Abstract
The accuracy of bearing fault diagnosis is of great significance for the reliable operation of rotating machinery. In recent years, increasing attention has been paid to intelligent fault diagnosis techniques based on deep learning. However, most of these methods are based on supervised learning with a large amount of labeled data, which is a challenge for industrial applications. To reduce the dependence on labeled data, a self-supervised joint learning (SSJL) fault diagnosis method based on three-channel vibration images is proposed. The method combines self-supervised learning with supervised learning, makes full use of unlabeled data to learn fault features, and further improves the feature recognition rate by transforming the data into three-channel vibration images. The validity of the method was verified using two typical data sets from a motor bearing. Experimental results show that this method has higher diagnostic accuracy for small quantities of labeled data and is superior to the existing methods.
Highlights
Three-Channel Vibration Images.Rotary machinery is widely used in helicopters, engines, turbines, and other mechanical equipment, and is a vital component in industrial applications
This multi-stage self-supervised learning method suffers from two problems: (1) it increases the complexity of the model, and (2) the pre-training features of the unlabeled data cannot be fully applied to the downstream tasks
A new fault diagnosis method based on self-supervised joint learning and threechannel vibration images is proposed; To improve the existing multi-stage self-supervised learning approach, an end-to-end self-supervised learning method is proposed that simplifies the model training process; Automatic fault feature extraction and learning with self-supervised learning can avoid the process of artificial feature extraction in traditional fault diagnosis methods, and improve the robustness of the model; A new method for constructing three-channel vibration images is proposed that has competitive performance compared with commonly used data processing methods; We combine self-supervised learning with fault diagnosis, and the obtained results are superior to those obtained with previous methods based on supervised learning and semi-supervised learning
Summary
The semi-supervised method combines the unsupervised and supervised learning techniques to train labeled and unlabeled data in a model, thereby improving the accuracy of recognition. Self-supervised learning is usually divided into two stages: the first stage is to learn the potential features of unlabeled data via auxiliary tasks, and the second stage is to transfer the features learned in the first stage to other downstream tasks, for example, image classification, target detection, and segmentation This multi-stage self-supervised learning method suffers from two problems: (1) it increases the complexity of the model, and (2) the pre-training features of the unlabeled data cannot be fully applied to the downstream tasks. A new method for constructing three-channel vibration images is proposed that has competitive performance compared with commonly used data processing methods; We combine self-supervised learning with fault diagnosis, and the obtained results are superior to those obtained with previous methods based on supervised learning and semi-supervised learning.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have