Unsupervised Hierarchical Denoising Guided Student-Teacher Framework for Battery Spot Welding Anomaly Detection

Accurate detection of battery spot welding quality is crucial for ensuring the safety and reliability of batteries. However, traditional visual detection methods are limited in their ability to identify complex defects. Furthermore, supervised deep learning approaches rely heavily on large labeled datasets, making them less adaptable to the challenge of scarce anomalous samples in real industrial scenarios. To address these issues, this paper presents an novel unsupervised hierarchical denoising guided student-teacher framework (HDST), which enhances anomaly detection performance by strengthening the feature differences between the student and teacher models. The proposed method comprises an anomaly generation, a hierarchical denoising guided student model, and a pre-trained teacher model. The student model denoises and reconstructs features from abnormal samples while utilizing feature distillation loss to guide the learning of normal representations. During the inference phase, anomaly detection and localization are accomplished by comparing the feature differences between the student and teacher models. Furthermore, this paper introduces a new battery spot welding anomaly detection dataset (BSW AD), which contains 8,148 images with pixel-level annotated defect samples. Experimental results demonstrate that HDST achieves AUROC scores of 97.05% at the image level and 93.78% at the pixel level, significantly outperforming existing mainstream methods. This approach offers an efficient solution for unsupervised anomaly detection in industrial settings and addresses the need for high-quality datasets in battery spot welding.