Adversarial Attack and Defense on Machinery Fault Diagnosis

In the field of mechanical fault diagnosis, deep learning models have demonstrated strong performance. However, limited focus has been directed towards the robustness of these models in resisting adversarial attacks. In this research, we concentrate on adversarial attacks aimed at deep learning-based machinery fault diagnosis systems, investigating both untargeted and targeted conditions. Experimental results reveal a significant reduction in the performances of typical deep learning-based models following adversarial attacks. Additionally, an effective defense scheme employing data normalization is proposed to protect victim models. This study provides a pioneering examination of adversarial examples in machinery vibration signals, contributing to the development of protection methods against adversarial attacks and the enhancement of deep learning-based model robustness.