Internal Thread Defect Generation Algorithm and Detection System Based on Generative Adversarial Networks and You...

Area of Science:

• Industrial inspection
• Mechanical engineering
• Computer vision

Background:

• Accurate detection of thread quality is vital for mechanical performance in industrial settings.
• Current machine vision methods struggle with internal thread defect detection due to geometric complexity and limited training data.
• Challenges include efficient detection and acquiring sufficient high-quality training samples.

Purpose of the Study:

• To develop an effective machine vision approach for internal thread defect detection.
• To address the limitations of insufficient training data in existing methods.
• To improve the accuracy and efficiency of industrial thread quality inspection.

Main Methods:

• Proposed a novel image acquisition structure for internal threads.
• Developed a data augmentation algorithm using Generative Adversarial Networks (GANs) to create high-quality training datasets.
• Employed a YOLO (You Only Look Once) algorithm for defect detection.
• Conducted multi-metric evaluation and compared results with external threads.

Main Results:

• Achieved high-similarity internal thread image generation using GANs.
• The YOLO algorithm demonstrated high detection accuracy for internal threads (94.27%) and external threads (93.92%).
• Effectively identified internal thread defects, overcoming previous limitations.

Conclusions:

• The proposed method, combining GANs for data augmentation and YOLO for detection, significantly enhances internal thread defect detection.
• This approach provides a robust solution for industrial inspection, improving mechanical performance assurance.
• The technique offers a viable and accurate method for quality control in manufacturing processes involving threads.