A Fast and Enhanced Ray Optical Propagation Model for Indoor and Urban Scenarios, Based on an Intelligent Preprocessing

of the Database
Gerd W ole, Reiner Hoppe, and Friedrich M. Landstorfer Institut f ur Hochfrequenztechnik, University of Stuttgart, Pfaffenwaldring 47, D-70550 Stuttgart, Germany e-mail: woele@ihf.uni-stuttgart.de WWW: http://www.ihf.uni-stuttgart.de
Ray-optical propagation models are often used for the prediction of the eld strength (and delay spread) in wireless mobile communication networks. But these models suffer from long computation times. For large areas in urban or indoor scenarios the computation times are in the range of hours which is too long for the planning of mobile radio networks. A new method for the acceleration of ray-optical models is presented in this paper. It is based on a preprocessing of the database and leads to acceleration factors near 1000, reducing the computation time to a few seconds on PCs. The propagation model is implemented for urban and indoor scenarios and very accurate results are obtained with the model. I. I NTRODUCTION Ray-optical propagation models are very often used for the prediction of the eld strength (and also for the delay spread) in indoor and urban scenarios [1], [2], [3]. They are very accurate because they consider waveguiding effects in street canyons (urban) or corridors (indoor) and they include diffraction at wedges (corners). There are two approaches to the ray optical prediction models: ray tracing and ray launching [4]. Both of them have their individual advantages and disadvantages. The ray tracing computes all rays for each receiver point individually and guarantees the consideration of each wall as well as a constant resolution. This individual computation is more time-consuming than the ray launching approach, where the rays are launched from the transmitter with a constant angle increment. But the ray launching might neglect a wall because it is very small and located between two rays. Different approaches to a better resolution with the ray launching were presented in the last years [5], but the ray launching has still kept the disadvantage of a variable resolution depending on the distance to the transmitter. The new approach presented in this paper combines the advantages of both ray optical models and neglects their disadvantages. The eld strength is computed with the Fresnel equations for the reection/transmission and with the GTD/UTD for the diffraction [6]. Also empirical diffraction models are available, because they can be calibrated with measurements [4]. Ray optical models are very time-consuming, because all possible rays must be determined and many reections and diffractions are computed. Especially 3D models consider thou-

Abstract |

sands of rays, but most of them can be neglected because their contributions are very small. Therefore implementations for 2D are also available, but they have a limited accuracy and the computation times are still in the range of hours an a standard PC. Several approaches to an acceleration of the models were presented in the last years and lead to acceleration factors up to 10 [7], [8]. But the computation times of ray optical models (2D and 3D) are currently still in the range of hours if many prediction points (large prediction areas) and interactions (especially diffractions) are computed. II. P REPROCESSING OF THE DATABASE The basic idea of the preprocessing is the reduction of identical operations, because the ray tracing algorithm determines nearly the same rays for neighboring prediction points and for each of these points the same computations are necessary (reection (diffraction) points are on the same walls (wedges) if the receiving pixels are adjacent to each other [9]). The evaluation of these possible rays reected at the same walls leads to similar results considering the same walls for the shadowing. So it is possible to accelerate the computation if the visibility relation between two walls is computed only once and stored in the memory [8]. If there is line of sight between the walls or between parts of the walls, the algorithm will always consider both walls for the computation of a potential reected (diffracted) ray. If there is no line of sight between the walls, no reected ray will be searched. So the computation of the reections (diffractions) depends on the visibility relation between the walls (or parts of the walls). This principle is well-known and very often used for the acceleration of the ray-optical algorithms [10]. In this paper an enhanced approach is presented. All walls are subdivided into tiles and all wedges are subdivided into segments as shown in gure 1. The visibility relations between each tile (segment) and all other tiles (segments) are computed in the preprocessing, because they are independent of the transmitter and receiver locations. For the decision about the visibility relation, the line of sight between the centers of the tiles (or segments) is evaluated. If there is line of sight between the centers, the rays from the center of the rst tile to the corners of the second tile are determined and the projection of the angles of the rays on the rst and second tile are stored together with the visibility relation.

Table II shows the dependency of the computation time on the size of the tiles and segments for two urban test scenarios. Resolution Nancy Stuttgart 5m 90 min 100 min 10 m 20 min 22 min 20 m 5 min 5 min

TABLE II: Computation times for the preprocessing of an area 600 m x 600 m of an urban data base depending on the resolution of tiles and segments (in all cases 5 m resolution for the receiving points) Fig. 1: Tiles and segments of a wall A similar computation for the visibility relations between tiles and segments and between segments and segments is done and also stored in the le of the preprocessing. The angles of the projection are very important, because they dene a range of possible reection (or diffraction) angles for the illuminated tile (or segment). The angle also continues on the neighboring tile, so a very accurate prediction of the rays is possible even if the tiles or segments are large (up to 5 or 10 meters for urban data bases) [9]. A further improvement is possible if the grid of the prediction points is also used in the preprocessing, because the prediction plane can be subdivided into tiles and the visibility relations between the tiles of the prediction grid and the tiles (and segments) of the walls represent the last part of the ray in the direction to the receiver. If different heights for the receiver locations should be computed in the prediction, individual prediction planes for each height must be considered in the preprocessing. If the receiver visibility relations are determined in the preprocessing, the only remaining visibility relations to be computed in the prediction are the ones from the transmitter to the tiles (of walls and prediction grid) and segments (see section III and gure 2). Area 600 x 600 1000 x 1000 1000 x 1000 1500 x 1500 2500 x 3200 Buildings 40 86 120 300 2000 File size 4 MB 6 MB 10 MB 20 MB 50 MB Time 8 min 12 min 25 min 60 min 10 h The computation times and le sizes for indoor data bases depend on the type of the data base (details, number of walls,...). The ofce buildings (IHF in Stuttgart and the building in Vienna, presented in [11]) lead to les smaller than 20 MB and computation times in the range of 2 hours. But these values depend on the size of the tiles and segments. Detailed information about the software features (les, times) can be found in [12]. Further information about the preprocessing, the subdivision of the walls and the parameters for the computation of the visibility relations are given in [9], [13]. III. P REDICTION WITH A PREPROCESSED DATABASE

Nancy Lille Stuttgart Stuttgart Munich

Fig. 2: Tree structure of the visibility relations Figure 2 shows the arrangement of all visibility relations computed in the preprocessing and in the prediction. Only the relations in the rst layer of the tree must be computed in the prediction, all other relations are determined in the preprocessing and can be read from a le. The visibility relations, determined in the preprocessing (all layers except of the rst layer of the tree) are independent of the transmitter location and can be used for all predictions with the same data base. Only the relations in the rst layer of the tree are depending on the location of the transmitter and must be computed in the prediction process for each transmitter location. Due to the small number of visibility relations in the rst layer of the tree, the computation times are very fast. Most time is spent on reading the visibility data from the les. If more than one transmitter is considered at the same time, the

TABLE I: Memory requirements and computation time for the preprocessing of different urban data bases depending on the size of the data base (Pentium II, 266 MHz) Tables I shows the memory requirements and computation times for different urban scenarios. The computation times given in table I are smaller than the computation time of a single prediction for the same area with the standard ray tracing (see table III), because each visibility relation is only computed once in the preprocessing while in the prediction the visibility relation might be considered and computed for many prediction points.

preprocessed data must only be read once and the prediction of the second transmitter is even faster than the prediction of the rst transmitter, because the visibility tree is already in the RAM of the PC. A further advantage of the new approach is obvious. The computation time for the prediction is nearly independent of the size of the prediction area, because the whole tree is computed once for each prediction and all receiver points are included in the prediction. Only the time for computing the eld strength is necessary for the evaluation of a prediction point and this time depends on the location of the point (inside or outside the prediction area). The rays to all preprocessed receiver points are always determined in each prediction. If the size of the prediction area is reduced, a smaller number of prediction points must be considered (more preprocessed prediction points are neglected) and the time for computing the eld strength is reduced. But this part of the computation time is very short compared to the time for determining the rays and therefore the total computation time is nearly independent of the size of the prediction area. The number of interactions inuences the computation time because each new interaction corresponds to a further layer in the visibility tree. Very good results are achieved with a maximum of six interactions (reections, transmission, and diffractions in different combinations with a maximum of two diffractions in each ray). In some cases a higher number of transmissions is very important for indoor computations and this will lead to a longer computation times. In contrast to the indoor model, the urban model does not consider the penetration of walls. But with this exception the same algorithms can be used for indoor and urban models. Area Nancy Stuttgart 400 m x 400 m 2s 2743 s 2s 3127 s 600 m x 600 m 8s 4607 s 6s 5134 s 800 m x 800 m 12 s 11232 s 11 s 13428 s 1000 m x 1000 m 47 s 29548 s 21 s 33541 s

IV. H YBRID PREDICTION MODELS Ray optical propagation models consider a maximum number of reections, transmissions, and diffractions. Due to that limited number, not all prediction points are reached with the ray optical algorithms (especially far away from the transmitter). This remaining part of the pixels is computed with empirical models, based on the direct ray between transmitter and receiver [12]. For urban scenarios the Walsch-Ikegami-COST 231 model is implemented [4], [14] and for indoor environments the model of Motley-Keenan is used [15]. A transition function between the empirical prediction and the ray-optical prediction leads to a smooth transition between the two models [4], [12]. An example for the transition function between the two models is shown in gure 3. Depending on the difference between the ray-optical and the empirical prediction, a weighted sum of both predictions is computed. The functions for the weight factor are either sin2 (x) or cos2 (x), so that the sum of both functions for all possible values of x is always 1. If the ray-optical prediction is obviously higher than the empirical prediction (because of the waveguiding), the ray-optical value is used. If the empirical value is higher, the transition is computed and if the empirical value is very high compared to the ray-optical value, the empirical value is chosen.
Weightfactor
q

Ray Tracing

@ R @

Empirical

0,5

Diffmin

Diffmax

Difference Emp. RT

Fig. 3: Transition between ray-optical and empirical prediction V. T RANSITION BETWEEN URBAN AND INDOOR MODELS If different data bases are used for the indoor and urban scenarios, a very simple interface can be implemented for the prediction model, because the tiles and segments at the walls surrounding the building dene the interface between the two data bases [12]. If the transmitter is located outside the building, all rays and their angles of incidence are stored for each tile of the surrounding walls of the building. The further computation of the indoor propagation is very simple with the indoor tool, because it uses the information of the incident rays on the surrounding tiles and follows these rays on their way through the indoor visibility tree. If the transmitter is placed inside the building, all rays reaching the tiles of the surrounding walls are stored and followed later with the urban tool on their way through the urban visibility tree.

TABLE III: Computation times for the prediction with preprocessing of the data base (written in italic) compared to the conservative 3-D Ray Tracing in two urban test scenarios The computation times for different urban scenarios are compared to an accelerated 3D model [10] and presented in table III. They are gained with a maximum of 4 interactions (all combinations of reections and diffractions with max. two diffractions). Indoor scenarios are computed with similar acceleration factors compared to rigorous 3D ray tracing models and the factors are in the range of 10000 for complex buildings (see gure 7) [12]. The new approach combines the accuracy of the ray tracing with the idea of the ray launching. Like the ray launching, the new model follows all rays from the transmitter to the receiver points. But in contrast to the ray launching, the accuracy and the resolution are very high, because all rays and their points of interaction are determined in the preprocessing with the ray tracing algorithm.

VI. C OMPARISON WITH MEASUREMENTS A comparison with measurements shows the performance of each propagation model. Many measurements in different scenarios (indoor and urban) were therefore used for the analysis of the performance of the new model [1], [9], [12], [16]. A. Urban Scenarios To show the accuracy for urban environments the wellknown scenario in Munich (Germany) was used [1]. Figure 4 presents the the hybrid prediction for the GSM network (900 MHz) and gure 5 shows the difference between the prediction and the measurement for the same scenario.

Route 0 Route 1 Route 2

Only Ray Tracing Mean StandardError deviation 0.5 dB 8.0 dB 0.1 dB 5.3 dB -0.6 dB 7.5 dB

Hybrid model Mean StandardError deviation 1.3 dB 6.7 dB 0.8 dB 5.0 dB -0.2 dB 6.8 dB

TABLE IV: Accuracy of the prediction in Munich

Fig. 6: Distribution of the error between prediction and measurement B. Indoor scenarios Fig. 4: Prediction for the scenario in Munich (Germany) with ray tracing (frequency 900 MHz) For indoor environments different benchmarks with measurement campaigns were carried out in different types of buildings. New ofce buildings like the University of Stuttgart [16], older ofce buildings like the University of Vienna [17] and very old buildings like the Marconi-Villa in Bologna [8] were used for the comparison. The results concerning accuracy and performance were compared to other indoor models [11]. Figure 7 shows the result for a new ofce building at the University of Stuttgart. The difference to the measurement (carrier frequency 1800 MHz, transmitted power 30 dBm) is shown in gure 8. The standard deviation is smaller than 6.5 dB and the mean error is 0.2 dB. Further examples and tables concerning the accuracy of the ray tracing model are given in [11].

Fig. 5: Difference between prediction and measurement Table IV shows the accuracy of the propagation model with the hybrid computation mode (as mentioned in section IV) and gure 6 shows the distribution of the error if the prediction is compared to the measurements. The total area (2000 x 2500 meters) is computed in less than 300 seconds on a standard PC [12]. Compared to other approaches [14], very fast and accurate predictions are possible. Further comparisons for the city of Stuttgart (Germany) [9] and Nancy (France) [10] conrm these results. The inuence and improvement due to the hybrid approach (as described in section IV) is also shown in table IV and compared to the standard 3D ray tracing.

Fig. 7: Prediction for an ofce building in Stuttgart

Fig. 8: Difference between prediction and measurement for the ofce building in Stuttgart (see gure 7) VII. C ONCLUSIONS A very fast and efcient propagation model was presented in this paper and the prediction results were compared to measurements. The comparison with measurements shows a very high accuracy of the new model. An empirical model, implemented in a hybrid approach, improves the accuracy and leads to even better results. A very simple interface between indoor and urban propagation models is dened using the visibility information of the preprocessed data bases. With this new prediction model it is possible to reduce the computation times for the planning of mobile radio networks to a few minutes and to increase the accuracy because more interactions can be considered with ray-optical models. R EFERENCES [1] K. Rizk, R. Valenzuela, S. Fortune, D. Chizhik, and F. Gardiol, Lateral, Full and Vertical Plane Propagation in Microcells and Small Cells, in 48th IEEE International Conference on Vehicular Technology (VTC), (Ottawa), pp. 9981003, May 1998. [2] T. Huschka, Ray Tracing Models for Indoor Environments and their Computational Complexity, in IEEE 5th International Symposium on Personal, Indoor, and Mobile Radio Communications (PIMRC), pp. 486 490, Sept. 1994. [3] S. S. Wand and J. D. Reed, Analysis of Parameter Sensitivity in a RayTracing Propagation Environment, in 47th IEEE International Conference on Vehicular Technology (VTC), (Phoenix, AZ), pp. 805 809, May 1997. [4] G. W ole, R. Hoppe, F. M. Landstorfer, and R. R. Collmann, Vergleich deterministischer und empirischer Ausbreitungsmodelle f ur die Planung von Mikrozellen, in ITGWorkshop Wellenausbreitung bei Funksystemen und Mikrowellensystemen, (Wessling), pp. 109116, May 1998. [5] G. Durgin, N. Patwari, and T. S. Rappaport, An Advanced 3D Ray Launching Method for Wireless Propagation Prediction, in 47th IEEE International Conference on Vehicular Technology (VTC), (Phoenix, AZ), pp. 785 789, May 1997.

[6] O. Landron, M. J. Feuerstein, and T. S. Rappaport, A Comparison of Theoretical and Empirical Reection Coefcients for Typical Exterior Wall Surfaces in a Mobile Radio Environment, IEEE Transactions on Antennas and Propagation, vol. 44, pp. 341351, Mar. 1996. [7] A. Glassner, ed., An Introduction to Ray Tracing. San Diego: Academic Press, 1989. [8] C. Carcio, A. Cortina, C. Passerini, and S. Salvietti, Fast Field Prediction Techniques for Indoor Communication Systems, in 2nd European Personal and Mobile Communications Conference (EPMCC), (Bonn), pp. 37 42, Nov. 1997. [9] R. Hoppe, G. W ole, and F. M. Landstorfer, Fast 3D Ray Tracing for the Planning of Microcells by Intelligent Preprocessing of the Database, in 3rd European Personal and Mobile Communications Conference (EPMCC), (Paris), Mar. 1999. [10] G. W ole, B. E. Gschwendtner, and F. M. Landstorfer, Intelligent Ray Tracing A new Approach for the Field Strength Prediction in Microcells, in 47th IEEE International Conference on Vehicular Technology (VTC), (Phoenix, AZ), pp. 790 794, May 1997. [11] G. W ole, P. Wertz, and F. M. Landstorfer, Performance, Accuracy and Generalization Capability of Indoor Propagation Models in Different Types of Buildings, in 10th IEEE International Symposium on Personal, Indoor and Mobile Radio Communications, (Osaka, Japan), Sept. 1999. [12] WINPROP, Software tool (incl. demo-version) for the Planning of Mobile Communication Networks and for the Prediction of the Field Strength in Urban and Indoor Environments. http://winprop.ihf.uni-stuttgart.de, Jan. 1999. [13] R. Hoppe, G. W ole, and F. M. Landstorfer, Schnelles 3 D Ray Tracing f ur die Planung von Mikrozellen durch intelligente Datenbankvorverarbeitung, in ITGWorkshop Wellenausbreitung bei Funksystemen und Mikrowellensystemen, (Wessling), pp. 127134, May 1998. [14] E. Damosso, ed., Digital Mobile Radio: COST 231 View on the Evolution towards 3rd Generation Systems. Bruxelles: Final Report of the COST 231 Project, published by the European Comission, 1998. [15] A. J. Motley and J. M. Keenan, Radio coverage in buildings, Bell System Technical Journal (BTSJ), vol. 8, pp. 19 24, Jan. 1990. [16] G. W ole and F. M. Landstorfer, Field Strength Prediction in Indoor Environments with Neural Networks, in 47th IEEE International Conference on Vehicular Technology (VTC), (Phoenix, AZ), pp. 82 86, May 1997. [17] G. W ole, F. M. Landstorfer, R. Gahleitner, and E. Bonek, Extensions to the Field Strength Prediction Technique based on Dominant Paths between Transmitter and Receiver in Indoor Wireless Communications, in 2nd European Personal and Mobile Communications Conference (EPMCC), (Bonn), pp. 29 36, Nov. 1997.