Rolling Bearing Fault Diagnosis Research Based On Deep Few Shot Learning

With the rapid development of modern industrialization,many large-scale machinery have become increasingly precise and complex.Rolling bearings,as indispensable components of machinery,play a crucial role.However,due to heavy workloads,failures often occur,which can result in incalculable economic losses and personnel injuries.Therefore,accurate and effective fault diagnosis is crucial for avoiding personnel injuries and ensuring work efficiency.In recent years,with the continuous development of deep learning technology,new ideas and methods have been brought to the field of fault diagnosis.However,due to the low frequency of failure occurrence and the difficulty of signal acquisition,how to achieve accurate diagnosis in small sample and zero sample situations has become a research challenge.This paper mainly studies the bearing fault diagnosis method based on few shot learning and generalized zero sample learning.First of all,in view of the limited data caused by the lack of a large number of labeled data in practical engineering applications and considering that only a small number of samples in each category in the fault data are likely to cause over-fitting problems,a metricbased feature extraction few shot fault diagnosis is proposed.method The method creatively combines the relative similarity information of sample groups with the supervision information of each specific category in the labeled source data.And to support this combination,a hybrid training strategy with global supervised training and episodic training in feature space is designed to solve the fault diagnosis problem.This method can achieve accurate fault diagnosis with only a small number of samples in each category.Secondly,considering that the lack of labeled data will hinder the development of deep learning in the field of fault diagnosis and collecting and labeling data is time-consuming and expensive,a small-sample fault diagnosis method based on feature refinement is proposed.In this method,unlabeled samples are used to optimize and adjust the class prototypes generated by labeled samples,and a combinatorial optimizer is designed to optimize the model effectively.This method can make full use of unlabeled samples,and has high fault diagnosis accuracy,and has strong adaptability to different situations.Finally,in order to solve the problem that the failure samples under all working conditions cannot be obtained in actual engineering,resulting in the lack of specific training data,resulting in unsatisfactory test performance,a fault diagnosis method based on generalized zero-shot learning is proposed.This method uses the relationship between the semantic attributes and non-semantic attributes of visible class samples,uses the semantic attributes of invisible class samples to infer the non-semantic attributes of invisible classes,and then combines the semantic attributes and non-semantic attributes into dual attributes to generate different visible class samples.This method achieves accurate classification of unseen samples and visible samples.The model methods proposed in this paper are verified on multiple public datasets,and relevant ablation experiments are conducted on the main parameters of the model.The effectiveness of the proposed method is verified through comparison with commonly used methods.