Adaptive diagnosis method based on gearbox unbalanced fault data

Abstract

ObjectiveThe existing intelligent fault diagnosis methods face challenges, such as model training relying on a large amount of labeled data, difficulty in obtaining fault data with different occurrence probabilities, and insufficient consideration of the impact of operating conditions. To address these challenges, a novel gearbox diagnosis method for adaptive inter-class and intra-class unbalanced fault data under varying working conditions was proposed.MethodsFirstly, a gated local connection network was utilized to reduce the reliance on the labeled data and extract intrinsic features directly from the original data. Secondly, a parallel mechanism of external and internal attention was designed to consider the distribution differences among inter-class and intra-class faults under different working conditions, adjusting the weights of extracted features accordingly. Finally, focal loss function was employed to focus on minority and challenging samples, enabling high-quality mining of unbalanced diagnostic information.ResultsThe proposed method is demonstrated by six unbalanced gearbox datasets, which shows great effectiveness and superiority in identifying unbalanced fault data.

Keywords

• Fault diagnosis
• Inter-class and intra-class imbalances
• Gated local connection network
• Attention parallel mechanism
• Focal loss