IIoT and Edge Analytics on the Factory Floor

Factory data becomes useful when a consumer can determine what a value represents, when it was observed, whether it is trustworthy, which production state produced it, and who owns its definition. Connectivity alone does not provide those properties.

Divide protocol roles

OPC UA is well suited to structured address spaces, typed data, methods, events, security, and information models close to industrial assets. MQTT is a lightweight publish/subscribe transport that decouples producers and consumers through a broker. They can be complementary.

That distinction matters. MQTT does not define your plant semantics, and OPC UA does not guarantee that a model is useful simply because it is structured. The namespace, information model, security profile, broker policy, and consumer contract still require engineering.

An edge application might read a controller through OPC UA, attach production context, and publish governed events through MQTT. Avoid flattening rich source semantics into ambiguous topic strings. Preserve units, data type, quality, source timestamp, asset identity, and the transformation version.

Design topics as contracts

An MQTT namespace should remain stable while payloads evolve under explicit versioning. A practical hierarchy could include enterprise, site, area, line, asset, and message class, but physical hierarchy alone is not sufficient. Consumers also need to distinguish state, event, alarm, metric, command, and configuration traffic.

Document retained-message behavior and session expiry. A retained “running” state can mislead a new subscriber after the machine disconnects unless birth/death certificates or freshness rules expose the stale state.

Quality of Service is not business-level exactly-once processing. QoS 1 can deliver duplicates. Consumers should be idempotent using message identity, source sequence, or event keys. QoS 2 adds protocol overhead and still does not solve every downstream duplication path.

Preserve time and quality

Source timestamps should be created as close to measurement as practical. Gateway receipt time is valuable, but it is not interchangeable with PLC sample time. Maintain synchronized clocks and monitor clock health.

Every value should carry or inherit:

• engineering unit and scaling;
• source and asset identity;
• measurement or calculation timestamp;
• quality or validity state;
• sample mode and expected interval;
• transformation lineage;
• production mode, recipe, and product context where relevant.

Without production state, an anomaly detector may learn that planned changeover, sanitation, manual mode, and bearing failure are variations of the same pattern.

Engineer store and forward

Network interruption is normal. The edge layer needs bounded buffering, durable queues where loss is unacceptable, ordering rules, disk monitoring, and an explicit overflow policy. After reconnection, replay traffic must not overwhelm the broker or downstream historian.

Test long outages, clock changes, broker failover, certificate expiry, full disk, duplicate replay, corrupted records, and abrupt power loss. Observe whether the system loses data silently or raises an actionable condition.

Put analytics at the right layer

Edge analytics is useful when low latency, bandwidth, local autonomy, or data volume makes central processing impractical. Examples include vibration feature extraction, image pre-processing, state-based aggregation, or a local quality interlock.

Keep safety and deterministic base control independent from analytics infrastructure unless the complete function is engineered and validated for that role. A cloud score should not become an undocumented permissive in a PLC.

Analytics deployments need versioned code, model identity, input schema, confidence or validity behavior, rollback, and monitoring. If inputs disappear or move out of the calibrated region, the service should abstain or degrade predictably.

Acceptance evidence

A production-ready pipeline can demonstrate:

1. traceability from source tag to consumer field;
2. correct units, timestamps, and quality propagation;
3. bounded behavior through outage and replay;
4. least-privilege identities and certificate rotation;
5. observable queue, disk, and transformation health;
6. consumer handling of duplicates and late data;
7. documented owners for namespace and schema changes.

IIoT architecture earns trust when data meaning and failure behavior are engineered as deliberately as connectivity.

A small acceptance test with high yield

Disconnect the edge node for longer than its in-memory queue, change a tag value several times, restart the broker, and reconnect. Verify which values arrive, in what order, with which timestamps and quality. Then repeat with a duplicate message and a clock offset.

If the consumer cannot distinguish a late historical event from a live state change, the pipeline is not safe to use for operational decisions—regardless of how attractive the dashboard looks.

Sources and further verification

The sources below support the factual frame of this article. Vendor case studies are treated as attributed claims, not independent performance validation.

1. MQTT Version 5.0 — OASIS Open
2. OPC UA Online Reference — OPC Foundation
3. OPC Unified Architecture — Overview and Concepts — OPC Foundation
4. NIST SP 800-82 Rev. 3 — Guide to Operational Technology Security — National Institute of Standards and Technology

Read the archive’s sourcing, correction, and evidence policy.