Prediction of f-CaO Multi-head DSC LSTM model for cement clinker

Establishing a cement clinker f-CaO prediction model based on convolutional neural networks or LSTM can achieve soft sensing of f-CaO content in cement clinker. However, in cement production, there are many auxiliary variable parameters, and the impact of different auxiliary variables on f-CaO varies at different time periods; And when using the model to extract features, the original time series of all auxiliary variables will be treated as valid inputs for extraction, and the coupling between auxiliary variables will interfere with the training of the model. To address this problem, a Multi head DSC-LSTM cement clinker f-CaO prediction model considering multiple time series is proposed. Firstly, calculate the retention time of the material in each device; Extract time series variables of different lengths according to the device; Finally, multiple parallel working DSC networks were used to extract features from auxiliary variable time series data from multiple devices, and the extracted features were weighted and integrated before being sent to LSTM for prediction. Experiments showed that the Multi head DSC-LSTM model had better prediction accuracy than LSTM and DSC-LSTM models.