Updated July 2026 · 8 min read · Adente Vision Engineering Team
Why is Modbus TCP the neutral integration path for AI inspection?
Modbus TCP is the lowest-friction way to carry an inspection result because it is the one protocol nearly every controller, gateway and SCADA package can read, regardless of brand. On a line that mixes a Siemens PLC, a Rockwell controller and an older machine with its own gateway, there is often no single native fieldbus they all share, but there is almost always Modbus TCP. It is an open, published protocol with no licensing gate, which is exactly why it survives as the common denominator on brownfield and mixed cells.
That neutrality is the whole argument. Instead of standardising the cell on one vendor's real-time bus to add inspection, you expose the result over Modbus TCP and let each controller read the registers it needs. Adente Vision is an edge-AI visual inspection unit built by ADENTE Advanced Engineering Technologies, part of the Aden Group, sold through automation system integrators, and Modbus TCP is one of five industrial protocols it exposes, alongside PROFINET, EtherNet/IP, EtherCAT and OPC UA, so the neutral path is available without giving up the option to move to a native bus later.
How do you carry an inspection result over a Modbus register map?
The result of an inspection is small data, so the register map that carries it is small too. A pass or fail decision is a single bit; a confidence or anomaly score is one value; a running part-count and a status word cover the rest of what a controller needs to act and to log. That is the entire payload for most cells, which is why Modbus, a register-and-coil protocol with no rich type system, is enough.
A representative map keeps each piece in a predictable place so the PLC program can read it every cycle. The exact addresses are set at commissioning to fit the controller's existing memory layout, which is a mapping choice rather than a fixed part of the protocol. What matters is that the controller polls the result register to drive its reject, reads the confidence register if it wants to apply its own threshold, and watches the status word for the handshake. Because the unit decides on the edge in the same enclosure, the register values are the output of a completed on-device inspection, not a request the controller has to wait on a network round-trip for.
Two Modbus details are worth agreeing before wiring. A single pass or fail bit fits a coil, while a value such as confidence or a count fits a holding register, so the map usually mixes both areas rather than forcing everything into one. And a value that spans two registers, a 32-bit count for example, has a word order the two ends must agree on, since Modbus does not fix it; a mismatch there reads as a wildly wrong number rather than a subtle one, which at least makes it easy to catch during commissioning. Neither point is unique to inspection, they are how any Modbus device is integrated, but settling them up front keeps the result clean.
A light handshake over the status word stops the controller from reading a stale result. The unit sets a busy or new-result flag as it publishes each part, and the controller reads the result only when that flag says the data is for the current part. Without it, a controller polling faster than the part rate can read the same result twice, or read across the moment the values change. This is the same ready-and-acknowledge discipline used on discrete I/O, carried into the register map so the neutral path is as reliable per part as a wired one.
A representative Modbus register map for an inspection result
| Register or area | Carries | Example use |
|---|---|---|
| Result coil or word | Pass or fail bit for the current part | PLC reads it and drives the reject |
| Confidence register | Per-part confidence or anomaly score | PLC applies its own accept threshold |
| Part-count register | Running inspected and reject counts | Line totals and OEE feed |
| Status word | Ready, busy, fault and active mode | Handshake and interlock |
| Mode or recipe register | Selected inspection mode or recipe | Changeover and variant selection |
The map above is a pattern, not a fixed specification. The point is how little it takes: five registers cover a full inspection handshake, which is why Modbus TCP is enough for a large share of cells that never need a heavier protocol.
How does Modbus TCP compare to PROFINET or EtherNet/IP here?
Modbus TCP trades determinism for simplicity, and that trade is the deciding factor. It runs over standard TCP/IP with no guaranteed delivery window, so it is a polled, best-effort exchange rather than a cyclic, time-bounded one. For carrying a pass/fail and a count to a controller that polls every few milliseconds, that is completely adequate and far easier to commission. PROFINET and EtherNet/IP add cyclic, deterministic timing and native integration with Siemens and Rockwell environments, which matters when the result has to be aligned tightly to motion or when the plant standard demands it.
So the choice is not which protocol is better in the abstract, it is what the result has to do. If the inspection result drives a reject a comfortable distance downstream and feeds counts to a dashboard, Modbus TCP is the neutral, low-effort path. If the cell is fast and tightly synchronised, or the site mandates a specific bus, move up to a real-time fieldbus. The unit exposes both, so this is a per-cell decision made at commissioning, not a purchase-time lock-in. For the related decision of when a rule-based check versus an AI model produces the pass/fail bit that rides these registers, see the rule-based versus AI machine vision guide.
What do the four inspection modes put on the wire?
Each of the four inspection modes, Anomaly, Defect, Counting and Quality, produces a result that fits the same small register set. Anomaly and Defect emit a pass or fail decision plus a confidence or anomaly score, which map to the result and confidence registers. Counting increments the part-count register directly, so a running total and a reject total are readable at any moment. Quality mode, which combines classical computer vision measurement with AI judgement, can publish a graded result rather than a bare pass or fail, still inside a register the controller reads on its normal scan.
Throughput sets whether polling keeps up. The unit is built to decide at 100+ parts per minute, so the controller has to poll the result register fast enough not to miss a part between decisions. On most lines a Modbus poll every few milliseconds clears that easily, but on the fastest cells it is worth checking the poll rate against the part rate, and moving to a cyclic bus if the margin is thin. As with any per-cell timing, confirm the poll-to-part margin on your own line rather than assuming it.
This post is a spoke of the pillar guide on AI visual inspection; to see the full protocol and I/O list the unit exposes, take the system overview.