Convolutional Neural Networks Based Rolling Bearing Fault Classification Under Variable Operating Conditions

Abstract

Rolling bearings are key machine elements used in various fields such as automotive, machinery, aviation, and wind turbines. Over time, faults may occur in bearings due to variable operating speeds and loads, contamination, etc., and this may cause a severe reduction in performance. In the future, an undetected bearing fault can lead to a fatal breakdown and substantial economic losses or even human casualties. Thus, bearing early fault diagnosis emerges as a critical and up-to-date topic. It is possible to obtain vibration, acoustic, motor current, etc., data that contain crucial diagnostics information regarding the health conditions of mechanical systems with various sensor technologies. With the era of big data, artificial intelligence (AI) algorithms have started to be utilized frequently in industrial applications. In this regard, convolutional neural networks (CNN) are increasingly popular with their capability to capture fault information without expert knowledge. This paper deals with a bearing fault diagnosis method based on one-dimensional convolutional neural networks (1D CNN) using vibration data. A multi-class classification problem was solved by examining different operating conditions for three health classes. Therefore, healthy state, inner raceway, and outer raceway faults were detected under variable operating speeds (900 and 1500 rpm) and loads (0.1 and 0.7 Nm). The effectiveness of the proposed 1D CNN method was evaluated with the Paderborn University (PU) dataset. As a result, rolling bearing early fault diagnosis was performed with an accuracy of 93.97%. It was observed that the proposed method was suitable for bearing fault diagnosis and can be utilized to optimize the rotary machinery maintenance costs by early fault detection.