Physically Enriched Product Adaptive Deep Networks for Industrial Metal Detection

This study addresses a longstanding metal-detection problem (classifying a product as clean or contaminated) using two-frequency electromagnetic fields by analyzing complex temporal data. Traditional contamination detection methods rely on analyzing permittivity and conductivity signals obtained when products pass through the detector's electromagnetic field. While reliable, these methods struggle with product effect - a phenomenon where conductive product signals overshadow contaminant signals, leading to false negatives. Advances in deep learning (DL) have enhanced detection accuracy but often require large training datasets, which are impractical in many industrial metal detection scenarios where the cost of data acquisition is high. To address these challenges, we propose physically enriched DL architectures that integrate domain knowledge with state-of-the-art feature extraction models. Our contribution centers on two key innovations: adaptive models capable of countering product effect and improving contamination detection accuracy compared to both traditional signal processing and machine learning (ML) methods, and second, an enhanced ability to classify contaminant (metal) type, enabled by a data preprocessing module that we call learnable denoisers. Extensive experiments on challenging real-world data demonstrate that our product adaptation modules and a newly designed training strategy that extracts pure metal signatures from contaminated products enable scalable performance across diverse products. The proposed models successfully mitigate both false positives and negatives (missed metal detection) breaking a stiff tradeoff. Further, the benefits of our approach are most pronounced when training data are limited indicating superior generalizability.