earticle

영어

We present an unsupervised anomaly detection framework for aluminum molten metal processes that trains exclusively on normal operational data. The extreme temperatures and dynamic luminance variations characteristic of molten metal environments make it practically infeasible to collect representative samples of anomalies—such as foreign material contamination, thermal irregularities, or flow inconsistencies—and manual annotation remains prohibitively labor-intensive. Our approach employs a Convolutional Autoencoder to capture the visual and thermal signatures of normal process states. Anomalies are detected by computing reconstruction errors and comparing them against thresholds derived from the upper percentiles of the normal error distribution. Experimental validation across multiple production lines demonstrates that our method achieves high detection accuracy even with limited abnormal samples, offering a practical solution for automated quality control and real-time process monitoring in industrial casting operations.