Wide Residual Relation Network-Based Intelligent Fault Diagnosis of Rotating Machines with Small Samples

Area of Science:

• Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning (DL) fault diagnosis methods typically require extensive fault samples for training.
• Industrial rotating machines (RMs) have limited fault samples due to long normal operational periods.
• This scarcity hinders the development of effective DL-based diagnostic models for RMs.

Purpose of the Study:

• To propose a novel Wide Residual Relation Network (WRRN) for intelligent fault diagnosis of RMs.
• To address the challenge of limited fault samples in industrial rotating machinery diagnostics.
• To develop a method capable of accurate fault identification with minimal or even single-sample data.

Main Methods:

• A Wide Residual Network (WRN)-based module extracts representative fault features from input samples.
• A relation module calculates similarity scores between sample pairs to determine fault categories.
• The WRRN model is trained on tasks with sufficient samples and then transferred for few-shot fault diagnosis.

Main Results:

• Extensive experiments on two RMs validated the WRRN method's effectiveness.
• The WRRN accurately identified fault types even with very few or single fault samples.
• The proposed WRRN significantly outperformed existing popular fault diagnosis methods.

Conclusions:

• The WRRN offers a robust solution for intelligent fault diagnosis in rotating machines with limited data.
• This approach overcomes the data scarcity limitation inherent in many industrial diagnostic scenarios.
• WRRN demonstrates superior diagnostic performance compared to conventional methods, enabling reliable RM health monitoring.