Özet:

Knowledge-based fault diagnosis methods have become more preferred as they do not need precise model and signal patterns required in model-based and signal-based diagnosis methods, respectively. Machine learning techniques provide remarkable results on fault diagnosis by mapping information from raw signals to health condition. However, their vulnerabilities against adversarial attacks arise as in the other industrial applications employing machine learning methods. In this study, the vibration signals containing 10 different health condition in the public CWRU bearing health condition dataset are converted into 2D images and Fast Gradient Sign Method (FGSM), Basic Iterative Method (BIM), Projected Gradient Descent (PGD) and Carlini and Wagner (CW) white box adversarial attacks are applied into the deep residual learning (DRL) network model which classifies the images of rolling bearing. The robustness of the DRL model is analyzed to examine the effect of the implemented adversarial machine learning attacks. According to the obtained results, the adversarial attacks fooled the DRL model, causing it to produce misclassification results and so decrease the bearing fault diagnosis classification accuracy. As a result of injecting quite small perturbation to 2D images, the classification accuracy, which was 99.98%, is decreased to 68.38%, 61.75%, 61.88% and 63.31% by FGSM, BIM, PGD and CW attack methods, respectively. The achieved results show that the adversarial machine learning attack methods have great potential to reduce the accuracy of bearing fault diagnosis classification.

Anahtar Kelimeler: