AI & Automation5 min readNetray Engineering TeamUpdated March 4, 2025

Real-Time Manufacturing Analytics: Architecture and Payback

Real-time manufacturing analytics delivers shop floor and order status within seconds to minutes of the physical event, instead of the next morning after a nightly batch. It is built by streaming machine signals and ERP transaction changes into a low-latency store that dashboards and alerts read continuously. The discipline is knowing where it is worth the cost. Downtime response, constraint utilization, and expedite decisions genuinely change with sub-minute data; monthly margin analysis does not. Build real-time for the decisions that happen within a shift, and keep everything else on batch.

Decide Which Decisions Justify Real Time

Latency requirements should be derived from decision cadence, not ambition. Machine downtime response and andon escalation need 10 to 60 seconds, because a stopped constraint costs money by the minute. Constraint work center utilization and short-interval control benefit from 1 to 5 minute latency. Order status, expedite decisions, and material shortages work fine at 5 to 15 minutes. Inventory accuracy, supplier performance, and financial reporting are perfectly served by nightly batch. Applying a streaming architecture uniformly multiplies cost and operational burden for no gain; a good design usually streams two or three data flows and batches the other twenty. Write the latency target next to each use case in the design document so the decision is explicit and reviewable later.

Sub-minute: machine state, downtime events, andon alerts, and cycle counts on constraint assets
1 to 5 minutes: constraint utilization, live OEE, and short-interval control boards
5 to 15 minutes: job status, material shortages, and at-risk order lists for expediting
Nightly batch: margin, supplier scorecards, inventory valuation, and executive trend reporting

Getting ERP Changes Out in Near Real Time

Polling the ERP every 30 seconds is the usual first attempt and the usual first failure, because repeated scans of jobtran or transaction tables load the production database exactly when it is busiest. Log-based change data capture is the correct mechanism: SQL Server CDC or a Debezium connector reads the transaction log and emits inserts, updates, and deletes with negligible impact on the OLTP workload. For Infor environments, ION and the Infor Data Lake can publish BODs for many entities, which is convenient for integration but usually carries higher and less predictable latency than direct CDC. Whichever path you choose, retain ordering guarantees per key so a status update never overtakes the insert that created the row.

Streaming Machine Data From the Floor

Machine data typically arrives over OPC UA from PLCs or through an MQTT broker using Sparkplug B, which provides birth and death certificates so you can distinguish a stopped machine from a lost connection. Normalize tags against an asset registry immediately, because raw tag names are plant-specific and unmaintainable in analytics. Compute derived states such as running, idle, changeover, and faulted in the stream rather than in the dashboard, so every consumer agrees on state. Buffer at the edge so a network interruption does not silently lose a shift of data. Reconcile stream-derived counts against ERP job transaction quantities daily, since operators will trust the board only if the parts it claims were produced match what the ERP eventually booked.

Prefer MQTT Sparkplug B or OPC UA over custom polling scripts; you get state and connection semantics
Map raw tags to an asset registry at ingest so downstream models never reference plant-specific names
Derive machine state and downtime reason in the stream, not in each dashboard
Buffer at the edge gateway for at least 24 hours to survive network or broker outages

Serving Layer, Alerting, and Reconciliation

Streams need a query-friendly destination. Options include a materialized-view database such as ClickHouse or a streaming engine like Flink writing aggregates to a fast store, with dashboards reading pre-aggregated rows rather than recomputing on every refresh. Alerting is where real time actually pays: an escalation triggered when a constraint asset is down more than five minutes changes behavior in a way a dashboard never will. Critically, reconcile the real-time store against the batch warehouse daily. Streaming systems drop and duplicate events, and if the live board and the month-end report disagree, users will stop trusting both. State clearly which store is authoritative for which decision, and keep the batch warehouse as the system of record for anything financial.

How Netray Delivers Real-Time Manufacturing Analytics

Netray starts by mapping decisions to required latency, which usually shrinks the streaming scope by 80 percent and the cost with it. We implement log-based CDC from SyteLine or LN so the ERP is never polled, normalize machine tags against an asset registry, and compute state and OEE in the stream. Our AI agents generate the transformation logic, the tag mapping, and the daily reconciliation tests between the real-time store and the warehouse. A typical first deployment covers one plant's constraint cells with sub-minute machine state and 5-minute order status in ten to fourteen weeks, running on-prem where controlled data requires it.

Frequently Asked Questions

How do I get real-time data out of my ERP without slowing it down?

Use log-based change data capture rather than polling. SQL Server CDC or a Debezium connector reads the transaction log and publishes inserts, updates, and deletes with minimal load on the production database, unlike repeated queries against high-churn tables such as job transactions. Enable CDC selectively on the twenty or thirty tables that drive your real-time use cases, and monitor log growth and retention after enabling it.

Is real-time analytics worth the cost for a single plant?

It depends on downtime cost and constraint structure. If an hour of unplanned downtime on your bottleneck costs several thousand dollars, sub-minute machine state and alerting typically pay back within months. If your operation is low-volume, high-mix with long cycle times and no hard constraint, hourly or shift-level refresh usually captures most of the value at a fraction of the engineering and operating cost.

Why do real-time dashboards disagree with month-end reports?

Streaming pipelines can drop, duplicate, or reorder events, and late-arriving corrections such as reversed transactions often are not replayed into the live store. The batch warehouse applies full reprocessing and business rules that the stream skips for speed. The fix is a daily automated reconciliation between the two, with variance thresholds and alerts, plus clear labeling so users know which source is authoritative for which decision.

Key Takeaways

1Decide Which Decisions Justify Real Time: Latency requirements should be derived from decision cadence, not ambition. Machine downtime response and andon escalation need 10 to 60 seconds, because a stopped constraint costs money by the minute.
2Getting ERP Changes Out in Near Real Time: Polling the ERP every 30 seconds is the usual first attempt and the usual first failure, because repeated scans of jobtran or transaction tables load the production database exactly when it is busiest. Log-based change data capture is the correct mechanism: SQL Server CDC or a Debezium connector reads the transaction log and emits inserts, updates, and deletes with negligible impact on the OLTP workload.
3Streaming Machine Data From the Floor: Machine data typically arrives over OPC UA from PLCs or through an MQTT broker using Sparkplug B, which provides birth and death certificates so you can distinguish a stopped machine from a lost connection. Normalize tags against an asset registry immediately, because raw tag names are plant-specific and unmaintainable in analytics.

Put this into numbers

Free interactive tools for exactly this problem. No signup to use them.

Free Tool

Manufacturing AI Readiness Assessment

Score your manufacturing operation's readiness for AI across data, systems, people, and governance, and get a prioritized roadmap for closing the gaps.

Free Tool