A novel few-shot fault diagnosis model for addressing nonstationarity in the ironmaking process

Abstract

The fourth industrial revolution is a green industrial revolution represented by artificial intelligence, clean energy, and other fields, which is both a challenge and an opportunity for the blast furnace ironmaking process (BFIP). Considering the dynamics, nonlinearity, nonstationarity, few shots, and many outliers in BFIP fault diagnosis, we proposed a novel method called Slow Feature Constrained-Least Squares Improved Generative Adversarial Network (SFC-LSIGAN). First, the sliding window is used to explore the process dynamics, while the deep learning model could better extract the deep nonlinearity between variables. Secondly, aiming at the properties of few shots and nonstationarity in BFIP, a new model was proposed based on the similar training process of Auxiliary Classifier GAN (ACGAN) and Deep Slow Feature Analysis (DSFA). Therefore, while completing the task of few-shot fault diagnosis, the proposed method further extracts the nonstationarity to improve the accuracy of the model. Furthermore, many outliers in the BFIP data are likely to have an impact on the quality of the generated samples. The least squares loss form function was introduced to enhance the quality of the generated samples and alleviate the mode collapse problem during the proposed model training process. Experiments on a real BFIP showed that, compared with the state-of-the-art methods, our SFC-LSIGAN method achieved superior performance in both data enhancement and fault diagnosis.