Deep learning-based fault diagnosis of planetary gearbox: A systematic review

Abstract

Planetary gearboxes are popular in many industrial applications due to their compactness and higher transmission ratios. With recent developments in the area of machine learning, Deep Learning-based Fault Diagnosis (DLFD) has become the preferred approach over traditional signal processing methods, physics-based models, and shallow machine learning techniques. This paper presents a systematic review that identifies key research questions for fault types, datasets used, challenges addressed, approaches applied to address the challenges and comparison of the methods using diagnosis accuracies, computation load, and model complexity. The review highlights that the researchers have focused on several challenges, including fault diagnosis under varying operating conditions, imbalanced data, noisy data, limited labeled fault samples, and zero faulty samples. To address these issues various methods have been proposed in the literature, such as incorporating signal processing, data augmentation, transfer learning using domain adaptation, adversarial learning, and integrating physics-based models. Enhancing the industrial applicability of DLFD methods requires validating these methods under multi-problem scenarios, improving transfer learning accuracy for cross-machine fault diagnosis, enhancing interpretability and trust, optimizing for lightweight implementation, and utilizing industrial datasets. Addressing these areas will enable DLFD methods to achieve greater reliability and wider adoption in industrial maintenance practices.