Sequence-Aware Vision Transformer with Feature Fusion for Fault Diagnosis in Complex Industrial Processes.

Area of Science:

• Engineering
• Computer Science
• Data Science

Background:

• Industrial fault diagnosis presents challenges with high-dimensional, long time-series data and complex couplings.
• Traditional methods and existing deep learning models like Vision Transformer (ViT) struggle with capturing both global temporal patterns and local information crucial for accurate diagnosis.

Purpose of the Study:

• To propose a novel global and local feature fusion sequence-aware ViT (GLF-ViT) for enhanced industrial fault diagnosis.
• To modify feature embedding in ViT to retain sampling point correlations and preserve local information.

Main Methods:

• Developed a GLF-ViT model that fuses global features from the classification token with local features from the encoder.
• Modified feature embedding to preserve sampling point correlations and local information.
• Conducted experiments analyzing data segment length, network depth, feature fusion, and attention head receptive field.

Main Results:

• Demonstrated that shallower encoder networks are more suitable for high-dimensional time-series fault diagnosis in complex industrial processes.
• The proposed GLF-ViT method significantly enhances complex fault diagnosis by effectively fusing global and local features.
• Outperformed state-of-the-art algorithms on the Tennessee Eastman (TE) dataset and a power transmission fault dataset.

Conclusions:

• The GLF-ViT model offers a significant advancement in industrial fault diagnosis, particularly for complex, high-dimensional time-series data.
• The findings suggest that a shallower network architecture combined with effective feature fusion is key to improving diagnostic accuracy.
• The approach shows strong generalizability and effectiveness across different industrial datasets.