VEXPLOR
TECHNOLOGY DEEP DIVE — DARK FACTORY IMPLEMENTATION

This is howa dark factory gets built.

Based on the floor of Korea's government Autonomous Factory Program site (Speefox, selected 2025), this page opens up the full closed loop — from data collection to production operations, stage by stage. Not a concept: the actual process we are implementing on a verified site.

KOREA MSS AUTONOMOUS FACTORY PROGRAM · PLC 243 · 186·25·6 SYNC VERIFIED · DES ACC 85% VERIFIED · AI→PLC VERIFIED
00THE LOOPCPS 5-LAYER · BI-DIRECTIONAL

The physical factory and the virtual one
drive each other.

The state of the physical plant streams into its digital twin in real time, and AI decisions made in the virtual plant return — after verification — as control commands on real machines. This bi-directional closed loop is the skeleton of a dark factory, and the four stages below cycle through it without pause.

A closed ribbon of light connecting machines and a decision nodeFIG — THE LOOP SIMPLY RUNS
STAGE 01 — PERCEPTION

Data collection —
every signal, turned into standard data.

Signal strands from the press line and data from server racks converging into an edge cabinetFIG — TWO SOURCES, ONE STREAM

A dark factory starts with data. On the Speefox program site we connect 243 PLCs — 186 finishing presses plus semi-finished, heat-treatment, de-oiling, washing and drying equipment — and collect the floor's signals in three levels: from field PLCs and sensors to process edge nodes, and from the edge to the plant data warehouse and lake.

Heterogeneous equipment is the real obstacle: every machine speaks a different protocol. An edge gateway unifies everything from direct-wired RS-485/RS-232 PLCs to OPC-UA and TCP/IP, and the collected data is organized into AAS (IEC 63278) standard asset models — so every value carries meaning: which machine, which parameter.

Machine signals alone are not enough — what to build, and by when, lives in the business systems. So we also connect orders, production plans, BOMs and work instructions from MES and ERP, putting the floor's current state and production's targets on one data foundation.

FIELD [ PLC 243 · SENSORS ] + BUSINESS [ MES · ERP — ORDERS · BOM · WORK INSTRUCTIONS ] EDGE GATEWAY [ RS-485 · RS-232 · OPC-UA · TCP/IP ] KAFKA EVENT BUS TSDB(INFLUXDB) · DATA LAKE AAS(BASYX) STANDARD ASSET MODELFIG — 3-LEVEL COLLECTION · 100,000+ RAW SENSOR RECORDS PER SECOND AT THE EDGE
  • Three-level collectionField PLCs and sensors, then process edge, then the plant data warehouse and lake. First-pass processing at the edge cuts network load and latency.SCALE100,000+ SENSOR RECORDS/SEC · PLC 243 (186 FINISHING PRESSES + 30 SEMI-FINISHED + HEAT-TREAT · DE-OIL · WASH · DRY)
  • Business systems — what to buildOrders, production plans, BOMs and work instructions come from MES and ERP. If machine signals say "what state we are in," this data says "what must be built, and by when."SOURCEEXISTING MES INTEGRATION · ERP ORDERS/DUE DATES/BOM · WORK INSTRUCTION HISTORY · MATERIAL DATA
  • What we collect — four data typesMachine state (run status, stop reasons, control parameters), process (production speed, cycle time, temperature), quality (vision inspection images, defect types) and logistics (position, inventory, transfer history).SIGNALSSTOP REASONS · MATERIAL REQUEST SIGNALS · PLATEN TEMPERATURE · QR TRACE HISTORY · DEFECT IMAGES · VISION UNITS 80
  • Heterogeneous protocol unificationMachines of different vintages and makers, unified behind one edge gateway. Older equipment is not left behind.PROTOCOLSDIRECT PLC (RS-485 · RS-232) · OPC-UA · TCP/IP · MQTT
  • Event bus, tiered storageData flows straight onto a Kafka event bus for microservices to divide and process, then lands in whichever store fits the data type.STACKKAFKA · INFLUXDB (TIME SERIES) · MONGODB (DOCUMENT) · POSTGRESQL (RELATIONAL) · S3/MINIO (DATA LAKE)
  • AAS standard asset modelsMachines, sensors and products modeled as international-standard asset shells (Shell · Submodel) — meaningful data instead of raw tags.STANDARDAAS IEC 63278 · ECLIPSE BASYX · WACE-AAS MODELER (3 DAYS PER MODEL · 57% FASTER, AAS MODELING ONLY)
STAGE 02 — DECISION

Analysis and simulation —
virtually first, always.

A holographic twin above the physical line branching into three scenariosFIG — THE FUTURE IS TESTED FIRST

On top of the collected data, the AI makes its decisions. The starting point is the production plan: taking orders, due dates and BOMs from ERP and MES as input, it optimizes what to build, when, and in what sequence — then validates that plan virtually before anything runs. One principle governs everything: no decision touches a machine before it has been verified in the virtual plant.

A hybrid simulation engine combining DES (discrete-event) and ABS (agent-based) compares What-If scenarios in parallel — machine breakdowns, material delays, rush orders, bottlenecks — to pre-verify each decision. Simulation accuracy was certified at 85% by an accredited testing body (a composite of time, resource and output accuracy), and ML-based parameter calibration keeps closing the gap with the real floor. The whole plant is visualized as a browser-based Unity 3D digital twin, synchronized with the floor at the 100ms level (in rollout).

DATA [ MACHINE STATE + MES·ERP ORDERS·BOM ] AI DECISIONING [ NSGA-II MULTI-OBJECTIVE · RL DISPATCHING (DQN·PPO) · CSP · MONTE CARLO ] DES + ABS SIMULATION [ PARALLEL WHAT-IF · ACC 85% VERIFIED ] PASS / FAILFIG — VALIDATE BEFORE EXECUTE · NO CONTROL WITHOUT VERIFICATION
  • DES+ABS hybrid simulationThe whole line runs virtually, scenario against scenario, results in numbers. Production plans and rush-order responses are verified before execution.PROOFSIMULATION ACCURACY 85% — ACCREDITED TEST, VERIFIED
    WHAT-IFDUE-DATE-FIRST VS UTILIZATION-FIRST VS MIN-CHANGEOVER · 1–3 MACHINE FAILURES · MATERIAL DELAY 30M/1H/2H · SENSITIVITY ±10/±20%
  • Production-plan optimization and AI decisioningWith orders, due dates and BOMs as input, it computes production sequence and material timing. Multi-objective optimization balances utilization, lead time and energy, while reinforcement-learning dispatching picks its own rules by situation.ALGONSGA-II · DQN · PPO · FALLBACK RULES SPT/EDD/CR/ATCS · CSP · MONTE CARLO
    FEEDBACKLEARNS OPTIMAL SPEED·TEMP·CYCLE PER MACHINE — SLOWS DOWN WHEN DEFECTS RISE, SPEEDS UP WHEN THERE IS ROOM
  • Predictive maintenance — equipment failure forecastingLearning from operating data and inspection images together, it flags equipment heading toward failure before it breaks and recommends maintenance timing. Wear-part life forecasting — press dies, for example — works the same way.MODELLSTM AE+CNN ANOMALY FORECASTING · TIME-SERIES LIFE PREDICTION + COX SURVIVAL ANALYSIS · PER-MACHINE MAINTENANCE GUIDANCE
  • Quality-process correlationWhen a defect appears, the data traces back to the process conditions that caused it. A defect whose cause is known does not come back.METHODPCA · SHAP FEATURE IMPORTANCE · RANDOM FOREST · SPC/CPK LIVE MONITORING
  • Unity 3D WebGL digital twinThe whole plant opens in a browser — no installation. CAD drawings become 3D models through an automated conversion pipeline.TWINUNITY URP · CAD (STEP/IGES)→FBX AUTO-CONVERSION · LOD ×4 · GPU INSTANCING · WEBSOCKET 100MS SYNC (IN ROLLOUT) · 30+ DYNAMIC OBJECTS AT 20FPS
  • Vision AI quality judgmentIt catches 0.1mm defects on glare-heavy metal surfaces. Low-confidence samples route to a person; when judgment drift is detected, retraining kicks in automatically.VISIONCNN ENSEMBLE (EFFICIENTNET-B3 · RESNET50 · MOBILENETV2) · TENSORRT INT8 <100MS · POLARIZED CAMERA + 2-STAGE GAN REFLECTION REMOVAL · 80 UNITS IN ROLLOUT
STAGE 03 — EXECUTION

Machine control —
only verified decisions reach the machines.

A verified beam of light passing through gate rings down into a PLC cabinetFIG — ONLY WHAT PASSES MAY COMMAND

This is where a simulation-approved decision reaches real equipment — through a double safety layer. Commands that would violate physical constraints are blocked by rule checks before execution, and control authority itself is handed over stepwise through the AAF framework (L0–L4), only as far as it has been verified. First the AI only judges (Shadow Mode); once its agreement rate is proven, it moves to approve-then-execute; only then to bounded autonomous execution.

This AI-to-PLC direct-control path has been demonstrated on physical equipment, and on the Speefox site we demonstrated simultaneous synchronization of 186 presses, 25 AMRs and 6 EMS units (the AMRs and EMS are the customer's existing equipment, integrated under AI supervision).

VALIDATED DECISION CONSTRAINT CHECK [ PHYSICS-VIOLATING COMMANDS BLOCKED PRE-EXECUTION ] AAF GATE [ SHADOW → APPROVE → AUTONOMOUS ] EDGE NODE [ COMMAND VALIDATION ] PLC EXECUTE AUDIT LOGFIG — CONTROL PATH · AI→PLC DIRECT CONTROL VERIFIED ON PHYSICAL EQUIPMENT
  • Large-scale heterogeneous control186 presses, 25 AMRs and 6 EMS units synchronized as one line. Not a machine at a time — the whole line, orchestrated.PROOF186 · 25 · 6 SIMULTANEOUS SYNC — VERIFIED · AI→PLC DIRECT CONTROL — VERIFIED
  • Per-equipment control scopePresses take run/stop, die-change signals and downstream-linked speed adjustment; heat treatment takes per-product parameters with load balancing; de-oiling and washing take energy-aware scheduling; EMS supplies material preemptively from depletion forecasts; AMRs run traffic-aware dynamic routing with collision avoidance; cranes balance FIFO against urgency.SCOPEPRESS RUN/STOP · DIE-CHANGE SIGNAL · LINKED SPEED CONTROL · HEAT-TREAT PARAMS · ENERGY SCHEDULING · EMS PREEMPTIVE SUPPLY · AMR DYNAMIC ROUTING · CRANE FIFO-URGENCY BALANCE
  • Machine-to-machine linked controlOne machine's failure must not cascade down the line — on breakdown, an alternate machine is assigned automatically; at a bottleneck, upstream and downstream speeds synchronize.LINKEDFAILURE → ALTERNATE ASSIGNMENT · BOTTLENECK → UP/DOWNSTREAM SYNC
  • Double verification and emergency stopThe system is designed to generate only constraint-satisfying solutions (CSP, penalty functions), with remote emergency stop and restart plus segmented industrial-network security.SAFETYCONSTRAINT-SATISFYING SOLUTIONS ONLY · REMOTE E-STOP/RESTART · NETWORK SEGMENTATION · FULL CONTROL AUDIT LOGGING
  • Stepwise authority handoverWe do not rush autonomy. In Shadow Mode the AI's judgments run in parallel with experienced operators, and only a verified agreement rate opens the next level.GOVERNANCESHADOW MODE → APPROVE-THEN-EXECUTE → AUTONOMOUS CONTROL · AAF L0–L4
STAGE 04 — OPERATION

Operations —
the loop runs; people handle exceptions.

A dark factory hall with AMR light trails and one person in a glass control roomFIG — THE FACTORY RUNS IN THE DARK

When the four stages connect, the plant runs like this: an anomaly is detected, the AI analyzes the cause, validates a response in simulation, reschedules production and redeploys machines and logistics. The whole cycle plays out live on the digital twin in the control room.

The human role changes — from a watcher of every machine to a decision-maker who handles the exceptions and approval requests the AI raises. New equipment and policy changes are tested by virtual commissioning before anyone touches the real line, eliminating the cost of failure. When judgment accuracy wavers, drift detection triggers MLOps retraining automatically.

DETECT ANALYZE [ ROOT CAUSE ] SIMULATE [ VALIDATE RESPONSE ] RESCHEDULE [ PLAN · DISPATCH ] EXECUTE DETECT …FIG — THE LOOP IN OPERATION · PEOPLE HANDLE EXCEPTIONS AND APPROVALS ONLY
  • Live control roomThe whole plant in one browser-based 3D twin — machine states, logistics flow and anomaly alerts moving on a single screen.VIEWUNITY 3D WEBGL · LIVE STATE · ALERT ROUTING
  • Anomaly response scenariosOn a breakdown, an alternate machine is assigned and the plan recomputed. A rush order gets a What-If simulation to find its optimal insertion point. Quality anomalies trigger severity-tiered alerts and adjust the linked processes together.CASESFAILURE → ALTERNATE ASSIGNMENT · RUSH ORDER → WHAT-IF INSERTION · QUALITY ANOMALY → TIERED ALERTS + MAINTENANCE TIMING GUIDANCE
  • Virtual commissioningEquipment additions, layout changes and dispatch-policy swaps are trialed in the twin, not on the live line.USENEW-EQUIPMENT VALIDATION · POLICY PRE-VERIFICATION · TRAINING · EMERGENCY DRILLS
  • AI that maintains itselfThe floor keeps changing. Simulation forecasts are compared against actual production continuously to retrain the models, and drift in judgment distributions triggers retraining automatically.MLOPSFORECAST VS ACTUAL, ALWAYS ON · KL DIVERGENCE DRIFT DETECTION · AUTO-RETRAIN · AUTO-CALIBRATION
  • Honest autonomyWe do not promise a lights-out plant overnight. Autonomy widens only within verified bounds, and the entire process leaves an audit trail.PRINCIPLESTEPWISE AUTONOMY · FULL AUDIT LOGGING · PROGRAM TARGET FIGURES (PLANNED) ARE NOT SHOWN HERE
05PORTABILITYSAME STACK · DIFFERENT PROCESS TYPES

Not one factory's success —
a technology stack that moves from plant to plant.

None of the above is a one-off build. The 2026 program site (Chexcar, used-car remanufacturing) is being built by transplanting the same stack from Speefox — proof that one stack handles both high-volume repetitive production and high-mix, per-vehicle judgment.

VERIFIED85%DES simulation accuracy — accredited test
VERIFIED186 · 25 · 6Presses, AMRs and EMS synchronized simultaneously
VERIFIEDAI → PLCDirect control demonstrated on physical equipment
Program target figures (planned) are not shown · VERIFIED = accredited test or field demonstration · AMRs and EMS are the customer's existing equipment integrated under AI supervision

How would this apply to your plant?

We diagnose your current equipment and data, then design the starting stage with you.

Autonomous Factory Service Request a Diagnosis