Chao Sun

At present, double-hormone artificial pancreas is a hotspot in the research of regulating blood glucose level; however, there are few studies on switching mechanism of the subsystems for double-hormone artificial pancreas, and most of the existing studies rely on the prior knowledge of application fields. A model predictive control algorithm for double-hormone artificial pancreas based on optimal three-branch decision switching rules is proposed to solve this problem. The blood glucose controllers of insulin subsystems, zero infusion subsystems, and glucagon subsystems are constructed using model predictive control algorithms. Then, using the three-branch decision theory, switching rules between the three subsystems are designed, and the optimal switching threshold value can be solved by the mixed fruit fly algorithm. The algorithm can realize automatic acquisition of the threshold in the three-branch decision switching rules and gets rid of the dependence on the prior knowledge in the field of diabetes research. By the end, the UVa simulation platform is used to conduct experiments on 33 virtual patients and the simulation results verify that the proposed algorithm can control blood glucose effectively within the normal range and can achieve satisfactory performabce in error of tracking set point, safety, operation efficiency, group control quality, and system stability.

One of the key indicators for evaluating finished cement products is the cement specific surface area. The soft sensor model for cement specific surface area serves as the foundation for scheduling cement production and is critical for increasing cement quality. To solve the issue of soft sensor models of cement specific surface area caused by the non-linearity, time lag and strong coupling of the cement industry’s big data, the soft sensor model of Time Convolution Network (TCN) and Attention Simple Recurrent Unit (ASRU) Network is proposed.TCN suppresses the issue of feature redundancy due to data coupling. Meanwhile, an improved network structure ASRU is established considering the long time series of process industrial data. ASRU can quickly screen out the higher value information from a vast amount of information to enhance the sensitivity to information. TCN-ASRU can capture the spatiotemporal features in the input data and the dynamic response relationship on the time sequence to solve the problems including time-varying time delay. The model was trained and validated on a cement specific surface area data-set collected. The results demonstrated that the model has the higher prediction accuracy and good generalization ability.

Due to the prominent time-delay problem in the data characteristics of cement grinding process, traditional soft-sensing models are difficult to effectively deal with this problem, leading to low measurement accuracy. We propose a soft-sensing method based on Multi-step Time Extraction Network (MTENet). This method can effectively handle the time-delay problem, thus realizing the accurate soft-sensing of the cement specific surface area. The encoder contains a new time feature extraction unit we proposed, which can effectively solve the time-delay problem of feature data through a clever combination with ConvLSTM. We also introduce parallel units to make the network aware of the surface area values. In the decoder part, we adopt the method of concatenating parallel features to ensure the accuracy of the final decoding result. Compared with some representative models on real data, the experimental results show that MTENet has a higher accuracy in cement specific surface area soft-sensing.

As the most critical equipment in the pre-calcination process of dry cement production, the temperature of the precalciner is an essential factor affecting the quality of cement. However, the cement calcination system is time-delayed, nonlinear, and multi-disturbance, which makes it difficult to predict and control the precalciner temperature. In this study, a deep learning-based Hammerstein model is proposed, and a model predictive control system is built to predict and control the precalciner temperature. In the prediction model, the CNN-GRU network architecture is used to extract the operating states of the precalciner, and an attention mechanism is employed to find and emphasize the important historical information in the extracted states. Then, an ARX model is built to predict the temperature of the precalciner using the extracted operating state information. The complex nonlinear model solution in the control system is formed into a linear control problem and an inverse solution problem. The generalized predictive control (GPC) is used for linear control, and the improved sparrow search algorithm (ISSA) is used for the problem of an inverse solution. Tested with data from a cement plant in Hebei, China, the prediction accuracy of the model proposed in this paper is 99%, and the established control algorithm has less overshoot compared to PID and better stability in anti-disturbance tests. It is demonstrated that the prediction model developed in this study has better accuracy and the control strategy based on this model has good robustness.

Due to the time lag, non-linearity and variable coupling in the process sector, production variables are hard to measure directly and require accurate soft sensor models. In light of this, this research proposes a soft sensor Convolution Fast Recurrent Unit (CNN-FRU) Network model. Convolution is used to get spatial features and get rid of redundancy data in variables. FRU changes the linear coupling constraint within GRU and solves the problem of blocked information flow in GRU. It can be used to extract time-varying delay characteristics from complex processes industry’s data. FRU is compared with GRU and LSTM on a standard data-set to demonstrate the advantages of FRU in terms of training speed and accuracy. As a specific application in the process industry, the proposed soft sensor model is trained and validated on the cement specific surface area data-set. The experimental results demonstrate that the soft sensor model can generalize better.