DeepRefine: Open Industrial Reference Architecture & Testbed

Architectural Positioning: DeepRefine is an open-source reference architecture and simulation testbed for Purdue Model Level 3 Supervisory Real-Time Optimization (RTO), Physics-Informed Digital Twins (PINN), and API 670 Machinery Diagnostics in continuous brownfield chemical and petroleum refining plants.

---

📌 Architectural Scope & Safety Boundaries

Important

Safety & Operational Boundaries (IEC 61511 / IEC 62443 / OSHA 1910.119 PSM):

1. Phase 1 Read-Only Operator Advisory Mode: DeepRefine possesses zero direct write authority over physical Level 1 actuators (control valves, variable frequency drives, or heaters). All AI/RTO recommendations are presented as bounded advisory targets requiring human operator sign-off and DCS validation.
2. Level 3 Supervisory Subordination: DeepRefine operates strictly at Level 3, generating target vectors for classical Level 2 Dynamic Matrix Controllers (DMC). The DMC enforces hard slew-rate limits and multivariable dynamic coupling.
3. Hardware-Enforced SIS Priority: Emergency shutdown (ESD) elements and safety instrumented functions (SIF) reside entirely in independent physical safety systems (e.g., Triconex / HIMA TMR 2oo3 logic solvers). A 1500 ms hardware dry-contact watchdog forces instant fallback to safe park mode if supervisory communications lapse.
4. HIL Benchmark Testbed: All plant dynamic responses and telemetry are driven by verified first-principles thermodynamic and kinetic benchmark simulators designed for rapid what-if testing and algorithm verification.

---

🏗️ Master Architectural Schema (Purdue / ISA-95)

Level 4: Enterprise Network & ERP
┌───────────────────────────────────────────────────────────────────────────────┐
│ • Streamlit ISA-101 Unified Operations Console (5 Dedicated Tabs)             │
│ • SAP PM / IBM Maximo Maintenance Dispatch (ISO 14224 Equipment Taxonomy)     │
│ • FastAPI Production Services & Prometheus Metric Telemetry Scrapers          │
└───────────────────────────────────────▲───────────────────────────────────────┘
                                        │  IEC 62443-3-3 Level 3.5 IDMZ
                                        │  (mTLS x.509 + Unidirectional Telemetry)
Level 3: Supervisory Advisory & Digital Twin
┌───────────────────────────────────────────────────────────────────────────────┐
│ • Pillar 1: Predictive Maintenance (API 670 Orbit Vectors, ISO 14224, Bi-LSTM)│
│ • Pillar 2: Thermodynamic Digital Twin (20-Tray VLE, 4-Lump FCCU, VDI 2048)   │
│ • Pillar 3: Supervisory RTO (PPO Continuous Policy + KKT Constrained QP DMC)  │
│ • Inferential Soft Sensors: FCC RON/MON, Jet Flash/Freeze, Diesel D86 95%    │
│ • LIMS Closed-Loop Feedback: Exponential bias update with 3.5σ Chauvenet gate │
└───────────────────────────────────────▲───────────────────────────────────────┘
                                        │  Read-Only Hardware Permissives &
                                        │  1500 ms Heartbeat Watchdog Relay
Level 1-2: Regulatory DCS & Independent SIL-3 SIS
┌───────────────────────────────────────────────────────────────────────────────┐
│ • Level 2 Dynamic DMC (FOPDT FIR Step-Response Convolution over P=25, M=5)    │
│ • Level 1 PID Regulatory Control Loops (Flow, Pressure, Level, Temperature)   │
│ • Level 0 Field Actuators & SIL-3 TMR 2oo3 Critical Slide Valve dP Interlock  │
└───────────────────────────────────────────────────────────────────────────────┘

---

💎 Core Technical Innovations

1. Rotating Machinery Reliability & Diagnostics (Pillar 1)

• API 670 Proximity Orbit Analysis: Orthogonal 90° X/Y eddy-current probe vector decomposition ($S_{\max}$), tracking 1X unbalance, 2X misalignment, and 0.43X oil whirl hydrodynamic instability on hydrodynamic journal bearings.
• API 617 & 610 Machine Hydraulics: Continuous ASME PTC 10 dynamic distance-to-surge monitoring for centrifugal compressors and NPSHa cavitation margin tracking for API 610 pumps.
• ISO 14224 Failure Taxonomy: Standardized mapping of mechanical failure modes, severity levels, and automated SAP Maintenance Notification (Type M2) work orders.
• Multi-Task Bi-LSTM Network: Simultaneous classification of 5-day mechanical failure horizon and continuous Remaining Useful Life (RUL in hours) with serialized TorchScript SCADA edge binary (< 10 ms inference).

2. Coupled Thermodynamic Digital Twin & Soft Sensors (Pillar 2)

• 20-Tray Dynamic VLE Distillation Engine: Multi-component state-space vapor-liquid equilibrium (Naphtha, Kerosene, Diesel, Atmospheric Residue) with tray hydraulic flooding checks via Fair's correlation.
• Coupled 4-Lump FCCU Reactor: Heavy gas oil catalytic cracking kinetics with reactor-regenerator thermal balance and coke burn emissions modeling.
• VDI 2048 Data Validation & Reconciliation (DVR): Constrained Weighted Least Squares (WLS) mass conservation engine with global $\chi^2$ Gross Error Detection (GED) and normalized residual penalty tests ($\tau$-test) to isolate faulty plant flowmeters.
• True Mass-Conserving PINN Surrogate: High-speed neural surrogate (< 1 ms) enforcing metric ton/hour ($\text{tph}$) mass balance using petroleum cut specific gravities ($\text{SG} \times \rho_{\text{water}}$).
• Inferential Soft Sensors with LIMS Reconciliation: Real-time virtual property predictors (RON/MON, Pensky-Martens flash point, freeze point, D86 95%) with automated lab assay feedback and exponential bias updating.

3. Supervisory RTO & Constrained Dynamic DMC (Pillar 3)

• Level 3 Supervisory RTO: Reinforcement learning continuous setpoint policy optimizing refinery crack margins and energy consumption under safe operational bounds.
• Level 2 Constrained Dynamic DMC: True Karush-Kuhn-Tucker (KKT) constrained Quadratic Programming (QP) solver optimizing over control horizon $M=5$ and prediction horizon $P=25$, enforcing actuator rate limits ($-\Delta u_{\max} \le \Delta u \le \Delta u_{\max}$) and absolute MV boundaries.
• FMEA Runtime Watchdog & Bumpless APC Transfer: Multi-tier fault detection with automatic failover to traditional linear APC if sensor sentinel checks fail.

4. Operational Technology (OT) Security & ISA-101 HMI

• IEC 62443 Level 3.5 IDMZ: Authenticated reverse-proxy staging architecture separating Level 4 IT applications from Level 2/3 plant control, blocking any direct valve commands.
• ISA-101 High Performance HMI: Greyscale SVG P&ID plant mimic designed per ANSI/ISA-101.01 ergonomics to reduce operator fatigue and surface abnormal situations instantly.

---

⚡ Quickstart Guide

1. Clone & Install Dependencies

git clone https://github.com/your-org/DeepRefine.git
cd DeepRefine
pip install -r requirements.txt

2. One-Command Master Launcher

Launch pre-flight integrity checks and the ISA-101 Streamlit Command Center:

python run_quickstart.py

• Dashboard URL: http://localhost:8501
  • Tab 1: Plant Overview & P&ID Mimic: ISA-101 high-performance console with live process telemetry.
  • Tab 2: Asset Health & API 670 Diagnostics: Orbit plots, 1X/2X harmonics, and ISO 14224 work orders.
  • Tab 3: Digital Twin & VDI 2048 DVR: Mass balance reconciliation, gross error detection, and what-if simulation.
  • Tab 4: Quality Inferentials & LIMS: Soft sensor property predictions and lab assay calibration updates.
  • Tab 5: Supervisory RTO & Dynamic DMC: Level 3 setpoint policy, constrained QP moves, and SIS status.

3. Run Automated Industrial Verification Suites

Execute the full council hardening and acceptance test matrix:

# 1. Council Hardening Verification (5 Core Industrial Pillars)
python src/test_council_hardening.py

# 2. 10-Point Factory & Site Acceptance Test Protocol (FAT / SAT)
python src/test_fat_sat_protocol.py

---

📊 AI/ML vs Classical Engineering Taxonomy

Architectural Subsystem	Implementation Methodology	Classification	Standard Compliance
API 670 Vibration & Orbits	Vector combination ($S_{\max}$), 1X/2X harmonic decomposition	Classical Deterministic	API 670 / ISO 10816-3
Machinery Anomaly & RUL	Multi-Task Bidirectional LSTM with TorchScript edge compilation	Deep Learning (AI)	ISO 14224
Sensor Sentinel	Chauvenet 3.5$\sigma$ outlier gate, ADC frozen channel detector	Statistical Quality Control	ASTM E178
VLE Column Distillation	Dynamic 20-tray MESH state-space VLE differential equations	First-Principles Physics	AIChE / GPSA
FCCU Cracking Reactor	4-lump Arrhenius reaction network + regenerator thermal balance	First-Principles Kinetics	UOP / Kellogg Kinetics
VDI 2048 DVR Engine	Constrained Weighted Least Squares (WLS) + Global $\chi^2$ Test	Analytical Optimization	VDI 2048 / ISO 5168
Digital Twin Surrogate	Deep Residual MLP with metric tons/hr ($\text{tph}$) mass penalty	Physics-Informed ML (PINN)	AIChE / VDI 2048
Inferential Soft Sensors	Empirical polynomial correlations with LIMS exponential bias	Semi-Empirical Inferential	ASTM D86 / D93 / D2386
Supervisory RTO	Continuous PPO Actor-Critic Setpoint Guidance	Reinforcement Learning (AI)	ISA-95 Level 3
Dynamic Matrix Control	KKT Constrained Quadratic Programming (QP) with FIR convolution	Advanced Process Control (APC)	Cutler & Ramaker (1979)
Safety Interlock Gateway	Emulated TMR 2oo3 majority voting + 1500 ms dry-contact watchdog	Hardware Interlock Emulation	IEC 61511 (SIL-3)
OT Network Security	IDMZ Reverse Proxy with mTLS certificate inspection	Cyber-Physical Security	IEC 62443-3-3

---

📜 License

Distributed under the MIT License. See LICENSE for more information.