عنوان مقاله
دیدگاه شبکه عصبی دینامیکی در مدل سازی فرایند غیرخطی
فهرست مطالب
مقدمه
ساختار شبکه عصبی دینامیکی
نتایج شبیه سازی
طراحی کنترلر برپایه مدل
نتیجه گیری
بخشی از مقاله
دیدگاه های گذشته در رابطه با شبکه های عصبی دینامیکی
یکی از معروف ترین کاربردها از فهم بیولوژیکی نسبت به مهندسی، شبکه ی عصبی مصنوعی می باشد. شبکه های عصبی مصنوعی استاتیک یا فیدفروارد (FANNs) به عنوان یک ابزار مفید در کاربرد های سیستم های مهندسی شیمی پدیدار شده است که می توان به این موارد اشاره کرد: (i) مدل سازی جریان حالت پایا، (ii) نقشه کشی جریان حالت پایا، (iii) بهینه سازی جریان حالت پایا.
کلمات کلیدی:
Pergamon Computers chem. Engng Vol. 21, No. 4, pp. 371-385, 1997 Copyright © 1996 Elsevier Science Ltd Printed in Great Britain. All rights reserved PII: S0098-1354(96)00281-5 0098-1354/97 $17.00+0.00 A DYNAMIC NEURAL NETWORK APPROACH TO NONLINEAR PROCESS MODELING ANDRE M. SHAW ~, FRANCIS J. DOYLE III ° and JAMES S. SCHWABER b School of Chemical Engineering, Purdue University; West Lafayette, IN 47907-1283, USA b E.I. DuPont de Nemours and Co., Inc., Wilmington, DE 19880, USA (Received 7 July 1995; accepted 28 December 1995) Abstract--The use of feedforward neural networks for process modeling has proven very successful for steadystate applications, but suitable applications for dynamic systems are still being studied. A novel approach is presented in this paper which uses intrinsically dynamic neurons inspired from biological control systems as the processing elements in network architectures. This results in a network which incorporates dynamic elements with continuous feedback. Two case studies show that the recurrent dynamic neuron network (RDNN) does an excellent job of predicting nonlinearities such as asymmetric dynamic response. In addition, the RDNN significantly outperforms linear models and more traditional neural network models for open-loop simulations. Finally it is shown how this RDNN model can be used in model-based control architectures, such as internal model control. Copyright © 1996 Elsevier Science Ltd Keywords: dynamic neural networks; nonlinear systems; process identification I. INTRODUCTION The opportunities for novel process engineering methodologies inspired by biological control systems has received increasing attention. There remains a great incentive to exploit the highly efficient and robust computational mechanisms of natural neuronal processing systems in an effort to improve the tools of process modeling and control. The motivation for this goal follows from an analysis of the numerous parallels in the dynamic attributes of process and biological systems, as well as the parallels in the performance requirements (Henson et al., 1994; Stark, 1993). Although the details are not well understood, it is acknowledged that biological control systems provide robustly stable control of highly nonlinear plants (Henson et al,, 1994). Furthermore, these control systems function properly under such adverse conditions as major sensor damage and/or loss. It is observed that the control actions involve many regulatory mechanisms operating on different, but relatively short, time scales to achieve the desired response. The central nervous system (CNS) serves as the main control center for these regulatory mechanisms and coordinates the various control activities of biological systems. The fundamental component of the central nervous system is the neuron, which is responsible for the rapid and accurate transfer of information between the CNS and the other parts of the body. Hence, a detailed analysis of neuron functionality--how it encodes and transmits information--will lead to increased insight into the nature of these highly efficient biological computational systems, The coupling of these concepts with engineering principles can enable a reverse engineering of the biological computational elements tbr process modeling and control applications. 1.1. Previous approaches to dynamic neural networks One of the more popular applications of biological understanding to engineering has been the artificial neural network. Static or feedforward artificial neural networks (FANNs) have emerged as useful tools in chemical engineering systems applications including: (i) steady state modeling; (ii) steady state planning and; (iii) steady state optimization (for representative references see MacGregor et al. (1991) on the connection to standard statistical regression tools; Pollard et al. (1991) on process identification and control). All of the above represent highly nonlinear but static (steady state) problems. In reality, chemical process operations are highly nonlinear as well as highly dynamic, and thus networks structures must be modified in other to properly model dynamic systems. FANNs have been used as the static nonlinearity in Hammerstein and Wiener models (Montague et al., 1991; Narenda and Parthasarathy, 1990) to model highly nonlinear, dynamic systems. The FANN is placed in series with a linear dynamic element to capture nonlinear dynamics. There are, however, processes in which the dominant nonlinearities cannot be separated as a distinct static element, thus there is a motivation to pursue a more general approach. One approach for introducing dynamics borrows from classical time series analysis (Morris et al., 1994). The idea is to replace static input/output data with appropriate time histories over a window of discrete times.
