Classification of Radar Emitters using Semi-Supervised Contrastive Learning

Sammanfattning: Radar is a commonly used radio equipment in military and civilian settings for discovering and locating foreign objects. In a military context, pilots being discovered by radar could have fatal consequences. As such, countermeasures to radar has been developed which intercepts the transmitted signal with the ability to identify, disturb and neutralize the source. This thesis will focus on the identification part, also known as Electronic Support (ES). Various machine learning models using supervised learning have previously been attempted achieving high accuracy and F1-Score. These supervised models however depend on a large set of labeled data which may not always be available. It is therefore of interest to develop models which are able to train on datasets with limited amounts of labeled data. This thesis explores a method using contrastive learning to pre-train models using unlabeled data in order to improve overall performance. The model consists of a 1-D CNN encoder and a contrastive transformer model. To perform the contrastive pre-training, three different classes of augmentations were used: Dynamic Time Warping Auto-Encoders, Data-Specific and Domain-Specific. The pre-trained models were then fine-tuned on various degrees of labeled data ranging from 1% - 100%. When compared to models without pretraining, it was found that the pre-training has an effect when the data contained less than 2% labeled data on a dataset containing 79 emitters. These results were amplified when noise in the form of spurious pulses were introduced in the testing set. The model without pretraining required at least 10% labeled to outperform the pretrained models under such circumstances.