Progressive Timestamp-Guided Soft Sensor: Breaking the Irregular Sampling Challenge in Molten Iron Silicon Content Modeling

As a high-temperature and high-pressure sealed reaction apparatus, the Blast Furnace (BF) poses challenges to the direct measurement of molten iron silicon content. Historically, it has been reliant on offline laboratory testing. Nevertheless, laboratory testing necessitates a substantial amount of time for sample collection, transportation, and analysis. This time-consuming assay is inherently irregular and fails to meet the requirements of molten iron quality optimization. Consequently, real-time soft sensor of molten iron silicon content is of critical importance for the BF ironmaking process. However, diverse previous deep learning-based modeling methods, which rely on temporal feature extraction networks, all overlook the temporal irregular characteristic of silicon content acquisition. To tackle this problem, a novel Time Stamp Linear Network (TStampLinear) is proposed, which employs timestamp information at multiple scales for the soft sensor modeling of silicon content. Firstly, a progressive two-stage pre-training framework for timestamp-guided prediction is utilized to augment the model’s sensitivity to timestamp information. Different methods of leveraging timestamp data are adopted in the two-stage task to assist the model in comprehensively utilizing timestamp information. Secondly, a drift-adapted timestamp-integration fine-tuning framework is put forward to offer an additional form of timestamp information and mitigate the distribution drift of the time series. The soft sensor experimental results of the silicon content in BF molten iron show that TStampLinear outperforms the optimal baseline model, achieving an approximate 7.08% improvement in the R2and a 2.13% improvement in MAPE.