Model-Agnostic Meta-Learning for Fault Diagnosis of Industrial Robots

The success of deep learning in the field of fault diagnosis depends on a large number of training data, but it is a challenge to achieve fault diagnosis of multi-axis industrial robots in the case of few-shot. To address this issue, this paper proposes a method called Model-Agnostic Meta-Learning (MAML) for fault diagnosis of industrial robots. Its goal is to train an effective industrial robot fault classifier using minimal training data. Additionally, it can learn to recognize faults in new scenarios with high accuracy based on the training data. Experimental results based on a six-axis industrial robot dataset show that the proposed method is superior to traditional convolutional neural network (CNN) and transfer learning, and that the diagnostic results with the same amount of data in few-shot cases are better than existing intelligent fault diagnosis methods.