Automatic modulation recognition plays a vital role in electronic warfare. Modern electronic intelligence and electronic support measures systems are able to automatically distinguish the modulation type of an intercepted radar signal by means of real-time intra-pulse analysis. This extra information can facilitate deinterleaving process as well as be utilized in early warning systems or give better insight into the performance of hostile radars. Existing modulation recognition algorithms usually extract signal features from one of the rudimentary waveform characteristics, namely instantaneous frequency (IF). Currently, there are a small number of studies concerning IF estimation methods, specifically for radar signals, whereas estimator accuracy may adversely affect the performance of the whole classification process. In this paper, five popular methods of evaluating the IF–law of frequency modulated radar signals are compared. The considered algorithms incorporate the two most prevalent estimation techniques, i.e., phase finite differences and time-frequency representations. The novel approach based on the generalized quasi-maximum likelihood (QML) method is also proposed. The results of simulation experiments show that the proposed QML estimator is significantly more accurate than the other considered techniques. Furthermore, for the first time in the publicly available literature, multipath influence on IF estimates has been investigated.
Techniques for automatic modulation classification (AMC) of radar signals are crucial for spectrum sensing, passive bistatic radars and electronic intelligence (ELINT). Most of existing AMC algorithms were evaluated using solely synthetic data. Meanwhile, signals intercepted in real-life scenarios are extra distorted due to the multipath propagation. This paper presents the novel, pattern recognition AMC method for frequency modulated radar signals, with improved resistance to noise and multipath. This has been achieved by a multi-stage feature selection process, involving over a dozen of popular radar signal metrics. Based on the feature selection results, the two advanced waveform features were utilized in the designed algorithm, i.e. quasi-maximum likelihood (QML) instantaneous frequency estimate and fractional Fourier transform (FrFT) profile. The proposed AMC method has been evaluated using both synthetic and real data. Obtained results show that proposed classification framework achieves overall accuracy of 93.6% for a set of real-life signals acquisitions, corrupted both by noise and multipath influence.