A knowledge fusion-driven domain generalization diagnosis framework for unseen operating conditions

In engineering applications, deep learning-based fault diagnosis must balance generalization and model complexity amid dynamic conditions. High generalization often demands complex models, yet excessive complexity reduces computational efficiency and hinders deployment. To address this, we propose Multi-knowledge Domain Generalization with Dynamic Temperature Distillation (MDG-DTD), a knowledge fusion-driven framework. First, we introduce a distillation method that fuses multiple specialized teacher models via confidence weighting and feature alignment to derive a generalized teacher model. Second, a dynamic temperature distillation strategy integrates attention transfer and adversarial temperature regulation to adaptively adjust the distillation process, optimizing student model learning and efficiency. Experimental results on the motor current and industrial robot vibration datasets validate MDG-DTD’s superiority over existing domain generalization methods in diagnosing faults under unseen conditions. Notably, the student model, using only 0.15% of the teacher model’s parameters and 0.36% of its floating point operations, retains at least 96.33% and 97.78% of its performance, respectively.