Session: 14-06-01: Machine Learning and Deep Learning in Safety, Reliability, and Maintenance I

Paper Number: 144598

144598 - Gearbox Fault Diagnosis With Deep Learning Under Variable Operating Conditions 

The gearboxes are indispensable universal components to transfer power and handle precision functionalities in modern factories and have been broadly used in military and industrial services. According to a survey, around 60% of transmission system faults are caused by individual gear errors. If exemplified on a sector basis, the literature also confirms that 19.1% of helicopter powertrain system malfunctions originate from gear failures. Typically, the gear failure modes encountered within industrial applications include pitting, tooth root cracks, surface wear, and spalls. Among these failure modes, diagnosing tooth cracks is of primary importance in terms of health management, as cracks may propagate along the rim or tooth and cause rapid failure. However, in most cases, sensor data being overwhelmed by random interference noise could also lead to subduing the characteristic frequencies in the vibrational signal. From the engineering point of view, a machine learning-based method that can automatically extract fault features from the vibrational signals would be precious since failing in early diagnosis of root cracks may result in a tooth being broken rapidly. The challenges of the state-of-the-art artificial intelligence algorithms are no longer just reaching higher classification accuracies with less training but are now focused on reducing the computational burden and model complexity due to a large amount of raw sensor data. In this regard, deep learning (DL) is increasingly popular due to its modeling and representational capabilities. From this standpoint, the present experimental research study developed a convolutional neural network-based method for diagnosing different levels of tooth root cracks (50%-100%) for spur gears. A series of vibration experiments were performed on a single-stage spur gearbox by using tri-axial accelerometers under variable working conditions. The effects of different levels of tooth cracks on vibration amplitudes were also investigated within experiments. Firstly, the location where maximum stress occurred at the tooth root and the crack propagation paths were determined by employing finite element analysis. Afterward, different levels of tooth root cracks were intentionally seeded to gear test samples with wire-cut electric discharge machining. The obtained findings were then evaluated and interpreted in time and frequency domains. To this end, the Fourier transform was applied to the time-sequence acceleration data in the time domain. To address the effect of sensor location on the vibration responses of healthy and cracked spur gear pairs comparatively, a total of six sensor locations were determined and tested in the research work. The primary consideration of the present study is to test the influence of (1) shaft speed, (2) working load, and (3) measurement axis on the overall classification accuracy of the proposed DL-based model. Within the vibration experiments, the influence of two shaft speeds (300-600 rpm) and three loading conditions (no-load, 10%, and 20%) on the vibration amplitudes and fault recognition accuracy were comparatively evaluated. The experimental findings demonstrated that the developed DL-based algorithm classified different crack levels under variable operating conditions between the overall accuracies of 95.015% and 99.955%

Presenting Author: Onur Can Kalay Texas Tech University

Presenting Author Biography: He received Ph.D. in 2023 at Bursa Uludag University

Authors:

Onur Can Kalay Texas Tech University
Fatih Karpat Bursa Uludag University
Esin Karpat Bursa Uludag University
Ahmet Emir Dirik Bursa Uludag >University
Stephen Ekwaro-Osire Texas Tech University