A novel digital twin method based on diffusion models for imbalanced fault diagnosis of rotating machinery

The operating time of rotating machinery under fault conditions significantly falls short compared to its normal functioning state, and procuring fault signals proves exceptionally challenging due to concerns related to safety and costs. Consequently, an imbalance between normal and fault status samples in rotating machinery frequently arises, gravely restricting the performance of intelligent diagnostic techniques. This paper aims to develop a novel digital twin method based on diffusion models to alleviate the data imbalance issues. When faced with a scarcity of fault samples, the digital twin model can create virtual signals to accurately mimic the rotating machinery's fault conditions. Initially, a non-autoregressive network, Idfwave, is built to execute the denoising task within the diffusion model, enabling the stable production of high-fidelity fault vibration signals. Additionally, integrating bidirectional dilated convolution, cutting-edge convolutional modulation techniques, and fast sampling strategies substantially enhances the quality of the generated signals. Subsequent fault diagnosis experiments were conducted across diverse datasets and various data imbalance scenarios to verify the generated samples' availability. The results indicate that the proposed model consistently produces signals indistinguishable from real ones, even when trained on a limited set of fault signals, and boosts the accuracy of fault identification and diagnosis in classification models. Furthermore, the proposed model surpasses other prevalent generation models in generating vibration signals, confirmed through comparative analyses.