A Soft-Sensing Model for Predicting Cement-Specific Surface Area Based on Inception-Residual-Quasi-Recurrent Neural Networks

The cement specific surface area is a critical indicator of cement quality. Most of the relevant detection instruments are now used for offline detection, which has high cost and poor real-time performance. For this reason, a soft sensing model can be developed to achieve the purpose of predicting the quality of cement. However, due to the data features in the process of cement grinding is complexity and coupling, the prediction accuracy of the universal model is not high and can not achieve good prediction results. Based on the above problems, an Inception-Residual-Quasi-recurrent Neural Networks-based cement specific surface area prediction model is proposed in this paper. The model uses the inception module’s parallel structure to extract a wider variety of data characteristics and enhance the reuse rate of critical features, successfully avoiding poor network prediction caused by discrepancies in convolutional kernel selection in superimposed convolutional layers. In the meantime, distinct convolutional modules use independent spatial aggregation kernels from different channels to improve spatial aggregation and reduce data coupling. Then, the traditional QRNN convolutional layers at the previous moment outputct-1of the hidden layer and the inputxtare replaced by the residual modules. By integrating the inception and residual modules with QRNN, a novel Inception-R-QRNN neural network is constructed. Overfitting is effectively avoided, and generalization ability is improved. The validity of the model is verified by experiments.