16 July 2024 in IEEE Transactions on Instrumentation and Measurement
The precise prediction of tumbler strength is of great significance to provide high-quality sinter products for the downstream blast furnace ironmaking process. However, most of the existing soft sensor models exclusively depend on the data-driven learning paradigm in the sintering process, ignoring the knowledge extraction and fusion. Building a pure data-driven neural network from the samples usually degrades its performance. How to extract and incorporate knowledge into data-driven models is always a very challenging problem in the industrial field. To remedy the weakness of the data-driven models, we propose a knowledge and data dual-driven graph neural network (KDGNN) for the tumbler strength prediction in an end-to-end fashion. First, we propose a knowledge representation method to mine the spatial coupling relationships among process variables by calculating their interactive influence under the guidance of prior knowledge. Then, to capture spatial–temporal correlations, we propose to design a gated spatial–temporal convolutional (GSTC) block to extract the knowledge-based and data-based features using a dual-branch network, where each branch performs its own duty individually. Meanwhile, the locality-preserving projection is developed to fuse these two kinds of features to obtain more discriminative and powerful features for soft sensor modeling. Finally, the experimental results on a real-world sintering process demonstrate that the proposed KDGNN outperforms the existing popular methods in the tumbler strength task.
6 May 2026 in Control Engineering Practice
15 January 2025 in Applied Energy
15 November 2025 in Knowledge-Based Systems
20 November 2024 in IEEE Transactions on Instrumentation and Measurement
Sintering represents a critical link in the steel production process, with the produced sinter ore serving as the core raw material for blast furnace ironmaking. Characterized by harsh working environment and multiparameter, the sintering process causes many obstacles to academic research on measurement and prediction of key indicators. To address this problem, this article introduces a multistep prediction strategy, named the state feature extractor-based multitarget networks delay updates algorithm (SFE-MDA), to predict burn-through point (BTP) in sintering process. Initially, considering the characteristics of measurable interrelated variables in the sintering process, initial variables are selected and categorized. Subsequently, following the input standards of reinforcement learning (RL) framework, the SFE is employed to derive latent variables and to model the Markov decision process (MDP). Finally, a novel multitarget networks delay updates algorithm is designed. The proposed approach is validated on real sintering data, with results indicating its superiority over existing deep learning methods in terms of long-term multistep prediction accuracy, thereby confirming its feasibility and superiority.
21 November 2025 in Expert Systems with Applications