Classification of battery laser welding defects via enhanced image preprocessing methods and explainable artificial intelligence-based verification

This study focuses on the precise classification of normal and defective laser welding based on image data. A deep-learning algorithm with data augmentation and transfer learning is employed to enhance the classification accuracy. A comprehensive comparison is performed to evaluate the traditional preprocessing methods against the newly proposed and fused methods across three categories to discern their impact on the classification. Welding defects in the automotive battery industry pose significant cost issues during practical operation, necessitating compensation for quality problems. These issues raise concerns regarding trust and business relationships. Our research implications include substantial cost reduction and safety enhancement in the automotive battery industry by facilitating faster and more accurate laser welding quality assessments. To address the "black-box" nature of deep-learning algorithms, we use gradient-weighted class activation mapping (Grad-CAM) as an explainable artificial intelligence technique that provides intuitive validation and quantitative criteria for classification. We observe how different preprocessing methods influence the classification performance, identifying techniques that maximize accuracy. This study presents a straightforward yet highly accurate method for evaluating laser welding quality using only image data, achieving an accuracy of up to 98%. This value surpasses the accuracy obtained using raw data by 45%. Thus, our method serves as an effective tool for defect assessment across industries, thereby enhancing overall quality management. The innovative and impactful nature of this study underscores its significance for advancing the quality assessment of laser welding.