Blast Furnace Condition Recognition Based on Improved GCN and Multi-Source Data Feature Fusion

Abstract

To address the problem of imbalance of blast furnace furnace condition data categories, this study uses category-weighted cross-entropy loss function and synthetic minority oversampling technique in the training process of graph convolutional network, which improves the performance of graph convolutional network model, and effectively mitigates the negative impact of imbalance of furnace condition data categories on the training of the model. In order to comprehensively recognize the blast furnace furnace conditions, a multi-source data feature fusion method is proposed, which further improves the model performance by fusing the furnace condition numerical data, the air outlet image features and the furnace top image features. The experimental results show that the model recognizes the furnace conditions with an accuracy of 94.25% and an F1 score of 94.8%, realizing the recognition of blast furnace conditions based on multi-source data feature fusion.