Semi-Supervised Condition Monitoring and Visualization of Fused Magnesium Furnace

This paper introduces a novel practice of using image based condition classification and visualization system to augment operators in the task of monitoring the working condition of fused magnesium furnace. The system implements two functions: working condition detection and remote visually reconstruction of the furnace flame. For the problem of working condition detection, we combine the image features and the smelting electrical currents to train the classifier under semi-supervised learning framework. We also introduce a highly efficient cross-entropy based optimization method for training. For the visualization task, we propose a practical end-to-end solution which can visually simulate the dynamic furnace flame at remote monitoring consoles according to the visual feature of the monitoring video. Finally, we introduce the distributed structure of the monitoring system which consists of a private cloud, an edge server and remote monitoring consoles. The proposed solution can be applicable for various monitoring tasks in industry. Note to Practitioners—This paper introduces an image based solution to the practical problem of condition monitoring of fused magnesium furnace. The solution includes two parts: working condition detection and visualization. First, for the abnormal condition detection problem, most existing approaches use the process monitoring video as the condition predictor. It also assumes that the video data are fully labeled with condition category, which often is not available in practice. This paper suggests a novel approach combining the monitoring video and the smelting electrical currents under semi-supervised learning framework to the development of condition classifier. Experiments show that the detection accuracy is remarkably improved comparing to the approach using monitoring video only. Second, this paper presents a remote visualization solution which enables instant reconstruction of the furnace flame at monitoring consoles using only essential features of the flame. This technique is well suited for situation of real-time remote visual monitoring with very limited network bandwidth. Finally, the condition monitoring system adopts a distributed structure including a private cloud, an edge server and remote monitoring consoles. The presented techniques in this paper can be applicable for various monitoring tasks in industry.