An unsupervised fault monitoring framework for blast furnace: Information extraction enhanced GRU-GMM-autoencoder

Abstract

Fault monitoring of the blast furnace ironmaking process (BFIP) is challenging due to the nonlinearity and dynamic characteristics. The lack of fault labels and the existence of outliers further increase the difficulty of fault monitoring. In particular, these outliers are considered normal operations in BFIP. In this case, an unsupervised autoencoder (AE)-based monitoring framework is proposed. Firstly, the gate recurrent unit (GRU) is used as an encoder to adaptively capture the nonlinear dynamic features of data. The extracted features are proved to be directly related to monitoring. Second, the variance constraint of latent variables (LVs) is added to the loss function to enhance the ability of information extraction. Third, an approach that combines AE structure and GMM to deal with outliers is proposed. Based on the probability clustering results of GMM, monitoring thresholds can be calculated adaptively for different samples. Meanwhile, the fault variables can also be located according to the increment of reconstruction error of GRU-AE. Finally, the proposed method is validated by a numerical example and the practical data gathered from the BF of a Chinese steel group. Compared to unsupervised statistical methods and deep learning methods, the proposed algorithm achieves the lowest false alarm rate (FAR) + missing alarm rate (MAR).