A Deep-Convolution-Generative-Adversarial-Networks-based Missing Data Filling Method for Blast Furnace Gas System in Steel Industry
Canguang Yang, Feng Jin, Jun Zhao, Wei Wang · 2022 IEEE 11th Data Driven Control and Learning Systems Conference (DDCLS) · 2022
The integrity of monitoring data is of great significance to ensure the accuracy of data analysis and reliable operation in blast furnace gas system of steel industry. In this study, a Deep-Convolution-Generative-Adversarial-Networks (DCGAN)-based data filling method is proposed for high proportion missing in time series, and the corresponding network structure is designed. Time series are transformed non-destructively into time domain images by a Gram matrix, and their temporal characteristics are preserved. The authenticity constraint and the context similarity are established to optimize the hidden variables, and the high-precision time domain image is generated through the DCGAN. The completed time series is obtained by inversely transforming the Gram matrix. The simulation results based on the actual operating data of a steel enterprise indicate that the proposed method is capable of has high padding accuracy in the case of a high proportion of random missing data and continuous missing data.