A Self-Healing Control Framework for Blast Furnace Based on Latent Space Feature Transformation

Xujie Zhang, Chunjie Yang, Siwei Lou, Yuelin Yang, Yi Li, Ping Wu

Abstract: The blast furnace ironmaking process (BFIP) involves a highly intricate charge composition, making it susceptible to various faults arising from abnormal charge distribution. Given the complexity of BFIP, there is an urgent demand for a self-healing control method to assist field engineers in fault diagnosis and recovery operation implementation. In response, this article presents a self-healing control method based on a generative adversarial network (GAN), termed the multivariable decoupled self-healing control (MVDSHC) method. To begin, expert knowledge is employed to select control variables as sample tags, which are then hierarchically organized to divide different subdatasets. Subsequently, a latent space transformation self-healing control model is constructed, and multiple training paths are utilized to establish an objective function for model training. Furthermore, a dual evaluation strategy is implemented to provide real-time assessment of abnormal conditions in the BFIP. When a fault is detected, the model generates a list of schemes illustrating the working conditions of transformed samples under different operation schemes. An operation scheme is deemed viable if the corresponding transformed sample is evaluated as healthy. The experimental results conducted on BFIP data demonstrate that the proposed method not only mitigates the limitations of few fault samples in the fault diagnosis task, but also effectively restores the BFIP to a healthy state.

External IDs:doi:10.1109/tim.2025.3635339