| Abstract Scope |
Carbon anode quality is vital for efficient aluminum smelting, yet manual inspection fails to detect 15–20% of defects, resulting in significant financial losses. This study develops a Keras-based Convolutional Neural Network (CNN) for automated binary classification of defective anodes. Utilizing a dataset of 440 industrial images augmented to 5,720 samples, the CNN was trained via a Keras backend and benchmarked against manual inspection and five traditional classifiers (including Random Forest and SVC).
The CNN achieved 96.63% accuracy and 100% recall for defects on unseen data, significantly outperforming the best traditional model (Random Forest: 75%). Retrospective application to three years of production data from the Volta Aluminium Company (VALCO) suggests the model could have identified 95% of 7,549 premature removals, potentially saving $2.3–$5 million. This research demonstrates that accessible deep learning frameworks can provide industrial-grade quality control, tripling defect detection rates and enabling 100% inspection throughput |