Result Quality Checks

Base Factor Weights

Bonus Modifiers

• +0.15 Cyclostationary detection — Positive cyclostationary feature extraction confirms periodic signal structure consistent with a digital datalink.
• +0.25 Multi-band correlation — Signal detected simultaneously on control and video downlink bands from the same emitter.

SNR Gate

When the measured signal-to-noise ratio is below 10 dB, the composite score is hard-capped at 0.35 regardless of factor weights or bonuses. This prevents low-SNR signals from producing high-confidence detections.

Provenance Fields

Verifying Provenance via API

Query the detection endpoint and inspect the provenance object in each result:

curl http://localhost:8080/api/detections/latest | python -m json.tool

Expected provenance block:

{
  "provenance": {
    "signature_db_version": "sig-v3.12.0-a4f8c2e",
    "band_config_version": "bands-v2.1.0",
    "analyzer_version": "2.4.1",
    "sensor_firmware_hash": "e3b0c44298fc1c14..."
  }
}

To verify a specific sensor's firmware against the expected hash:

curl http://localhost:8080/api/sensors/{sensor_id}/firmware-check

Common FP Sources

• WiFi access points — 2.4 GHz and 5.8 GHz WiFi can trigger frequency and bandwidth matches. Usually rejected by modulation mismatch but high-power APs near the sensor may pass initial screening.
• Bluetooth / BLE devices — FHSS pattern in the 2.4 GHz ISM band can produce hop rate matches. Typically rejected by bandwidth check (1 MHz vs expected 10+ MHz).
• Microwave ovens — Wideband noise bursts at 2.45 GHz. These produce short-duration detections that are filtered by the persistence gate (minimum 200 ms continuous signal).
• Surveillance cameras — Analog video transmitters on 5.8 GHz can mimic drone video downlinks. Distinguished by constant-on pattern vs burst transmission.
• Amateur radio — Narrowband signals near 433 MHz or 900 MHz may match some legacy drone control frequencies.

Benign Rejection Pipeline

Before a detection reaches the output, it passes through the benign rejection pipeline:

1. Known-source filter — Signals matching registered benign source profiles (site-specific WiFi MACs, known fixed transmitters) are suppressed.
2. Persistence gate — Signals shorter than 200 ms are dropped.
3. Bandwidth sanity check — Signals with bandwidth outside the expected range for any known drone protocol are flagged.
4. Modulation classifier — CNN-based modulation classification rejects non-drone modulation types.

Reviewing Suppressed Detections

Suppressed (rejected) detections are logged for audit purposes. To review them:

# View suppressed detections from the last hour
curl "http://localhost:8080/api/detections?status=suppressed&since=1h"

# View rejection reason for a specific detection
curl http://localhost:8080/api/detections/{detection_id}/rejection-reason

Filing a False Positive Report

When you confirm a detection is a false positive:

1. Note the detection_id and timestamp from the detection payload.
2. Record the suspected real source (WiFi AP model, Bluetooth device, etc.).
3. Submit via the API or the Max UI:

curl -X POST http://localhost:8080/api/quality/fp-report \
  -H "Content-Type: application/json" \
  -d '{
    "detection_id": "det-2026-03-01-00142",
    "actual_source": "Ubiquiti UAP-AC-PRO WiFi AP",
    "notes": "Mounted 3m from sensor antenna"
  }'

Common Causes

• SNR too low — The drone signal is below the noise floor at the sensor location. Check snr_db in raw scan data. Below 6 dB, detection probability drops sharply.
• Frequency not in band config — The drone operates on a frequency that is not included in the active band configuration. Verify with GET /api/config/bands.
• Signature not in database — New or rare drone model without a matching signature entry. The CNN classifier may still catch it, but heuristic scoring will be low.
• DJI Mini series — DJI Mini 2/3/4 use OcuSync with very low power and narrow bandwidth that can evade cyclostationary detection. Known limitation documented in release notes.
• Sensor saturation — Strong nearby transmitters can saturate the ADC, masking weaker drone signals. Check adc_clip_count in sensor health.

Diagnostic Steps

1. Pull the raw scan data for the time window of the missed detection:

curl "http://localhost:8080/api/scans?from=2026-03-01T10:00:00Z&to=2026-03-01T10:30:00Z" \
  --output scan_dump.json

2. Check if the expected frequency band was being scanned:

curl http://localhost:8080/api/config/bands | python -m json.tool

3. Verify the signature database includes the drone model:

curl "http://localhost:8080/api/signatures?search=dji+mini+4"

4. Check sensor health during the time window for ADC clipping or calibration warnings:

curl "http://localhost:8080/api/sensors/{sensor_id}/health?from=2026-03-01T10:00:00Z"

Performance Summary

• Holdout accuracy: 90.3% across 14 classes
• Noise false positive rate: reduced from 89.7% to 9.6% after dedicated noise class training
• Classes: DJI-OcuSync, DJI-Lightbridge, FrSky-ACCST, Futaba-FASST, Spektrum-DSMX, ExpressLRS, Crossfire, analog-video-5.8G, WiFi-drone, custom-FHSS, custom-DSSS, LTE-relay, noise, unknown

Checking Deployed Model Version

The CNN model can be deployed as ONNX (CPU inference) or TensorRT (GPU-accelerated on Jetson). To check which is active:

curl http://localhost:8080/api/classifiers/cnn/status

# Example response
{
  "model_id": "varta-cnn-v3.2.1",
  "runtime": "tensorrt",
  "input_shape": [1, 1, 128, 128],
  "classes": 14,
  "inference_ms_p95": 4.2,
  "onnx_path": "/opt/varta/models/cnn_v3.2.1.onnx",
  "tensorrt_path": "/opt/varta/models/cnn_v3.2.1.engine",
  "loaded_at": "2026-03-01T06:00:12Z"
}

ONNX vs TensorRT Paths

• ONNX (CPU): Used on Mini v2 (Pi 5) and any deployment without a supported GPU. Inference latency is typically 35–80 ms depending on hardware.
• TensorRT (GPU): Used on Pro (Jetson Orin) deployments. Inference latency is 3–5 ms. Requires TensorRT engine rebuild when the model version changes.

Performance Summary

• Overall accuracy: 88.9%
• Shahed-series recall: 92.3% — Distinctive turbine signature provides strong acoustic fingerprint.
• Quadcopter recall: 90.9% — Multi-rotor blade-pass frequency is reliable in low-wind conditions.
• Fixed-wing recall: 81.2% — Propeller noise is more easily masked by ambient wind.

Microphone Placement Effects

• Mount microphone arrays at least 2m above ground to reduce ground-reflection interference.
• Avoid placement near HVAC units, generators, or road traffic — continuous broadband noise degrades SNR.
• Wind screens are mandatory for outdoor deployment. Without them, wind noise above 15 km/h saturates the low-frequency bins.
• For multi-mic arrays (bearing estimation), maintain calibrated inter-element spacing within 1mm tolerance.

Background Noise Impact

Acoustic detection performance degrades with ambient noise level:

Calibration Profiles

• Calibration profiles are stored on-sensor at /opt/varta/config/calibration.json and synced to Max on registration.
• Profiles include: antenna gain table, cable loss compensation, noise floor baseline per band, and DF array phase offsets (if applicable).
• Calibration is performed during initial deployment and should be re-run after any hardware change (antenna replacement, cable rerouting, firmware update).

Uncalibrated Penalty

When a sensor's calibration profile is missing or expired (older than 90 days), a 50% penalty is applied to all confidence scores from that sensor. This is a deliberate safety mechanism to prevent uncalibrated sensors from producing high-confidence false results.

# Calibration status in detection output
{
  "sensor_id": "pro-04",
  "calibration_status": "expired",
  "calibration_age_days": 112,
  "confidence_penalty": 0.50,
  "original_confidence": 0.78,
  "penalized_confidence": 0.39
}

Running a Calibration Check

To verify calibration status across your sensor fleet:

# Check calibration status for all sensors
curl http://localhost:8080/api/sensors/calibration-status

# Example response
{
  "sensors": [
    {"id": "pro-01", "calibrated": true, "age_days": 23, "status": "valid"},
    {"id": "pro-02", "calibrated": true, "age_days": 87, "status": "expiring_soon"},
    {"id": "pro-03", "calibrated": false, "age_days": null, "status": "uncalibrated"},
    {"id": "pro-04", "calibrated": true, "age_days": 112, "status": "expired"}
  ]
}

Triggering Recalibration

Initiate a calibration run on a specific sensor:

curl -X POST http://localhost:8080/api/sensors/{sensor_id}/calibrate \
  -H "Content-Type: application/json" \
  -d '{"mode": "full", "duration_minutes": 15}'

The sensor will enter calibration mode, scan all configured bands, establish noise floor baselines, and update its calibration profile. The sensor is temporarily offline during calibration.

Detection History API

Retrieve paginated detection history for analysis:

# Fetch last 100 detections
curl "http://localhost:8080/api/detections?limit=100"

# Fetch detections in a time range with confidence filter
curl "http://localhost:8080/api/detections?from=2026-03-01T00:00:00Z&to=2026-03-02T00:00:00Z&min_confidence=0.5"

# Fetch detections grouped by sensor
curl "http://localhost:8080/api/detections/by-sensor?from=2026-03-01T00:00:00Z"

CSV Export from Max

The Max UI provides a one-click CSV export from the detection history view. The export includes all detection fields, provenance metadata, and confidence breakdowns.

For programmatic CSV export:

# Export detections as CSV
curl "http://localhost:8080/api/detections/export?format=csv&from=2026-03-01&to=2026-03-07" \
  --output detections_week.csv

# Export with specific columns
curl "http://localhost:8080/api/detections/export?format=csv&fields=timestamp,confidence,class,sensor_id,snr_db" \
  --output detections_summary.csv

Confidence Distribution Analysis

Use the statistics endpoint to analyze confidence score distributions over a given period:

curl "http://localhost:8080/api/quality/confidence-stats?from=2026-03-01&to=2026-03-07"

# Example response
{
  "period": "2026-03-01 to 2026-03-07",
  "total_detections": 847,
  "confidence_distribution": {
    "0.0-0.2": 12,
    "0.2-0.4": 38,
    "0.4-0.6": 156,
    "0.6-0.8": 412,
    "0.8-1.0": 229
  },
  "mean_confidence": 0.67,
  "median_confidence": 0.71,
  "false_positive_reports": 14,
  "fp_rate_percent": 1.65,
  "snr_gate_triggered_count": 38
}

Periodic Quality Review

Recommended quality review cadence:

• Daily: Check /api/quality/confidence-stats for anomalies in FP rate or mean confidence shifts.
• Weekly: Export CSV and review low-confidence detections. Cross-reference with any known drone activity or FP reports.
• Monthly: Full calibration status audit, model version consistency check across fleet, and signature database update verification.