A Diffusion-Based Unified Framework for Open-World Dynamic Wheel Recognition System Construction and Maintenance With...

Area of Science:

• Artificial Intelligence
• Computer Vision
• Manufacturing Technology

Background:

• Wheel recognition systems are crucial for manufacturing but face challenges like data incompleteness and novel class detection.
• Dynamic production environments and scarce training data hinder the construction and maintenance of these systems.

Purpose of the Study:

• To propose a unified diffusion model-based framework to address data challenges in wheel recognition systems.
• To enhance the construction and maintenance of intelligent wheel recognition systems in dynamic manufacturing environments.

Main Methods:

• A style and triple structure guided diffusion model synthesizes high-quality wheel images for dataset creation.
• A generative out-of-distribution (OOD) mixture novelty detection method uses synthesized OOD data for precise boundary establishment.
• Self-supervised style-consistency learning bridges the domain gap between synthetic and real images.

Main Results:

• The proposed framework demonstrates superior performance over state-of-the-art methods in wheel recognition and model updating tasks.
• Diffusion models effectively generate sufficient training data, overcoming scarcity issues.
• The framework successfully enhances downstream task performance by addressing domain gaps.

Conclusions:

• This work represents the first application of diffusion models to intelligent system construction and maintenance in industrial manufacturing.
• The proposed framework offers a robust solution for data incompleteness and novelty detection in wheel recognition.
• Artificial General Content (AIGC) technology shows significant potential for advancing industrial manufacturing processes.