R-CTGAN: A method for cement clinker fCaO data augmentation and prediction based on cascade temporal GAN

Abstract

To improve the accuracy of fCaO content predictions, which is often hindered by sparse temporal features in cement manufacturing data, we propose a data augmentation and prediction model based on R-CTGAN (Regression-Cascade Temporal GAN). This model integrates a dual-layer cascade GAN with a coordinate attention mechanism and a regression prediction network. By incorporating the Wasserstein distance and multi-dimensional dynamic time warping into the GAN loss function, we enhance the temporal consistency and detail fidelity of the generated data. This process expands the data scale and feature space, leading to better training of the regression network. The regression network employs an efficient channel attention mechanism to extract features, thereby increasing sensitivity and prediction accuracy during the cement clinker calcination process. Experimental results on cement production data confirm the model’s superior accuracy and robustness in predicting fCaO content.