DDPM‑LGFormer: a local and global transformer denoising diffusion probabilistic model for damage diagnosis with imbalanced data

Annals of Forest Science · Springer Nature · 2026-10-29 · DOI: 10.1007/s10845-026-02995-0

Аннотация

Accurate damage diagnosis is critical for engineering system reliability, yet data imbalance often degrades model performance. To tackle this issue, we introduce a local‑global transformer denoising diffusion probabilistic model (DDPM‑LGFormer). Unlike generative adversarial networks (GANs), our diffusion framework employs a Markov chain that incrementally adds noise to data and learns to reverse the process for high‑fidelity sample generation. The LGFormer module captures both fine‑grained local features and comprehensive global damage patterns, enabling realistic synthetic samples. Extensive experiments on bearings, gears, and airship envelopes show that DDPM‑LGFormer consistently outperforms GAN‑based methods in generating authentic damage samples, demonstrating strong generalization across heterogeneous objects. Moreover, the proposed model substantially boosts the diagnostic accuracy of downstream classifiers, offering a practical and innovative solution to imbalanced damage diagnosis in real‑world scenarios.

Читать у издателя →

Похожие статьи