BatterySurAD: A Dataset for Anomaly Detection on Pouch-Type Reflective Battery Surfaces with Spatial Zone Annotations

Battery surface defects pose critical safety risks in lithium-ion manufacturing, yet existing anomaly detection datasets lack the spatial granularity needed for reflective battery surfaces with position-dependent defect distributions. We present the first large-scale battery surface anomaly detection dataset, BatterySurAD, with fine-grained 15-zone annotations across dual-view surfaces and 4 defect types. Through comprehensive statistical analysis of feature distributions across 720 independent tests, we demonstrate: (1) substantial domain shift from general-purpose benchmarks with 418/720 tests (58.1%) showing statistically significant distributional differences ( \(p < 0.05\) ), establishing BatterySurAD as a distinctly more challenging evaluation setting, (2) position-dependent detection difficulty varying 3 to 5 times between boundary and central zones, (3) model-specific failure patterns with reconstruction-based methods showing catastrophic degradation (DFR: \(\mu \) =7.59 on denting vs. \(\mu \) =0.14–0.18 for memory-based models), and (4) view-dependent complexity with back surfaces exhibiting 35–45% higher feature distances. Zero-shot evaluation of six state-of-the-art models reveals consistent failure patterns attributable to geometry, reflective surfaces, and spatial heterogeneity—characteristics absent in flat-surface benchmarks. We further include CPU-feasible foundation-model baselines (DINO/CLIP) with light-weight adaptation (calibration and linear probing), offering a practical reference point for rapid deployment-oriented evaluation. BatterySurAD enables systematic investigation of spatially-structured anomaly detection in safety-critical manufacturing contexts.