Semi-supervised fault diagnosis method based on dynamic data extension

The accuracy of traditional deep learning models may be affected by labeled sample size. It is significant to design a semi-supervised learning method to make full use of massive unlabeled samples when there is limited data and the data quality is poor due to noise. However, traditional semi-supervised learning method is unable to guarantee the effectiveness of model since it only focuses on increasing labeled sample size but sample dimension, which also has an important impact to performance of model. This paper proposes a semi-supervised learning method based on dynamic extension not only can increase sample size but also increase sample dimension to improve the effectiveness of model. Dynamic extension is used to increase labeled sample dimension, and high-dimensional samples can use large models with stronger feature extraction capabilities to serve semi-supervised learning, which can effectively extract weak fault features submerged by noise. Experimental validation for benchmark dataset of rolling bearing shows that 30.3% improvement of fault diagnosis accuracy can be achieved in the case when only few labeled data with poor quality is available.