Least decision information-based dual-anchor framework for cross-domain fault diagnosis

In this study, a new concept termed ε -least decision information ( ε -LDI) is proposed to characterize the essential decision-making information of cross-domain diagnostic models. By focusing on the minimum information required for prediction, ε -LDI provides a novel perspective on domain-invariant representation learning and offers a promising approach to improving model generalization. The theoretical analysis demonstrates that the lower bound of ε -LDI can be estimated using a method inspired by the probability of necessity and sufficiency (PNS). Motivated by this finding, a dual-anchor framework (DAF) is developed to extract the minimum decision information required for accurate classification. Within DAF, features are first decomposed into positive and negative components based on the characteristics of SoftMax and the formulation of PNS. The positive components are further selected according to a relaxed ε -LDI criterion, and their statistical moments are aligned through a moment-matching mechanism, whereas the negative components are directly aligned with zero. Moreover, a negative-direction weight perturbation mechanism is designed to emulate adverse feature variations by introducing perturbations into the classifier parameters, thereby enhancing the robustness and generalization capability of the model. Extensive experiments conducted on three datasets validate the effectiveness and superiority of the proposed method.