Fault Diagnosis of Rotating Machinery Based on Improved Self-Supervised Learning Method and Very Few Labeled Samples

Area of Science:

• Machine Learning
• Artificial Intelligence
• Signal Processing

Background:

• Convolutional Neural Network (CNN) methods excel at feature extraction for machine fault diagnosis but require extensive labeled data, risking overfitting with limited samples.
• Existing methods struggle with insufficient labeled data, hindering accurate fault mode classification in real-world scenarios.

Purpose of the Study:

• To develop a novel ResNet-based fault diagnosis method (DTC-SimCLR) capable of achieving high accuracy with minimal labeled samples.
• To address the overfitting issue in Convolutional Neural Network (CNN) models when training data is scarce.

Main Methods:

• A novel method, DTC-SimCLR, is proposed, combining data transformation combinations (DTCs) designed via mutual information with a self-supervised learning approach (1-D SimCLR) for feature encoder optimization.
• The method utilizes a 1-D ResNet architecture and a fully-connected layer classifier, where the feature encoder is pre-trained using unlabeled data and its parameters are fixed during classifier training.
• DTCs are randomly applied to batch training data, enabling rapid training of the 1-D ResNet without augmenting the dataset size.

Main Results:

• The DTC-SimCLR model demonstrated superior performance and diagnostic accuracy on cutting tooth and bearing fault datasets, even with a very limited number of labeled samples.
• The self-supervised learning component (1-D SimCLR) effectively optimized the feature encoder using unlabeled data, enhancing the model's generalization capabilities.
• The integration of DTCs facilitated quick training of the ResNet model without necessitating an increase in labeled training data.

Conclusions:

• DTC-SimCLR offers a highly effective solution for machine fault diagnosis in low-data regimes, significantly outperforming traditional methods.
• The proposed approach successfully mitigates the need for large labeled datasets, making advanced fault diagnosis more accessible and practical.
• This research highlights the potential of combining data augmentation strategies with self-supervised learning for robust machine diagnostics.