Rolling bearing impact feature extraction based on masked self-supervised learning
The existing mechanical fault intelligent diagnosis methods generally require a large number of reliable samples for model training.However,in application scenarios,the labeled training data are scarce.To address the problem,this paper proposes a method for extracting local impact fault feature of rolling bearings based on masked self-supervised learning.It employs random mask to perform Boolean operations on the original signal of the faulty bearings,and obtains samples for self-supervised feature extraction training.Then,the masked signal is entered into the masked-self-supervised learning network,based on loss function including the difference between the input and output kurtosis of the network and the random mask self-supervised learning,and the ability of the network to extract impact fault feature from the original fault signal is obtained.Our simulation signal analysis indicates the proposed method rebuilds the impulse sequences in the original signal with a reconstruction accuracy of 96.68%when the mask occlusion ratio is 95%and the training rounds 250.The rolling bearing fault experiments further show the proposed method effectively extracts the fault impact sequences from noisy signals without additional training data and has a huge potential for applications.On condition that the effects are superior to the best results of the other methods in comparison,the proposed method has a computing time of less than 20 seconds,far better than MCKD-type methods,demonstrating its huge application potentials.
国家科技期刊平台
NSTL
万方数据
