Hide/Show Apps

Emitter identification with incremental learning using symbolic representations

Erol, Aybüke
Radar receivers collect mixed signals from all electromagnetic sources in the environment. The ultimate goal of electronic intelligence is to find the types of these sources with the help of a priori information, known as emitter identification. Emitter identification system aims to find a representative for each emitter in the environment and update them over time. Hence, such a non-stationary and continuous flow of data is of this thesis concern which is beyond the scope of traditional –offline or batch- machine learning systems. Another challenge is that the system can not know all possible emitter types and does not have a priori knowledge about the number of emitters. Therefore, incremental or online learning methods should be considered for the update of emitter representatives. After obtaining a representative for each emitter in a typical incremental learning algorithm, these representatives should be compared with a list of previously available emitter types. This part requires symbolic data analysis since the radar parameters generally operate interval-based. During simulations, among incremental learning algorithms, fuzzy ART, Bayesian ART, SOM and KDESOINN are examined and several extensions are proposed for the selected online learning networks. An ART-based structure based on Jaccard index is also proposed and tested for symbolic classification. The results indicate that the proposed symbolic data analysis method has outperformed other distance metrics and that the proposed algorithmic extensions enhance the performance of the selected online learning algorithms, while KDESOINN is observed to perform the best in terms of accuracy.