Conditional FrFT-Enhanced Reconstruction With Global Consistency for ADS-B Trajectory Anomaly Detection

Abstract

With the digital transformation of air traffic management (ATM) systems, automatic dependent surveillance–broadcast (ADS-B) has become an important data source for aviation surveillance. However, its open-broadcast nature and reliance on external navigation links make it vulnerable to anomalies caused by data-quality fluctuations, equipment failures, and malicious manipulation, which may degrade surveillance trustworthiness and introduce safety risks. To address the insufficient characterization of tunable time–frequency structures in existing ADS-B anomaly detection methods, as well as the limited capability of sequential models to aggregate global trajectory features, this article proposes a fractional Fourier transform-guided conditional autoencoder (FrFT-CAE). The proposed method employs a stacked long short-term memory sequence autoencoder to learn local temporal dynamics and introduces a multilayer FrFT–Gabor convolutional branch to extract global conditioning representations. The conditioning information is then adaptively injected into the latent space via cross-attention, thereby implicitly modeling the consistency between local temporal dynamics and global time–frequency structures during reconstruction. This improves normal-trajectory reconstruction capability and anomaly separability. Anomalies are scored using reconstruction error, and the detection threshold is determined by the $\mathbf {\mathit {\textit {n}} \sigma }$ rule on a validation set. Experiments are conducted on a self-collected ADS-B dataset (SCD) and a public ADS-B benchmark dataset. The results show that FrFT-CAE achieves stable overall detection performance under different data sources and anomaly definitions, with average $F1$ -scores of 0.9309 and 0.9250 on the self-collected dataset and the public benchmark dataset, respectively. Ablation experiments further demonstrate that the FrFT–Gabor conditioning branch and the cross-attention fusion module effectively enhance the collaborative modeling between global structural representation and local temporal reconstruction, thereby improving the overall performance of ADS-B trajectory anomaly detection.