Optimization of Neural Network Architecture for Classification of Radar Jamming FM Signals

Radar jamming signal classification is valuable when situational awareness of radar systems is sought out for timely deployment of electronic support measures. Our Thesis shows that artificial neural networks can be utilized for effective and efficient signal classification. The goal is to optimize an artificial Neural Network (NN) approach capable of distinguishing between two common radar waveforms, namely bandlimited white Gaussian jamming noise (BWGN) and the ubiquitous linearly frequency modulated (LFM) signal. This is made possible by creating a theoretical framework for NN architecture testing that leads to a high probability of detection (PD) and a low probability of false alarm (PFA). It is assumed that some signal processing is required to improve the odds of correct signal classification. Therefore, our approach is to train the NN with a matrix populated with M spectra, where each spectrum is a sequence of N samples. Following standard procedure, 70% of the spectra are used for training, 15% for validation, and the remaining 15% for testing. A multilayered feedforward network topology is chosen for neuron arrangement and interconnection. Extensive experimentation is required to reveal the optimal number of hidden neurons, the optimal number of hidden layers, and lastly, the most efficient learning algorithm. Results show that a 98.5% PD and a 1.7% PFA of classifying between signal and noise are achieved when utilizing 10 or more hidden neurons. Changing the number of hidden layers has no significant effect in performance but the use of one hidden layer does reduce training time. Furthermore, the effect of using the Scaled Conjugate Gradient learning algorithm nearly yields a 100% PD. An architecture with at least ten hidden neurons, one hidden layer, and a Scaled Conjugate Gradient for training algorithm encapsulates an optimal architecture for our application