Skip-CBAM Based Unsupervised Anomaly Detection of Fused Magnesia Furnace

Abnormal ultra-high temperature semi-molten conditions in a fused magnesia furnace (FMF) can lead to furnace leakage. To ensure high detection performance, existing models require sufficient expert-labeled training data for supervised learning. However, the above model has limited ability to identify unknown abnormal working conditions, and label production is time-consuming and expensive. This paper suggests an unsupervised detection method based on skip-connected generative adversarial network (GAN) to address this issue. Specifically, to adapt to the strong interference caused by irregular changes in images caused by irregular water mist and flicker scenes, existing methods remove interference through multi-step processing for detection, which is difficult to apply in practice. In contrast, we use a reconstruction model of convolutional block attention module (CBAM) with skip connections for end-to-end computation, which improves performance while being more concise. This method determines abnormal areas through reconstruction and distinguishes abnormal working conditions by comparing abnormal scores. Finally, this method is compared with existing unsupervised and supervised methods on actual FMF images, and the application results demonstrate the effectiveness of this method.