Dominant Path 1

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 5

DOMINANT PATHS FOR T D STRENGTH PREDICTION

G. Wolfle and E M. Landstorfer


Institut fir Hochfrequenztechnik, University of Stuttgart,
Pfaffenwaldring 47, D-70550 Stuttgart, Germany
e-mail: woelfle@ihf.uni-stuttgart.de
W.http:Nwww.ihf.uni-stuttgart.de

Abstract - An algorithm for the determination of the more interactions, they can generally be neglected because of
dominant paths for indoor wave propagation is presented. their higher attenuation. The new approach presented in this
The algorithm computes a tree of the relations between the paper introduces dominant paths to describe all rays passing the
rooms inside the building and the branches of the tree are same rooms and walls. The dominant paths for the scenarios in
used for the determinationof the dominant paths. Based on figure 1 are given in figure 2.
these dominant paths, three different prediction models are
presented and compared with one another and with mea-
surements. Two of the three models are based on neural
networks, trained with measurements and the third model
is an empirical model. With the neural prediction models a.
good generalizationis achieved and they are very accurate
in buildings not used for the training of the neural network.

I. INTRODUCTION
Fig. 2: Multipath propagation: Representation of multi-
II I
paths by a single dominant path
There are basically two different approaches to the predicti
of the electric field strength inside buildings, both of which have The second effect of multipath propagation is shown in figure
their individual disadvantages. Empirical models, based on the 3. There are different dominant paths, passing different rooms
regression of data gained in measurement campaigns, are very and penetrating different walls. These paths do not contribute
fast but not very accurate [ 13. Deterministic models (ray-optical equally to the total field strength and cannot be represented by
models) depend heavily on the accuracy of the data base [3] and a single path.
are very time-consuming [2].
A new approach is given by prediction models based on ar-
tificial neural networks. In these models, the neural networks
are trained with measured data and adapt their parameters to
approximate the measured field strength [4]. Especially for in-
door environments some successful improvements of empirical
models were developed [ 5 ] , but they are based on the direct ray
and do not consider multipath propagation.

11. ANALYSISOF MULTIPATH PROPAGATION Fig. 3: Multipath propagation: Different dominant paths
The difference of the path losses between two dominant
paths for the same receiver point is often greater than 6 dB.
These paths with higher attenuation can be neglected for the
computation of the over-all field strength, because their con-
tribution is very small as compared to the main path. In most
cases it is only necessary to determine the main dominant path
to get an accurate prediction. More than one or two dominant
paths are necessary in very particular situations, but the error by
Fig. 1: Multipath propagation: Possible rays
neglecting the third or fourth path is generally very small.
If multipath propagation with indoor scenarios is analyzed, A further improvement of the approach with dominant paths
two different effects must be considered. As shown in figure 1, as compared to deterministic models consists in the fact that
different rays may reach the receiver passing the same sequence small tolerances in the locations of the walls have only a lim-
of rooms and penetrating the same walls. The contributions of ited influence on the paths, because in contrast to the ray-optics
those rays which offer the same number of interactions to the no exact location of the reflection and diffraction points is nec-
over-all field strength are very similar and if other rays show essary for the determination of the path.

0 1998 EEE
0-7803-4320-4/98/$5.00 552 VTC '98
111. DETERMINATION
OF THE DOMINANT PATHS ing rooms lead to a new branch in the tree, except if the wall is
already used in this path in a layer above.
A. Location of rooms After the determination of the tree, the dominant paths be-
tween the transmitter and the receiver can easily be computed,
The data base of a building generally contains only infor- because the tree represents in its branches all possible domi-
mation about the location and the material of the walls but no nant paths and the sequence of rooms passed. If the receiver is
information about rooms. For determining the dominant paths,
a detailed information about the location of the rooms is manda-
located in room
m.
a the fxee must only be examined for room
The corresponding dominant path can be determined by fol-
tory, however.
A new algorithm for the computation of the location of rooms
a
lowing all branches from room back to the root of the tree.
Further information abont the tree and the determination of the
was implemented in 2D and 3D. While the faster 2D-algorithm paths is given in [ 6 ] .
can be used in cases where only a single floor in considered, the
3D-algorithm is necessary for multi-floor buildings. C. Combination of rooms
In a first step the intersections of lines symbolizing the walls
are computed, Thus the corners of rooms are found (2D). In 3D For determining dominant paths the same algorithm should
the intersecting lines define wedges. The second step combines be applicable to situations where a path passes through a se-
walls (or wedges) to rooms. quence of rooms or where: transmitter and receiver are located in
The determination of the rooms includes an analysis of the the same room. If a path passes through a sequence of rooms, all
neighboring rooms. For each room all neighboring rooms and rooms considered are combined to form a new single room by
walls coupling to these rooms are determined. The space around erasing all coupling walls in the branch representing the dom-
the building is also considered as a room. inant path in the room-tree (see figure 5 ) . This combination
of different rooms to a new room is shown in figure 6 for the
B. Treefor the room-relationship two examples given in figure 1. Now the path inside the new
room can be computed by using the algorithm described in the
The information about the neighboringrooms is used to com- following section.
pute a tree of the room-structure of the building. For the build-
ing presented in figure 4 the tree of the room-structure is shown D. Determination of the path inside a room
in figure 5.
If transmitter and receiver are located in the same room, the
dominant paths inside this room are determined with the infor-
mation about the convex comers as shown in figure 6 .

Fig. 4: Room-numbers (left) and coupling walls (right)

Fig. 6: Convex comer!; (or wedges) of a room and their


influence on the determination of the paths
Layer 2
Each room is described by its surrounding walls. For the sake
Layer 3 ............ of simplicity the following explanations refer to the 2D-cas.e.
In the 2D-case all convex comers of the room get a different
number.
Fig. 5 : Tree of the room-structure presented in figure 4 There are two different scenarios for the determination of the
path between the transmitter and the receiver: Line of sight and
The root of the tree corresponds to the room in which the obstructed line of sight. The first case is very easy because the
transmitter is located. The first layer contains all neighboring dominant path is the direct ray between transmitter and receiver.
rooms and if there is more than one coupling wall between the In obstructed line of sight the dominant path must lead via con-
room of the transmitter and the neighboring rooms, the neigh- vex comers to the receiver. For the determination of the path
boring room is placed in the first layer as many times as there a tree with the convex comers is computed. All comers visi-
are coupling walls between the two rooms. All further layers are ble from the examined comer are new branches in the tree. As
determined in a similar way. All coupling walls to the neighbor- shown in figure 7, the corner-tree starts with the comers visible

0-7803-4320-4/98/$5.00 0 1998 JEEE _551


__ VTC '98
from the transmitter. The receiver is also included in the tree. Different paths are determined for each receiver point and the
Each time the receiver is found in the tree, the comers along the path with the smallest total attenuation L is chosen and called
path can be determined by following the branches back to the minimum-loss dominant path (MLDP). It is possible to deter-
transmitter. mine different types of dominant paths by adjusting the three
weighting factors WFS, W T and WI. In contrast to the model
described in [6], only one criterion ( W F S = W T = W I = 1) is
chosen for the models presented in this paper.
Two of the three prediction models presented in this paper
need only the minimum-loss dominant path (MLDP) for the
prediction of the field strength. Alternative paths with higher
attenuation values are only considered in the third model, if
the difference between their attenuation L and the attenuation
Fig. 7: Relations between the convex corners of the room LMLDPof the MLDP is smaller than a definable threshold.
Iv. PREDICTION OF THE FIELD STRENGTH
E. Selection of the minimum-loss dominant path (MLDP)
Three different prediction models are presented in this pa-
The algorithm for the determination of the dominant paths per and described in the following sections. The first two mod-
leads to more than one solution [6]. But in most cases only one els are based on the minimum-loss dominant path (MLDP), the
solution with the smallest path loss is necessary for an accurate last one needs further information about alternative paths (ADP
prediction. This most important path is called minimum-loss model).
dominant path (MLDP). The MLDP is chosen by utilizing the
user-definable criterion L: A . Empirical prediction of the field strength with MLDP

L = W F S * LFS + W T * LT + W I * LI (1)
The computation of the path loss of the empirical MLDP
model is given in equation (5).
L represents the empirical determination of the total attenuation
along the path and consists of three different parts: LE = LFS + L T + LO (5)

Free space attenuation LFS The offset LO can be used for the calibration of the model.
The free space loss LFS depends on path length 1 and fre- This model is nearly similar to that of Motley and Keenan [I],
quency f: but in contrast to the latter this model is based on dominant
paths and not on direct rays. The improvement gained with us-
1 ing dominant paths can be seen by comparing with results ob-
LFs/dB = -27.56 dB + 20 log f
MHz
+ 20 log -
m tained with the model of Motley-Keenan, as shown in figure
12.
Transmission loss LT
The accumulated transmission loss of all walls passed is re- B. Parameters of the dominantpath for the neural networks
presented by LT. If the path intersects n walls with their
individual transmission losses La, the total transmission loss For the computation of the field strength with neural net-
LT is computed from works, the parameters of the minimum-loss dominant path must
be determined. Because the dominant paths represent a group
Nr of nearly similar rays between transmitter and receiver (see sec-
LT = CL^ (3) tion 11), all relevant parameters of these rays governing propa-
i=l gation should be considered in the description of the dominant
path. Very good results have been obtained with the following
Interaction loss LI parameters:
Changes in the direction of the path as given by angle a, at
point i represent an additional loss. All angles CY( are accu- 0 Free space attenuation L F Salong the path
mulated and normalized with a factor (YLto give the interac- As described in equation (2).
tion loss LI.
. NI e Transmission loss LT
The transmission loss is computed with equation (3).
(4)
0 Interaction loss L I
Best results are gained with Q L = 5" . . .8". If some walls The accumulated angles cy2 of changes in direction are nor-
are supporting the changes in the direction (waveguiding) at a malized with equation (4).
specific point, the contribution at this point to the path loss is e Waveguiding along the path
reduced, depending on the support of the walls as described Each dominant path represents different rays. All of them
in [5] for the transmitter environment. are guided by reflections at the walls or by diffractions at the

0 1998 IEEE
0-78034320-4/98/$5.00 554 VTC '98
corners of wedges in the same direction (see section 11). To input neurons represent thle normalized difference of the param-
include all these rays in the prediction, they are described eters (path length, transmission loss, interaction loss and wave-
by the new parameter waveguiding. The waveguiding of the guiding) between the a1te:mative and the minimum-loss domi-
walls depends on their material (reflection loss), their orien- nant path (MLDP).
tation (reflection angle) and the distance between the walls In contrast to the model described in [6], the relative and
and the path [4]. These three parameters are combined to not the absolute parameters of the altemative paths are used. A
give the parameter waveguiding, as described in [4]. relative description is much better because the neural network
0 Local reflectors and shielding effects at the transmitter site must be trained with measurements and if the alternative path
The computation of these local effects is described in [5]. is described with absolute values, more combinations of input
values (measurements) for the neural network are necessary to
0 Local reflectors and shielding effects at the receiver site achieve a good generalization.
The advantage of this model is obvious, if the results in fig-
For each dominant path all mentioned parameters are gained ure 11 are compared with the results in figure 10. In shadowed
from a vector oriented data base and are normalized for the input regions often two paths halve nearly the same contribution to the
of the neural network. The prediction of the field strength is field strength, but the first and the second model use only the
based only on these parameters. MLDP for the prediction. This model considers also the alter-
C. Neural network prediction of thejeld strength with MLDP native path and the predicltion is more accurate. But the compu-
tation time for the ADP model increases because all alternatives
In this model an artificial neural network is used for the com- and their relative description to the MLDP wmust be computed.
putation of the field strength. The parameters of the dominant
path (see section 1V.B) are combined with a multi-layered feed- E. Training patterns for the neural networks
forward perceptron. The backpropagation algorithm was used
The neural network models were trained with measurements
to train the network with measurements. Figure 8 shows the
structure of the neural network. obtained at the University of Stuttgart. The field strength of
more than 5000 receiver points and 15 transmitter positions in
R.- w. w.
~ . e r m . waw-r~nallon~ . l l ~ d different environments was measured to get enough data for the
lb”#. mm. Spnm mkekm #We An#s W .R W .
1 1 1 1 I 1 1 1 training and the validation of the neural networks.
To improve the generalization capability, not all possible pat-
tems were used. If each pattern is interpreted as a vector in
vector space, the Euclidian distance between the patterns can
be computed. As patterns with very small distances contain
nearly the same information only one of them is necessary for
the training. The details of the algorithms for the selection of
the pattems are described in [5] and [6].

‘YP-
.t
v. PRElDICTlON RESULTS
A comparison between the new dominant path models and
Fig. 8: Topology of the neural network for the prediction measurements is shown in figures 9 to 1 1 . In addition to pre-
senting the accuracy of the three models, these figures also show
All training patterns of the network are gained from measure-
the dominant paths for different scenarios.
ments at the University of Stuttgart. For each measured point
Compared to standard empirical models [l], the new mod-
inside the building all parameters of the dominant paths (input
els are more accurate and compared to deterministic models [2]
values of the neural network) are determined and stored in the
they are very fast because 1 he computation time for whole build-
training pattern together with the measured field strength.
ings (2000 pixels) is smaller than 10 minutes on a workstation.
The improvement of the neural networks compared to the
A comparison between the measurements and the new
empirical model are obvious from figure 12, especially in shad-
owed regions and in long corridors with waveguiding effects. MLDP-empirical-predictlion model is given in figure 9 (Mean
error = 5.8 dB, standard deviation = 9.9 dB). In contrast to the
D.Neural network prediction of thejeld strength with ADP standard empirical predict.ion models (Motley-Keenan model:
mean error = 24 dB, std-dav = 26 dB), the dominant path model
In this model again a neural network is used for the predic- is more accurate in nearly every situation.
tion of the field strength, bu%the prediction is based not only on The prediction of the neural MLDP model (section I W )
the MLDP. Alternative dominant paths (ADP) are also consid- leads to a mean error of -0.4dB and to a standard deviation
ered in the prediction modell. of 3.3 dB for the scenario given in figure 10.
The neural network is similar to the network described in The performance of the alternative path model (ADP model,
figure 8, the only difference is the four additional input neu- see section 1V.D) is presented in figure 11 (mean error = 0.2 dB,
rons for the description of the alternative path. These additional std-dev = 3.6 dB).

0-7803-4320-4/98/$5.00 8 1998 IEEE 555 VTC ‘98


W
1
W

‘I.

Fig. 9: Difference between measurement and prediction


with the new empirical model Fig. 12: Prediction and measurement in corridor 1 of the
IHF office building

VI. CONCLUSIONS
Three different prediction models for the prediction of the
electric field strength inside buildings were presented in this pa-
per and compared with measurements. All of them are based on
dominant paths, a new and very fast approach to the determina-
tion of multipath indoor propagation. The results are highly
accurate, even in buildings not previously used for the training.
ACKNOWLEDGMENTS
The authors want to thank Deutsche Forschungsgemein-
schaft (DFG) for supporting their work.
Fig. 10: Difference between measurement and prediction
with the neural network model REFERENCES
[l] J. M. Keenan and A. J. Motley, “Radio coverage in build-
ings,” In: BTSJ, vol. 8, pp. 19 - 24, Jan. 1990.
[2] T. Huschka, “Ray tracing models for indoor environments
and their computational complexity,” in IEEE 5th Intema-
tional Symposium on Personal, Indoor and Mobile Radio
Communications (PIMRC) 1994, pp. 486 - 490,1994.
[3] V. Degli-Esposti, C. Caciofi, M. Frullone. G. E v a “Sensi-
tivity of ray tracing indoor field prediction to environment
modeling,” in Internat. Conference on Electromagnetics in
Advanced Applications (IECAA), Turin, Italy, Sept. 1997.
[4] G. Wolfle and F. M. Landstorfer, “Field strength predic-
tion with dominant paths and neural networks,” in MIOP
Fig. 11: Difference between measurement and prediction 1997, Sindeljingen, Germany, pp. 216 - 220, Apr. 1997.
with the alternative path model
[SI G. Wolfle and F. M. Landstorfer, “Field strength predic-
A comparison between the three different models and the tion in indoor environments with neural networks,” in
Motley-Keenan model is given in figure 12. For this analysis IEEE 47th VehicularTechnology Conference (VTC) 1997,
the scenario of figure 10 was used and the three predictions for Phoenix, AZ, pp. 82 - 86, May 1997.
corridor 1 are compared with one another. Figure 12 also shows [6] G. Wolfle, F. M. Landstorfer, R. Gahleitner, E. Bonek
the improvement gained by the use of the neural networks for “Extensions to the field strength prediction technique
the dominant path models. based on dominant paths between transmitter and receiver
The dominant path models were also used for the prediction in indoor wireless communications,” in 2nd European
in different buildings not used for the training of the neural net- Personal and Mobile Communications Conference 1997,
work and the models tumed out to be very accurate even in these Bonn, Germany, pp. 10 -18, Sept. 1997.
buildings [6].

0-78034320-4/98/$5.00 0 1998 IEEE 556 VTC ‘98

You might also like