id: 02203880
dt: b
an: 02203880
au: Sitompul, Erwin Parasian
ti: Identification of nonlinear dynamic systems using neural networks with
    online parameter adjustment.
so: Berichte aus der Automatisierungstechnik. Aachen: Shaker Verlag;
    Kaiserslautern: Univ. Kaiserlautern, Fachbereich Elektrotechnik und
    Informationstechnik (Diss.) (ISBN 3-8322-4066-7/pbk). vi, 115~p.
    EUR~45.80 (2005).
py: 2005
pu: Aachen: Shaker Verlag; Kaiserslautern: Univ. Kaiserlautern, Fachbereich
    Elektrotechnik und Informationstechnik (Diss.)
la: EN
cc: 
ut: learning algorithm; Levenberg-Marquardt method
ci: 
li: http://www.shaker.de/de/content/catalogue/index.asp?lang=de&ID=8&ISBN=978-3-8322-4066-0
ab: Summary: Neural networks were found to be very powerful when implemented in
    the identification of nonlinear dynamic systems. In such
    implementations, a neural network constructs a nonlinear mapping
    between the input and the output of the system without needing deep
    understanding of the physical interrelationships that describes the
    system. A neural network, as a universal approximator, can model any
    relationship to any degree of accuracy, provided that it has a suitable
    structure and appropriate parameters. The neural network acquires the
    knowledge of the system through a learning process by using measurement
    data from the system and stores the knowledge in its parameters. The
    estimation of these parameters can be considered as an optimization
    problem in an attempt to minimize a given cost function with the
    parameters as the free variables. Thus, there are two factors that can
    be exploited to increase the performance of a neural network in
    fulfilling the task assigned to it: formulating a suitable cost
    function and devising an efficient optimization method to minimize the
    cost function. In the presence of a system change, the knowledge stored
    in the parameters is not up to date and cannot be used to describe the
    system anymore. Devising a way to overcome the performance
    deterioration of neural networks due to system change is found very
    inspiring. In view of these, the main objectives of this thesis are the
    utilization of cost functions to improve the performance of neural
    network as a model of nonlinear dynamic system and the design of
    capable online parameter adjustment scheme that enables the network to
    cope with system change. Prior to doing them, a learning algorithm for
    neural network structure which is suited for multi-step-ahead
    prediction is devised. Beginning with a brief description of the
    approach in nonlinear system identification using neural networks,
    typical structures of neural networks used for the purpose are also
    presented. As the result of surveying for appropriate network
    structures, a specific network structure, the MLP network, is chosen
    for further investigation. The devised learning algorithm combined with
    the implementation of the Levenberg-Marquardt method yields a powerful
    learning algorithm which can produce a network with good
    multi-step-ahead prediction, the MLPER network. Provided with the
    network structures and an effective learning algorithm, the influence
    of cost functions is explored. Two cost functions are analyzed. A new
    exponential quadratic cost function is thoroughly formulated and is
    integrated to the learning algorithm. The utilization of this cost
    function results in a slight performance improvement. A linear
    quadratic cost function originated from robust control theory is taken
    over and the effect on learning result is investigated. The resulting
    networks are proved to be able to deliver rigid responses when the
    learning processes are conducted by using measurement data which are
    contaminated by outliers. The design of an online parameter adjustment
    scheme is done by dividing the network parameters into two parts, the
    long-term and the short-term memory part. The long-term memory part
    contains parameters which are nonlinear to the network output. The
    short-term memory part includes parameters which are linear to the
    network output. Only the parameters of the short-term memory part are
    adjusted. Their linear relationship to the network output allows the
    use of linear optimization methods for recursive estimation. The
    advantage of this approach is that the estimation process is free from
    optimization problem regarding to local minima. By not changing the
    whole network parameters, the stability of the network in delivering
    output while the parameters are adjusted is improved. A parameter
    expansion measure is conducted to increase the degree of freedom
    available for the recursive estimator. At the same time, it also
    increases the dynamic modeling ability of the network, since the new
    network has additional memory to save the data from the past. This
    measure results in two other structure variants, the expanded MLP
    network and the expanded MLPER network. Finally, experimental tests are
    conducted on a mathematical model and two real systems which deliver
    promising results. By using the online parameter adjustment scheme, the
    models can keep delivering accurate output in the presence of system
    change and in the shortcomings of measurement data.
rv: