RF Spectrum Sensing & AI Classification
The identification layer. Wideband receivers hear drone control and video links from kilometers away, direction-find the aircraft — and the pilot — and on-device AI turns raw pixels into recognized targets.
Spectrum monitoring
A stationary wideband monitoring station covers 30 MHz to 6 GHz — the entirety of common drone control, telemetry and video bands — out to a 5 km detection radius. It detects signals below 0 dB signal-to-noise, separates overlapping emissions within 50% spectral overlap, and direction-finds with ≤5° RMS accuracy in urban multipath environments. Blacklist/whitelist management keeps friendly aircraft quiet while unknown emitters raise alarms.
The same network localizes the pilot: cross-bearing on the control link positions the operator within 16° at 1 km — often the most actionable intelligence in the entire engagement, because the drone is only the messenger.
On-device AI
Two AI module families run the detection and recognition stack directly at the sensor. The export-class module delivers 21 TOPS (INT8) on a Jetson Xavier NX-class platform; a fully domestic variant delivers 16 TOPS on a BM1688-class NPU with −40…+60 °C operation and RS-422/485/232 interfaces for direct servo and vehicle integration. Both detect 2×1-pixel targets in air backgrounds, recognize infrared signatures at ≤10×10 pixels and visible ones at ≤24×24 pixels, hold ≥20 targets simultaneously, and re-acquire tracks after up to 10 seconds of occlusion.
RF spectrum monitoring station — parameters
| Frequency range | 30 MHz–6 GHz |
|---|---|
| Detection radius | ≥5 km |
| Direction finding accuracy | ≤5° (RMS), urban environment |
| Sensitivity | ≤10 dBµV/m; detects SNR <0 dB signals |
| Overlapping-signal handling | Separates emissions with ≤50% spectral overlap |
| Pilot localization | 1 km range, ≤16° bearing |
| Black/white lists | Yes |
| Installation / interfaces | Ø650 × 450 mm, RJ45 data, AC 220 V |
| Weight | ≤12 kg |
AI image-processing modules — parameters
| Detection capability | ≤2×1-pixel targets (air background), moving & stationary, complex terrain |
|---|---|
| Recognition | IR ≤10×10 px; visible ≤24×24 px; ≥85% accuracy at ≥50% contrast |
| Multi-target / blind track | ≥20 targets / ≥3 s blind track, ≥10 s re-acquisition after occlusion |
| Codecs | H.264 / H.265 |
| Module A (export class) | 21 TOPS INT8, 6-core ARM, 8/16 GB RAM, ≤80 g module |
| Module B (fully domestic) | 16 TOPS INT8, BM1688-class NPU, 8-core ARM, ≤10 W, −40…+60 °C, shock ≤80 g |
| Interfaces | Gigabit Ethernet ×1–2, RS-422/232/485, MIPI, BT.1120, TTL |
| Video / data protocols | RTSP / UDP |
RF & AI FAQ
What does spectrum monitoring miss?
Radio-silent aircraft — pre-programmed autonomous flights, home-built machines on non-standard links, or emitters that stay quiet until the last moment. That gap is exactly why the architecture is layered: spectrum identifies what speaks, EO/IR sees what does not.
Why on-device AI instead of server-side?
Latency and bandwidth. Detection at the sensor means the track exists before the video stream leaves the device; a 2×1-pixel target never needs to be transported anywhere to be found.
Designing the identification layer?
Tell us the RF environment — legal bands, friendly traffic, urban or rural — and we will propose a monitoring layout.
Ask technically