Model-based sensitivity analysis of dual-channel software-defined radio direction finding for unmanned aerial vehicle control signals

Abstract

Relevance. Direction finding of drone-control emissions supports spectrum monitoring when protocol identifiers are unavailable, but open datasets provide neither phase-coherent dual-channel captures nor independent devices. Aim. This study evaluates whether a reproducible radio-frequency gate followed by an amplitude-comparison direction-finding proxy is internally consistent under an explicit two-channel antenna model and identifies the assumptions that bound its use. Objectives. The work separates tuning from final evaluation and examines conditioning, calibration, multipath, and negative-class composition without constructing an unsupported total error budget. Methods. Single-channel in-phase/quadrature Tampere recordings feed burst detection and fingerprinting; the second channel and directional responses are simulated. Complete recordings are divided chronologically into train, validation, and untouched test blocks with guard intervals. The threshold is selected only on validation data. Test rates receive Wilson confidence intervals, and the receiver-operating-characteristic area uses a stratified bootstrap. A circular moving-block bootstrap supplies the median interval used by the null detector. Multipath sensitivity is summarized over independent nuisance draws at each model point. Scientific novelty. The contribution is reproducible integration with falsifiable boundaries: sufficient differential-response slope and absence of saturation define the well-conditioned set, while fingerprint claims are tied to recording and negative-class coverage. Practical significance. The workflow prevents test-threshold leakage and supplies calibration and conditioning checks before hardware trials. Results. On the untouched within-recording test block, the validation-selected threshold yielded high precision and recall, but the pre-specified precision criterion was not met. All matched-model proxy scenarios and all angles of the independent acceptance sweep passed the software checks; neither result is a field-accuracy estimate. The edges of the angular sector were excluded from the well-conditioned set because one channel saturates there and the local slope vanishes. Conclusions. The pipeline is computationally coherent under its model, whereas cross-device recognition, field-clutter performance, and joint field accuracy remain unverified. Independent sessions, field-relevant negatives, antenna calibration, and live coherent two-channel measurements are required.