Convolutional Recurrent Neural Network to Determine Whether Dropping Slag Dart Fills the Exit Hole During Tapping in a Basic Oxygen Furnace

This paper presents a system that uses a convolutional recurrent neural network (CRNN) model to determine automatically whether or not a slag dart has plugged the exit hole of a basic oxygen furnace (BOF) during the tapping operation. The CRNN model uses a convolutional neural network (CNN) model to analyze image characteristics, and a long short-term memory (LSTM) model to analyze sequential characteristics in real-time video collected using a closed-circuit television (CCTV). The output result of the last pooling layer of the neural network is stored using a CNN model for each frame of image data; this process obtains vectors that store feature values of each frame, and recognizes six conditions during the tapping operation in the BOF. These feature values of the frames are input to the LSTM as sequential data; the LSTM classifies whether the darting operation succeeded or failed. The integrated CNN-LSTM model, can consider spatial and temporal characteristics simultaneously in the analysis of slag-dart hitting. Deep learning related to the determination of the slag dart plugging was performed by applying the proposed model to 226,800 images collected from using the CCTV; classification accuracy of 99.45 pct was obtained. We built a pilot system that provides real-time information on the tapping operation status and the dart operation results from video. This system can increase the accuracy and decrease the human workload required to identify dart-hitting during the tapping operation. Compared to the previous system that relies on the operator's visual judgment, the accuracy of dart input judgment increased by 10 pct, and the workloads related to the dart judgment task and the monitoring dart input were reduced by 30 pct.