Högskolan i Skövde

This report studies automated visual inspection for battery assembly using three deep learning methods with different strengths: a supervised object detector (YOLOv7), a reconstruction-based autoencoder, and EfficientAD. A common pipeline was used to ensure a fair comparison. All images were resized to 256 × 256, normalised, and augmented using brightness changes, flips, colour variations, and mild geometric transforms. Performance was evaluated at the image level using precision, recall, F1 score, and ROC–AUC, supported by confusion matrices and visual heatmaps. Stress tests were also performed under changes in background, lighting, and cell orientation.The results show that YOLOv7 is effective for enforcing clear assembly rules such as component presence, count, and left–right terminal polarity, making it suitable for stop or rework decisions. The autoencoder provides clear heatmaps that highlight missing or extra structures and is useful for monitoring data drift, but it is sensitive to background and pose changes. EfficientAD achieves the best overall anomaly detection performance and robustly highlights both texture anomalies and spatial deviations. Although EfficientAD performs best in anomaly detection, it does not produce bounding-box outputs needed for explicit rule checking. Based on this, a layered inspection strategy is proposed: YOLOv7 is used for rule-based checks, EfficientAD acts as a high-sensitivity anomaly filter, and a lightweight autoencoder monitors data drift over time. This approach improves reliability and practical deployment in real battery production lines.