NOTQIN — OT Track

NOTQIN Physics

4 min read

NOTQIN Physics (Physics-Informed Industrial AI)

1. Project Summary

A physics-informed AI engine that grounds every maintenance recommendation in governing engineering equations, ODE simulations, and cost scenarios — not just statistical ML. Operators ask engineering questions about machines; the agent uses SymPy/SciPy to solve physics equations and verify recommendations against hard safety constraints before answering.

2. Architecture

POST /ask  or  POST /ask/stream (SSE)
  ↓
Claude LLM (via OpenRouter or Google AI Studio)
  │ 15 tool calls
  ↓
┌──────────────────────────────────────────────┐
│ fetch_telemetry        → InfluxDB / CSV       │
│ list_equations         → equations.yaml       │
│ solve_equation         → SymPy symbolic       │
│ evaluate_equation      → SymPy numerical      │
│ simulate_motor_thermal → SciPy ODE (RK45)     │
│ simulate_dye_bath      → SciPy ODE            │
│ simulate_pressure_decay → SciPy ODE           │
│ classify_failure       → failure-modes.yaml   │
│ estimate_scenario_costs → action-scenarios.yaml│
│ verify_recommendation  → constraints.yaml     │
└──────────────────────────────────────────────┘
  ↓
ProofObject (PASS / WARN / BLOCK verdict)
  ↓
AnswerCard (answer + equations + confidence + cost scenarios)

3. Tech Stack

LayerTechnology
APIFastAPI 0.115 + Uvicorn
LLMClaude via OpenRouter or Google AI Studio
Physics SolverSymPy 1.13 (symbolic + numerical)
ODE SimulatorSciPy 1.13 (solve_ivp, RK45)
OptimizationCasADi 3.6 (optional, nonlinear)
Unit RegistryPint 0.24 (dimensional analysis)
TelemetryInfluxDB client + CSV fallback
ConfigYAML (equations, assets, constraints, scenarios, failure modes)
MetricsPrometheus client

4. Physics Engine

Equations Library (50+ equations in config/equations.yaml)

  • Bearing: RMS velocity, ISO 281 L10 fatigue life, BPFO/BPFI fault frequencies
  • Motor: Thermal ODE (winding temp transient), efficiency equation
  • Compressor: Specific power, pressure decay ODE
  • Weaving HVAC: Psychrometric dew point, saturation
  • Dyeing: Heat transfer, time-to-setpoint ODE
  • Each has: formula, variables (name, unit, description), tags, ISO reference

ODE Simulators

# Motor thermal winding temperature transient
SimulationEngine.motor_thermal(current, ambient_temp, thermal_resistance, thermal_capacity)
→ SimResult(time_series, units, steady_state)
 
# Dye bath temperature ramp (heating + hold)
SimulationEngine.dye_bath_ramp(initial_temp, target_temp, ramp_rate, hold_duration)
 
# Compressor pressure decay (leak detection)
SimulationEngine.compressor_pressure_decay(initial_pressure, leak_rate)

Asset Library (7 families in config/assets.yaml)

  • bearing, motor, compressor, weaving_hall_hvac, dyeing_machine, autoclave, cnc_machine
  • Per asset: state variables, inputs, hard limits (warn/alarm thresholds), preferred solver

Constraint Library (config/constraints.yaml)

  • Thermal: winding max °C, discharge max pressure bar
  • Load: motor current imbalance % max
  • Safety interlocks: autoclave pressure, vibration thresholds

Verification Engine

ProofObject.verdict: PASS | WARN | BLOCK
ProofObject.block_reason: "Temperature exceeds hard limit: 185°C > 180°C max"

Failure Mode Taxonomy (config/failure_modes.yaml)

  • Bearing: outer race spall, inner race spall, cage wear, lubrication starvation
  • Motor: winding short, overtemp, current imbalance
  • Compressor: discharge temp high, pressure low, leaking
  • Autoclave: pressure overshoot, cure undertime, steam isolation failure

Cost Scenario Engine (config/action_scenarios.yaml)

Action: replace_bearing_now
  downtime: 4h = 8,000 MAD
  spares: 2,500 MAD
  benefit: extends MTTF by 6 months

Action: delay_to_scheduled
  risk: 35% chance of unplanned failure = 40,000 MAD average cost

Action: do_nothing
  risk: progressive failure, CBAM audit evidence gap

5. Role-Aware Responses

The agent adapts its answer framing per user role:

RoleFocus
operationsProduction throughput, downtime cost
maintenanceTechnical root cause, repair procedure
financeMAD cost, payback period, CBAM exposure
engineeringPhysics equations, ODE simulation
operatorSimple instructions, safety warnings

6. Current Status

Stage: MVP — ready for integration testing

Last commit: Apr 23, 2026

API Endpoints:

  • POST /ask — blocking Q&A (JSON in, AnswerCard out)
  • POST /ask/stream — SSE stream (thinking → tool calls → verification → final)
  • GET /health — liveness + config summary
  • GET /equations — equation catalog (filterable by tag)
  • GET /roles — supported user roles
  • GET /failure-modes — failure taxonomy
  • GET /metrics — Prometheus metrics

Tests: tests/test_symbolic.py, tests/test_direct.py, tests/test_agent_debug.py

What’s complete (85%):

  • Full agent loop with 15 tools
  • SymPy solver (50+ equations)
  • SciPy ODE simulator (3 simulators)
  • Pint unit registry
  • Failure-mode classifier
  • Cost scenario estimator
  • Constraint verifier
  • Prometheus + SSE streaming

What’s incomplete (15%):

  • 7 of 22 asset types scaffolded (others follow same pattern)
  • CasADi optimization not fully wired
  • No Dockerfile (run from source)
  • Real telemetry caching (re-queries InfluxDB each call)
  • No per-user auth on /ask endpoint
  • State Service integration (doesn’t poll Redis digital twin)

7. Integration Position

Physics Service ← called by →
  IEIA (operational questions needing physics validation)
  Analyst Service (complex anomaly explanations)

Physics Service reads from →
  NOTQIN InfluxDB (live telemetry)
  NOTQIN CSV (offline dev)

8. Next Actions

  • Create Dockerfile for physics-service
  • Add Redis cache for telemetry (don’t re-query InfluxDB per call)
  • Integrate with IEIA: IEIA calls physics-service for recommendation validation
  • Integrate with State Service for “current” machine condition
  • Complete remaining 15 asset type schemas
  • Add rate limiting + auth on /ask
  • Load test with 10 concurrent requests

Graph View

Backlinks