Variable Speed Bearing Anomaly Detection via Order Tracking and Cascaded Memory-Augmented Autoencoder

Bearings are essential and crucial components of rotating machinery. Bearing faults can jeopardize the safe operation of mechanical equipment and diminish production efficiency. Anomaly detection (AD) has been recognized as a solution, but its effectiveness is still hindered by lack of abnormal samples and uncertainties resulted from variable speed. In this study, an AD scheme is proposed for bearings under variable speed conditions. First, order envelope signals are extracted via angular domain resampling and demodulation to obtain more representative vibration features of variable speed bearings. A cascaded memory-augmented autoencoder (CMAE) is designed to memorize multiscale normal patterns for reconstruction and prevent the decoder from overgeneralizing. By this means, sensitivity to anomalies can be enhanced, even in the absence of sufficient abnormal samples. In turn, anomalies can be detected more efficiently and accurately. Finally, extensive experimental results validate the superiority and applicability of the proposed scheme.