On a Hybrid CNN-Driven Pipeline for 3D Defect Localisation in the Inspection of EV Battery Modules.

Area of Science:

• Materials Science
• Computer Vision
• Manufacturing Engineering

Background:

• Reliable electric vehicle (EV) battery production necessitates accurate surface defect detection and precise localization for automated processes.
• Conventional 2D computer vision struggles with real-world coordinate mapping on curved EV battery surfaces, hindering digital twin and root cause analysis.

Purpose of the Study:

• To develop a hybrid 3D inspection pipeline for precise, millimetre-level defect localization on EV battery modules.
• To overcome the limitations of 2D methods in mapping defects to real-world coordinates on complex battery geometries.

Main Methods:

• Integration of calibrated dual-view multi-view geometry for defect point projection and triangulation.
• Utilizing single-image neural 3D shape inference to complement areas with limited multi-view coverage.
• Employing generative AI (GenAI) for data augmentation with physically informed, synthetic defect variations.

Main Results:

• Achieved millimetre-scale localization accuracy for surface defects on EV battery modules.
• Outperformed a per-view convolutional neural network (CNN) baseline in both defect segmentation and 3D continuity.
• Generated traceable, real-world defect coordinates on complex battery surfaces.

Conclusions:

• The hybrid 3D approach provides accurate as-built localization of defects on EV battery modules.
• This method significantly enhances automated inspection, repair, and quality control processes for EV batteries.
• The pipeline offers improved utility for digital twins and root cause analysis in battery manufacturing.