Interpretation of Deep Learning Models in Bearing Fault Diagnosis

Abstract

In recent years, data-driven techniques such as deep learning (DL) have been widely represented in the literature in the field of bearing vibration condition monitoring. While these approaches achieve excellent performance in classifying bearing faults on controlled laboratory data sets, there is little information available about their applicability to more realistic working conditions. As a first step towards revealing the generalizability of DL models, we aim to understand the underlying representations that DL networks use to classify bearing defects. An interpretable DL model can give us hints on how to increase its transferability by, e.g., using data augmentation, changing input representations and/or adapting model architectures. We use the Grad-CAM algorithm along with signal transformations to identify the elements of the input spectrogram that contribute to class attribution. The results show that removing time-domain information from the spectrogram has a minor impact on its performance. Instead, the network learns distinct average frequency profiles. We therefore conclude that the networks learn signal features very specific to the physical properties of the specific test setup, such as the frequency response function, rather than more general features related to bearing defects.