Vilović, Ivan
Predviđanje jakosti elektromagnetskog polja u bežičnim lokalnim mrežama zasnovano na neuronskom modelu i optimizaciji rojem čestica : doktorska disertacija / Ivan Vilović ; [mentor Robert Nađ] - Zagreb : I. Vilović ; Fakultet elektrotehnike i računarstva, 2008. - iii, 155 str. : ilustr. u bojama ; 30 cm + CD
Bibliografija str. 148-151
Ova disertacija se bavi problemom rasprostiranja elektromagnetskog polja u zatvorenom prostoru, gdje je vrlo teško
dobiti toènu razdiobu polja. Kvalitetnu vezu s okosnicom komunikacijskog sustava omoguæuju pristupne toèke, koje
trebaju biti pažljivo rasporeðene kako bi prostor bio pokriven odgovarajuæom snagom signala. Opæenito razlikujemo
jednostavne i složene prostore. Složeni prostori imaju neparalelne i hrapave zidove s nepoznatom dielektrièkom
konstantom u odnosu na geometrijski pravilne, jednostavne prostore.
Razdioba polja kod jednostavnih prostora se može utvrditi nekom od empirijskih ili deterministièkih metoda.
U našem sluèaju korištena je Motley-Keenan i metoda slijeðenja zrake u svrhu predviðanja snage polja u bilo kojoj
toèki jednostavnog prostora. Bolji rezultati se postižu, ako se elektromagnetski parametri zidova dobiju mjerenjem.
U tu svrhu razvijena je nedestruktivna mjerna metoda zasnovana na mjerenjima koeficijenata refleksije i prijenosa
u slobodnom prostoru. Iz dobivenih rezultata izluèena je kompleksna dielektrièka konstanta.
Navedene metode, praktièki, nije moguæe primijeniti na složeni prostor, pa je u našem sluèaju primijenjen neuronski
model za predviðanje snage signala. Kao rezultat istraživanja upotrijebljen je višeslojni perceptron za konfiguraciju
mreže. Ulazi u neuronsku mrežu su koordinate položaja pristupnih i prijamnih toèaka, a izlaz je odgovarajuæa snaga polja.
Levenberg-Marquardt algoritam s Bayesovom regulacijom je odabran za uèenje neuronske mreže, kao rezultat istraživanja tri
razlièita algoritma uèenja. Neuronski model je testiran na stvarnom složenom prostoru, èija geometrijska i konstrukcijska
složenost onemoguæuje primjenu bilo koje druge metode. Neuronska mreža je obuèavana i testirana s izmjerenim snagama polja
na raznim toèkama prijama. Dobiveni rezultati potvrðuju ispravnost pristupa.
Osim za predviðanje razdiobe polja, neuronska mreža je upotrebljena i za odreðivanje optimalnog položaja pristupne toèke.
Optimizacijski postupak je proveden algoritmom zasnovanim na roju èestica (PSO). Rezulati su usporeðeni s vrijednostima
dobivenim algoritmom zasnovanim na mravljoj koloniji i genskim algoritmom. Algoritam zasnovan na roju èestica daje toènije
rezultate i brže se izvodi na raèunalu od algoritma zasnovanog na mravljoj koloniji, a jednako je toèan kao i genski algoritam.
Ključne riječi:
Jednostavni prostor, složeni prostor, bežièna lokalna mreža, rasprostiranje elektromagnetskog polja,
pristupna toèka, slijeðenje zrake, koeficijent refleksije, koeficijent prijenosa, kompleksna dielektrièka konstanta,
neuronska mreža, Levenberg-Marquardt algoritam s Bayesovom regulacijom, optimizacijski algoritam roja èestica,
optimizacijski algoritam mravlje kolonije, genski algoritam. This dissertation deals with an indoor propagation problem where it is difficult to rigorously obtain the field
strength distribution. Access points need to provide good link to the communications backbone of the system.
They need to be positioned carefully so that they cover the building with appropriate signal level. Commonly
environments can be distinguished as simple or complex ones. The complex environments include non-parallel
and non-smooth walls with unknown permittivity.
The field strength distribution in the simple environments with parallel walls can be determined by some empirical
or deterministic method. Motley-Keenan and ray tracing methods are used to predict field strength at any receiving
point of the simple environment. The better results are obtained with ray tracing method when the values of electromagnetic
parameters of the walls are obtained by measurements. A free space non-destructive method is introduced for measurement of
reflection and transmission coefficients and complex dielectric constant extraction.
Application of these methods to the complex environment is very difficult with not accurate results. Absence of real accurate
method for the signal strength prediction in indoor environment enables usage of the neural network methods in this area.
The neural modeling process includes theoretical and experimental investigations that result in the model based on multilayer
perceptron. Inputs are the positions (coordinates) of the access points and of the receiving points, while the output has one
neuron to obtain relevant signal strength level. As a training rule we have selected the algorithm that updates the weight and
the bias values according to Levenberg-Marquardt optimization model with Bayes regularization. This choice is a result of
extensive investigation where neural network architecture and three different learning algorithms have been analyzed.
The selected model is tested at particular building environment, such that it's geometrical and construction complexity
makes the application of any analytical method to be very difficult. The neural network is trained and tested with measured
field strength at various receiving points. The results are very promising.
Such trained neural network is used for predicting the field strength distribution as well as for prediction of the optimum
access point position. The optimization process for optimal access point position is performed with the PSO algorithm which
results are compared with results of Ant Colony Optimization (ACO) and Genetic algorithm. The results show PSO as faster and
more accurate algorithm in comaprison with ACO algorithm, but equal accuarte as genetic algorithm.
Keywords:
Simple environment, complex environment, Wireless Local Area Network (WLAN), electromagnetic propagation,
access point, ray tracing, reflection coefficient, transmission coefficient, complex dielectric constant,
neural network, Levenberg-Marquardt algorithm with Bayes regularization, particle swarm optimization algorithm,
ant colony optimization algorithm, genetic algorithm.
621.39 537.8
Predviđanje jakosti elektromagnetskog polja u bežičnim lokalnim mrežama zasnovano na neuronskom modelu i optimizaciji rojem čestica : doktorska disertacija / Ivan Vilović ; [mentor Robert Nađ] - Zagreb : I. Vilović ; Fakultet elektrotehnike i računarstva, 2008. - iii, 155 str. : ilustr. u bojama ; 30 cm + CD
Bibliografija str. 148-151
Ova disertacija se bavi problemom rasprostiranja elektromagnetskog polja u zatvorenom prostoru, gdje je vrlo teško
dobiti toènu razdiobu polja. Kvalitetnu vezu s okosnicom komunikacijskog sustava omoguæuju pristupne toèke, koje
trebaju biti pažljivo rasporeðene kako bi prostor bio pokriven odgovarajuæom snagom signala. Opæenito razlikujemo
jednostavne i složene prostore. Složeni prostori imaju neparalelne i hrapave zidove s nepoznatom dielektrièkom
konstantom u odnosu na geometrijski pravilne, jednostavne prostore.
Razdioba polja kod jednostavnih prostora se može utvrditi nekom od empirijskih ili deterministièkih metoda.
U našem sluèaju korištena je Motley-Keenan i metoda slijeðenja zrake u svrhu predviðanja snage polja u bilo kojoj
toèki jednostavnog prostora. Bolji rezultati se postižu, ako se elektromagnetski parametri zidova dobiju mjerenjem.
U tu svrhu razvijena je nedestruktivna mjerna metoda zasnovana na mjerenjima koeficijenata refleksije i prijenosa
u slobodnom prostoru. Iz dobivenih rezultata izluèena je kompleksna dielektrièka konstanta.
Navedene metode, praktièki, nije moguæe primijeniti na složeni prostor, pa je u našem sluèaju primijenjen neuronski
model za predviðanje snage signala. Kao rezultat istraživanja upotrijebljen je višeslojni perceptron za konfiguraciju
mreže. Ulazi u neuronsku mrežu su koordinate položaja pristupnih i prijamnih toèaka, a izlaz je odgovarajuæa snaga polja.
Levenberg-Marquardt algoritam s Bayesovom regulacijom je odabran za uèenje neuronske mreže, kao rezultat istraživanja tri
razlièita algoritma uèenja. Neuronski model je testiran na stvarnom složenom prostoru, èija geometrijska i konstrukcijska
složenost onemoguæuje primjenu bilo koje druge metode. Neuronska mreža je obuèavana i testirana s izmjerenim snagama polja
na raznim toèkama prijama. Dobiveni rezultati potvrðuju ispravnost pristupa.
Osim za predviðanje razdiobe polja, neuronska mreža je upotrebljena i za odreðivanje optimalnog položaja pristupne toèke.
Optimizacijski postupak je proveden algoritmom zasnovanim na roju èestica (PSO). Rezulati su usporeðeni s vrijednostima
dobivenim algoritmom zasnovanim na mravljoj koloniji i genskim algoritmom. Algoritam zasnovan na roju èestica daje toènije
rezultate i brže se izvodi na raèunalu od algoritma zasnovanog na mravljoj koloniji, a jednako je toèan kao i genski algoritam.
Ključne riječi:
Jednostavni prostor, složeni prostor, bežièna lokalna mreža, rasprostiranje elektromagnetskog polja,
pristupna toèka, slijeðenje zrake, koeficijent refleksije, koeficijent prijenosa, kompleksna dielektrièka konstanta,
neuronska mreža, Levenberg-Marquardt algoritam s Bayesovom regulacijom, optimizacijski algoritam roja èestica,
optimizacijski algoritam mravlje kolonije, genski algoritam. This dissertation deals with an indoor propagation problem where it is difficult to rigorously obtain the field
strength distribution. Access points need to provide good link to the communications backbone of the system.
They need to be positioned carefully so that they cover the building with appropriate signal level. Commonly
environments can be distinguished as simple or complex ones. The complex environments include non-parallel
and non-smooth walls with unknown permittivity.
The field strength distribution in the simple environments with parallel walls can be determined by some empirical
or deterministic method. Motley-Keenan and ray tracing methods are used to predict field strength at any receiving
point of the simple environment. The better results are obtained with ray tracing method when the values of electromagnetic
parameters of the walls are obtained by measurements. A free space non-destructive method is introduced for measurement of
reflection and transmission coefficients and complex dielectric constant extraction.
Application of these methods to the complex environment is very difficult with not accurate results. Absence of real accurate
method for the signal strength prediction in indoor environment enables usage of the neural network methods in this area.
The neural modeling process includes theoretical and experimental investigations that result in the model based on multilayer
perceptron. Inputs are the positions (coordinates) of the access points and of the receiving points, while the output has one
neuron to obtain relevant signal strength level. As a training rule we have selected the algorithm that updates the weight and
the bias values according to Levenberg-Marquardt optimization model with Bayes regularization. This choice is a result of
extensive investigation where neural network architecture and three different learning algorithms have been analyzed.
The selected model is tested at particular building environment, such that it's geometrical and construction complexity
makes the application of any analytical method to be very difficult. The neural network is trained and tested with measured
field strength at various receiving points. The results are very promising.
Such trained neural network is used for predicting the field strength distribution as well as for prediction of the optimum
access point position. The optimization process for optimal access point position is performed with the PSO algorithm which
results are compared with results of Ant Colony Optimization (ACO) and Genetic algorithm. The results show PSO as faster and
more accurate algorithm in comaprison with ACO algorithm, but equal accuarte as genetic algorithm.
Keywords:
Simple environment, complex environment, Wireless Local Area Network (WLAN), electromagnetic propagation,
access point, ray tracing, reflection coefficient, transmission coefficient, complex dielectric constant,
neural network, Levenberg-Marquardt algorithm with Bayes regularization, particle swarm optimization algorithm,
ant colony optimization algorithm, genetic algorithm.
621.39 537.8