbalance. In the FOWT use case, the records generated for unseen sea states can incorporate the diversity and complexity of the majority ones, hence decreasing overfitting for the majority of sea states in downstream binary classification, highlighting the efficacy and generalization of HVAE.