Few-Shot Cross-Domain Fault Diagnosis of Bearing Driven by Task-Supervised ANIL

Meta-learning has effectively addressed the limit of deep learning fault diagnosis models that demands a large number of samples. However, existing meta-learning models lack the capacity of feature reuse and task adaptability. To address the cross-domain fault diagnosis tasks with small samples, the feature reuse capability and task adaptability of existing meta-learning models need further improvements. To achieve this goal, this paper introduces a new approach built upon the task-supervised Almost No Inner Loop (ANIL). The proposed approach adopts a residual network to optimize the backbone structure of the inner loop, enhancing the feature reuse capability of the meta-learning in the unknown domain. An auxiliary term is introduced to define a supervised task-adaptive loss function, further updating the weight parameters of the inner loop meta-learner by monitoring the states of all meta-diagnostic tasks. The proposed method is used to analyze vibration signals from various bearings. The results demonstrate its superiority over traditional meta-learning methods in multiple sets of cross-domain fault diagnosis tasks with small samples.