A Deep Transfer Learning Approach to Few-Shot Fault Diagnosis in Underwater Manipulators

This paper presents a deep transfer learning-based approach for diagnosing multiplicative faults in underwater manipulators using limited operational data. Given the limited availability of data, transfer learning is utilized to enhance model performance. A pre-trained model from conventional manipulators is adapted to the underwater domain through model-based transfer learning. The convolutional neural networks (CNNs) and long short-term memory (LSTM) networks are used to retain both local and temporal fault characteristics, improving fault feature extraction. The source domain model is fine-tuned using a small sample dataset from the target domain, where lower layers are frozen and the top layers are fine-tuned for fault diagnosis, achieving improved accuracy. The results from the case study demonstrate that the proposed approach delivers high accuracy in diagnosing actuator faults in underwater manipulators.