Radar Emitter Identification Based on Co-clustering and Transfer Learning

In recent years, artificial intelligence and machine learning have been used to identify radar emitters. However, there are two basic assumptions in traditional machine learning: (1) the data of training set and test set are independent and identically distributed. (2) only a large number of labeled data can train a classification model which can effectively classify test set data. In other words, without enough training samples, it is impossible to learn a classifier that performs well in new radar emitters. In transfer learning, the existing radar data from related fields can help to learn such a classifier, which is called knowledge transfer. This can be achieved by eliminating the distribution difference between the original data and new radar emitter data. Therefore, we choose a new domain adaptation method called Manifold Embedding Distribution Alignment (MEDA) to solve the above challenges. MEDA firstly maps the data to Grassmann manifold, and dynamically aligns marginal distribution and conditional distribution in this space. Finally, based on the mapped data, a domain invariant classifier is obtained. Yet, this method is sensitive to the differences between data. If the similarity between data is too low, the knowledge learned from the original data will have a negative impact on the identification of new data. Thus, we introduce the co-clustering algorithm to discover the partial “instance-feature” structure over both instances and features. Simultaneously, considering the related knowledge of “instance-feature” and related new instances, the original data can be reconstructed. The reconstructed original data can be more relevant with the new radar emitter data. Empirical results demonstrate the effectiveness of our method.