A Model-Agnostic Meta-Learning Fault Diagnosis Method Based on Dynamic Weighting

Motor is an important component in industrial production, so it is very important to detect and diagnose the motor fault in time. The performance of traditional deep learning methods is often limited by the amount of data, which leads to the problem of unreliable model performance in the face of few shot fault diagnosis problem. Meta-learning can use previous few shot learning experience to quickly guide a new few shot problem to build a reliable model. However, traditional meta-learning methods only focus on training the meta model by accumulating loss functions for different tasks, without considering the specific contribution from different tasks. Therefore, when some data is polluted by noise, the construction of meta model may be affected, thus affect the accuracy of fault diagnosis. Therefore, this paper aims to propose a meta-learning fault diagnosis method based on dynamic weighting to overcome the above problems. Different weights are assigned according to the size of the training tasks’ loss function, and smaller weights are assigned to tasks with larger loss functions, so as to realize dynamic weakening of the negative effects of unreliable tasks and build a meta model that can quickly adapt to powerful fault diagnosis performance. Experimental verification for motor fault diagnosis shows that the fault diagnosis accuracy of this method is improved significantly.