EL-DenseNet: a novel method for identifying the flame state of converter steelmaking based on dense convolutional neural networks

The identification of flame status in converter steelmaking is of great significance for steel smelting and molten steel quality. It can monitor the converter smelting process, strictly control the smooth progress of the production process, and effectively avoid personal injury. However, converter steelmaking is located in a production environment with high temperatures, high smoke, and strong physical and chemical reactions. The results of traditional manual fire observation are influenced by various factors such as experience and environmental conditions, showing unstable classification results, resulting in low accuracy in identifying the flame state of converter steelmaking. In order to improve the accuracy of flame state recognition in converter steelmaking, this article fully utilizes the flame image information provided by a certain steelmaking company during the converter steelmaking blowing process to classify the converter steelmaking flame images. This article proposes a novel dense convolutional neural network (EL-DenseNet) model for flame state recognition in converter steelmaking. Firstly, the efficient channel attention mechanism (ECAM) was introduced into DenseBlock to enhance the model’s attention to different channels, locate relevant useful information, and suppress useless information, thereby improving the model’s ability to capture key features in flame images; Then, in the model training and validation stage, LabelSmoothing was used to replace the original cross entropy loss function, smoothing the real labels and reducing overfitting of the model to the training data. Through experiments, it has been shown that the training accuracy of the EL-DenseNet model proposed in this article is 99 % , and the testing accuracy is 96.7 % . This improves the accuracy of flame state recognition in converter steelmaking, saves manpower and material resources, and reflects the effectiveness and superiority of the model.