Counterfactual Threshold Learning for Drone RF Signal Classification Under Interference Conditions

Drones utilize radio frequency (RF) signals for communication, providing pertinent information about their activities. The RF-based drone wireless monitoring system is expected to effectively achieve precise identification of types and operational modes in real-world scenarios, thereby providing more valuable information for drone regulation. The realization of the these objectives primarily confronts two challenges: mitigating the impact of noise and interference on feature learning and simultaneously accomplishing coarse-grained and fine-grained classification tasks of RF signals. Building upon the aforementioned considerations, we propose a neural collapse
$(\mathcal {NC})$
-inspired Counterfactual Threshold Learning (CTL) framework. The proposed framework leverages counterfactual causality to guide the neural network (NN) in accurate threshold learning, thereby more effectively eliminating noise-related features. Meanwhile, based on the
$\mathcal {NC}$
, we keep the classifier’s simplex Equiangular Tight Frame (ETF) structure unchanged during training and solely train the intermediate layers. This strategy guarantees consistent class prototypes, preventing performance degradation caused by the proximity of prototypes in fine-grained tasks. We integrate the aforementioned techniques into CTL framework and thoroughly validate it using two open-source datasets.