Self-adaptive selection graph pooling based fault diagnosis method under few samples and noisy environment.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Industrial Systems

Background:

• Neural network (NN)-based methods are widely used for intelligent fault diagnosis.
• Limited faulty samples and noise interference hinder the performance of existing NN-based methods.

Purpose of the Study:

• To propose a self-adaptive selection graph pooling method to address the limitations of existing NN-based fault diagnosis techniques.
• To improve diagnosis performance with limited data and enhance robustness against noise interference.

Main Methods:

• Graph encoders with shared parameters extract local structure-feature information (SFI) from sensor-wise sub-graphs.
• Temporal continuity of SFI is maintained via concatenation, forming a global sensor graph.
• A self-adaptive node selection mechanism alleviates noise interference by focusing on fault-relevant nodes.
• Multi-scale graph features are extracted using local max pooling and global mean pooling for a multi-layer perceptron.

Main Results:

• The proposed method achieves superior diagnosis performance with limited data.
• The method demonstrates strong anti-interference ability in noisy environments.
• The self-adaptive node selection mechanism provides good interpretability through visualization.

Conclusions:

• The self-adaptive selection graph pooling method offers an effective solution for intelligent fault diagnosis in industrial systems.
• The approach enhances diagnostic accuracy and robustness, particularly in data-scarce and noisy conditions.
• The method's interpretability facilitates understanding and trust in the fault diagnosis process.