Every component of the A.U.R.A. system — from raw sensor ingestion to AI-generated recommendations
FastAPI backend + WebSocket streaming + single-page frontend — no installation, browser only
A.U.R.A. runs as a FastAPI web server with an asyncio background loop that generates and ingests sensor data every second, runs the full ML pipeline, and broadcasts results to all connected browsers via WebSocket simultaneously — no polling.
Data is persisted to SQLite (WAL journal mode, foreign keys enforced) with a normalized schema: faults → locations → time-series readings. The frontend is a vanilla JS SPA with zero framework overhead.
An SVG-based schematic of the International Space Station rendered live in the browser. Each of the 7 modules displays a color-coded health indicator driven by real-time WebSocket data: green = nominal, amber = warning, red = critical fault active.
Clicking any module navigates to Sensor Detail pre-filtered to that location. A Three.js 3D mode provides an alternative spatial view of the ISS structure.
The Dashboard aggregates all 20 sensor parameters from all 7 ISS locations into a unified view, organized by ECLSS subsystem so operators can scan by system rather than raw parameter name.
Each cell displays the current value, unit, nominal range, and a status indicator. The historical queue buffers 10,000+ rows for fast trend lookups without database round-trips.
Sensor Detail provides a deep-dive into a single location: all 20 parameters with current value, nominal range, deviation score, and ML classification confidence. The Trends tab plots parameter history as interactive Chart.js line charts with nominal range overlays.
A.U.R.A. uses two distinct alert levels to prevent alarm fatigue. Critical alerts fire only when the ML pipeline is highly confident (Random Forest confidence >85%, sustained for configurable consecutive ticks). Warning alerts fire when parameters drift toward limits.
Active fault detected. RF confidence above threshold for N consecutive readings. Immediate action required.
Parameter approaching limit. Fault not yet confirmed. Monitor closely, schedule inspection.
The AI Analyst integrates a large language model to synthesize all model outputs — Isolation Forest scores, Random Forest classification and confidence, LSTM failure probability, LSTM Remaining Useful Life, and DQN action recommendation — into natural-language situational reports.
Operators get a plain-English explanation of what's wrong, why the system believes it, and what to do next. The analyst can be queried interactively for deeper investigation.
The Maintenance tab surfaces the DQN Recommender's output: a ranked list of corrective actions for the current system state. Each recommendation includes the target fault, the suggested procedure, and the model's confidence score.
When Random Forest confidence exceeds 92%, the system bypasses the DQN and applies the direct fault-to-action mapping from the knowledge base — highest-confidence path first.
Each model passes its output downstream — raw sensor noise becomes an actionable recommendation in under a second
failure_prob (0–1) and rul (remaining useful life in hours). Enables proactive scheduling.Every threshold that drives alert behavior is configurable at runtime through the Settings page. Changes persist to a JSON config file and take effect on the next server tick — no restart needed.
alert_min_consecutive — ticks of anomaly before alert fires (default: 10)alert_cooldown_seconds — minimum time between repeat alerts (default: 600)alert_critical_rf_gate — RF confidence threshold for critical alert (default: 0.85)latch_threshold — RF confidence to latch fault state (default: 0.95)latch_min_consecutive — ticks to confirm fault latch (default: 3)