Multi-grain Cascade Recurrent Neural Network for Nonlinear Time-varying Process Soft Sensor Modeling

Soft sensor provides accurate online estimation of primary variables based on historical data of secondary variables. Recurrent neural network exhibits better performance on sequence data. Deep forest shows quite robust performance across nonlinear data from different domains. In this paper, a multi-grain cascade recurrent neural network(gcRNN) is put forward to soft sensor modeling for nonlinear time-varying processes. Based on deep forest framework, LSTM and GRU are chosen as individual learners in multi-grain moving window scanning and cascade structure to deal with sequence data. Performance of the proposed method is tested and verified on two industrial cases with different process characteristics. Compared to the PLS, RF, LSTM and GRU soft sensing approaches, the root mean squared errors (RMSE) of gcRNN are the smallest. The prediction results also illustrated that gcRNN achieves better performance for primary variable quality prediction in nonlinear time-varying processes.