WO2023187447A1

WO2023187447A1 - Method for determining the correct placement of an antenna with radiation pattern prediction

Info

Publication number: WO2023187447A1
Application number: PCT/IB2022/053002
Authority: WO
Inventors: Tiago Manuel ENCARNAÇÃO PEREIRA; Filipa GOMES RAIMUNDO; Miguel LOURENÇO NUNES; Paulo Mateus Mendes; Rui Guilherme COELHO DA SILVA; Franciele CICCONET
Original assignee: Bosch Car Multimedia Portugal, S.A.; Universidade Do Minho
Priority date: 2022-03-29
Filing date: 2022-03-31
Publication date: 2023-10-05

Abstract

The present invention describes a method for determining the correct placement of an antenna in a support structure, for example, of a vehicle. The technology herein disclosed, consists on the position optimization of an antenna located in a fixed platform resorting to the use of Bayesian Optimization in conjunction with an electromagnetic solver and a fully convolutional neural network (U-Net), which will determine the fitness function through electromagnetic simulations of the behaviour of the antenna in a given position.

Description

DESCRIPTION

"Method for determining the correct placement of an antenna with radiation pattern prediction"

Technical Field

The present application describes a method for determining the correct placement of an antenna in a support structure .

Background art

For a Global Navigation Satellite System ( GNSS ) proper use in a vehicle , the antenna must be positioned at a point where its radiation pattern is optimal and with the least number of possible interferences .

Manually analysing the behaviour of the antenna in all possible positions can be a waste of computation and time , since within the viable space there are infinite points . Applying optimi zation models with evolutionary algorithms increases the accuracy of the search path, directing the solver to regions with the best radiation pattern . As these algorithms require many evaluation functions in space , it can make their use limited when the time per evaluation is signi ficant .

The positioning of a set of antennas has a direct e f fect on the accuracy of a vehicle ' s location system [ l ] - [ 2 ] and it is then necessary to look for ways to minimi ze interference between them and optimi ze their performance . A comparison of four stochastic models to search for the optimal position of antennas on a platform was carried out in [ 3 ] , showing that evolutionary models , although slower, were more accurate in the optimal location. Genetic algorithms have also been used to optimize the positioning of multiple antennas on a ship's platform in [4] , reducing the amount of time needed to simulate a position according to the superiority of one position at the expense of another, comparing the fitness functions of the chromosomes. For the case of an antenna, characteristic models were applied to determine the position of a monopole antenna on a metal roof of a vehicle, which achieves the highest gain in the horizontal plane [5-6] .

However, evolutionary optimization algorithms usually are used in problems where there is enough time for many model evaluations. But if there is not a closed-form expression for the objective function (black-box) , the optimizing objective functions take a long time, namely minutes, hours, or even days to evaluate each localization. The determination of the correct localization depends on two steps: one is to compute the radiation pattern at each guess position, and the other is the decision on which test positions should be performed the calculations in order to arrive at the desired position. In case we want to avoid many iterations, if there is not access to derivatives and if the search space is not too big, like this antenna placement problem, the Bayesian optimizer is more indicated [7-8] .

This technique was already used for antenna array design as an alternative to other techniques in [9-11] , but not in antenna placement problems yet.

In [6] the position of an antenna is examined using a genetic algorithm and with the XFdtd software, a Finite Difference Time Domain (FDTD) solver. In order to avoid the intense and time-consuming calculations to determine the radiation diagram at each position for the full vehicle model , a calculation based on the prediction of such radiation pattern and on a vehicle subset model is used that , supported by a database of previous radiation patterns , is able to predict the radiation diagram of the full model .

Summary

The present invention describes a method for determining the correct placement of an antenna in a support structure , said placement providing the best radiation pattern of the antenna, comprising the following steps of : random initiali zation of an antenna placement optimi zation with an antenna placement ( Pn) , the antenna placement optimi zation comprising :

- determining which are the ideal candidates for the solution based on the antenna placement ( Pn) ;

- determine the s impli fied Reali zed Gain and Axial Ratio of the antenna placement ( Pn) ;

- predict a Reali zed Gain and Axial Ratio ;

- obtain a fitness function;

- i f a positive solution found then an output result is provided, else , i f the result is not adequate the lookout for ideal candidates for solution will be repeated .

In a proposed embodiment of present invention, the radiation pattern prediction comprises the steps of :

- obtaining a model fitting result based on an acquired segmentation map from a U-Net convolutional neural network; - obtain a embed output training based on the obtained model fitting;

- obtain a predicted reali zed Gain (RG) and Axial Ratio (AR) .

Yet in another proposed embodiment of present invention, the model fitting is configured to determine the machine learning model generali zation rate of similar data from which it was trained based at least in one of an : Kernel filter, a Recti fied Linear Unit activation function, a max pooling operation with stride and Feature Maps .

Yet in another proposed embodiment of present invention, the embed output training comprises the initial data set used for training the U-Net convolutional neural network .

Yet in another proposed embodiment of present invention, the predicted reali zed Gain (RG) and Axial Ratio (AR) comprises simpli fied data ( SD) as input and the precise data ( RD) as output , for each point defined by the antenna placement optimi zation .

Yet in another proposed embodiment of present invention, the segmentation map comprises data from at least one of an elevation angle value ( 0) and az imuth angle ( C|J ) for each antenna placement ( Pn) .

Yet in another proposed embodiment of present invention, the U-Net convolutional neural network comprises input radiation patterns of antenna placements ( Pn) .

Yet in another proposed embodiment of present invention, the input radiation patterns comprise at least a set of magnitude data points that characterise the antenna in each of antenna placements ( Pn) .

Yet in another proposed embodiment of present invention, the U-Net convolutional neural network is trained with the input radiation patterns .

Yet in another proposed embodiment of present invention, the Reali zed Gain and Axial Ratio prediction is obtained from an embed input and output training and predicted reali zed Gain (RG) and Axial Ratio (AR) of a radiation pattern prediction .

General Description

The present invention describes a method for determining the correct placement of an antenna, for example on a support structure , for example , of a vehicle . This method includes a methodology to decrease the time required to compute the radiation pattern at each test position, and a methodology to determine the antenna position that meets the antenna radiation requirements .

An antenna' s radiation pattern depends , besides the antenna characteristics , on its location and the materials of its platform / support structure . This behaviour can be determined by solving Maxwell ' s equations in electromagnetic solvers , where one creates a simulation model as close as possible to the real one . The accuracy of these models depends on many factors such as the details and material composition of the structure where this antenna will be installed . To achieve the most accurate representation, this simulation model should be complex and representative of the real thing, which in turn implies more computational time will be needed to perform the simulation . The correct placement of an antenna on a support structure may also be supported by simulation tools , however electromagnetic simulations of full , or relevant , vehicle models may be very time consuming, and a full simulation may take weeks or months to complete .

To overcome these limitations , one of the solutions herein considered is to simulate a simpli fied model of the vehicle and use this simulation data as input of a convolutional neural network (U-Net ) with a certain accuracy to estimate a precise radiation diagram of the full model of the vehicle , becoming it possible to analyse many more positions and radiation patterns in a much shorter time than solely using a solver . Along with this methodology, further optimi zers are posteriorly added so the search for the best location is found in the fewest possible steps .

The technology herein disclosed, consists on the optimi zation of the position of an antenna on a fixed platform / support structure using Bayesian Optimi zation (BO) as a solution to the problem of maximi zing reali zed gain and minimi zing axial ratio in a GNSS antenna placement , in conj unction with an electromagnetic solver and a fully convolution neural network (U-Net ) , which will determine the fitness function through electromagnetic simulations of the behaviour of the antenna in a given position . The correct definition of the fitness function is important for the optimi zation algorithms to work ef ficiently, which implies that the response provided by the electromagnetic s imulator must reflect , as much as possible , the desired radiation patterns of the antenna, to ensure the ef fectivenes s of the search algorithms . The Bayesian optimizer, supported on the Bayes Theorem, provides a conditional probabilistic method that seeks to draw conclusions using incomplete information, providing a relationship between data/points ( P) and a fitness model/ function (F) . In cases when the objective fitness function F is unknown, a general approach is to create a surrogate model F' that creates an image of the objective function F as more samples are observed. This can be done by different approaches, but the most common is to apply the general Bayesian inference principles to Gaussian Processes (GPs) .

The herein proposed method ensures the required accuracy of the resulting calculations, without spending huge amounts of time in a full car model electromagnetic simulation. It allows to obtain the antenna behaviour in some specific placements, for the full and simplified model of the vehicle, and, based on this, to use machine learning tools to predict precise radiation patterns for the full model, starting from given simplified entries, such as the roof model.

The biggest challenge related to the simulation of the electromagnetic behaviour of an antenna in a vehicle is to model the geometry and define the materials of the vehicle and the antenna. This modelling can entail simplification of the geometry (removing, altering or joining components) ; geometry healing (automatic process that removes CAD incongruities) ; changes in spatial orientation and position (moving or rotating the model according to what is necessary) ; defining the materials for each element (for each component, defining proper and realistic materials) . Since testing the antenna behaviour in all possible locations of the vehicle roof would be nearly impractical , i f not impossible due to the required amount of time needed, the proposed method to obtain the antenna behaviour in some speci fic placements for the complete and simpli fied model of the vehicle , uses machine learning tools to predict precise ( full model ) radiation patterns given simpli fied entries ( roof model only) . Thus , using a search optimi zation algorithm and a simpli fied model of the vehicle geometry allows to obtain faster response times , and then use machine learning to approximate the radiation pattern to obtain a more accurate solution .

Therefore , the antenna placements on the roof of car, both for full and simpli fied vehicle model , are entered into a U- Net Convolutional Neural Network, and a reali zed gain (RG) and axial ratio (AR) are obtained according to the elevation angle 0 value and azimuth angle c|), being the RG and AR expressed in dBi and dB, respectively . The herein used U-Net convolutional neuronal network architecture was modi fied and extended the work with a reduced number of image data and with more accurate image segmentation . The use of U-Net Convolutional Neural Network allows to predict a precise simulation of the full vehicle model , using a partial model together with information from data obtained from the full model , either from simulations or from measurements . The accuracy of the sub-model simulations will tend to increase as the database with full model simulations or measurements are expanded as the methodology is in place . In the proposed embodiment , the U-Net is used to obtain the precise simulated results from a simpli fied model , but other algorithms could also be used, i f they are able to use a combination of results from simpli fied models and data from full models or measurements .

Brief description of the drawings

For better understanding of the present application, figures representing preferred embodiments are herein attached which, however, are not intended to limit the technique disclosed herein .

Figure 1 illustrates the initial data set , comprising a set of thirteen sampled antenna placement points P1~P13 , i . e . , Pn, located at di f ferent distances within a limited area of a vehicle , in this particular case of 1600x1200 centimetres .

Figure 2 illustrates a global overview of the stages comprised in the antenna prediction and positioning optimi zation .

Figure 3 illustrates several stages of a Bayesian Optimi zation method, illustrated by an Unknown function and prior mean function ( a ) ; an Acquisition function and mean update (b ) ; an Algorithm evaluation one ( c ) and an Algorithm evaluation two ( d) .

Description of Embodiments

With reference to the figures , some embodiments are now described in more detail , which are however not intended to limit the scope of the present patent application . The present invention discloses a methodology for predicting radiation diagrams of antennas on fixed surfaces of a vehicle using machine learning techniques . The proposed model allows to solve magnetic simulation problems , both in small or larger areas , particularly suitable for vehicles , which demand a lot of simulation time and resources , especially when combining the radiation pattern calculation with searching for the best antenna position . While the optimi zation model may find optimal point in the smallest number of obj ective function calls , the U-Net further helps reduce this time by predicting a precise model from a simpli fied model faster than a simulator precise model . After this prediction, which is based on a simpli fied model of the vehicle to speed up the calculation time , the exact location for the placement of the antenna on the vehicle is determined in order to satis fy a given radiation characteristic .

The preferred embodiment steps will comprise an electromagnetic simulation, an algorithm control stage and an arti ficial intelligence algorithms stage that will find the correct position . Said embodiment may also use 3D vehicle models and 3D antenna models . The use of supervised machine learning methods aiming to speed-up computation of antenna radiation pattern to infer the outputs on the remaining points with more accuracy and in a faster way that the frequency solver results .

A Bayesian optimi zation model will be added over the U-Net as a way of finding the best position for an antenna . Thus , the U-Net Convolutional Neural Network is intended to replace the electromagnetic simulation of antennas in complete models that require a lot of time , and instead make the prediction from the simpli fied models . The simpli fied simulation results are sent to the U-Net and the prediction results are sent to the Bayesian optimization model for calculation of the objective function.

The prediction model of the precise radiation patterns (full vehicle model) is integrated in a Bayesian optimization (BO) model. This strategy allows to use a less precise antenna model for faster simulation, and with a convolutional neural network the radiation patterns can be predicted, bringing them closer to more accurate models' results (longer simulation) . The criterion chosen to define the best antenna placement is given by the maximization of the difference between the average realized gain, RGi, and the average axial ratio, ARi, bellow the elevation angle 0 < 0max in the position 1, for all azimuth angles predicted by neural network .

Illustrated in Figure 1 is one example of an initial data set, comprising a set of sampled antenna placement points located at different distances within a limited area of a vehicle. For this particular illustration, there are thirteen possible antenna placement points P1~P13, P_n, located at different distances within a 1600x1200 limited area of the vehicle.

The global overview of the stages comprised in the antenna positioning optimization, supported by Figure 2, comprises two major stages, a radiation pattern prediction (200) and an antenna placement optimization (300) .

Generically speaking, the method for determining the correct placement of an antenna, comprises three stages, a training stage, a prediction stage (200) , and an optimization stage

(300) .

In the radiation pattern prediction stage (200) , it is possible to predict, from simplified values, all the precise values of realized gain (RG) and axial ratio (AR) within all the possible existing points of the search surface.

Finally, in the optimization stage (300) , the resulting RG and AR values can be included in the optimization model (e.g. Bayesian Optimization) to search for the best antenna location .

The proposed method for determining the correct placement of an antenna simulates a simplified model of the vehicle with less computational time and makes a prediction of a more complex vehicle model, which presents a much lower simulation computational time than if the most accurate vehicle models were simulated at each searching point.

The radiation pattern prediction (200) is comprised by three stages, a model fitting (201) , a embed input and output training (202) and a predicted realized Gain (RG) and Axial Ratio (AR) (203) .

In an initial phase, interpreted as a training stage, the U- Net convolutional neural network, comprised in the radiation pattern prediction (200) , is data fed by input radiation patterns (101) of antenna location placements (Pn) . The training input antenna P_n radiation patterns (101) of the roof top of the vehicle, for full and simplified vehicle model, are submitted into the U-Net convolutional neural network, e.g., an electromagnetic simulator, in order to obtain a resulting output segmentation map based on elevation angle value ( 0) , ranging from 0 to 180 , and azimuth angle ( c|) ) ranging from 0 to 360 , in units of 1 degree , for each location point ( P_n) .

The U-Net network is a fully convolutional network end to end, containing only convolutional training layers separated into two paths and no dense layer, the reason for which it is able to accept images of any si ze .

When a suf ficient amount of data ( 101 ) is not available to carry out the training and testing of the U-Net convolutional neural network, a possible alternative is to use the data augmentation technique which consists of generating new data using simple techniques such as rotation, reflection, displacement and zoom over the original images , increasing the set of initial samples and avoiding over- fitting . This technique contributes to the model being used in cases where it is not possible to obtain a database large enough for training and testing . After obtaining the RG and AR data of both the simpli fied and precise vehicle models at the points described in Figure 2 , data augmentation techniques are applied . In the U-Net ' s learning process metrics , unlike RG, the AR has little correlation between the simpli fied and the precise models , which af fects the prediction and accuracy of the model a little .

The model fitting ( 201 ) of the radiation pattern prediction ( 200 ) is configured to determine the machine learning model generali zation rate of similar data from which it was trained . The model fitting ( 201 ) configuration depends on the used machine learning model that will be used, which for the proposed U-Net Model is configured with : a 3x3 Kernel filter, a Rectified Linear Unit (ReLU) activation function, and at the end of each step, a 2x2 max pooling operation with stride equal to 2, stride equal to 2, Feature Maps equal to [32, 64, 128] . The model fitting (201) allows therefore to adjust model parameters to improve the precision results.

The embed input and output training (202) of the radiation pattern prediction (200) comprises the initial data set used for training as illustrated on Figure 1. In the proposed example, the data set comprises thirteen sampled antenna placement points P1~P13, i.e., Pn, located at different distances within a limited area of a vehicle, in this particular example 1600x1200, and is configured accordingly the proposed machine learning model, which in this particular case for the U-Net is defined with

Number of Channels = 2 (corresponds to Realized Gain and Axial Ratio)

Number of Outputs = 2 (corresponds to Realized Gain and Axial Ratio)

Batch Size = 32 (number of training examples used in one iteration)

Number of Epochs = 0 (number of complete passes through the training dataset.)

The embed input and output training (202) is therefore configured to train and test the model to predict the next steps .

The predicted realized Gain (RG) and Axial Ratio (AR) (203) of the radiation pattern prediction (200) comprises simplified data (SD) as input and the precise data (PD) as output, for each point defined by the Bayesian Optimizer. It is configured to receive the results of the electromagnetic simulator for the simplified vehicle model and predict the values of the precise vehicle model.

The proposed method to predict the radiation patterns (200) over a surface of a vehicle allows to reduce electromagnetic simulation time that are so time-consuming as the complexity of the used vehicle model or the simulation parameters. Therefore, a simplified model is simulated with less computational time, and the prediction of a complex model is obtained that, if solved integrally for all the location points ( Pn) , would require huge amounts of computational simulation time.

The antenna placement optimization (300) , i.e., the optimization model, is comprised by a random initialization step (303) , a candidate for solution step (304) , a simplified Realized Gain and Axial Ratio step (305) , a complete Realized Gain and Axial Ratio step (306) , a Fitness function step

(307) and solution found testing step (308) . In the testing

(308) step, two possible results may arise, if a positive solution found (309) then an output result is provided (310) ; if the result is not adequate (311) , the procedure to follow will be again the lookout for candidates for solution (304) . The random initialization step (303) comprises the definition of required parameters, and the candidates for solution step (304) comprises definition of the new point (P_n) to be analysed. The simplified Realized Gain and Axial Ratio step (305) is the simulation step, where the obtained simplified vehicle model is simulated with the new point (P_n) defined. The complete Realized Gain and Axial Ratio step (306) comprises the resulting precise values of RG and AR predicted by the radiation pattern prediction (200) referring to the new point (P_n) . The fitness function step (307) , or objective function (F) , comprises the maximization of the RG and the minimization of the AR. Finally, the solution found testing step (308) will provide the best solution (310) based on the best position if the performed tests are affirmative (309) , or if not (311) , the optimization process (300) will conduct a new analysis (304) based on said point (P_n) .

For the optimization model (300) , it is proposed the Bayesian Optimization. It works by building a probabilistic model of the objective function (F) using the Bayesian machine learning technique, called the surrogate function. As illustrated in Figure 3, the optimization comprises at least an unknown function and prior mean function (a) , an acquisition function and mean update (b) , algorithm evaluation one (c) and algorithm evaluation two (d) . In the unknown function and prior mean function (a) of an input point or antenna placement where p_n = (x,y) and n=0, the unknown function F( ) (400) is illustrated, as well as the Gaussian Process (GP) mean value (401) and the Gaussian Process standard deviation (404) . In the acquisition function and mean update (b) it is illustrated the antenna placement l for n=l, and the minimum acquisition function (402) . Within a similar approach, in algorithm evaluation one (c) and algorithm evaluation two (d) , p2 and p3 for n=2 and n=3 respectively are represented, as well as an acquisition minimum (403) . Subsequently, using a Gaussian regression process, it is then efficiently researched with an acquisition function before the candidate samples (402) be chosen for evaluation in the real objective function (F) . This is the most efficient approach in terms of the number of function evaluations required. Everything the user knows about the objective function (F) can be used to estimate the cost of different candidate samples (304) that we may want to evaluate the search. This estimation of objective functions is made with a surrogate function. To select the next sample from the search space, Bayesian antenna placement optimization (300) uses an acquisition function that measures the value that would be generated by evaluation of the objective function (F) at a new point (P_n) , based on the current posterior distribution over the fitness value, and these estimates and credible intervals are obtained using a Gaussian process regression. This optimization technique has the property to minimize the number of objective function evaluations and performs well even in settings where the objective function has multiple local maximums .

The improvements of the utilization of the BO are mainly related with the optimal location finding within fewer iterations compared to the evolutionary algorithms (GA and PSO) , but it also managed to be more accurate in its position prediction .

The below described method shows how Bayesian optimization (300) evaluates the objective function in an initial space and allocates the remainder of a budget of N function evaluations .

The Basic Bayesian Optimization method (300) , comprises the steps of

-Place a Gaussian Process (GP) prior function on an objective function of the antenna placement F' (P) where by default the mean ] (P)=0 (deterministic) ; -Observe the objective function F at n position operating on the current prior distribution and while n < N, where N is the totality of the n positions:

-Update the posterior probability distribution on the objective function F using all available data; -Let P_n be a maximizer of the acquisition function over P, where the acquisition function is computed using the current posterior distribution;

-Get the radiation pattern, composed by realized gain (RG) and axial ratio (AR) from the simplified model and send it to the U-Net;

-the U-Net predicts the precise values of radiation pattern, RG and AR, and send it back to BO;

- Observe F ( Pn) ;

-Update the Gaussian Process (GP) prior distribution with the new data to produce a posterior, which will become the prior in the next step;

-Increment n;

-Return either the point evaluated with the best F(P_n) , or the point with the largest posterior mean.

References

[1] E. Adegoke, J. Zidane, E. Kampert, P. A. Jennings, C. R. Ford, S. A. Birrell, and M. D. Higgins, "Evaluating machine learning & antenna placement for enhanced gnss accuracy for cavs," in 2019 IEEE Intelligent Vehicles Symposium (IV) . IEEE, 2019, pp. 1007-1012.

[2] M. Kirkko- Jaakkola, S. Feng, Y. Xue, X. Zhang, S. Honkala, S. S"oderholm, L. Ruotsalainen, W. Ochieng, and H. Kuusniemi, "Effect of antenna location on gnss positioning for its applications," in 2016 European Navigation Conference (ENC) . IEEE, 2016, pp . 1-7.

[3] A. Jauhri, J. D. Lohn, and D. S. Linden, "A comparison of antenna placement algorithms," in Proceedings of the Companion Publication of the 2014 Annual Conference on Genetic and Evolutionary Computation, 2014, pp . 1223-1230.

[4] T. H. O'Donnell, R. Haupt, K. Lysiak, and D. J. Jacavanco, "Issues involved in developing a genetic algorithm methodology for optimizing the position of shipboard antennas," in Evolutionary and Bio-Inspired Computation: Theory and Applications III, vol. 7347. International Society for Optics and Photonics, 2009, p. 73470U.

[5] J. Yang, J. Li, and S. Zhou, "Study of antenna position on vehicle by using a characteristic modes theory, " IEEE Antennas and Wireless Propagation Letters, vol. 17, no. 7, pp. 1132-1135, 2018.

[6] J. M. K. Infantolino, M. J. Barney, and R. L. Haupt, "Using a genetic algorithm to determine an optimal position for an antenna mounted on a platform," in MILCOM 2009-2009 IEEE Military Communications Conference. IEEE, 2009, pp . 1- 6.

[7] Frazier, Peter I. "A tutorial on Bayesian optimization." arXiv preprint arXiv : 1807.02811 (2018) .

[8] Brochu, Eric, Vlad M. Cora, and Nando De Freitas. "A tutorial on Bayesian optimization of expensive cost functions, with application to active user modeling and hierarchical reinforcement learning." arXiv preprint arXiv: 1012.2599 (2010) .

[9] M. J. Inman, J. M. Earwood, A. Z. Elsherbeni, and C. E. Smith, "Bayesian optimization techniques for antenna design," Progress In Electromagnetics Research, vol. 49, pp . 71-86, 2004. [10] H. Gazzah and J.-P. Delmas, "Optimization of the antenna array geometry based on a bayesian doa estimation criterion, " in 2011 IEEE International Conference on Acoustics, Speech and Signal Processing

(ICASSP) . IEEE, 2011, pp . 2544-2547.

[11] A. Ghani, F. Keyvani, and S. H. Sedighy, "Antenna array placement on limited bound for isotropic and optimal direction-of-arrival estimation, " IET Signal Processing, vol. 12, no. 3, pp . 277-283, 2018.

Claims

1. Method for determining the correct placement of an antenna in a support structure, said placement providing the best radiation pattern of the antenna, comprising the steps of:

- random initialization (303) of an antenna placement optimization (300) with an antenna placement (Pn) , the antenna placement optimization (300) comprising:

- determining which are the ideal candidates for the solution (304) based on the antenna placement (Pn) ;

- determine the simplified Realized Gain and Axial Ratio (305) of the antenna placement (Pn) ;

- predict a Realized Gain and Axial Ratio (306) ;

- obtain a fitness function (307) ;

- if a positive solution found (309) then an output result is provided (310) , else, if the result is not adequate (311) the lookout for ideal candidates for solution (304) will be repeated.

2. Method for determining the correct placement of an antenna according to the previous claim, wherein the radiation pattern prediction (200) comprises the steps of:

- obtaining a model fitting (201) result based on an acquired segmentation map from a U-Net convolutional neural network;

- obtain a embed output training (202) based on the obtained model fitting (201) ;

- obtain a predicted realized Gain (RG) and Axial Ratio (AR) (203) .

3. Method for determining the correct placement of an antenna according to any of the previous claims, wherein the model fitting (201) is configured to determine the machine learning model generalization rate of similar data from which it was trained based at least in one of an: Kernel filter, a Rectified Linear Unit activation function, a max pooling operation with stride and Feature Maps.

4. Method for determining the correct placement of an antenna according to any of the previous claims, wherein the embed output training (202) comprises the initial data set used for training the U-Net convolutional neural network.

5. Method for determining the correct placement of an antenna according to any of the previous claims, wherein the predicted realized Gain (RG) and Axial Ratio (AR) (203) comprises simplified data (SD) as input and the precise data (PD) as output, for each point defined by the antenna placement optimization (300) .

6. Method for determining the correct placement of an antenna according to any of the previous claims, wherein the segmentation map comprises data from at least one of an elevation angle value (0) and azimuth angle ( C|J ) for each antenna placement (Pn) .

7. Method for determining the correct placement of an antenna according to any of the previous claims, wherein the U-Net convolutional neural network comprises input radiation patterns (101) of antenna placements (P_n) .

8. Method for determining the correct placement of an antenna according to any of the previous claims, wherein the input radiation patterns (101) comprise at least a set of magnitude data points that characterise the antenna in each of antenna placements (P_n) .

9. Method for determining the correct placement of an antenna according to any of the previous claims, wherein the U-Net convolutional neural network is trained with the input radiation patterns (101) .

10. Method for determining the correct placement of an antenna according to any of the previous claims, wherein the Realized Gain and Axial Ratio prediction (306) is obtained from an embed input and output training (202) and predicted realized Gain (RG) and Axial Ratio (AR) (203) of a radiation pattern prediction (200) .