Agenda
Signal Processing Seminar
- Friday, 24 May 2019
- 14:30-15:30
- HB 17.150
Adaptive classification of radar emitters
Aybuke ErolMETU, Turkey
Radar receivers collect interleaved signals from all electromagnetic sources in the environment. The ultimate goal of electronic intelligence is to separate these sources (deinterleaving) and find their types (emitter identification). Knowing the type of a source, it is possible to comment on its mission and operation. All in all, deinterleaving and emitter identification together build a system that solves an adaptive classification problem. One of the biggest challenges in this problem is that the system does not know all emitter types in the world since a great part of this information is confidential within each country. In addition, radar receivers sequentially provide radar pulses to the system. Therefore, the classifier should be able to increase its number of classes whenever an unfamiliar emitter type is encountered. What’s more, it should be able to distinguish between the unfamiliar emitter types, which enforces online learning.
The proposed system solves deinterleaving using fuzzy ARTMAP due to several reasons. First, it is supervised which makes the system able to start with a priori information or data. Secondly, it works with sequential input and enables online learning. Last but not least, it can increase its number of classes. After fuzzy ARTMAP, radar clusters are formed. Next, a representation for each cluster should be found, to be compared with the representations of already known emitter types. The challenge here is that describing an emitter type by single numeric values would not be fair as radar features are generally interval based. For example, emitters today do not operate on a single frequency, they rather have a frequency range in which they can operate. Hence, the representation and comparison of emitter types and radar clusters are considered under symbolic data analysis. Both parts, solved with fuzzy ARTMAP and symbolic data analysis, are improved in terms of classification accuracy from their baseline methods with the use of Jaccard index.
