Task-adaptive learning with gradient weighting for bearing fault identification under small sample conditions

Deep learning has demonstrated remarkable success in fault identification tasks; however, its performance often degrades under small sample conditions due to its dependence on large-scale labeled data. Meta-learning (defined as a learn-to-learn paradigm) offers a promising solution for small sample learning, yet it is challenged by task imbalance caused by disparities in task importance, complexity, and quantity, which impairs generalization. In this paper, a meta-gradient weighting meta-learning (MGWML) approach is proposed for small sample rolling bearing fault diagnosis. A meta-optimization framework is proposed to dynamically adjust task weights using meta-gradient information, with a dual-loop optimization strategy implemented to enhance convergence and adaptability across diverse task distributions. Concurrently, a base learner based on parallel attention is constructed to extract both local features and global dependencies from vibration signals. The effectiveness and robustness of MGWML is validated on two bearing fault datasets. In cross-working condition fault diagnosis, MGWML achieves a maximum accuracy of 99 %, and it maintains over 90 % accuracy under 5 dB noise interference.