Frequency diffusion-based vibration signal generation for machinery fault identification under limited data

Accurate machinery fault identification remains challenging under data scarcity, where conventional generative models struggle to synthesize signals with precise fault-related spectral characteristics. To address this, we propose a frequency domain-based signal generative-identification framework that enhances both signal fidelity and localized spectral feature representation. The framework integrates the entire pipeline from sample generation to fault identification, coupling the score-based diffusion approach with transformer-based identification models through the discrete Fourier transform. It accomplishes a unified investigation of spectral feature-driven sample synthesis, ensuring minimal deviation from fault patterns. Focusing on two critical rotating machinery types, deep groove ball bearings and centrifugal compressor blades. This framework generates high-fidelity fault signals that reflect real-world frequency resonance bands. Experimental validation demonstrates that synthesized signals reduce sliced Wasserstein distance by an average of 43% compared to generative adversarial network and time-domain diffusion-based approaches, while improving fault identification accuracy up to 7% per fault category. Ablation studies verify the framework’s capability to preserve transient features and suppress spectral mode collapse, achieving a maximum identification rate of 99.36% across variable fault types. This study provides a robust data augmentation solution for intelligent fault identification under limited-data scenarios, particularly for situations requiring localized spectral fidelity.