Few-shot cross-domain fault diagnosis via adversarial meta-learning

This paper introduces a fault diagnosis approach, MLAML, which reconstructs data reconstruction, meta-learning, and adversarial learning for cross-domain fault diagnosis in small-sample scenarios. To enhance signal quality, an improved sparse denoising autoencoder regularized by maximum mean discrepancy is proposed within the data reconstruction stage, effectively suppressing noise and retaining critical information. Following this, discriminative fault features are captured through a lightweight multi-scale feature extraction module, while a meta-learning network facilitates transfer learning across domains with limited labeled samples. Adversarial learning is used for domain adaptation, reducing pseudo-label noise, and improving diagnostic accuracy. Experiments on two-bearing datasets show that MLAML outperforms traditional methods, achieving superior fault diagnosis accuracy even with minimal labeled data, demonstrating its robustness and efficiency in cross-domain applications.