Enhancing power equipment defect identification through multi-label classification methods.

Area of Science:

• Electrical Engineering
• Computer Science

Background:

• Accurate power equipment defect identification is crucial for maintenance.
• Traditional methods are inefficient due to subjective manual records and limited scope.
• Current approaches often oversimplify status assessment based solely on defect grade.

Purpose of the Study:

• To develop a multi-label classification dataset for power equipment defects.
• To evaluate and compare various machine learning and deep learning methods for defect classification.
• To enhance the precision of power equipment defect identification and classification.

Main Methods:

• Compiled historical defect records into a novel multi-label classification dataset.
• Assessed 11 established multi-label classification methods (traditional ML and deep learning).
• Employed balanced loss functions to address sample imbalance and segmented the task into label recall and ranking stages.

Main Results:

• Methods considering label correlations demonstrated significant performance advantages.
• Balanced loss functions effectively mitigated sample imbalance issues.
• Segmenting the classification task into recall and ranking stages improved overall performance.

Conclusions:

• The developed dataset and evaluated methods offer a more precise approach to power equipment defect classification.
• Considering label correlations and addressing sample imbalance are key to improving accuracy.
• The created dataset is publicly available to advance research in power equipment health assessment.