Asymmetric convolution and multi-head self-attention based meta-transfer learning network for fault diagnosis of underwater thrusters under few-shot and multi-condition scenarios

With the continuous advancement of ocean resource exploration and exploitation, the reliable operation of underwater thrusters is crucial for operational safety, making their fault diagnosis capability a key factor in ensuring system effectiveness. However, most existing few-shot fault diagnosis methods are confined to single operating conditions, whereas underwater thrusters operate under multi-condition environments. This complexity causes identical faults to exhibit distinct characteristics across conditions, leading single-condition models to generalize poorly to new operating scenarios. To address this challenge of few-shot fault diagnosis under multi-operational conditions, this paper proposes a Meta-Transfer Learning Network with Asymmetric Convolution and Multi-Head Self-Attention (AC-MHSA-MTL) for underwater thrusters. Leveraging meta-learning's key advantages for few-shot adaptation, the method constructs a feature extraction network. Asymmetric convolution is introduced to overcome the scale limitation inherent in standard convolutional kernels, while multi-head self-attention is employed to bolster the model's ability to discern extended-range relationships in time-series signals. Furthermore, an optimized metric learner is designed to improve the flexibility of similarity assessment. After training, the feature extractor and metric learner are frozen and transferred to the target domain for accurate fault diagnosis in new operating conditions. Finally, the effectiveness is validated with a multi-condition underwater thruster fault dataset.