Abstract: Health monitoring and early fault detection of rotating machinery are crucial for ensuring equipment operational safety. Traditional fault feature extraction methods often rely on preset parameters, limiting their effectiveness in practical applications. Although the Ramanujan Periodic Transform can effectively separate periodic components in signals by constructing orthogonal subspaces, its performance heavily depends on high-fidelity fault simulation models, which typically require prior knowledge of fault features, making it difficult to generalize in scenarios lacking fault samples. To address this issue, a digital twin architecture integrating conditional diffusion models and the Ramanujan Periodic Transform (Diffusion-Ramanujan) is proposed. This method employs a classifier-free guided conditional diffusion model to transform unknown-type fault samples into healthy samples. By comparing the distribution differences in the envelope spectra before and after transformation, potential fault features are automatically mined. Based on the extracted features, a phenomenological simulation model is constructed as a digital twin to drive the Ramanujan Periodic Transform, enabling parameter-independent fault feature separation from monitoring data and the establishment of health indicators for equipment health monitoring. Experimental results on the IMS bearing dataset demonstrate that the proposed method can effectively identify early faults without prior fault knowledge, achieving earlier fault detection time and higher monitoring accuracy compared to baseline models such as Improved Envelope Spectrum and Anomaly-Transformer. This provides an effective pathway for rotating machinery health management that does not rely on preset parameters or prior knowledge.

Key words: rotating machinery, early fault detection, diffusion model, Ramanujan periodic transform, digital twin