An inspection system earns its place through the handshake, not the model. A trigger arrives, a verdict must return within a fixed window, the verdict must reach the correct physical part at a reject station downstream, and the result must attach to a batch record that survives an audit. Each of those is a specific engineering decision.
The PLC detects presence, or an encoder count indicates the part has reached the inspection position. This is the trigger source; a free-running timer is not acceptable on a line that stops.
The camera is triggered and the strobe fires. Exposure is short enough that motion blur does not consume the resolution the optics provide.
Capture, transfer and inference complete within the allocated window. Worst-case latency is what matters, because the tail is what misses the next trigger.
Pass, fail or fault is returned over digital I/O or fieldbus, in a form the existing controller already understands.
The verdict travels with the part through a shift register indexed by encoder count, so it arrives at the reject station attached to the right unit.
Air blast, diverter, gate or robot pick fires at the reject station against the tracked part.
Verdict, image reference and model version are written to MES or a historian against the batch or serial record.
| Interface | Typical latency | When it is used |
|---|---|---|
| Digital I/O | Sub-millisecond | Simple pass and fail with a fault line; the most robust option |
| Fieldbus (PROFINET, EtherNet/IP, EtherCAT) | 1–10 ms | Where defect class or measurement values are needed, not just a verdict |
| OPC UA | 10–100 ms | Supervisory and MES level rather than in the control loop |
| MQTT to historian | Not in the control path | Traceability, analytics and long-term retention |
A verdict in the control loop belongs on digital I/O or fieldbus. OPC UA and MQTT carry the record, not the decision.
This is agreed with the controls engineer at design time, and it is the question that separates a commissioned system from a pilot.
Options are: reject the part by default, pass the part by default, stop the line, or fall through to manual inspection. Which is correct depends entirely on the consequence of an escaped defect against the cost of a stoppage, and it is a business decision informed by engineering rather than the reverse.
The system also has to distinguish between fault types. A camera that has failed, illumination that has degraded past a threshold, and inference that has exceeded its time budget are three different conditions that may warrant three different responses.
Degradation is the harder case. A hard failure is obvious. Illumination that has dimmed by fifteen per cent over a year is not, which is why a reference target imaged on a schedule belongs in the design rather than in a maintenance wish list.
Over digital I/O for a simple pass or fail with a fault line, or over a fieldbus such as PROFINET, EtherNet/IP or EtherCAT where defect class or measurement values are needed. OPC UA and MQTT carry the traceability record rather than the in-loop decision.
Through part tracking — typically a shift register indexed by encoder count — so the verdict travels with the part to a reject station downstream and survives line stops and restarts. Without it, a correct verdict can be applied to the wrong unit.
That is defined by an agreed fault protocol: reject by default, pass by default, stop the line, or fall through to manual inspection. The choice depends on the consequence of an escaped defect against the cost of a stoppage, and it is agreed at design time.
Written for engineers. Share it with one.
Tell us the specification, the constraint and the deadline. Programmes that cross silicon, radio, embedded and AI are where Faststream is strongest.