Learning transferable features in deep convolutional neural networks for diagnosing unseen machine conditions

Area of Science:

• Artificial Intelligence
• Machine Learning
• Mechanical Engineering

Background:

• Deep learning, particularly CNNs, shows promise in mechanical fault diagnosis.
• Existing methods often fail due to distribution shifts between training and testing data in real-world scenarios.

Purpose of the Study:

• To develop a transfer learning framework for intelligent fault diagnosis that addresses data distribution challenges.
• To leverage knowledge from large datasets to improve diagnosis accuracy on new, similar tasks.

Main Methods:

• A pre-trained CNN is developed on extensive datasets to learn hierarchical features.
• The pre-trained CNN's architecture and weights are transferred and fine-tuned for new diagnostic tasks.
• Three transfer learning strategies are investigated to assess feature transferability.

Main Results:

• The proposed framework successfully transfers learned features from a pre-trained CNN.
• Diagnosis performance is significantly boosted for unseen machine conditions, including diverse working conditions and fault types.
• The study demonstrates the effectiveness of transfer learning in overcoming distribution shifts.

Conclusions:

• The presented transfer learning framework offers a robust solution for intelligent fault diagnosis under varying conditions.
• Pre-trained CNNs provide a valuable foundation for adapting to new diagnostic tasks, enhancing reliability and accuracy.
• This approach is crucial for real-world applications where data distributions are not static.