The 400-Millisecond Gap on the Molding Line
Northfield Plastics runs 22 injection-molding machines on one shop-floor VLAN. Nine older presses expose only Modbus RTU over a daisy-chained RS-485 bus; thirteen newer presses offer OPC UA. A single edge gateway polls all machines every 500 ms, converts values to MQTT, and stamps each sample with its own arrival time. The historian powers an OEE dashboard that now reports 6.2% unexplained downtime, which shift supervisors dispute. At shift change, the 100 Mbps copper uplink reaches 68% utilization and gateway queue depth spikes. Engineering also wants to add four 4K inspection cameras to the same switch. The plant cannot stop production for a network redesign and has a 10-week window before the next audit. The controls engineer proposes simply polling every 100 ms to 'get better data'.
1. The controls engineer wants to cut the polling interval from 500 ms to 100 ms. Evaluate that proposal and list the evidence you would collect before changing the architecture.
Read the worked answer
Faster polling does not fix timestamp provenance and can make the data worse. Nine presses share one RS-485 bus; at low baud rates a multi-register read can take well over 100 ms per device, so a 100 ms poll for all nine is physically impossible. The gateway will either queue requests, overwrite samples, or report stale values as fresh. The 6.2% discrepancy is more likely caused by gateway arrival timestamps rather than machine event time, plus OEE state mapping that does not match PLC fault codes.
Evidence to collect:
- Compare gateway clock to each PLC clock and measure offset and drift over 24 hours.
- Log request-to-response latency and queue depth per device, not just aggregate uplink utilization.
- Capture change-of-state events and sequence counters at the controller and at the gateway.
- Reconcile a sample of disputed downtime events against PLC fault words and operator reason codes.
- Verify shift calendars, run-rate denominators, and whether minor stops are being rounded away.
Only after proving the loss is latency or timestamp-related should polling change, and then per device based on process dynamics, not globally.
2. Recommend a target connectivity design for the 22 presses and the four 4K cameras. Address segmentation, time synchronization, and data buffering.
Read the worked answer
Keep control traffic deterministic and separate from video. Place presses and edge gateways in an OT cell or zone, and connect to the historian through an industrial firewall or DMZ conduit. Never bridge the cameras into the control VLAN.
For the nine Modbus presses, use protocol gateways in small groups with local store-and-forward. Poll each device at a rate justified by its dynamics, and expose data northbound with both source and gateway timestamps. For the thirteen OPC UA presses, use subscriptions with publishing intervals matched to the process, not a single global poll.
Deploy PTP, or at minimum an OT-served NTP source, and record source and gateway timestamps. For cameras, create a separate VLAN with a 1 Gbps uplink and QoS or rate limits. Four 4K streams at 15 fps can exceed the existing 100 Mbps copper uplink, so the shared switch is a capacity and availability risk. Buffer at the edge for 24 to 72 hours so production data survives network outages.