Multimodal Adaptive Signal Fusion for Domain Generalization in Imbalanced Few-Shot Rotating Machinery Fault Diagnosis

Domain generalization-based fault diagnosis has emerged as a promising approach for addressing cross-domain challenges in rotating machinery condition monitoring. However, practical industrial scenarios are characterized by severe class imbalance and limited fault samples. These factors significantly impede the generalization capability of diagnostic models across different operational conditions. Therefore, this article proposes a Multimodal Adaptive Signal Fusion for Domain generalization (MASFD) framework for imbalanced few-shot fault diagnosis in rotating machinery. The framework incorporates an adaptive signal mode decomposer for frequency-specific signal decomposition. Lightweight parallel mode enhancers are employed for efficient multimodal feature extraction. An adaptive fusion module with dynamic weighting mechanisms is integrated to combine multimodal features. A domain separation network explicitly disentangles domain-invariant features from operational variations. Fast meta-learning enables rapid adaptation to unseen working conditions through episodic training strategies. Extensive experiments are conducted on two benchmark datasets under three imbalance ratios, achieving average accuracies of 80.64%, 73.25%, and 65.85%, outperforming baseline methods by 3-15%. The proposed framework demonstrates superior generalization capability and computational efficiency, providing a practical solution for real-world industrial fault diagnosis under challenging imbalanced few-shot scenarios.