🥈 2nd Place Winner — UBC Defence Tech Hackathon 2026

A high-performance RF Signal Classification and Geolocation Common Operating Picture (COP) developed for the LockedIn challenge.

• Deep Learning: 1D-CNN (PyTorch) for raw IQ feature extraction
• Machine Learning: Scikit-Learn Hybrid Ensemble (HistGradientBoosting + MLP + OC-SVM)
• Geolocation: Multi-receiver TDoA & RSSI Trilateration with Kalman Filter smoothing
• Backend: Flask + Socket.IO for real-time data orchestration
• Frontend: Leaflet.js + custom tactical HUD (Glassmorphism & Neon UI)

# 1. Set up virtual environment
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

# 2. Set your API key
echo "API_KEY=your-team-api-key" >> .env

# 3. Place training data in data/ directory
#    Download the HDF5 file and put it at data/training.h5

# 4. Train the Hybrid Classifier
python3 main.py train

# 5. Start the tactical dashboard
python3 main.py server --port 5050
# → Open http://localhost:5050

LockedIn_FindMyForce/
├── classifier/
│   ├── signal_classifier.py    # Hybrid Engine: Deep 1D-CNN + Expert Features + Ensemble
│   └── __init__.py
├── pipeline/
│   ├── geolocator.py           # RSSI & TDoA trilateration + Kalman filtering
│   ├── track_manager.py        # Persistent track lifecycle (Tentative -> Confirmed)
│   ├── associator.py           # Multi-receiver pulse grouping (Temporal/IQ Similarity)
│   ├── feed_consumer.py        # Real-time SSE stream consumption
│   └── __init__.py
├── dashboard/
│   ├── index.html              # Tactical HUD (Tailwind + Custom CSS)
│   ├── style.css               # Premium "Find My Force" HUD styling
│   └── app.js                  # Socket.IO & Leaflet map orchestration
├── data/                       # RF training data (HDF5)
├── models/                     # Saved model artifacts (.joblib + CNN weights)
├── server.py                   # High-concurrency Flask/Socket.IO backend
├── main.py                     # Unified CLI entry point
├── requirements.txt
└── .env                        # Configuration (Team API Key)

# Start dashboard server
python3 main.py server [--port 5050] [--debug]

# Train classifier on HDF5 data
python3 main.py train

# Stream live feed observations (debug)
python3 main.py stream

# Fetch team score
python3 main.py score

• Deep 1D-CNN: Extracts 64-dimensional latent features directly from raw 256-sample IQ snapshots. Architecture includes multiple Conv1D layers, BatchNorm, and Dropout for robust representation learning.
• Feature Extraction (Manual): 42 expert-engineered features (spectral entropy, kurtosis, Phase Histogram, LPC coefficients, etc.).
• Hybrid Ensemble: Concatenates deep CNN features with manual features, followed by a Voting Ensemble of HistGradientBoosting and MLP.
• One-Class SVM: Trained on high-dimensional hybrid features of friendly data — detects hostile/civilian signals as anomalies.

1. Load HDF5 data → extract raw IQ and 42 manual feature vectors.
2. Train 1D-CNN on raw IQ to minimize Cross-Entropy loss for known classes.
3. Freeze CNN and use as a deep feature extractor.
4. Scale combined manual + deep features (Quantile Transformer).
5. Train the Multi-class Ensemble on the hybrid feature space.
6. Calibrate anomaly threshold at 15th percentile of friendly OOD scores.
7. Save hybrid model bundle to models/classifier.joblib.

• Converts RSSI to distance via path-loss model: d = d_ref × 10^((RSSI_ref - RSSI) / (10 × n))
• Nonlinear least-squares optimization (scipy least_squares)
• SNR-weighted receiver contributions
• Fallback to 2-receiver and single-receiver modes

• Uses time-of-arrival differences to compute hyperbolic position fixes
• Linearized initial estimate, then nonlinear refinement
• Requires ≥3 receivers with valid ToA

• 70% TDoA + 30% RSSI weighted position average when both methods available
• GDOP (Geometric Dilution of Precision) estimation

• State: [x, y, vx, vy] (position + velocity)
• Constant velocity motion model
• Measurement update weighted by geolocation uncertainty

• TENTATIVE → CONFIRMED after 2 updates
• CONFIRMED → COASTING after 20s without updates
• COASTING → LOST after 60s
• Kalman-smoothed position estimates
• Classification confidence EMA fusion across observations

Groups independent receiver observations into single-emitter groups using:

1. Temporal gate: observations within 500ms window
2. Classification gate: same signal type (or at least one "unknown")
3. IQ similarity gate: cosine similarity > 0.5

• Classification (40%): Correctly identifies friendly types AND detects hostile/civilian as anomalies
• Geolocation (30%): RSSI + TDoA hybrid with Kalman smoothing minimizes CEP
• Novelty (30%): OC-SVM flags unknowns; use intelligence briefing labels for hostile sub-types

• Airborne-detection — Airborne surveillance radar
• Airborne-range — Airborne range-finding radar
• Air-Ground-MTI — Air-to-ground moving target indicator
• EW-Jammer — Electronic warfare / broadband jammer
• AM radio — Commercial AM radio broadcast

• Leaflet map with dark tactical overlay
• Real-time track markers color-coded by affiliation (green=friendly, red=hostile, orange=unknown, blue=civilian)
• Uncertainty radius circles around each track
• Track history path polylines
• Receiver station markers with tooltip details
• Live observation feed panel
• Track detail panel with classification confidence, GDOP, velocity
• Track filter by affiliation
• Score display with per-component breakdown
• One-click eval submission button
• Socket.IO real-time updates (<100ms latency)