Lightweight cost-sensitive multi-expert dual knowledge transfer network for imbalanced fault diagnosis

Data imbalance remains a persistent challenge in industrial intelligent fault diagnosis. Traditional diagnostic models often favor majority classes, while lacking sufficient recognition capabilities and predictive confidence for minority classes. Current solutions for data imbalance often adopt isolated strategies. They often fail to transfer rich diagnostic knowledge from the majority class to the minority class, resulting in only marginal improvements in minority class performance. Furthermore, they cannot sufficiently improve prediction confidence for classes prone to confusion. Therefore, this paper proposes a lightweight cost-sensitive multi-expert dual knowledge transfer network that effectively combines ensemble learning and a cost-sensitive mechanism. Specifically, it combines ensemble learning with multi-depth diagnostic knowledge and an improved dual knowledge distillation method to achieve the integration and transfer of imbalanced diagnostic knowledge. It effectively suppresses confusing classes, significantly improving the knowledge acquisition ability and prediction confidence of minority classes. The adaptive cost-sensitive loss dynamically adjusts the classification weights based on the posterior probability, enhancing focus on minority classes. Additionally, the framework incorporates multiple lightweight network designs, including multi-scale feature fusion, ghost transformation operations, and spatial and channel refinement convolution, to achieve robust knowledge encoding at extremely low computational costs. Extensive comparison and ablation experiments on the gearbox and two hydraulic pump datasets confirm the framework's superior performance across diverse imbalance types and ratios, offering a highly promising solution for industrial imbalanced fault diagnosis.