EFLightCaps: an efficient feature-focused lightweight capsule network with frequency-domain regularization for rotating machinery fault diagnosis

Abstract The prevalent issues of small samples and heavy noise in industrial settings have severely limited the application effectiveness of intelligent diagnostic methods. To address these challenges, an efficient feature-focused lightweight capsule network (EFLightCaps) is proposed. Firstly, a plug-and-play lightweight Ghost multiscale convolution (LGMC) block is designed to integrate multiscale convolution with gating mechanisms through a multi-branch architecture, enabling multi-scale feature fusion and enhancing the model’s noise robustness. Secondly, an efficient feature-focused routing mechanism is proposed to optimize information transfer between capsules through multi-step feature reconstruction and feature focusing strategies, which substantially enhances generalization capabilities in small-sample scenarios. Finally, a dual-norm frequency-domain regularization loss function is designed to leverage the complementary advantages of first-order and second-order norms, enhancing both feature extraction capability and reconstruction quality. Extensive experiments on two distinct datasets demonstrate that EFLightCaps achieves superior diagnostic performance and computational efficiency in scenarios with small samples and heavy noise.