Speed-guided diffusion probabilistic model integrated with variational autoencoder for fault diagnosis under limited data

Deep learning-based fault diagnosis has shown great potential in intelligent condition monitoring. However, its performance heavily depends on large amounts of labeled data, which are often scarce in real-world industrial scenarios, especially under varying speed conditions. The data limitations hinder the generalization ability of fault diagnosis models. To address this issue, we propose a novel speed-guided denoising diffusion probabilistic model integrated with variational autoencoder (SDPM-VAE) for generating high quality data. The speed signal is incorporated as a conditional prior through a cross-attention mechanism to preserve speed-dependent characteristics in the generated data. Additionally, a hybrid VAE and DDPM framework is proposed to enhance data quality while reducing computational cost. The coarse reconstructions and latent representations derived from the VAE are integrated into the reverse diffusion process. Experimental results on two gearbox datasets demonstrate that the SDPM-VAE outperforms four state-of-the-art methods in both the quality of the generated data and fault diagnosis accuracy, thereby validating its effectiveness for data augmentation under limited data conditions.