Intelligent compound fault diagnosis under imbalanced data: a dynamic sampling and hybrid attention based approach

Abstract Rotating machinery often operates under complex working conditions where compound faults frequently occur. However, diagnosing such compound faults under data imbalance scenarios remains a significant challenge in industrial applications. This study presents a groundbreaking computational model to tackle these challenges through three key innovations. The proposed framework improves diagnostic capabilities from data and model angles through three main innovations. Specifically, a unique dynamic class-aware sampler is proposed to optimize the data distribution, which effectively reduces model learning bias at its origin without the complications of synthetic data creation. Besides, a hybrid feature learning structure that combines the short-time Fourier transform with a convolutional block attention module augmented convolutional neural network is developed. This construction employs channel and spatial attention mechanisms to disentangle intricate signals and concentrate on essential time–frequency characteristics. A hierarchical multi-label classifier with a label-adaptive attention mechanism is designed to adaptively detect single and compound faults by learning the relations between heterogeneous fault embeddings. Comprehensive experiments on the publicly available bearing dataset reveal that the proposed model delivers exceptional diagnostic accuracy, particularly in conditions of high data imbalance, confirming its applicability in real-world industrial settings.