A Robust Feature Extraction for Automatic Fault Diagnosis of Rolling Bearings Using Vibration Signals

Bearing faults are one of the main reasons for rotary machine failure. Monitoring bearing vibration signals is an effective method for diagnosing faults and preventing catastrophic failures in rotary mechanisms. The state-of-the-art vibration monitoring algorithms are mainly based on frequency or time-frequency domain analysis of rotary machines that are operating in steady state. However, the steady state assumption is not valid in applications where the loads and speeds are time-varying. Finding a method for capturing the variability in vibration signals, which are caused by varying loads and speeds, is still an open research problem with potentially many applications in emerging areas such as electric vehicles. In this paper, we address the problem of vibration signal monitoring by applying a feature extraction algorithm to rotary machine signals measured by accelerometers. The proposed method, which is based on the wavelet scattering transform, achieves overall high accuracy while being computationally affordable for real-time implementation purposes. In order to verify the effectiveness of the proposed methodology, we apply our technique to a well-known vibration benchmark dataset with variable load. Our algorithm can diagnose various faults with different intensities with an average accuracy of 99% and thus effectively outperforming all prior reported work on this dataset. INTRODUCTION MOTIVATION A rolling bearing carries load by placing rolling elements such as balls or rollers between two adjacent bearing rings, which are also known as races. The relative motion of the races causes the rolling element to roll with low friction. Due to the low friction, rolling bearings are among the most frequently used components in rotary machines. On the other hand, a fault in bearings can result in catastrophic rotary machine failures. Accordingly, real-time monitoring of bearings for safety critical applications such as transportation and aviation is of crucial importance. PRIOR WORK Prior fault diagnosis algorithms fall within two broad categories, namely, model-based and data-driven approaches. Model-based fault diagnosis methods, when there exists a precise knowledge of the underlying governing dynamics of the system, generate reliable results [1–3]. However, accurate physical modeling of rotary machines is a very challenging and demanding task. Furthermore, model-based fault diagnosis algorithms are difficult to generalize to systems with different models [4, 5]. On the other hand, data-driven fault diagnosis methods are considered as pattern recognition problems. In the data-driven approaches, signal processing techniques are applied to the measured vibration signals in order to extract important system features. The health state of the machine is then diagnosed by classifying the extracted features [6, 7]. The key difference between prior data-driven methods is Proceedings of the ASME 2017 Dynamic Systems and Control Conference DSCC2017 October 11-13, 2017, Tysons, Virginia, USA