Frequency domain guided latent diffusion model for domain generalization in cross-machine fault diagnosis

Abstract

Most existing cross-domain fault diagnosis methods focus on extracting domain-invariant features for fault monitoring by employing regularizers. However, these methods often struggle to manage significant domain shifts encountered in industrial environments. Additionally, regularizers tend to overemphasize features from the source domain, leading to negative transfer. To address these challenges, a novel data augmentation-based approach is proposed, termed the Frequency Domain Guided Latent Diffusion Model for Domain Generalization in Cross-Machine Fault Diagnosis. This approach involves the construction of an advanced latent diffusion model with an encoder–decoder architecture, which generates diverse and reliable samples by training the forward noise addition and backward denoising processes. To enhance feature generalization, a frequency domain reconstruction module is introduced, which employs frequency filtering and frequency resampling techniques. This module guides the latent diffusion model in generating domain-invariant samples by leveraging the correlation between the frequency domain and feature generalization. By utilizing data augmentation in the frequency domain, this method mitigates domain shift and avoids the computational risks associated with regularizers. Finally, a fault diagnosis model based on empirical risk minimization is developed, and experimental results demonstrate that the model outperforms the state-of-the-art methods by 4% in cross-domain fault diagnosis accuracy.