Neural Network Applications in Control
2: Department of Cybernetics, University of Reading, Reading
3: DaimlerBenz AG, Berlin
The aim is to present an introduction to, and an overview of, the present state of neural network research and development, with an emphasis on control systems application studies. The book is useful to a range of levels of reader. The earlier chapters introduce the more popular networks and the fundamental control principles, these are followed by a series of application studies, most of which are industrially based, and the book concludes with a consideration of some recent research.
Inspec keywords: chemical engineering; fuzzy neural nets; image colour analysis; approximation theory; process control; medical control systems; nonlinear control systems; power control; nonlinear dynamical systems; fuzzy control; haemodynamics; machine control; speech processing; motor drives; health care; state estimation; neurocontrollers; computer vision
Other keywords: neurocontrol; fuzzyneural control; speech application; approximation theory; state estimation; chemical process application; vision application; neurofuzzy adaptive modelling; realtime drive control; intensivecare blood pressure management; system identification; dynamic artificial neural network; nonlinear process control; electric power application; colour application; digital neural networks; nonlinear dynamical process
Subjects: Neurocontrol; Fuzzy control; Control applications in chemical and oil refining industries; Biological and medical control systems; Control of electric power systems; Neural computing techniques; Nonlinear control systems
 Book DOI: 10.1049/PBCE053E
 Chapter DOI: 10.1049/PBCE053E
 ISBN: 9780852968529
 eISBN: 9781849193498
 Page count: 309
 Format: PDF

Front Matter
 + Show details

Hide details

 + Show details


1 Neural networks: an introduction
 + Show details

Hide details

p.
1
–16
(16)
In this chapter, neural networks are discussed. The present field of neural networks links a number of closely related areas, such as parallel distributed processing, connectionism and neural computing, these being brought together with the common theme of attempting to exhibit the method of computing which is witnessed in the study of biological neural systems. A fundamental aspect of artificial neural networks is the use of simple processing elements which are essentially models of neurons in the brain. These elements are then connected together in a wellstructured fashion, although the strength and nature of each of the connecting links dictates the overall operational characteristics for the total network.
 + Show details


2 Digital neural networks
 + Show details

Hide details

p.
17
–32
(16)
When McCulloch and Pitts first studied artificial neural networks (ANNs), their neuron model consisted of binary signals contributing to a sum which was then thresholded to produce the output of the neuron. This model quickly evolved to the well known 'function of weighted sum of inputs' model.The definition 'an interconnected system of parameterised functions' covers many types of ANNs and neuron models.The functions are respectively, 'output a function of the weighted sum of inputs', and 'output a function of the contents of the addressed memory location'.
 + Show details


3 Fundamentals of neurocontrol: a survey
 + Show details

Hide details

p.
33
–66
(34)
The articel gives a concise overview of fundamentals of neurocontrol. Both feedforward and recurrent networks are considered and the foundations of approximation of nonlinear dynamics with both structures are briefly presented. A continuoustime, statespace approach to recurrent neural networks is presented; it gives valuable insight into the dynamic behaviour of the networks and may be fast in analogue implementations. Recent learning algorithms for recurrent networks are surveyed with emphasis on the ones relevant to identification of nonlinear plants. The generalisation question is formulated for dynamic systems and discussed from the control viewpoint. A comparative study of stability is made discussing the CohenGrossberg and Hopfield approaches.
 + Show details


4 Selection of neural network structures: some approximation theory guidelines
 + Show details

Hide details

p.
67
–90
(24)
Control engineers have not been slow in making use of recent developments in artificial neural networks: a pioneering paper was written by Narendra and Partnasarathy and more recent developments are surveyed in this book. Neural networks allow many of the ideas of system identification and adaptive control originally applied to linear (or linearised) systems to be generalised, so as to cope with more severe nonlinearities. Such strong nonlinearities occur in a number of applications e.g. in robotics or process control. Two possible schemes for 'direct' adaptive and 'indirect' adaptive control are shown and other schemes will be found elsewhere in this book, but in this chapter we shall concentrate on the modelling to be carried out by the artificial neural networks.
 + Show details


5 Electric power and chemical process applications
 + Show details

Hide details

p.
91
–108
(18)
If current research efforts worldwide on neural networks are to gain recognition and continue at their current levels, it is essential that theoretical advances are accompanied by industrial applications where the advantages/disadvantages of this new technology can be properly assessed against more conventional techniques. The past few years have seen a marked shift towards practical, as opposed to purely simulation, studies in the field of control and systems engineering, which is a healthy sign both of a maturing technology and successful technology transfer. A good selection of such applications are included in this book and the present chapter contains the results of two studies, one concerned with nonlinear modelling of a 200 MW boiler in an electrical power station [1] the other with inferential estimation of viscosity, a key quality indication in a chemical polymerisation reactor [2, 3].
 + Show details


6 Studies in artificial neural network based control
 + Show details

Hide details

p.
109
–140
(32)
Artificial neural networks can be used as a representation framework for modelling nonlinear dynamical systems. It is also possible to incorporate these nonlinear models within nonlinear feedback control structures. Several possibilities for modelling and control of nonlinear dynamical systems are studied in this chapter. We present case studies illustrating the application of these techniques. A more detailed coverage of the material in this chapter may be found in the reviews.
 + Show details


7 Applications of dynamic artificial neural networks in state estimation and nonlinear process control
 + Show details

Hide details

p.
141
–160
(20)
Since the mid1980s interest in a major area of strategic research has emerged: that of artificial neural networks. The much wider availability and power of computing systems, together with new theoretical research studies, is resulting in expanding areas of application. It is particularly significant in these circumstances that the extremely important aspects involved in developing complex industrial process applications is emphasised, especially where safetycritical perspectives are prominent. Additionally, in complex processes it is important to understand that conventional feedforward networks imply that the manipulated process inputs directly affect the plant outputs. This is not true in complex processes where some manipulated inputs affect internal states that go on to affect the system outputs. A further complication in complex industrial processes is the display of direction dependent dynamics. The studies described here focus on the application of a dynamic network topology that is capable of representing the directional dynamics of a complex chemical process. An application of neural networks to the online estimation of polymer properties in an industrial continuous polymerisation reactor is presented. This approach leads to the implementation of an inferential control scheme that significantly improves process performance to marketdriven grade changes. The generic properties of the approach are then demonstrated by transferring the technology to a totally different plant. The application is to the nonlinear predictive control of the pressure of a highly nonlinear, high purity distillation tower.
 + Show details


8 Speech, vision and colour applications
 + Show details

Hide details

p.
161
–178
(18)
In this chapter some examples are presented of systems in which neural networks are used. This shows the variety of possible applications of neural networks and the applicability of different neural network techniques. The examples given also show some of the varied work done in the Cybernetics Department at the University of Reading.
 + Show details


9 Realtime drive control with neural networks
 + Show details

Hide details

p.
179
–202
(24)
Drive control is an important task of the DaimlerBenz subsidiary company AEG DaimlerBenz Industrie. In order to increase the performance of modern drive systems it is advantageous to exploit nonlinear control techniques, e.g. neurocontrol. Within the nonlinear control framework bounds on control, state and output variables can be taken into account. Different control objectives may also be pursued in different operating regions. This is especially important if safety requirements must be met. Furthermore, nonlinear dependencies such as friction, hysteresis or saturation can be included in the mathematicalphysical modelling process and control scheme, if known. On the other hand these effects are very often neglected in drive control in order to design linear controllers, but this often results in poor control. If the mathematical structure (equations) representing a process is not known, very difficult to obtain or just too timeconsuming to evaluate, learning systems may be engaged to improve the modelling process.
 + Show details


10 Fuzzyneural control in intensivecare blood pressure management
 + Show details

Hide details

p.
203
–228
(26)
This chapter addresses three fundamental and important issues concerning the implementation of a knowledge based system, in particular a fuzzy rule based system, for use in intelligent control. These issues are knowledge acquisition, computational representation, and reasoning. Our primary objective is to develop systems which are capable of performing selforganising and selflearning functions in a realtime manner under multivariable system environments by utilising fuzzy logic, neural networks, and a combination of both paradigms with emphasis on the novel system architectures, algorithms, and applications to problems found in biomedical systems. Considerable effort has been devoted to making the proposed approaches as generic, simple and systematic as possible.
 + Show details


11 Neural networks and system identification
 + Show details

Hide details

p.
229
–252
(24)
Neural networks have become a very fashionable area of research with a range of potential applications that spans AI, engineering and science. All the applications are dependent upon training the network with illustrative examples and this involves adjusting the weights which define the strength of connection between the neurons in the network. This can often be interpreted as a system identification problem with the advantage that many of the ideas and results from estimation theory can be applied to provide insight into the neural network problem irrespective of the specific application.
 + Show details


12 Neurofuzzy adaptive modelling and construction of nonlinear dynamical processes
 + Show details

Hide details

p.
253
–292
(40)
This chapter addresses a range of neurofuzzy algorithms that automatically construct parsimonious models of nonlinear dynamical processes. The process dynamics are typically unknown and complex (i.e. multivariate, non linear and time varying) making the generation of accurate models by conventional methods. In these instances more sophisticated (intelligent) modelling techniques are required. Weight identification, known as learning, is achieved by optimising the weights with respect to some error criteria across a set of inputoutput pairs. This set is known as a teaching set and must adequately represent the systems dynamic behaviour. Typically this type of modelling is termed black box modelling where the internal representation does not reflect the behaviour of the physical system.
 + Show details


Back Matter
 + Show details

Hide details

p.
293
(1)
 + Show details
