Semi-supervised high-uncertainty deep canonical variate analysis for fault diagnosis in blast furnace ironmaking

Abstract

Blast furnace ironmaking process (BFIP) is of paramount importance in the steel industry. Reliable fault diagnosis for BFIP is crucial to ensure production safety, improve efficiency and quality, reduce costs, and maximize resource utilization. However, establishing effective fault diagnosis models is hindered by challenges including non-linearity, dynamics, widespread noise, and the scarcity of labeled data alongside an abundance of unlabeled data. To address these issues, this paper proposes a new fault diagnosis method for BFIP called semi-supervised high-uncertainty deep canonical variate analysis (SHDCVA). The proposed algorithm consists of three main parts, including (1) high-uncertainty nonlinear dynamic feature capture, (2) robust semi-supervised framework construction, and (3) model solving and parameter optimization. Firstly, a high-uncertainty deep canonical variate representation method is proposed from a probabilistic perspective, which can capture high-uncertainty nonlinear dynamic characteristics. The high-uncertainty property can effectively deal with data noise and enhance the reliability of downstream fault diagnosis model. Moreover, this paper proposes a robust semi-supervised classification framework that can efficiently utilize limited labeled samples and a large amount of unlabeled samples. The supervised part controls the release of labeled samples by training signal annealing method (TSA) to prevent overfitting, while the unsupervised part enforces model smoothing by applying adversarial perturbations to enhance robustness. Subsequently, an efficient computational method is devised to generate adversarial perturbations and the overall objective is constructed. Finally, the effectiveness of SHDCVA is confirmed through a practical case study utilizing genuine BFIP data.