Session: 15-03-02: Machine Learning and Deep Learning in Safety, Reliability, and Maintenance II

Paper Number: 166857

A CNN-LSTM-Based Approach for Fault Diagnosis in Asymmetric Gears Under Random Speed Variation 

Gearboxes are widely utilized in strategic applications, including wind energy systems, aero engines, military vehicles, and rail transportation. However, harsh operational conditions, material defects, high temperatures, and inadequate lubrication often lead to gear failures, such as surface wear, pitting, and cracks. Among these, the latter is typically responsible for massive damage, as cracks account for a more rapid failure (tooth broken) than pitting and wear. The gearbox condition monitoring task is occasionally carried out with the help of vibration measurements, as the manifestation of a crack would noticeably affect the signal’s normal spectrum. In this respect, comprehending the meshing and resonant modulations requires substantial field experience in engineering practice regarding signal processing and gear dynamics, which is far from providing a readable rendition for average users. To address such drawbacks, deep learning has been introduced and has become increasingly popular in equipment fault diagnostics. A gearbox monitoring algorithm should diagnose different crack levels using a segment of vibration data measured under random speed variations (RSVs) since the vibration pattern continuously changes if the time-varying operating conditions are considered. A few studies have been reported on this matter. Still, some either reported lower classification accuracies or made the assumption that the input shaft speed does not change noticeably within a rotational cycle, failing to present an extensive analysis covering both RSV and tooth cracks. What is more, these studies all focus on gear pairs with symmetric tooth profiles. The evidence in the available literature proves that gear designs with asymmetrical teeth have the potential to provide crucial advantages compared to their symmetric counterparts, including fatigue life and mesh stiffness. From this standpoint, the present research developed a six-degree-of-freedom dynamic model of a single-stage gearbox to simulate its vibration characteristics in the presence of tooth root cracks (50%-100%) for gear pairs with symmetric (20°/20°) and asymmetric (20°/30°) teeth under RSV. To this end, the input shaft speed was set to 10 Hz with a variation of ± 20%. Further, a methodology that combines a one-dimensional convolutional neural network (1-D CNN) and long short-term memory (LSTM) was adopted to solve a multi-stage classification problem. The performance of the developed 1-D CNN- LSTM model was then optimized through the Hilbert-Schmidt independence criterion-based global sensitivity analysis. In this regard, an attempt was also made to reduce the epistemic uncertainty that arises due to not knowing the best hyper-parameter configuration to enhance the model’s explainability, interpretability, and trustworthiness. The primary consideration was understanding whether using a gear tooth with an asymmetric profile could enhance the classification accuracy of the proposed 1-D CNN-LSTM model. In the second portion of this study, we benefited from signal segmentation in diagnosing gearbox's health status independent of the number of faulty teeth and fault location. To this end, the simulated signals were divided into multiple segments based on the interval of gear teeth. Further, a new vibration dataset was created by randomly varying the location of the crack, considering it may occur in any gear teeth in practice. Ultimately, the proposed 1-D CNN- LSTM architecture was trained with the original data but tested using artificially created data. A signal-to-noise ratio of 5 was added to signals to complicate the fault diagnosis task and approximate reality for all cases evaluated within the scope of this work. The performance of the proposed approach was compared with those of traditional CNN and LSTM. The findings revealed that using pairs of gear with asymmetric teeth could facilitate crack diagnosis under RSV with the help of a machine learning-based approach. 

Presenting Author: Onur Can Kalay Texas Tech Univeristy

Presenting Author Biography: Dr. Onur Can Kalay is a mechanical engineer and researcher specializing in power transmission, dynamic modeling, deep learning applications in fault diagnosis, fatigue analysis, and uncertainty quantification. He has extensive experience in gear dynamics, particularly in the design and analysis of asymmetric gears, and has worked on experimental and computational methods for condition monitoring in mechanical systems.

Dr. Kalay received his Ph.D. from Bursa Uludağ University, where he contributed to multiple TÜBİTAK, TUSAŞ, and industry-funded projects focused on gear fatigue, machine learning-based fault detection, and hybrid material applications. His research led to novel findings on the impact of asymmetric tooth profiles on fatigue life, impact resistance, and AI-based diagnostics.

Currently, Dr. Kalay is a Postdoctoral Research Associate at Texas Tech University, where he expands his expertise in probabilistic design, remaining useful life prediction, Bayesian optimization, and sensitivity analysis. His recent collaborations with TU Dortmund involve developing advanced machine learning models for predictive maintenance and reliability engineering.

Dr. Kalay has published over 23 journal articles, 25 conference papers, and a book chapter. His research has been recognized with best paper awards, and he actively collaborates with multidisciplinary teams from engineering, biomechanics, and applied mathematics across multiple institutions. His future work focuses on AI-driven diagnostics, digital twin modeling, and novel gear materials for aerospace and automotive applications.

Authors:

Onur Can Kalay Texas Tech Univeristy
Fatih Karpat Bursa Uludag University
Stephen Ekwaro-Osire Texas Tech University