Marija Agatonovic

Neural networks-based DOA estimation of multiple stochastic narrow-band EM sources

Localization of multiple stochastic narrow-band electromagnetic sources in the far-field is considered in the paper. Artificial neural networks-based approach is proposed to allow for an efficient direction of arrival (DOA) determination of electromagnetic signals radiated from stochastic sources as one of the key steps in the source localization procedure. It uses correlation matrix, obtained by signal sampling via antenna array in far-field scan area, to train an appropriate model based on MLP (Multi-Layer Perceptron) neural network. Proposed approach is validated on the example of a neural model performing accurate and fast one-dimensional (1D) DOA estimation of the position of three stochastic sources placed at fixed angle distance in azimuth plane.

Hybrid ANN model for accurate 2D DOA estimation of a radiating source

Given that simulation models may often suffer from reduced accuracy when applied to real environmental conditions, in this paper we propose a hybrid model for two-dimensional direction of arrival (2D DOA) estimation of a radiating source. The model is based on artificial neural networks (ANNs), and its development is conducted in two phases. Initially, an ANN is trained to predict angular positions of a simulated radiating source in a certain range of azimuth and elevation angles. The second phase includes development of a corrective empirical ANN aimed to improve the accuracy of the simulation-based network. Finally, the hybrid ANN model is able to account for real environmental conditions and physical aspects of the receiving antenna array. The performance of the model is verified by measurements for several positions of the transmitting antenna.

Artificial Neural Networks for ranging of passive UHF RFID tags

Ranging of passive Ultra High Frequency (UHF) Radio Frequency Identification (RFID) tags in indoor environments is a topical issue nowadays. Due to complexity of such an environment, there is no effective solution to this problem. In this paper we investigate application of Artificial Neural Networks (ANNs) in indoor localization of passive UHF RFID tags. Namely, we estimate distance between a reader antenna and a couple of tags attached to an item, using nonlinear mapping that ANNs perform between measured values of the Received Signal Strength Indicator (RSSI), turn on power and phase on the one hand, and the distance on the other. The proposed ANN model calculates distance with an average error of 7.31 cm.

Application of neural networks in spatial signal processing (invited paper)

Neural networks (NNs) have proven to be a very powerful tool both for one-dimensional (1D) and two-dimensional (2D) direction of arrival (DOA) estimation. By avoiding complex and time-consuming mathematical calculations, NNs estimate DOAs almost instantaneously. This feature makes them very convenient for real-time applications. Further, unlike the well known MUSIC algorithm, neural network-based models provide accurate directions without additional calibration procedure of antenna array and a priori knowledge of the number of sources. In this review paper, the results achieved by the research group at the Faculty of Electronic Engineering in Nis are presented. The problem of DOA estimation of narrowband signals impinging upon different configurations of antenna arrays is addressed. Both Multi-Layer Perceptron (MLP) and Radial Basis Function (RBF) neural networks are considered, and their advantages and disadvantages are discussed. To improve the resolution of DOA estimates, sectorization model is introduced. As shown in this work, neural network-based models demonstrate high-resolution localization capabilities and much better efficiency than the MUSIC.

Empirical ANN models for 2D direction of arrival estimation

Empirical Artificial Neural Network (ANN) models are developed for Two-Dimensional Direction of Arrival (2D DOA) estimation of a source signal. For that purpose, experimental data obtained from measurements in an anechoic chamber are utilized. Performance of ANN models are compared to 2D MUSIC algorithm in regard to estimation accuracy and speed of calculations. It is demonstrated that the proposed models outperform MUSIC in cases when small number of snapshots are utilized for DOA estimation and at the same time, are more suitable for real-time implementation.

DOA estimation using radial basis function neural networks as uniform circular antenna array signal processor

In this paper, estimation of direction-of-arrival (DOA) of source signals employing radial basis function neural networks (RBF-NNs) is presented. The signal model used in the algorithm is based on the circular antenna array geometry. RBF-NNs are trained and tested to estimate DOAs of different number of signals in azimuth plane. The performance of the RBF-NNs is evaluated in noisy environment for various values of signal-to-noise (SNR) ratio.