Enhancing cement clinker quality prediction through integrated convolutional and recurrent neural networks with adversarial data augmentation

Abstract In the cement calcination process, predicting free calcium content is often hampered by data imbalances and inadequate extraction of temporal information. Addressing this issue,this paper presents a Convolutional Recurrent Neural Network (CRNN) model enhanced by a Distribution-Aware Encoding Generative Adversarial Network (DAE-GAN) to improve free calcium content prediction in cement clinker. Utilizing a sliding window technique for feature extraction and a GAN for data augmentation, the model efficiently learns temporal relationships. An embedded encoder within the GAN and a convolutional transformation layer above a Bidirectional Long Short-Term Memory (BiLSTM) network deepen the model’s understanding of sequential data. Experimental results demonstrate the model’s effectiveness in addressing data imbalances and enhancing time-series precision, supporting improved monitoring and control of cement calcination.