MHFCF-HSKD: noise-resistant lightweight fault diagnosis based on multi-modal heterogeneous feature collaborative fusion and hierarchical structured knowledge distillation

To address poor diagnostic performance of existing gear fault diagnosis approach in strong noise environment, insufficient multiscale feature fusion of multisensor information, a conflict between massive parameters/calculations of deep model and the computing resources of equipment, a novel noise-resistant lightweight fault diagnosis based on multi-modal heterogeneous feature collaborative fusion and hierarchical structured knowledge distillation network (MHFCF-HSKD) is proposed. First, multi-dimensional data is normalised by MHFCF, and the designed SPCA module is used for feature fusion, thereby transferring 1D signals into 2D RGB images. And then the improved S2CGNet framework is applied to enhance model noise resistance capability, and the dynamic loss factor balance mechanism in different dimensions enables the lightweight Student model to efficiently learn the knowledge from the well-learned Teacher network. Whereafter, numerous experiments are carried out on the self-built NEUG dataset and public SEU dataset, which show that the diagnostic performance of the proposed method is significantly improved with lower parameters compared to other approaches. The parameter count of Student network reduces by approximately 6 times of the Teacher network while maintaining a similar level of accuracy, which further demonstrates the effectiveness, robustness and generalisation ability of the improved MHFCF-HSKD network.