A zero-shot attribute-embedded model with a feature difference mapping sigmoid function for compound fault diagnosis...

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Machinery compound fault detection is challenging due to the scarcity of training data.
• Existing methods often require extensive compound fault samples, which are impractical to collect.

Purpose of the Study:

• To develop a zero-shot learning model for diagnosing unseen compound faults without requiring compound fault data.
• To improve the accuracy of attribute recognition in multi-label classification tasks.

Main Methods:

• Proposed a zero-shot attribute-embedded model for compound fault diagnosis (ZSAECFD).
• Constructed attribute prototypes for single and compound faults using only single fault data.
• Introduced a novel activation function, feature difference mapping sigmoid (F-sigmoid), to enhance feature differences and alleviate gradient vanishing.

Main Results:

• The ZSAECFD model achieved 81.82% diagnostic accuracy for unseen bearing compound faults.
• The model reached 88.17% diagnostic accuracy for unseen gearbox compound faults.
• Demonstrated superior performance compared to classical and advanced zero-shot learning methods.

Conclusions:

• The proposed ZSAECFD model effectively diagnoses unseen compound faults using only single fault data.
• The F-sigmoid activation function enhances diagnostic accuracy by amplifying feature differences.
• The approach offers a practical solution for compound fault diagnosis in real-world engineering scenarios.