System Overview
Jope.SMB is a Simulated Moving Bed (SMB) chromatography control system that uses real-time Raman spectroscopy and PLS-based concentration soft-sensing to replace traditional UV single-wavelength detection. It operates under FDA 21 CFR Part 11 electronic records and signatures compliance.
Purpose of this document
This document gives a high-level understanding of what Jope.SMB does, who uses it, and what quality attributes matter. For visual topology see High-Level Topology. For the Console ↔ Inference wire contract see ZMQ Integration Protocol.
Scope
In scope:
- 6-column SMB chromatography (extendable to N columns)
- Raman spectrum acquisition via RS2000 dual-probe (2048 wavenumbers, 785 nm)
- PLS + Ridge soft-sensor for EPA / DHA / DPA concentration (fish oil case)
- MPC (with embedded DANN) + PID hierarchical control
- Batch lifecycle FSM (
idle → running → held → completed), backed by Stateless - Electronic signature + append-only audit trail (21 CFR 11.10 / 11.100)
- Model hot-swap + on-demand training via REST management channel
- Multilingual operator UI (en / zh-TW / zh-CN / ja), runtime-switchable
Out of scope (this document):
- Business logic beyond chromatography process
- Recipe contents and SOPs
- Operator training materials (see Operator Manual, coming soon)
Stakeholders
| Role | Concern |
|---|---|
| Plant Operator | Daily batch operation · alarms · electronic signatures |
| Plant QA | Audit trail · signature validity · compliance reports |
| Plant IT | User management · backup · deployment · patching |
| Jope engineering | Jope.Core + Jope.UI + Jope.SMB development |
| AI research (off-site) | PLS model training · validation · delivery |
| FDA / Auditor | CSA (Computer System Assurance) evidence |
Key Quality Attributes
| Attribute | Target | Notes |
|---|---|---|
| Inference round-trip | ≤ 20 ms p95 | See ZMQ Protocol · Performance |
| Batch FSM transitions | deterministic | Stateless-backed, transition log recorded |
| Audit integrity | append-only · tamper-evident | Hash-chained events, 21 CFR 11.10(e) |
| Data retention | 7 years | Batch records + audit events + Raman spectra |
| Localization | 4 locales · runtime-switchable | en / zh-TW / zh-CN / ja |
| Deployment | Plant IPC (Console + Historian) + dedicated Inference Host on plant LAN | No cluster dependency |
| Availability | 99.9% | Uptime target during production runs |
High-Level System Boundaries
- Plant IPC: Operator Console + Historian (TimescaleDB). Console owns all hardware I/O.
- Inference Host (dedicated, separate machine): Python Inference Server runs on its own plant-LAN host — keeps inference, model registry, and training jobs fully isolated from Plant IPC. Linux + systemd is the recommended environment.
- Hardware tier: 6 chromatography columns + rotary / pneumatic valves + pumps + UV / Raman detectors, connected to Plant IPC via serial buses (Modbus / ASCII / APAX)
- Off-site: ML research team dev machines for model training — no direct network link to plant; model files delivered via signed package
Document Map
- High-Level Topology — full runtime diagram and component descriptions
- ZMQ Integration Protocol — wire contract between Console and Inference Server
- Hardware Interfaces (coming soon) — device-level communication protocols
- Compliance Mapping (coming soon) — FDA 21 CFR Part 11 controls mapped to system layers
- Deployment Topology (coming soon) — installer, process manager, backup, upgrade paths