TL;DR
This paper introduces a novel two-step methodology for deinterleaving unknown RADAR signals, a crucial task in electronic warfare. By combining HDBSCAN clustering in the Frequency-Pulse Width (F-PW) plane with hierarchical agglomerative clustering using optimal transport distances, the authors effectively separate and group pulses from different RADAR emitters, even in complex, interlaced signals.
What Problem Does This Paper Address?
In the increasingly sophisticated world of electronic warfare, accurately identifying and classifying RADAR emitters is paramount. This involves "deinterleaving" signals, which means separating individual RADAR pulses from a complex stream that may contain signals from multiple sources. Traditional methods often rely on basic pulse parameters like time of arrival (TOA) and pulse repetition interval (PRI), but modern RADARs exhibit agility in these parameters, making them difficult to distinguish. Furthermore, obtaining large, labeled datasets of real-world RADAR signals is challenging, hindering the application of data-hungry deep learning techniques. This paper tackles the problem of deinterleaving and clustering unknown RADAR pulses from simulated data, aiming to develop a robust and adaptable methodology.
Methods
The proposed methodology consists of two main stages:
- Clustering in the F-PW Plane:
- Pre-processing: Pulse data (frequency and pulse width) is normalized using quantile normalization to ensure consistent scaling and preserve data distribution.
- HDBSCAN Clustering: The HDBSCAN algorithm is employed to group pulses in the F-PW plane. HDBSCAN is chosen for its ability to handle noise and outliers, and its hierarchical nature allows for flexible parameter tuning. The goal here is to initially separate pulses that likely belong to the same RADAR system.
- Hierarchical Agglomerative Clustering with Optimal Transport Distance:
- Cluster Agglomeration: Since a single RADAR might be represented by multiple clusters in the F-PW plane (due to variations in its parameters), a second stage of clustering is performed. This stage uses hierarchical agglomerative clustering.
- Optimal Transport Distance: Instead of standard distance metrics, the authors use optimal transport distances to measure the similarity between these initial F-PW clusters. This approach is particularly useful for comparing probability distributions (represented by cluster pulse characteristics) and is robust to variations in pulse counts and timing. The optimal transport distance helps to group clusters that originate from the same RADAR, even if their F-PW characteristics are not perfectly identical.
Datasets
The authors utilized simulated RADAR data generated by a certified simulator. This simulator provided a large diversity of signals representing typical RADARs, with characteristics drawn from a database of 60 RADAR systems. The simulated data was labeled, allowing for the application of supervised metrics to validate the proposed methodology. The paper emphasizes the necessity of simulated data due to the difficulty of obtaining real-world labeled datasets.
Key Insights and Ideas
- F-PW Plane as a Primary Feature Space: The F-PW plane proves to be a powerful initial space for separating RADAR pulses, effectively grouping them based on fundamental characteristics.
- HDBSCAN's Robustness: HDBSCAN's ability to handle noise and outliers is crucial for real-world RADAR signals, which often contain imperfections.
- Optimal Transport for Cluster Fusion: Using optimal transport distances for hierarchical clustering is a novel and effective way to merge initial clusters. It allows for a more nuanced comparison of cluster characteristics than traditional metrics, especially when dealing with the probabilistic nature of RADAR pulse distributions.
- Deinterleaving as a Multi-Stage Process: The paper highlights that deinterleaving is not a single-step problem. A two-stage approach, first separating individual pulses and then grouping them by emitter, is more effective.
- Simulated Data is Key: The reliance on simulated data underscores the challenges in acquiring and labeling real RADAR signals, a common hurdle in this field.
Key Results
The methodology was validated on labeled simulated data. The results demonstrated that the two-step approach successfully:
- Identified the number of transmitters present in the signal.
- Effectively grouped pulses belonging to the same RADAR emitter.
- Reconstructed lobes in the DTOA-TOA plane, indicating accurate deinterleaving.
- The hierarchical agglomerative clustering with optimal transport successfully fused initial clusters, leading to a final grouping that correctly identified the 4 transmitters in the example signal.
Limitations and Future Work
- Assumption of Single Transmitter per Initial Cluster: The current HDBSCAN step assumes that initial clusters in the F-PW plane correspond to a single transmitter. This assumption can fail when RADARs have very similar characteristics (frequency, pulse width), leading to potential confusion.
- Handling of Low-Pulse Clusters: Some initial clusters may contain too few pulses to be reliably used in the optimal transport calculation.
- Future Work: The authors plan to investigate methods for separating similar RADARs in frequency and pulse width. They also aim to develop metrics for evaluating their method on more signals and compare it with existing approaches. Additionally, exploring the use of optimal transport for more challenging problems and developing better metrics for cluster fusion are future directions.
Summary updated