NeSyLAD: A Neuro-Symbolic Approach for Unsupervised Logical Anomaly Detection

Detecting logical anomalies in industrial settings remains a significant challenge for conventional anomaly detection methods. Unlike structural anomalies such as scratches or dents, logical anomalies involve violations of component relationships, quantities, and arrangements that require reasoning about complex constraints. In this paper, we propose a NeuroSymbolic Logical Anomaly Detection (\methname) framework that combines deep learning-based component segmentation with symbolic rule extraction and logical reasoning. Our approach extracts interpretable rules from neural network activations and applies logical reasoning to detect and explain anomalies in industrial images. We evaluate our approach on the MVTec LOCO dataset, demonstrating its effectiveness in detecting logical anomalies across various product categories including breakfast boxes, juice bottles, screw bags, pushpins, and splicing connectors. Experimental results show that our approach not only achieves high detection accuracy but also provides human-readable explanations of detected anomalies, making it particularly valuable in industrial quality control settings.