High-Level Topology

This page shows the full runtime topology — where each software process lives, what hardware it talks to, which protocols connect them, and where off-site actors sit relative to the Plant IPC.

Diagram

Layout

Inference Server always runs on its own dedicated plant-LAN host. This isolates the inference workload, leaves Plant IPC focused on hardware I/O + compliance, and gives Python / model retraining its own CPU / RAM / (optional) GPU budget.

Components

Operator Console · Jope.SMB

• Process: WPF .NET 8 interactive Windows application
• Responsibilities: Batch control FSM · operator UI · audit trail writes · electronic signature · historian writes · hardware I/O
• State: Stateful — holds batch state machine, user session, in-memory alarm queue, recipe cache
• Built on: Jope.Core (Transport · Compliance · Identity) + Jope.UI (Shell · Login · Signature dialog · Theming · i18n)

Inference Server · Inference Host (dedicated)

• Process: Python 3.11 running as a long-lived daemon on a dedicated plant-LAN host
• OS: Linux (recommended — systemd service) or Docker / Podman container
• Responsibilities: Raman spectrum → concentration prediction (PLS + Ridge) · model registry · training jobs · health endpoint
• State: Stateless per request — no batch memory. Model loaded once from disk registry at startup; hot-swappable at runtime.
• Channels: ZMQ REP for predict, HTTP REST for /model/* and /training/*
• Why dedicated host: isolates inference workload from real-time hardware I/O, allows larger models / parallel inference / optional GPU acceleration, keeps training jobs from interfering with Console responsiveness, Linux is the native environment for Python ML tooling

Historian · TimescaleDB

• Process: TimescaleDB 2.x on the Plant IPC (same machine as Console)
• Content: all Raman spectra · all predictions · all signatures · all alarms · all audit events · all batch records
• Retention: 7 years (FDA requirement)
• Access: Console writes directly; Inference Server may read for optional model QC audits

Hardware Cluster

Device	Role	Primary Protocol
RS2000 Raman Spectrometer	2048-wavenumber spectrum acquisition, dual probe	Modbus RS485 (gateway to Ethernet)
NP7000 / EPP / P3700 Pumps	Feed / eluent / recycle flow control	Serial ASCII 16-byte / binary 0x88 / binary 0xFF
SKS S3612 rotary valve	12-port inlet/outlet zone switching	ASCII 12-char over RS232/485
SKS S3203 switching valve	3-port feed routing	ASCII 4-char
APAX pneumatic valves (PV301-306)	On/off gas/liquid isolation	APAX module
NU3000 / UV1000D UV Detectors	Auxiliary absorbance readings	Serial ASCII 16-byte

See Hardware Interfaces (coming soon) for device-level register maps.

Off-Site Actor · AI Research

• Location: Academic / research partner dev machine (off plant network)
• Workflow: Trains PLS + Ridge models on historical Raman + HPLC-labeled data, exports .joblib model file + metadata JSON
• Delivery: Signed + encrypted package transferred to plant (sneaker-net or secure transfer). No direct network link to Plant IPC — keeps production air-gap intact.

Communication Channels

From	To	Protocol	Direction	Purpose
Raman	Console	Modbus RS485	→	Spectrum read
Console	Pumps	Serial ASCII / Binary	↔	Flow setpoint · status
Console	Valves	Serial / APAX	↔	Position command · feedback
Console	UV Detectors	Serial ASCII	←	Absorbance read
Console	Inference Server	ZMQ REQ-REP	↔	predict · hot path ≤ 20 ms
Console	Inference Server	HTTP REST	↔	Model load · training trigger · health (cold)
Console	Historian	TCP (TimescaleDB)	→	Writes: spectra, predictions, audit, batch
Inference	Historian	TCP (TimescaleDB)	← (optional)	Model QC audits, read-only
Off-site Dev	Plant	Signed package	→	.joblib model delivery

See ZMQ Integration Protocol for the Console ↔ Inference wire contract.

Deployment

Inference Host is always a separate machine on the plant LAN. Console reaches it via TCP using the endpoint address in its config file (example: tcp://10.0.1.42:5555).

Full deployment steps (installer, process manager, backup, upgrade) are in Deployment Topology.