Reliable fault diagnosis of bearings with varying rotational speeds using envelope spectrum and convolution neural networks

Determining the optimal features that are invariant under changes in the rotational speed variations of rolling element bearings is a challenging task. To address this issue, this paper proposes an acoustic emission (AE) analysis-based bearing fault diagnosis invariant under fluctuations of rotational speeds using envelope spectrums (ES) and a convolutional neural network (CNN). The ES extracted from the raw AE signals provides valuable information about the characteristic defect frequency peaks and variations to bearing rotational speeds when faults appear on a bearing. The proposed method employs CNN to automatically extract high quality features and classify bearing defects. In the experiment, a CNN trained on a dataset corresponding to one revolutions per minute (RPM) is used to detect patterns from datasets corresponding to other RPMs to verify that the classification is accurate and invariant under rotation speed fluctuations. The efficacy of the proposed method is verified on AE-based low-speed bearing data under various rotational speeds. The experimental results show that the proposed method is effective at detecting bearing failures, provides an average classification accuracy of about 86% under fluctuating RPM, and outperforms other state-of-the-art algorithms.