Machine learning for electronic intelligence

Electronic intelligence is concerned with gathering information about radar emitters by intercepting and analysing their signals. By collecting this information, characteristic features are found that can be exploited to recognise known emitters. Traditional analysis and identification approaches rely on databases, which contain descriptions of the radars' operational modes. In conventional radar systems, modes are designed to fulfil a certain function and exhibit constant patterns of the waveform parameters. Agile multifunction radars, however, optimise their waveform parameters for the specific situation and therefore, they do not exhibit constant patterns. Consequently, traditional databases cannot effectively represent their emission characteristics. Moreover, they are unable to efficiently capture the relationships between emissions and hence, they cannot model the emitters' behaviour. This thesis suggests a new emission model, which understands the radar emissions as a language with an inherent hierarchical structure of the five modelling levels letters, syllables, words, commands, and functions. Such an emission model allows exploiting machine learning methods, which are designed for natural language processing and in particular for the field of representation learning. Based on this approach, methods are developed for the four tasks of emission prediction, the identification of the emitter type, the learning of behavioural models, and the recognition of unknown emitters. To this end, predictive models are created that capture the behaviour of agile radar emitters. In addition, several architectures are investigated that combine multiple emitter models into an ensemble. Two contrasting approaches are compared for all four tasks throughout this thesis, namely the "memoryless" Markov chain and the Long Short-Term Memory recurrent neural network, which is designed to "remember" the past. The hierarchical structure of the emission model significantly increases the performance of all considered tasks in comparison to traditional signal analysis methods. These process radar pulses, i.e. letters, based on which emitter identification and recognition of unknown emitters are not successful. Moreover, it is shown that machine learning methods provide large benefits in comparison to conventional approaches in the majority of the settings