The widespread use of machine learning models in fault diagnosis introduces vulnerabilities to adversarial attacks, which can compromise equipment operation and pose significant security risks. Current security methods often depend on specific attacks and exhibit limited adaptability in fault diagnosis applications. In this study, we propose a dual-branch feature decoupling-driven auxiliary network (DFDAN) to enhance the security of fault diagnosis systems. This assistive network operates independently of specific attack types and integrates seamlessly with existing fault diagnosis models without necessitating any modifications. Specifically, DFDAN decouples invariant and perturbed features through dual branches and employs a detection-then-purification approach to improve adversarial robustness while maintaining standard accuracy. We introduce the boundary adversarial example generation (BAEG) method and feature decoupling to efficiently extract and separate adversarial-related features from classification-related features. Additionally, we propose a feature frequency domain learning-based purification (FFDP) method to address the ill-posed problem of reconstructing one-dimensional vibration signals, thereby improving purification accuracy and generalization. Extensive experiments on public datasets demonstrate the effectiveness and superiority of DFDAN in enhancing fault diagnosis system security, providing a viable solution for adversarial defense in this field.