VoltSentinel

AI-powered digital twin for electrical infrastructure that detects anomalies, evaluates operational risk, and simulates failure scenarios with explainable diagnostics.

Comment

Inspiration

Traditional electrical grid and industrial infrastructure monitoring is fundamentally reactive. In existing substations and factory plants, operators typically discover electrical faults only after a circuit breaker trips, machinery sustains costly damage, or severe arcing/fires erupt.

We wanted to rethink electrical monitoring from the ground up by asking: What if an AI-powered digital twin could detect the subtle, compound warning signs hours before failure occurs?

That vision became VoltSentinel—a real-time predictive digital twin combining physics-based thermal modeling, machine learning anomaly detection, and deterministic explainability.

What It Does

VoltSentinel monitors live telemetry across 12 electrical feeder circuits grouped into 4 distinct operational zones:

  • North Zone: Residential & Municipal infeed.
  • South Zone: High-load EV charging hubs & tech campuses.
  • Industrial Zone: Heavy induction motor banks & chiller plants.
  • Commercial Zone: Edge data center racks & logistics facilities.

Key Capabilities:

  1. Interactive 2D Digital Twin: A hierarchical topology visualization (Utility Substation → 4 Zones → 12 Feeders) with live status indicators, glowing busbars, and circuit health metrics.
  2. 3-Layer Hybrid AI Anomaly Engine:
    • Layer 1 (Statistical): Rolling $3\sigma$ Z-scores tracking voltage, current, and temperature deviations.
    • Layer 2 (ML Outlier Detection): Scikit-Learn Isolation Forest evaluating multidimensional feature vectors ($V, I, T, \text{PF}, P, \frac{\Delta T}{\Delta t}$).
    • Layer 3 (Engineering Rules): Deterministic IEC/IEEE thermal ($75^\circ\text{C}$), overcurrent, and power factor bounds.
  3. Deterministic Risk Scoring ($0 - 100$): Classifies severity into NORMAL, WATCH, WARNING, or CRITICAL with automated failure mode categorization (THERMAL_OVERLOAD, VOLTAGE_INSTABILITY, COOLING_FAILURE, SUSTAINED_OVERCURRENT, INTERMITTENT_FAULT).
  4. Offline Explainable AI: Every alert provides a structured root-cause explanation ("Why this alert?") and practical operator mitigation steps ("Recommended Action").
  5. Scenario Simulation Lab: An interactive control room allowing engineers to inject physical fault trajectories (Thermal Runaway, Feeder Overload, Voltage Surge, Cooling Failure, Micro-Arcing) with accelerated time multipliers ($1\times, 2\times, 5\times, 10\times$).
  6. Hardware-Ready Telemetry Ingestion: Canonical POST /api/telemetry contract built to interface with real ESP32 microcontrollers with non-invasive current transformers (CTs) and isolated thermistors.

How We Built It

  • Backend & ML Pipeline: Built with FastAPI, Pydantic v2, and SQLAlchemy on SQLite with bounded ring-buffer telemetry retention. The anomaly engine combines rolling Z-score baselines with a pre-trained Scikit-Learn Isolation Forest.
  • Real-Time Streaming: High-throughput WebSockets broadcast live 1 Hz multi-circuit telemetry frames (/ws/telemetry) and immediate incident alerts (/ws/events).
  • Frontend & Digital Twin: Developed with Next.js 14+ (App Router), TypeScript, React, and Tailwind CSS, featuring custom SVG topology layouts and Recharts multi-channel oscillographs.
  • Physics Simulator: Modeled Joule-heating thermal mass inertia ($T_t = f(I_t^2)$) and autoregressive electrical noise (AR(1)) to ensure measurements exhibit realistic temporal continuity.

Challenges We Ran Into

  1. Modeling Physics-Based Telemetry Continuity: Unlike synthetic dashboards with independent random numbers, real electrical systems have physical inertia—conductors take time to heat up and cool down based on current squared ($I^2R$). We implemented a discrete thermal time-constant differential equation to model gradual thermal buildup.
  2. Sub-Second Hybrid ML Inference: Combining statistical baselines, scikit-learn tree ensembles, and deterministic risk rules for 12 simultaneous circuits at 1 Hz required optimizing feature extraction vectors to prevent WebSocket event loop lag.
  3. Deterministic Explainability: Ensuring diagnostic root-cause text and actionable recommendations are 100% deterministic and available offline without relying on expensive, high-latency cloud LLMs.

Accomplishments That We're Proud Of

  • 100% Hardware-Independent Intelligence Layer: Real ESP32 microcontrollers can seamlessly stream to POST /api/telemetry using the exact same API contract without altering the intelligence pipeline.
  • Flawless End-to-End Demo Sequence: Injecting thermal runaway on Circuit 07 visibly demonstrates the entire pipeline in under 15 seconds: telemetry changes → ML flags anomaly → Risk crosses Critical → Digital Twin pulses red → Incident generates explainable advice → System recovers on reset.
  • 12/12 Automated Tests Passing: Comprehensive pytest coverage across API endpoints, anomaly detection thresholds, risk scoring, and simulator state transitions.

What We Learned

  • How to combine statistical process control with machine learning tree ensembles for high-reliability industrial IoT monitoring.
  • Designing high-density dark UI/UX for industrial control rooms that balances information hierarchy without cognitive overload.
  • Structuring hardware ingestion APIs to facilitate rapid prototyping before deploying physical split-core CT sensors.

What's Next for VoltSentinel

  • Physical ESP32 Pilot Deployments: Field-testing non-invasive split-core CT clamps (SCT-013) on real-world multi-phase distribution panels.
  • Transformer DGA (Dissolved Gas Analysis) Extension: Incorporating acoustic and gas sensor features into the ML isolation forest.
  • Edge ML Deployment: Porting the Isolation Forest model directly to microcontrollers using TensorFlow Lite for Microcontrollers (TFLite Micro).

Built With

Share this project:

Updates

Submission history