An adaptive conditional diffusion generative adversarial network for machinery fault diagnosis

Abstract

Machinery fault diagnosis is pivotal for ensuring reliable operation and health management of industrial equipment. However, obtaining sufficient balanced datasets remains challenging in real-world scenarios, significantly limiting the generalization performance of traditional deep learning methods. To address these limitations, this article proposes an adaptive conditional diffusion generative adversarial network (ACDGAN), aiming to enhance both data generation quality and training stability for machinery fault diagnosis. Specifically, the contributions of ACDGAN are threefold. First, a cascaded dual-path adaptive diffusion network is designed to effectively capture multiscale features via sequential adaptive diffusion processes. Second, a condition-guided adaptive category control mechanism is introduced to guide sample generation, significantly improving sample diversity while ensuring precise categorical alignment. Finally, an adaptive dynamic diffusion training strategy is developed, dynamically modulating diffusion levels and adaptive loss functions to optimize both rapid global feature capture and accurate local modeling. Experimental validations on three representative industrial case studies demonstrate that ACDGAN achieves superior performance in generating high-fidelity vibration signals with robust category consistency and training stability, especially under limited samples and high-noise conditions. The proposed method thus offers a novel and effective solution for intelligent maintenance in industrial applications.