Real-Time OEE Monitoring: The Architecture That Actually Works on the Shop Floor

Overall Equipment Effectiveness is the single most quoted, most misunderstood number in modern manufacturing. Half the OEE dashboards we see in Indian plants are mathematically wrong. The other half are mathematically right but practically meaningless because the inputs are gamed at source.

A working OEE system is an architecture problem, not a UI problem. This article is the reference architecture we deploy when we own the OEE stack end-to-end.

The Five Layers

• Acquisition — PLC / SCADA tags, energy meters, vision sensors, operator panels. Raw signals only.
• Edge — a local node per line normalising signals to canonical events. Buffers 72 hours on WAN outage.
• Stream — a message broker (Kafka in larger plants, MQTT in smaller) that durably stores every event.
• Compute — stateless services that compute availability, performance and quality on rolling windows.
• Surface — dashboards, ANDON boards, mobile push and ERP postings.

Where OEE Calculations Go Wrong

1. Planned downtime is misclassified as availability loss — inflates losses, masks real bottlenecks.
2. Ideal cycle time is set to nameplate — should be set to demonstrated best, not vendor brochure.
3. Quality is measured at end-of-line only — misses rework hidden in WIP.
4. Micro-stops below threshold are ignored — in high-velocity lines, these are the entire problem.

Canonical Event Schema

Every event flowing into the OEE engine carries the same five fields: machine ID, timestamp (UTC, millisecond), event type, value and source. Downtime reasons, scrap codes and quality categories are attached as enriched metadata downstream. Keep the event shape ruthlessly clean — every "just one more field" costs you for the next decade.

Real-Time vs Eventually Consistent

A live dashboard target of 200 ms is fine for the production manager. The ERP confirmation does not need to be sub-second — it needs to be accurate. Decouple the two paths. Live OEE runs on the stream. ERP posting runs on a settled view that triggers on confirmed material events.

Operator-Sourced Downtime Reasons

No PLC will ever tell you that the line is down because the trolley operator went for chai. Auto-detect the stop — then prompt the operator on a tablet with a short, hierarchical list (Mechanical → Bearing failure → Conveyor drive) plus a free-text field. Force a reason within a configurable grace window or the system promotes the event to "uncategorised — escalate to shift supervisor".

The "Andon Without Andon Boards" Pattern

Physical ANDON boards still matter — but in 2026 your real ANDON is the shift supervisor’s phone. Push the same event taxonomy to a mobile app with SLA-driven escalation: 5 minutes silence → shift lead, 10 minutes → plant head, 20 minutes → operations director. Audit every escalation as part of the weekly review.

Reference Stack

For a 200-machine plant, the stack we standardise on:

• Connectivity: OPC-UA from machines, MQTT from cheaper sensors, REST for legacy bolt-ons.
• Edge: a small industrial PC per line running our edge agent.
• Broker: Kafka in 3-broker HA across a small on-prem cluster.
• Time-series store: TimescaleDB. Relational store: PostgreSQL.
• Compute: Python services for analytics, Go services for the hot path.
• Surface: React dashboards, mobile push via FCM, ERP postings via IDoc/BAPI.