Unsupervised Insulator Defect Detection Method Based on Masked Autoencoder.

Area of Science:

• Engineering
• Computer Science
• Materials Science

Background:

• Maintaining high-speed rail operational safety relies on structural integrity of insulators.
• Conventional defect detection methods require large labeled datasets and struggle with class imbalance and anomaly detection.

Purpose of the Study:

• To develop an unsupervised framework for insulator defect detection.
• To address limitations of conventional methods in terms of data requirements and anomaly detection capabilities.

Main Methods:

• Utilized a masked autoencoder (MAE) architecture based on a vision transformer (ViT).
• Employed an asymmetric encoder-decoder structure with high-ratio random masking for representation learning.
• Implemented a dual-pass interval masking strategy for enhanced defect localization during inference.

Main Results:

• Achieved competitive image- and pixel-level performance on benchmark datasets.
• Demonstrated significant reduction in computational overhead compared to existing ViT-based methods.
• Enabled high-precision defect detection through image reconstruction without manual annotations.

Conclusions:

• The proposed unsupervised framework offers a scalable and efficient solution for real-time industrial inspection.
• This method is particularly effective under limited supervision scenarios.
• It facilitates robust representation learning and accurate defect localization for critical infrastructure.