We present a general approach for antenna design and optimization based on consensus of results from a number of independently trained deep neural networks (DNNs). The aim of using the consensus is to reduce the uncertainty of results from a single DNN. The approach leads to several orders of magnitude faster antenna optimization and design compared to the optimization based on a full-wave solver and allows a compromise between the analysis speed and its accuracy. The used DNNs are multilayer perceptrons (MLP) with multiple fully connected hidden layers. As an example, we consider the Yagi–Uda antenna with four design parameters and optimize it for the maximal forward gain. The training of neural networks is done on datasets of several sizes, up to 1 million antenna samples. The samples are generated either randomly or at a uniform grid over the design space using the method of moments.

In order to provide a constant and complete operational picture of the maritime situation in the Exclusive Economic Zone (EEZ) at over the horizon (OTH) distances, a network of high frequency surface-wave-radars (HFSWR) slowly becomes a necessity. Since each HFSWR in the network tracks all the targets it detects independently of other radars in the network, there will be situations where multiple tracks are formed for a single vessel. The algorithm proposed in this paper utilizes radar tracks obtained from individual HFSWRs which are already processed by the multi-target tracking algorithm at the single radar level, and fuses them into a unique data stream. In this way, the data obtained from multiple HFSWRs originating from the very same target are weighted and combined into a single track. Moreover, the weighting approach significantly reduces inaccuracy. The algorithm is designed, implemented, and tested in a real working environment. The testing environment is located in the Gulf of Guinea and includes a network of two HFSWRs. In order to validate the algorithm outputs, the position of the vessels was calculated by the algorithm and compared with the positions obtained from several coastal sites, with LAIS receivers and SAIS data provided by a SAIS provider.

MultiLayer perceptron (MLP)-based neural model for an efficient 1D DoA estimation of highly correlated stochastic EM sources is proposed. Model extracts this information from a spatial correlation matrix obtained by sampling stochastic signals by a receiving antenna array. Algorithm for optimal selection of correlation matrix elements from the upper triangle of the matrix is suggested to reduce the number of antenna array elements needed to obtain an accurate neural model with relatively simple architecture. This approach is verified by determining the azimuth position of two sources with the mutual correlation in the range [0.8–0.95] using antenna arrays with different number of elements.

Hierarchical neural model for an efficient one-dimensional direction of arrival (DoA) estimation of stochastic electromagnetic (EM) sources with a variable radiation power is proposed. Model is trained to provide an azimuth position of such sources based on a spatial correlation matrix obtained by a signal sampling at a reception point and then used as an input to a neural model. It consists of two hierarchical levels realized by the multilayer perceptron (MLP)-based neural networks. The first level is responsible to reduce the dimensionality of the considered DoA problem which allows for its easier solution at the second level. Accuracy and run-time of the proposed model is verified on an example of determining the azimuth position of two stochastic EM sources in noisy conditions through comparison with a standalone MLP model and a model based on the root MUSIC algorithm.

An efficient neural network-based approach for tracking of variable number of moving electromagnetic (EM) sources in far-field is proposed in the paper. Electromagnetic sources considered here are of stochastic radiation nature, mutually uncorrelated, and at arbitrary angular distance. The neural network model is based on combination of probabilistic neural network (PNN) and the Multilayer Perceptron (MLP) networks and it performs real-time calculations in two stages, determining at first the number of moving sources present in an observed space sector in specific moments in time and then calculating their angular positions in azimuth plane. Once successfully trained, the neural network model is capable of performing an accurate and efficient direction of arrival (DoA) estimation within the training boundaries which is illustrated on the appropriate example.