Federated Flight Anomaly Detection with Self-Supervised Forecasting: A Flight-Disjoint Evaluation

Reliable anomaly detection is essential for the safe operation of unmanned aerial vehicles (UAVs), particularly in distributed settings where flight telemetry is generated across multiple vehicles or operational sessions and cannot always be centrally collected. This work investigates UAV anomaly detection using models capable of generalizing to unseen flights while keeping raw telemetry within the simulated client partitions during federated training. A self-supervised forecasting framework is evaluated under a simulated federated partitioning of the ALFA corpus and benchmarked using local training, Federated Averaging (FedAvg), FedProx, personalized FedAvg, and centralized learning under a strictly flight-disjoint evaluation protocol. The evaluation comprises 47 complete flights, including 36 failed and 11 nominal flights, grouped into 13 heterogeneous session-based federated clients that serve as statistical partitions rather than physically independent UAVs. The resulting federated setting exhibits non-IID characteristics due to differences in local flight composition and data volume across clients. Federated training is performed for 20 communication rounds with two local epochs per round, and evaluation uses three-fold flight-disjoint cross-validation. Complete flights, including failed flights, are assigned exclusively to either the training or held-out test partition within each fold, ensuring that no samples or overlapping windows from a test flight contribute to model training. An LSTM learns nominal flight dynamics through multivariate next-step forecasting, and the resulting forecasting error is converted into an anomaly score. FedAvg achieves an area under the receiver operating characteristic curve (AUC) of 0.908 and detects 83.3% of failed flights, performing competitively with centralized learning, which obtains an AUC of 0.904 and an 84.4% detection rate. Beyond discrimination performance, calibration, threshold scope, temporal persistence, detection delay, and false-alarm burden are also evaluated. With rolling calibration and a global threshold, the detector maintains an 87.2% flight-level detection rate with a median detection delay of 0.72 s while reducing the false-alarm burden from 202.1 to 66.0 events per hour. However, this alarm frequency remains too high for unattended operational deployment. A more conservative configuration reduces the burden to 8.17 events per hour while retaining 82.8% failed-flight detection and achieving an event precision of 0.851, although the median detection delay increases to 3.38 s. Within this session-partitioned experimental setting, parameter aggregation recovers most of the performance loss associated with local-only training and approaches centralized performance without pooling raw telemetry. These findings should be interpreted as evidence for distributed model aggregation under statistical data partitioning rather than as validation of a fully deployed federated UAV network, while the operational analysis further shows that strong discrimination and flight-level detection do not necessarily translate into practically acceptable alarm behavior.