Few-Shot Fault Diagnosis Based on an Attention-Weighted Relation Network

Area of Science:

• Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Complex energy conversion systems face reliability issues due to pneumatic control valve anomalies.
• Traditional deep learning fault diagnosis requires extensive labeled data, which is often scarce.

Purpose of the Study:

• To develop a novel fault diagnosis method for pneumatic valves that overcomes the limitation of small sample data.
• To enhance the accuracy and efficiency of fault detection and classification in critical industrial systems.

Main Methods:

• Proposed an attention-weighted relation network (AWRN) integrating few-shot learning and attention mechanisms.
• Implemented the AWRN for feature extraction and fault classification in pneumatic valve systems.
• Validated the method using a constructed DA valve fault dataset and a benchmark PU rolling bearing fault dataset.

Main Results:

• The AWRN achieved 99.15% accuracy on the DA valve fault dataset.
• Attained an average accuracy of 98.37% on the PU rolling bearing fault dataset.
• Demonstrated superior performance compared to state-of-the-art methods, especially with limited training data.

Conclusions:

• The proposed AWRN method is highly effective for fault diagnosis with small sample sizes.
• This approach significantly improves system reliability by enabling accurate and efficient fault detection.
• The AWRN offers a promising solution for fault diagnosis in complex industrial systems with data scarcity.