Deep learning-based autonomous weld quality inspection in battery pack manufacturing using a two-stage model.

Area of Science:

• Materials Science
• Manufacturing Engineering
• Artificial Intelligence

Background:

• Laser welding is crucial for electric vehicle battery pack assembly, but current inspection methods are inconsistent and fail to detect latent defects.
• Existing quality control relies on manual, electrical, and physical checks, which are insufficient for ensuring high-performance and safety standards.

Purpose of the Study:

• To develop a robust, two-stage deep learning framework for automated laser welding quality inspection.
• To address the challenge of scarce defect data in industrial inspection scenarios.
• To improve the detection and classification accuracy of latent defects in laser welds.

Main Methods:

• A two-stage deep learning approach was proposed, starting with an autoencoder trained on normal data using a novel loss function.
• The second stage employed multi-class defect classification utilizing the encoder and bottleneck layers from the first stage.
• Data augmentation techniques were integrated to enhance model generalization capabilities.

Main Results:

• The proposed framework achieved 100% accuracy and an F1-score of 1.0 in defect detection and classification.
• It significantly outperformed the conventional rule-based system, which had 99.31% accuracy and a 0.995 F1-score.
• The system demonstrated perfect consistency across all tested samples, validating its effectiveness in data-constrained environments.

Conclusions:

• The developed two-stage deep learning framework provides a scalable and intelligent solution for automated quality inspection in industrial laser welding.
• This approach is particularly effective in scenarios with limited defect data, offering superior performance over traditional methods.
• The framework ensures high weld quality, directly impacting the performance and safety of critical components like electric vehicle battery packs.