LFC-DGNet: A likelihood feature compositional domain generalization network from single-fault to unseen multi-component compound fault diagnosis across machines

Due to the scarcity of fault data in industrial machines, the domain generalization of fault knowledge from laboratory source machines to industrial target machines has become a promising research of intelligent diagnosis. However, the simplicity of laboratory machines structures and their fault mode limited to single components pose significant challenges in recognizing compound faults in industrial machine with multi-component structures. To address this challenge, a Likelihood Feature Compositional Domain Generalization Network (LFC-DGNet) is proposed, which extracts cross-machine invariant fault features through likelihood of features distribution and adaptively integrates fault knowledge from single-component features to achieve unseen compound fault diagnosis. A likelihood regularization loss is proposed to match the feature distributions of each class across different machines, capturing invariance and improving discriminability while limiting model overfitting. Additionally, to fuse knowledge from single components, a compositional feature perturbation module is proposed, which extends the unseen compound fault feature space through adaptive feature synthesizing. Under the constraint of single-fault modes in the source domain, LFC-DGNet achieves the average accuracy of 86.5%, 89.76% 84.32% and 90.31% testing in four different target machines, and demonstrate its superiority and effectiveness.