Inference on the plant floor, never on the PLC

OPC-UA is the right choice when you need rich type information, historical access (HA service set), and explicit security contexts - we run Basic256Sha256 with mutual X.509 certs issued from the plant CA, with audit logging on the OPC-UA server. MQTT Sparkplug B is the right choice for wide-area, low-bandwidth, high-cardinality edge-to-cloud flows - its Birth/Death certificates and unified namespace (UNS) topology make it easier to handle thousands of intermittently-connected devices. We typically run OPC-UA south of the DMZ (level 2 to level 3) and bridge to Sparkplug B for cloud-side analytics, with TLS 1.2+ and mTLS on the broker (HiveMQ or EMQX) and tagged ACLs per edge gateway. Both modes are wrapped in IEC 62443 zone-and-conduit policy.

IEC 62443 zone-and-conduit thinking maps cleanly onto the Purdue model. We place training infrastructure at level 4 (or in the customer VPC), the inference and feature store at level 3, and a data-diode or one-way replication into the level 3.5 DMZ from level 2 historians. Conduits are firewalled, certificate-pinned, and monitored - typically Claroty or Nozomi on the OT side feeds anomaly alerts into the same SIEM. Models never call out to the public internet; updates arrive as signed offline bundles, scanned at the boundary, and registered in a private MLflow. Latency budget for advisory inference is in the 100-500ms range, which fits the segmentation without contortion.

Any AI output that contributes to a GMP decision - golden-batch pass/fail, in-process release, deviation flag - becomes a Part 11 electronic record. That means ALCOA+ (attributable, legible, contemporaneous, original, accurate, complete, consistent, enduring, available) with audit trails on every change, secure e-signatures, and validated change control. We pin model versions, hash the feature vectors, and write an immutable audit log into the eBR system - MasterControl, Veeva Vault QualityDocs, or a plant historian configured for Part 11. The model itself is validated as a Computer System under GAMP 5: risk-categorised, IQ/OQ/PQ executed, with a continued process verification (CPV) plan tied to FDA Process Validation Stage 3. Pharma never auto-acts; the system recommends and a qualified person approves.

Vibration-based bearing health and RUL prediction needs raw or near-raw waveform data, not aggregated historian tags - typically 1-25.6 kHz sampling from accelerometers, captured by an edge box (NI cDAQ, OSIsoft PI Connector for high-speed, or an Ignition Edge with the Sparkplug module). For motor-current signature analysis, 10-100 Hz is enough. The PI / Ignition tags that most plants log at 1 Hz with compression are fine for OEE and process anomaly detection, but useless for incipient bearing-fault detection - you cannot see a 2 kHz bearing pass frequency in 1 Hz data. We do an information-content audit before quoting a predictive-maintenance project: if the historian compression is dropping the relevant frequencies, we install or commission edge capture first and start the model later.

It is a Pareto curve, not a single number. We start by labelling 5-10k images per defect class with the plant QA team, train a two-stage detector (a fast classifier for present/absent, then a localiser for defect type and bounding box), and calibrate the threshold against a labelled holdout. On a typical electronics or food-packaging line we target recall tuned to the floor the line requires on safety-critical defects (cracked seal, foreign object) - that comes at the cost of a false-reject rate around 1-3%, which is acceptable because rejects go to a rework lane, not the bin. For cosmetic defects we tune the other way: lower recall, near-zero false-reject, because over-rejecting blows up the scrap rate and operators will disable the system within a week. The model goes live in shadow first - it flags, the human inspector still decides - for two to four weeks before we hand over the gate.

A golden-batch model learns the multivariate envelope of the recipe phases - temperatures, pressures, agitation, addition timings - from a curated set of historically good batches. We pull the batch genealogy from the ISA-88 batch executor (DeltaV Batch, Rockwell PlantPAx, Siemens SIMATIC BATCH) as B2MML or via the OPC-UA historical access service, segment by recipe phase, and fit either a multivariate PCA / PLS model (the regulated-pharma default - explainable, easy to validate under GAMP 5) or a more flexible variational autoencoder for non-regulated chemicals. At runtime each new batch is scored phase-by-phase; deviations beyond the Hotelling T-squared or SPE control limits surface to the batch-record reviewer with the exact tags driving the deviation. For commercial pharma the model is an aid, not an arbiter - release-by-exception still requires a qualified-person signature against the eBR.

A traceability graph links every finished serial number back through its work-order, the operator and machine that built it, the in-process parameters from the historian, the component lots consumed (per BOM), and from there to the upstream supplier lots - across however many tiers the customer can give us data for. When a defect is found in the field, the query is 'which serials share this combination of supplier lot, machine, shift, and process-parameter band?' - instead of recalling six months of production we recall the affected sub-population, often two orders of magnitude smaller. We build the graph in Neo4j or TigerGraph for relational manufacturers, or in a columnar store with materialised joins where the volumes are higher (automotive Tier 1, semiconductor). For regulated industries the graph is read-only from validated source-of-truth systems (MES, ERP, LIMS) - we do not become the system of record, we become the system of inference on top of it.