Spatio-Temporal Feature Enhancement for Recognizing Strongly Correlated Sequential Actions in Aircraft Assembly

Area of Science:

• Aerospace Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Aircraft assembly quality relies on operator consistency during positioning and clamping.
• Capturing long-term temporal correlations in industrial settings is challenging.
• Existing methods struggle with complex human-machine interactions in assembly.

Purpose of the Study:

• To develop an advanced action recognition network for aircraft assembly tasks.
• To address the challenge of long-term temporal dependencies in industrial environments.
• To improve the precision and real-time monitoring of assembly processes.

Main Methods:

• Proposed the Long-Term Strongly Associated Action Recognition Network (LTSA-Net) based on the C3D backbone.
• Incorporated SimAM attention and BN modules for spatiotemporal feature enhancement.
• Introduced LTSFEM for global temporal information extraction and CWSTB for parameter compression.
• Utilized a dedicated aircraft assembly dataset, AdamW optimizer, and Mixup data augmentation.

Main Results:

• LTSA-Net achieved 98.82% recognition accuracy on the custom LTSA-Dataset.
• The model demonstrated a per-frame inference time of 42 ms.
• Successfully balanced high precision with real-time performance requirements.

Conclusions:

• LTSA-Net provides a practical technical solution for intelligent monitoring in aircraft assembly.
• The network effectively captures long-term, strongly correlated features in complex industrial environments.
• Achieved high accuracy and real-time inference, meeting industrial demands.