A Multitask Crisscross Network for One-Shot Bearing Multiattribute Fault Diagnosis.

Area of Science:

• Engineering
• Machine Learning
• Data Science

Background:

• Data-driven fault diagnosis, especially for bearings, faces significant challenges due to limited sample availability.
• Operational constraints hinder the collection of bearing fault data, leading to poor model performance on unseen data, particularly with single-sample fault instances.

Purpose of the Study:

• To propose a novel multitask crisscross network (MTCCN) for effective one-shot fault diagnosis.
• To leverage transfer learning by treating fault semantic information as transferable attributes.
• To enhance diagnostic accuracy by integrating multiple information sources.

Main Methods:

• Developed a multitask crisscross network (MTCCN) utilizing transfer learning principles.
• Employed a multitask learning approach to predict multiple fault attributes simultaneously.
• Implemented a horizontal task-sharing network for global feature extraction and vertical attribute classifier chains (ACC) with directed acyclic graphs (DAG) to model attribute correlations.
• Integrated an information map to fuse heterogeneous data sources.

Main Results:

• The MTCCN demonstrated significant efficiency and robustness across various datasets.
• The transfer learning approach effectively utilized limited fault data for accurate diagnosis.
• Quantified transfer effects confirmed the model's capability in low-sample scenarios.

Conclusions:

• The proposed MTCCN is an efficient and robust solution for one-shot bearing fault diagnosis.
• The multitask learning and transfer learning strategies effectively overcome the challenge of limited sample data.
• The integration of heterogeneous information sources further boosts diagnostic prediction accuracy.