Radar Emitter Recognition (RER) is used in Electronic Warfare (EW) to avoid being detected during stealth operations, or to detect specific radar installations for targeting in an offensive campaign.Its role in both offensive and defensive measures make it a critical component within the scope of a larger Electronic Countermeasure (ECM) strategy. Radar spectrograms are captured from emissions, measuring frequency and power over time. The emission samples vary in the amount of time spent in each frequency, making a single approach to feature extraction ineffective. This thesis attempts to address RER challenges using modified radar signals from spectrograms sampled from five different classes, each containing 250 unique examples. A hierarchical approach is used where spectrograms with fewer time intervals have their features extracted from the magnitude spectrum, while signals with more time intervals have features extracted from the frequency domain. Both of these feature extraction methods are tested using k-Nearest Neighbor (kNN) and Support Vector Machine (SVM) algorithms. Experiments show that the hierarchical approach to feature extractionis a viable new way of thinking about spectrogram based RER.
Table of Contents
1 Introduction -- 1.1 Radar Emitter Recognition Beginnings -- 1.2 Traditional Features in Early Radar Detection -- 1.3 Extracting Features from Spectrogram Samples -- 1.4 Overview -- 2 Literature Review -- 2.1 Traditional Features in Early Emitter Recognition -- 2.1.1 Pulse Descriptor Words in Detail -- 2.2 Beyond Traditional Features -- 2.2.1 Intrapulse Information -- 2.2.2 Fractal Dimensions -- 2.2.3 Extents as features -- 2.3 Conclusion Summary of Literature - 3 Methods -- 3.1 Method Overview -- 3.2 Simulated Radar Spectrogram Data -- 3.2.1 Spectrogram Signal to Noise Ratio -- 3.2.2 Exploratory Data Analysis -- 3.2.3 Waveform Class Descriptions and Shape Geometry -- 3.2.4 Spectrograms Coherence -- 3.2.5 Confounding Data -- 3.3 Frequency Over Time Feature Extraction -- 3.3.1 Defining Class Geometry using Subsample Peak Interpolation -- 3.4 Power Over Frequency Feature Extraction -- 3.4.1 Defining Class Geometry using Max Smoothing -- 3.5 Classification Methods -- 3.5.1 Confusion Matrix -- 3.5.2 Monte Carlo Cross Validation -- 3.5.3 k-Nearest Neighbor -- 3.5.4 Support Vector Machine -- 3.5.5 One-versus-Rest SVM -- 3.5.6 Hierarchical SVM -- 4 Results -- 4.1 Testing Methods -- 4.2 Confusion Matrix -- 4.2.1 Definition of Metrics Used -- 4.3 Euclidean kNN Test Results -- 4.4 SVM Test Results -- 4.4.1 One-versus-Rest SVM Test Results -- 4.4.2 Hierarchical SVM Test Results -- 4.5 Analyses -- 5 Conclusion -- 5.1 Limitations -- 5.2 Future Work -- 5.3 Conclusion -- Appendix -- ReferencesNotes
Theoretical thesis.
Includes bibliographical referencesAwarding Institution
Macquarie UniversityDegree Type
Thesis MResDegree
MRes, Macquarie University, Faculty of Science and Engineering, Department of ComputingDepartment, Centre or School
Department of ComputingYear of Award
2019Principal Supervisor
Len HameyRights
Copyright Robert Newport 2019
Copyright disclaimer: http://mq.edu.au/library/copyrightLanguage
EnglishExtent
1 online resource (x, 66 pages) colour illustrationsFormer Identifiers
mq:71061
http://hdl.handle.net/1959.14/1270456
