Solution Architecture
This section details the technical architecture of the AI-powered multi-sensor intelligence platform that Trifork will deliver as part of the Aerotec-Trifork partnership for YPF's Unmanned Aerial Monitoring Service.
Contents
| Section | Description |
|---|---|
| Architecture | Platform design and component architecture |
| Sensors | Multi-sensor integration: methane, thermal, LiDAR, GPR |
| AI Use Cases | 17 mandatory detection and monitoring capabilities |
| Integration | FlightHub 2, SAP PM, SCADA, and Genetec connectivity |
| Alerting | Real-time alerting pipeline and prioritization |
Executive Summary
The Trifork AI Platform sits on top of DJI FlightHub 2, providing the multi-sensor intelligence layer that transforms raw drone data into actionable operational insights. This is not flight management - Aerotec handles autonomous operations through FlightHub 2. The platform focuses on what the data means.
Platform Concept
┌─────────────────────────────────────────────────────────────────────┐
│ ENTERPRISE INTEGRATION │
│ ┌──────────────┐ ┌──────────────┐ ┌──────────────┐ │
│ │ SAP PM │ │ Genetec │ │ RTIC │ │
│ │ Work Orders │ │ Security │ │ Dashboards │ │
│ └──────────────┘ └──────────────┘ └──────────────┘ │
└─────────────────────────────────────────────────────────────────────┘
│
┌─────────────────────────────────────────────────────────────────────┐
│ TRIFORK AI PLATFORM │
│ ┌──────────────────────────────────────────────────────────────┐ │
│ │ Multi-Sensor Intelligence │ │
│ │ • AI Use Cases • Event Correlation • Alerting │ │
│ └──────────────────────────────────────────────────────────────┘ │
│ │
│ ┌──────────────┐ ┌──────────────┐ ┌──────────────┐ │
│ │ Inference │ │ Correlation │ │ Triton │ │
│ │ Pipeline │ │ Engine │ │ Ensemble │ │
│ └──────────────┘ └──────────────┘ └──────────────┘ │
└─────────────────────────────────────────────────────────────────────┘
│
┌─────────────────────────────────────────────────────────────────────┐
│ DATA INGESTION LAYER │
│ ┌──────────────────────────────────────────────────────────────┐ │
│ │ Unified Multi-Sensor Data Pipeline │ │
│ │ • FlightHub API • Sensor Data • Geospatial │ │
│ └──────────────────────────────────────────────────────────────┘ │
└─────────────────────────────────────────────────────────────────────┘
│
┌─────────────────────────────────────────────────────────────────────┐
│ DJI FLIGHTHUB 2 │
│ ┌──────────────┐ ┌──────────────┐ ┌──────────────┐ │
│ │ Flight Ops │ │ Mission │ │ Media │ │
│ │ Control │ │ Planning │ │ Storage │ │
│ └──────────────┘ └──────────────┘ └──────────────┘ │
└─────────────────────────────────────────────────────────────────────┘
│
┌─────────────────────────────────────────────────────────────────────┐
│ AUTONOMOUS DRONE FLEET │
│ ┌──────────────┐ ┌──────────────┐ ┌──────────────┐ │
│ │ Dock 3 │ │ Matrice 4TD │ │ Matrice 400 │ │
│ │ Stations │ │ Thermal │ │ Heavy Lift │ │
│ └──────────────┘ └──────────────┘ └──────────────┘ │
└─────────────────────────────────────────────────────────────────────┘
Data Flow Architecture
Key Differentiator: Multi-Sensor Correlation
The platform's unique value is correlating signals across sensor modalities:
| Single Sensor | Limitation | Multi-Sensor Correlation |
|---|---|---|
| Methane detection | 300% error at wind >7 m/s | + Visual + thermal confirms source |
| Thermal anomaly | False positives from sun | + Visual context eliminates false alarms |
| Visual leak | Cannot confirm substance | + Thermal + methane identifies gas vs. liquid |
| GPR subsurface | Abstract data visualization | + Geospatial overlay with surface assets |
Technology Stack
| Component | Technology | Rationale |
|---|---|---|
| Inference Server | NVIDIA Triton | Multi-model ensemble orchestration |
| Vision AI | GroundingDINO + SAM2 | Zero-shot detection + precise segmentation |
| OCR Pipeline | AWS Rekognition | Text extraction with confidence scoring |
| Geospatial | PostGIS | Digital twin asset correlation |
| Message Queue | RabbitMQ + Kafka | Real-time event streaming |
| Time Series | InfluxDB | Sensor telemetry storage |
Production Proof: US Rail Yard
The platform architecture is production-proven in Trifork's autonomous rail yard monitoring deployment for a major US maintenance operator:
| Capability | Production Performance |
|---|---|
| Detection accuracy | 92-98% in industrial outdoor conditions |
| End-to-end latency | ~30 minutes from mission to analytics |
| Coverage | 100-200 acres with 4 autonomous docks |
| Operations | Night flights with spotlight, all weather |
The same GroundingDINO + SAM2 + OCR ensemble pattern, PostGIS digital twin, and micro-batch processing approach transfers directly to YPF.
Design Principles
1. AI Assists, Humans Decide
The platform provides confidence-scored recommendations, not autonomous actions:
- Detection events include confidence levels
- Operators review flagged anomalies
- Escalation based on severity tiers
- Human-in-loop for work order creation
2. Phased Accuracy Improvement
AI models improve through operational feedback:
- Phase 1: Baseline data collection and annotation
- Phase 2: Initial heuristics with operator validation
- Phase 3: Correlation logic refinement
- Phase 4: Production accuracy targets met
3. Graceful Degradation
System remains operational with partial sensor availability:
- Individual sensor failures don't block detection
- Reduced confidence without correlation
- Alerts on sensor health issues
- Manual override capabilities
SLA Compliance
| Requirement | Platform Design |
|---|---|
| Raw data upload ≤10 min | Direct FlightHub API integration |
| Analytics ≤20 min | Micro-batch processing pipeline |
| AI availability ≥90% | Redundant inference deployment |
| Detection accuracy ≥95% | Ensemble model + operator feedback loop |