Deep learning-based detection and severity classification of distributed pitting faults in helical gearboxes

Abstract

Gears, which play a vital role in the transmission of power and torque between shafts in industrial machinery, can be susceptible to faults such as pitting, cracking, wear, and corrosion, caused by harsh operating conditions. Among these faults, distributed pitting, characterized by the formation of small pits on all surfaces of the teeth, is more difficult to detect. The early detection and severity assessment of such faults are crucial in preventing further damage to machinery. Deep learning methods, particularly autoencoders (AEs) and convolutional neural networks (CNNs), have demonstrated significant potential for detecting gear faults in vibration signals. This study explores the use of variational AEs (VAEs) as feature extractors to classify the severity of distributed pitting faults in helical gears. For this purpose, experiments were conducted using a test rig equipped with a two-stage industrial helical gearbox under four distinct load conditions, employing healthy gears and gears with varying pitting severities. Vibration signals, captured by accelerometers, were combined into one-dimensional time-series data and subsequently transformed into two-dimensional representations using two approaches to leverage image classification capabilities of deep learning methods. Various configurations of CNN, AE, and VAE models were evaluated for fault severity classification using different input modalities. Of all the methods tested, VAE combined with a classifier performed best in almost all load conditions. In addition, because not all fault types may be represented during model development, the study also examined the behavior of the best-performing models when one fault class was excluded from training. This additional evaluation was used to assess the generalization to unseen fault classes under incomplete fault coverage, and the VAE–classifier combination showed more reliable behavior than the competing alternatives within this evaluation setting.