Connect computer vision to your manufacturing stack with Roboflow: edge inference plus PLC, SCADA, MES, and ERP integration over EtherNet/IP, MQTT, and OPC UA.
A computer vision model that detects defects but talks to nothing is a demo, not a production system. On a real factory floor, the value of vision AI depends on what happens after the detection. For example, does the reject gate fire? Does the operator see the alert? Does the MES hold the lot? Does the ERP finally learn the true cost of scrap?
Answering those questions means understanding where vision fits in the industrial software stack. Manufacturers already run a layered architecture, often described by the ISA-95 model, and vision AI must plug into it rather than sit beside it. In this post, we walk through each layer of that stack, show what vision contributes at every level, and explain how Roboflow Workflows connect to each one using the protocols your plant already speaks.
The CV Integration Stack: Automation Pyramid
The ISA-95 standard organizes plant systems into four levels:
- Levels 0 to 1: sensors and control. Physical processes, sensors, actuators, and the PLCs (Programmable Logic Controllers) that run the machinery in real time.
- Level 2: supervision. SCADA (Supervisory Control and Data Acquisition) systems and HMIs (Human-Machine Interfaces) that let operators monitor and control equipment across a line or plant.
- Level 3: execution. The MES (Manufacturing Execution System) that dispatches work orders, tracks work-in-process, enforces quality rules, and records what actually happened.
- Level 4: business planning. The ERP (Enterprise Resource Planning) system that plans what to build, manages inventory and costs, and answers to the balance sheet.
Information flows down as plans and up as results. Vision AI enters this picture as a new kind of sensor, one that perceives the physical product itself, not just temperatures, pressures, and part counts. This perception is useful at every level, but each level needs it delivered differently.
Vision and the PLC: Closing the Loop in Milliseconds
At the control level, vision earns its keep by acting, not reporting. A camera watches parts move down the line, a model classifies each one, and the result reaches the PLC fast enough to divert a bad part before it passes the reject gate. The PLC does not care about confidence scores or bounding boxes. It needs a clean digital signal: pass or fail, on this part, right now.
That requirement shapes the architecture. Inference must run at the edge, on hardware near the camera, because a round trip to the cloud introduces latency a moving conveyor will not forgive. Roboflow Inference runs custom-trained models on edge devices ranging from NVIDIA Jetsons to industrial PCs, and Workflows adds the logic layer. If the model detects a defect above a confidence threshold, that triggers the PLC. Otherwise, the model lets the part pass.
The connection itself uses the protocols automation engineers already trust. Workflows includes enterprise integration blocks for writing directly to PLCs, plus PLC Relay, an edge service that reads and writes PLC tags over Allen-Bradley EtherNet/IP, Modbus TCP, and Siemens S7. Your controls engineer maps a tag, your vision workflow writes to it, and the ladder logic takes it from there. No custom middleware required.
Vision and SCADA/HMI: Giving Operators Eyes on Quality
One level up, the audience changes from machines to people. SCADA systems and HMIs exist so operators see what the line is doing and intervene when something drifts. Vision results belong on those screens in the form of live defect counts, annotated images of the last failure, alarm states when reject rates climb.
MQTT is the workhorse protocol here. It is lightweight, publish-subscribe by design, and supported by virtually every modern industrial platform. A vision workflow publishes each inspection result to a topic, and any subscribed system picks it up. Our guide to broadcasting computer vision predictions over MQTT walks through the pattern, and our tutorial on connecting Roboflow to Ignition using MQTT shows a complete build: a custom-trained model tracking products on a line, with counts and annotations displayed live in an Ignition Perspective dashboard.
The publish-subscribe model carries a quiet architectural benefit. The vision system publishes once, and the HMI, the historian, and the MES can all consume the same stream. You never rebuild the integration when a new consumer shows up.
Vision and MES: Turning Detections into Production Records
The MES runs the factory floor, but it cannot see the product. It knows a work order is running on Line 3, for example, but it doesn’t know the parts coming off Line 3 have a scratch pattern that started 20 minutes ago. Vision fills this blind spot, and the MES gives vision results somewhere permanent to live.
The integration points are familiar: REST APIs, MQTT, OPC UA, or direct database writes, depending on the MES. Each inspection becomes a structured record: part serial or lot number, pass/fail verdict, defect classification, timestamp, and the image itself attached as objective evidence. From there, the MES does what it always does. It updates the unit's genealogy, places a quality hold on a suspect lot, dispatches rework, and feeds nonconformance and CAPA workflows in the quality system with data far richer than a hand-written inspection log.
We cover this relationship in depth in our post on Manufacturing Execution Systems and where vision AI picks up, including real-world patterns like defect detection results writing back to MES with images attached for traceability.
Vision and ERP: Quality Data the Business Can Price
The ERP sits at the top of the stack and thinks in dollars, dates, and quantities. It rarely wants individual inspection events. It wants the aggregates those events roll up into, such as first-pass yield by line, scrap cost by product, defect rates by supplier lot.
In most architectures, vision data reaches the ERP indirectly, flowing through the MES or a data platform that summarizes it. But the business impact is direct. When every unit is inspected and every defect is classified, the ERP's scrap and rework numbers stop being estimates. Purchasing confronts a supplier with defect rates tied to specific incoming lots. Finance sees the real cost of quality by product family. Planning adjusts yield assumptions with data instead of folklore. Our post on reducing scrap with computer vision explores how that visibility changes decisions well above the plant floor.
Design Principles for the Vision Layer
A few patterns show up in nearly every successful integration:
- Run inference at the edge, integrate everywhere. Keep the model close to the camera for speed and resilience, then distribute results upward through standard protocols.
- Match the payload to the layer. Send the PLC a boolean. Send SCADA a status message and an annotated image. Send the MES a full structured record. Send the ERP a rollup. One detection, four representations.
- Publish once, subscribe many. Use MQTT or OPC UA topics as the distribution backbone so new consumers never require new plumbing.
- Keep humans in the loop where it counts. Route borderline detections to an operator dashboard for review, and feed confirmed hard examples back into training so the model improves over time.
Building the Stack with Roboflow
Roboflow was built for exactly this kind of integration work. You train a model on images of your own parts and defects, then assemble the application in Workflows, a visual canvas with native blocks for MQTT, OPC UA, and PLC triggers alongside logic, branching, and image transformations. You deploy with Inference on the hardware of your choice, from air-gapped edge servers to cloud endpoints, and manage fleets of devices through the Deployment Manager. Manufacturers already use this stack for industrial inspection and machine vision applications that talk to PLCs and HMIs from day one.
Does vision AI replace my existing inspection sensors?
No. Photo eyes, proximity sensors, and gauges still handle presence and dimension checks well. Vision adds a layer of perception those sensors cannot provide: surface defects, assembly verification, label and print inspection, and anything that requires judging how a product looks.
Which protocol should I start with?
If the goal is real-time control, work with your controls engineer to write directly to PLC tags over EtherNet/IP, Modbus TCP, or S7. If the goal is visibility and data distribution, start with MQTT. Most plants end up using both.
Do I need an MES before I add vision?
No. Many manufacturers start with a standalone vision workflow that triggers a reject gate and displays results on an HMI, then integrate with MES and ERP as the system proves itself.
Start Where the Pain Is
The takeaway is simple. Vision AI does not replace your PLC, MES, SCADA, or ERP. It gives all four of them something they have never had: the ability to see the product. Start at the layer where the pain is sharpest, connect through the protocols you already run, and let the results flow up the stack.
Ready to build the vision layer for your plant? Get started with Roboflow for free, or explore our manufacturing solutions.
Cite this Post
Use the following entry to cite this post in your research:
Erik Kokalj. (Jun 3, 2026). The Computer Vision Manufacturing Integration Stack. Roboflow Blog: https://blog.roboflow.com/the-cv-integration-stack/