A Novel Labeling Method for Radar Pulse Data With Complex Modulations and Parametric Overlaps

This paper aims to solve a problem: using azimuthal information to label radar pulse data with complex inter-pulse modulations and overlaps on electromagnetic parameters. In these tough conditions, azimuthal information is the key for data labeling instead of parameter-based methods. According to the measurement principle, the azimuthal information from the identical emitter shares the same positional relationship between the radiation source and the receivers. We propose a transformation, the inv-Hough transform, to reveal this positional relationship. The data from the identical emitter show highly linearly correlated distributions in the transformed space. Through transformation, the problem becomes a clustering problem. Dirichlet Process Mixture Models (DPMM) are widely used to address clustering problems, as they can automatically estimate the number of clusters. However, the outliers deteriorate clustering performances. Therefore, two improvements based on DPMM are proposed: adding a pre-clustering step to improve the outlier adaptation and distributed inference to improve the federated learning capability. The experimental results show that our method outperforms parameter-based methods when complex interpulse modulations and overlaps on electromagnetic parameters occur. In addition, the pre-clustering step significantly improves outlier adaptation and parametric stability for our method. And the method can work in federated learning scenarios.