Study Of Soft Sensor Modeling Method Based On Recurrent Neural Network For Petrifaction Production

With more and more requests on control, computation, energy saving and promote benefits, and security of modern petrifaction corporation, all kinds of measure requirements are increasing day by day. The connotation and extension of production process inspection are more deepen, expanded compared with anciently.First, we need all kinds of parameters information of general control process, what is more important; we need the parameters information closely related to optimal control, such as product quality. On the other hand, the increasingly high requirements on precision, real time character of measurement make it develop from static to dynamic.There are two ways to solve the measure request of modern petrifaction production process: First, following the development idea of traditional inspection technology, realized process parameter's online measure by manufacturing new scale instrument. The other way is using soft sensor technique which is rushed from process control and inspection field recently years to structure soft instrument. This method adopts indirect measure idea, replaces the function of hard instrument with computer soft ware, to realize the estimation of parameters which is difficult to measure directly by calculation.Soft sensor technique is a powerful tool for automatic inspection and process optimal at present. Professor Mcavoy who is a famous expert on process control set it in one of the few general directions of control need to focus on. Soft sensor technique has expansive application foreground, so it achieved abroad attention and recognition in recent years.Based on the deeply analysis for the soft sensor technique and artificial neural network, the article advances a concept which is achieved soft sensor modeling for petrifaction production process using recurrent neural network, and proposes a modified neural network model named HF Elman neural network, moreover, the article constructs corresponding chaos training algorithm. Then, the article applied the modified neural network model to soft sensor modeling of Rectifying Column. Simulation results show that the modified model and the algorithm has improved learning efficiency availably, advanced training speed and forecast precision.Main results and contributions of this article are as follows:1. Several soft sensor models and neural network technique are analyzed and studied systemically, and then select soft sensor model based neural network modeling of Rectifying Column which is a continuous, complicated and time-delays process.2. Aimed at the shortcomings of static feedforward network in deal with dynamic problem, soft sensor method based on dynamic neural network is proposed. And advanced a new modified Elman neural network, named Hybrid Feedback Elman neural network, which is based on the studied of typical dynamic neural network, named Elman neural network, then proves the new model's stability from the mathematical.3. A concept which is used chaos mechanism to construct two chaotic optimize algorithm is proposed, and then training the weights to deal with the limitations of traditional Back Propagation algorithmic. Simulation results show that the algorithm is effective.4. The article studied craftwork flow of ethylene product line, and pretreated the data collected from the scene, and then founded the soft sensor models of Rectifying Column. Simulation results show that the modified model and algorithm can meet the requirements for production process real-time monitoring of product quality well, and then laid the foundation for the optimal control of enterprises.