Fusion layer architecture

DroneEAR AI is not a sensor. It is a software intelligence layer that sits above a customer's sensors and merges their outputs into one correlated picture. The design consumes normalised detection events — a source, a coarse zone, a confidence value, a bearing and a timestamp — rather than raw audio recordings or RF captures.

The five signal sources the layer processes:

RF — radio detection of control and telemetry links
Acoustic — acoustic vector detection capturing bearing and elevation at a single point, or a distributed mesh network utilising multilateration
Telemetry — direct telemetry messages from cooperative platforms
ADS-B and network Remote ID — live public feeds as context
Human reports — structured operator reports

Acoustic vector detection and processing

A sound field is described by two quantities: acoustic pressure, which is a scalar, and acoustic particle velocity, which is a vector. Classic microphones measure only pressure — they derive direction indirectly, from the time delay between multiple transducers.

An acoustic vector approach measures both quantities at once. It consists of pressure sensing combined with orthogonally oriented particle velocity measurement. The pressure component determines what kind of acoustic event is happening; the particle velocity vector points to where it comes from. The system applies real-time digital filtering directly at the edge nodes to suppress constant acoustic rotor noise.

Consequences for the fusion layer:

Direction in a single point or mesh. Captures bearing and elevation without requiring large transducer spacing, or links nodes into a 3D cross-correlation (TDOA) mesh for precise coordinate mapping.
Broadband. Covers the entire audio bandwidth — from low-frequency threats (engines, blasts) to high-frequency acoustic signatures.
Single-sensor blind spots. If a drone flies in a structural shadow (NLOS), radar misses it. If it flies on a pre-programmed route in radio silence, RF detection stays quiet. The acoustic layer catches the sound, allowing the fusion engine to maintain the contact track.
Non-collaborative and hybrid modes. Each node can operate independently, locked in place, or deployed on mobile platforms with minimal data transfer overhead.
Passivity. The system primarily listens passively, emits no RF signals, cannot be electronically jammed, and does not disclose its presence.
Low SWaP. Optimised software algorithms allow deployment on commercial off-the-shelf (COTS) hardware, minimising cost and environmental footprint.

Limits worth naming. Acoustic detection depends on distance and weather — wind, temperature and humidity affect sound propagation. In urban environments, there is a strong noise floor. The fusion layer addresses these physical limits precisely by requiring validation from other sensor classes.

RF detection

The RF layer detects control and telemetry links between drone and operator. It passively captures characteristic signatures in the bands drones use. RF detection is robust against darkness, fog and dust, but does not give a precise position — only a bearing or a presence.

Telemetry and ADS-B

Telemetry messages from cooperative platforms and public ADS-B data provide context — what is legally and identifiably in the air. This is important for separating registered traffic from an unknown contact.

Network Remote ID

Network Remote ID shows drones that identify themselves. In DroneEAR this feed is context only — it is never presented as a threat. A drone that does not identify itself will not appear here.

Human reports

Operators can enter structured reports into the system — visual contact, direction, time. The fusion layer treats these reports as one of the signals and correlates them with the other sources.

Two-tier correlation and escalation

The fusion layer produces a threat assessment only when independent sources agree. Tier 1 (Rule Engine) operates on strictly deterministic rules – no AI model is on the decision path, completely preventing hallucinations. The AI model (Tier 2) serves a purely structural role: it translates raw signal data into a clear sentence for the operator without the ability to alter the threat level. The output is strictly filtered by a validation gate.

The HIGH rule requires two independent sensors to agree and an unknown identity. Every assessment is explainable — the operator sees which sources contributed, what confidence they brought, and why the rules computed the result. The local model runs entirely on-premise, requiring no external API calls.

Stage note

DroneEAR AI is a working concept demonstrator. Drone-threat signals are simulated. Live public feeds are ADS-B and network Remote ID. The acoustic and fusion layer is designed and ready for integration with real sensors in a pilot via MQTT/Kafka interfaces — it is not part of the running public demo.