Iron-Tapping State Recognition of Blast Furnace Based on Bi-GRU Composite Model and Post-Processing Classifier

The recognition of blast furnace iron-tapping status is crucial for improving the production efficiency and quality of ironmaking plants. However, traditional manual observation methods face interference from a large amount of iron powder and other dust in the environment, as well as problems such as iron splashing during tapping, making it difficult to achieve accurate real-time recognition. In addition, the liquid-level detection required for automatic recognition of iron-tapping status is also affected by factors such as iron splashing and foam, surface oxidation, and slag skin, making it impossible to accurately identify the iron-tapping status. These issues highlight the urgency of developing innovative identification methods. To address this, this study collected 2 million continuous three-month field liquid-level data, and constructed a time-series dataset of iron-tapping liquid levels through data slicing. Based on deep learning, a hybrid neural network model of bidirectional gated recurrent units (Bi-GRUs) and attention mechanism was designed to extract the features of iron-tapping iron water status for state recognition, and then a post-processing classifier was designed and introduced for multipoint judgment. Finally, an iron-tapping status recognition accuracy of 98.1% was achieved in on-site applications. The application results show that the deep learning-based blast furnace iron-tapping state recognition technology proposed in this study can achieve high-precision and strong adaptability recognition under the harsh blast furnace iron-tapping environment in ironmaking plants. This technology not only improves production efficiency and quality but also avoids human interference, reduces labor costs, and is of great significance for intelligent metallurgical production.