Cochlea-inspired spectral representation with bi-branch ensemble learning for single-source domain generalization motor fault diagnosis

Abstract Single-source motor fault diagnosis often degrades under unseen speed and load conditions because a single operating condition provides limited spectral diversity, whereas unconstrained augmentation may introduce non-fault-related artifacts. To address both limitations, we propose CoMel-Brain, which couple fault-sensitive spectral representation with controlled pseudo-domain learning. The CoMel representation maps cochlear-decomposed vibration components onto the Mel scale to emphasize low-frequency fault patterns. Based on this representation, a phase-preserving magnitude-spectrum perturbation expands spectral diversity while maintaining structural information. A guidance–exploration architecture further aligns augmented representations with the fault-related semantics of the original samples and stabilizes knowledge integration through periodic parameter averaging. Experiments on a seven-class induction-motor dataset covering 15 speed–load conditions yield average target-domain accuracies of 93.60%, 95.38%, and 84.87% for source domains A, H, and O, respectively, outperforming the strongest competing methods by 7.23, 3.66, and 6.47 percentage points.