Automatic detection of signals by using artificial intelligence techniques
- Manuel Rosa Zurera Director
- María del Pilar Jarabo Amores Co-director
Universidade de defensa: Universidad de Alcalá
Fecha de defensa: 31 de xaneiro de 2012
- José Luis Sanz González Presidente/a
- Roberto Gil Pita Secretario
- Saturnino Maldonado Bascón Vogal
- Amerigo Capria Vogal
- Rafael Pérez-Jiménez Vogal
Tipo: Tese
Resumo
The automatic detection of signals (targets) in additive interference (clutter and noise) is not a problem completely solved nowadays. Many different approaches are reported every year in the specialized literature depending on the targets to be detected and the kind of interference present in the environment where the sensor is working. In this way, a detection approach that is able to work in different environments is searched in this thesis. This thesis tries to solve two detection problems: the detection of moving Swerling 0 targets in synthetic Weibull-distributed clutter and white Gaussian noise; and the detection of moving vessels in marine radar environments. A relationship between these two problems is found in the thesis, allowing to propose a unique detection scheme that works in both cases. According to the detection problems to be solved in the thesis, some premises are set. Synthetic radar scans are generated in simulated environments having: time-correlation between consecutive cells; and constant clutter properties (skewness parameter) inside a scan, but variable scan-to-scan. Targets of different sizes and shapes are included in the synthetic radar scans. Different radar environments have been considered in the thesis by using the statistical parameters of sea, sea-ice and ground clutters reported in the literature. From these environments, it is observed that the clutter statistics are different each other, making the problem of proposing a detector scheme able to work with high performance in different environments more complicated. For solving the detection problems this thesis deals with, Artificial Intelligence (AI) based detectors are proposed, and compared with commonly used detectors selected from the literature. The coherent detector set as reference is the target sequence known a priori (TSKAP) detector. The incoherent detector set as reference is based on constant false alarm rate (CFAR) techniques. From AI techniques, two feed-forward artificial neural networks (ANNs) strategies are selected: the multilayer perceptrons (MLPs) and the radial basis function ANNs (RBF-ANNs, also referred as RBFNs). By using these AI techniques, coherent and incoherent approaches are proposed. An additional contribution is made in the thesis by proposing new modes of selecting the cells to be processed. Thus, not only the commonly used non-delayed selection modes are used, but also additional delayed selection modes are studied. These proposed modes are based on 2-dimension selection templates, instead of the 1-dimension templates commonly used in CFAR detectors. Experiments considering the reference and AI-based coherent detectors have been carried out in simulated sea, sea-ice and ground environments. In these experiments, the influence of the following parameters in the design stage of the detectors is studied: the clutter properties of the data sets used to design the detectors (for training the MLPs and RBFNs, and for setting the detection threshold); the selection modes; the number of selected cells; and the number of hidden neurons in AI-based detectors. From these studies, the values for obtaining the highest performance, while maintaining a low computational cost, are selected. Once the reference and AI-based detectors are designed, they are tested using a set of radar scans never processed before (test data set). This data set is composed of radar scans with different clutter conditions (simulating real environments). The performance obtained for this data set is slightly lower than the one achieved in the design stage. Moreover, the performances achieved for each particular radar scan of the test data set, i.e. for different clutter conditions, present low variations, denoting high robustness of the detector against changes in clutter conditions. According to these low performance variations, we can infer the performance achieved by the detectors when processing new radar scans in the future with similar clutter properties as the ones used here. Similar studies have been made when using reference and AI-based incoherent detectors in synthetic sea, sea-ice and ground environments. From the analysis of incoherent detectors in synthetic sea clutter, we focus on the differences observed with respect to coherent detectors in synthetic sea clutter. First, a low performance decrease is observed, being expected because the incoherent detectors only process the amplitude of the cells. And second, a high computational cost decrease is observed in AI-based incoherent detectors because less information is used at their inputs. The results obtained when designing and testing incoherent detectors in synthetic sea-ice and ground environments are not reported in the thesis because similar performance losses and computational cost decreases as the ones presented above are obtained. Finally, incoherent detectors have been designed and tested when processing radar scans obtained by a standard marine radar. This marine radar was sited in the FINO-1 German research platform (North Sea, Germany). It has been statistically checked that the measurements of the clutter fit the Weibull distribution. In this way, CFAR and AI-based detectors were designed by tuning their parameters in order to obtain the highest performance, while maintaining a reasonable computational cost. Once designed, they were tested obtaining similar conclusions as for the synthetic case: high robustness against clutter condition changes and low performance loss when processing new radar scans. The computational costs of the proposed configurations of MLP and RBFN-based incoherent detectors are reported. The processing speed needed to process radar scans in real-time is also reported. And since commercial processing units can fulfill this processing speed, the proposed AI-based detector can process marine radar scans in real-time. From the analysis of the performance obtained in the different cases of study, this thesis finishes with the following conclusion: the proposed AI-based detectors outperform the reference detectors in all the cases of study presented in the thesis. This conclusion is obtained when processing radar scans from radars working at different frequencies, with different resolutions and with different receivers (coherent and incoherent). Moreover, this conclusion is independent of the radar environment under study (sea, sea-ice and ground).