MGMD: A novel cross-working remaining useful life prediction method for rolling bearings under small-sample conditions

To address heavy data reliance and poor cross-working condition generalization of rolling bearing remaining useful life (RUL) prediction models under small-sample scenarios, this paper proposes a novel method integrating meta-learning, parallel temporal modeling and the domain adversarial mechanism. Firstly, feature extraction is performed on vibration signals of rolling bearings under multiple working conditions to obtain a feature set including time domain, frequency domain, and trigonometric features. Then, a novel prediction network named MGMD under the Model-Agnostic Meta-Learning (MAML) framework is constructed. Through the parallel structure of Gated Recurrent Unit (GRU) and Mamba, the local dynamic changes and long-range acceleration trends of the degradation process are captured, respectively, forming comprehensive features with both details and global information. Domain Adversarial Neural Networks (DANN) are introduced to alleviate cross-working condition distribution shift, enabling efficient cross-domain adaptation under small-sample conditions. The initial parameters of the network are optimized by means of the inner-loop and outer-loop meta-update mechanisms of MAML to improve small-sample rapid adaptation ability. Experimental verification on the IEEE PHM 2012 bearing dataset shows that the proposed method achieves a significant improvement in prediction accuracy. It provides a feasible solution for cross-working RUL prediction of rolling bearings under small-sample conditions.