Large-Scale Pre-Training Improves Anomaly Detection in Fused Magnesium Smelting Process

Reliable anomaly detection is essential for the fused magnesium smelting process. Existing researches typically pursuing network structure to improve model performance, while overlooking the influence of data scale for training. To tackle the problem of limited labeled data in real industrial scenarios, this paper utilizes network pre-trained on large-scale public dataset, ImageNet1K, alleviating the issue of insufficient samples for fused magnesium furnace anomaly detection. This paper also uses a large-scale fused magnesium furnace anomaly dataset FMF-Big for a baseline comparison. The empirically study shows that, even though the data in ImageNet1K differs greatly from the fused magnesium furnace scenario, the pre-trained model brought a 1.81% increase in accuracy (97.59% detection accuracy). To our best knowledge, this is the first exploration to investigate the impact of large-scale pre-training on the anomaly detection in fused magnesium smelting process. We believe the study will have inspiring implications for subsequent researches.