Research On Meta Learning For Fault Diagnosis Methods With Limited Data

Intelligent fault diagnosis has become an important means to ensure the safe operation of equipment and prevent major disasters in the process of industrial production.However,the intelligent method relying on deep learning often needs massive data to predict a reliable model.It is usually difficult to collect sufficient failure data in practical industrial production,thus limits the application of intelligent diagnosis methods.How to solve the problem of fault diagnosis in few-shot scenario is still a key difficulty.Aiming at the problem,this work adopts meta learning method.There are two problems in how to apply meta learning to fault diagnosis and improve the adaptability of meta learning:(1)the random order of meta learning tasks may damage the knowledge generalization between adjacent tasks;(2)The small number of samples in meta tasks brings great fuzziness and uncertainty to the adaptation of meta knowledge in specific tasks.Therefore,focusing on the topic of few-shot fault diagnosis based on meta learning,this work proposes two algorithms from the aspects of task sequencing and uncertainty.The main contributions of this work can be summarized as follows.(1)This work proposes a task-sequencing meta learning algorithm for intelligent few-shot fault diagnosis with limited data.By sorting a large number of tasks in meta learning,the proposed method can avoid the damage of knowledge generalization of closely adjacent tasks.Therefore,it can learn a highly adaptive meta knowledge and have stronger generalization ability in practical industrial scenes.The proposed method uses the ability of meta learning,which can learn to learn,to find sufficiently sensitive and highly adaptive initialization parameters.Then it can quickly adapt to a new fewshot fault diagnosis task.Moreover,the model learns tasks from easy to difficult,so as to avoid the damage of generalization between tasks,improve the generalization ability of meta knowledge,and make meta learning perform better in fine-grained working conditions.(2)In this work,a novel meta learning method based on uncertainty for few-shot fault diagnosis is proposed.The uncertainty evaluation is introduced into the inner loop of meta learning to guide the optimization process.It can help prevent meta knowledge from being ill adapted because of the ambiguity of specific tasks.In the few-shot fault diagnosis,each task with a small number of samples has uncertainty.Considering this,the proposed method optimizes the inner loop of meta learning,and uses uncertainty to guide the first step gradient descent process of the meta knowledge.Therefore it can help meta learning be more robust and reliable in the inner loop.The inner loop is also fed back to the outer loop,and the inner parameters can help the outer layer learn more robust meta knowledge.