Worked Architectures
Four end-to-end examples, each walked through the authority ladder: which rung the evidence supports, and what stays independent when the model is wrong. One worked example beats three abstractions.
This page is the closing page of the AI & ML Integration section; read the authority gate first. It introduces no new claims — it composes the positions already established across the section into four concrete architectures. Review pending.
How to read each example
Every architecture answers four questions in the same order — the job, the rung (highest authority the evidence supports, and why not higher), what stays independent when the model fails, and the seam and the contract where the result crosses into control. The recurring shape is the envelope architecture: the learned component informs or proposes; a verified non-learned layer decides and protects; no safety function routes through the learned path.
| Architecture | Model family | Rung | What stays independent |
|---|---|---|---|
| Vision inspection cell | 2-D CNN | 2 · advisory | Reject/divert actuation, part-tracking, interlocks (PLC-owned) |
| Predictive-maintenance pipeline | 1-D CNN / temporal | 2 · advisory | Protection and process control; maintenance planning absorbs errors |
| PINN soft sensor | Physics-informed | 1–2 (0 if the loop closes on it) | Real instruments, deterministic estimators, the control loops |
| Read-only LLM copilot | LLM | 2 advisory · 0 for drafted code | Everything — the copilot cannot act |
1. Vision inspection cell
The job. A 2-D CNN inspects each part — presence/absence, defect class, or a measurement — so the cell can reject or pass it.
The rung — advisory (level 2), PLC owns the action. The CNN classifies; a deterministic rule or a person sets the accept/reject policy, and the PLC owns the reject-trigger timing and actuation. Not higher: a perception model’s field accuracy cannot be inferred from its bench accuracy, so a misclassification must be caught downstream, not trusted as a plant fact.
What stays independent. The reject/divert mechanism, part-tracking, and any interlock are deterministic and outside the model. A missed or false classification degrades yield or passes a bad part to a downstream check — it defeats no safety function, because none depends on the CNN.
The seam and the contract. The CNN runs at the cell (edge inference) and publishes a result — class, confidence, model version, timestamp, quality — that the PLC reads and gates. Low confidence or a stale result routes to the defined response (hold, divert to manual inspection), never a silent pass.
2. Predictive-maintenance pipeline
The job. A 1-D CNN or temporal model watches vibration (and possibly motor current) on a rotating asset and flags likely bearing or gearbox wear for maintenance planning.
The rung — advisory (level 2). It flags; a person schedules. Not higher: the fault-diagnosis literature’s near-100% accuracies are, as a class, an artefact of train/test leakage — under a leakage-free component-wise split the honest number can fall to near chance, and no source measures learned diagnosis on a real industrial fleet (see Model Families & Fit). A fleet operator whose training set holds many of its own bearings may earn more — but that is a claim the operator’s own validation must make. Sensor choice matters: on the reference rig, vibration beat motor current in every case.
What stays independent. Nothing trips, derates, or issues an automated work order on the model’s say-so above advisory; protection and process control are untouched. A missed or false flag is absorbed by the planning process.
The seam and the contract. Acquire and infer at the edge (a kHz waveform cannot be reconstructed by oversampling a decoupled tag); publish the verdict — class, confidence, model version, freshness — into the historian/CMMS, not the raw waveform. Drift monitoring decides when the flag is no longer trustworthy.
3. PINN soft sensor
The job. A physics-informed model estimates an unmeasured or hard-to-measure variable from sparse instruments — a soft sensor feeding operator awareness.
The rung — read-only to advisory (level 1–2); level 0 for anything the loop closes on. A PINN may support estimation and calibration, but it must not be the authoritative estimator in a control loop: a data-driven PINN can revert to a steady-state solution when the data flow stops (operationally, between measurement updates), and because the PDE residual can stay low while the solution is wrong, a residual-based health monitor can be fooled — silent, plausible, wrong. Compare against an established numerical solver, and validate at the deployment numerical precision — an FP32 edge target can break a PINN that was healthy in FP64.
What stays independent. The real instruments, the deterministic estimators (EKF/UKF/observers), and the control loops that would otherwise use them. The soft sensor is a second opinion cross-checked against measurements, not a replacement; the loop never closes on it.
The seam and the contract. The estimate crosses with an explicit uncertainty band and a freshness bound; when it drifts out of the validated regime or disagrees with an independent measurement, it is flagged and the operator falls back to the measured/deterministic value.
4. Read-only LLM copilot
The job. An LLM answers engineering and maintenance questions over the plant’s own documentation, alarms, history, and model evidence — and can draft procedures or code for a human to verify.
The rung — advisory (level 2) for the copilot; offline (level 0) for any code it drafts. It consumes curated, read-only context through allow-listed tools, never raw tag streams, and holds no write access. Drafted control code goes through grammar/compiler/formal verification with a human owner before it is anything more — raw output is not a control artefact. The NIST AI RMF Generative AI Profile is the relevant companion, and confabulation (fluent, plausible, wrong output) is the failure mode to design against.
What stays independent. Everything. The copilot cannot act; it can only inform. Its worst failure is a wrong answer a reviewing engineer must catch — which is why every output is reviewed and every drafted artefact verified before use.
The seam and the contract. Read-only, least-privilege access inside the OT security zones (NIST SP 800-82r3 placement); each answer cites the twin state, document, or model evidence it rests on, so a human can check it. No proposal reaches the plant without schema validation, authorization, and — where required — human confirmation.
What the four share
flowchart LR
subgraph LEARNED["Learned component — informs / proposes"]
M["Vision · vibration · PINN · LLM
result + confidence + version + freshness"]
end
subgraph GATE["Verified non-learned layer — decides / protects"]
ENV["Envelope: fresh? in-distribution?
confident? corroborated?"]
end
SAFE["Safety functions · interlocks · protection
(independent of the learned path)"]
ACT["Plant action"]
M -->|gated contract| ENV
ENV -->|"pass"| ACT
ENV -.->|"stale / low-confidence:
defined failure response"| ACT
SAFE --> ACT
classDef learned fill:#e8eff8,stroke:#2b5797,color:#1e1e1e
classDef verified fill:#fdf6e3,stroke:#8b6914,color:#1e1e1e
classDef safety fill:#faeeec,stroke:#a33327,color:#1e1e1e,stroke-width:2px
classDef io fill:#f2f2ef,stroke:#a0a09a,color:#1e1e1e
class M learned
class ENV verified
class SAFE safety
class ACT io
- The ladder, not the model, sets the authority. Three of the four top out at advisory; none holds a safety function; none closes a control loop on a learned estimate.
- The non-learned envelope always holds the veto, and the safety-function path never routes through the learned component.
- The result crosses as a gated contract — class/estimate, confidence/uncertainty, model version, timestamp, freshness — not a raw stream; a stale or low-confidence result triggers the defined response.
- Authority is perishable — each example depends on drift monitoring, honest (leakage-free, deployment-condition) validation, and a maintained model-evidence ledger to keep its rung.
Review status
Distilled from a corpus-owned Draft note and Review pending. This page introduces no new technical claims; it composes the positions verified across the section. Return to the authority gate, the method register, or any of Safety & Security Boundaries, Interfaces & Handshakes, Model Families & Fit, The Digital Twin, and Validation & Lifecycle.
This site is an independent educational reference for industrial automation standards. It paraphrases and organizes guidance in original language; it is not a substitute for the official published standards, professional engineering judgment, or legal review, and content has not been independently verified against current published editions.
Items marked Review pending await technical review against the identified edition. Items marked Planned have no reviewed source material behind them. Do not rely on this site for compliance determinations, safety-critical design decisions, or legal interpretation.