Adding AI vision to an existing line works best as a staged project: audit the line, choose the inspection point, plan the imaging and the PLC connection, train on real images, validate in parallel with the current process, and then hand the system over to the people who will run it. Skipping a stage is the most common reason a retrofit stalls after installation.
This guide covers each stage in plain terms for manufacturing and quality teams adding inspection to equipment that is already running, where space, wiring, and downtime are limited.
Key Takeaways
- A retrofit succeeds or fails on the physical details: available space, mounting, lighting, power, and the signal that triggers the camera. These should be reviewed before a system is chosen.
- The best inspection point is as early as a defect can reliably be seen, and before more value is added to the part. It also needs a place to act on the result, such as a reject gate or a stop signal.
- Integration is usually the largest source of delay. The trigger, the pass/fail output, and the data handshake with the PLC should be defined on paper before installation day.
- Training images should come from real production across shifts, materials, and lots, and should include borderline parts. A small, varied set usually works better than a large, uniform one.
- Running the new system in parallel with the current process, before it controls anything, builds a validation record and reduces risk. A VS Series smart camera suits simple retrofits, and the VS-G Series suits lines that need several cameras or stored images.
What Does Adding AI Vision to an Existing Line Involve?
Adding AI vision means installing cameras, lighting, and a processing unit on a line that was not designed for them, then connecting the result to the equipment that moves, sorts, or stops parts. The vision software is only one piece. The rest is mechanical mounting, electrical wiring, communication with the PLC, and a plan for what happens when a part fails.
Retrofits differ from new-line builds in one important way. There is no room to redesign the layout, so the system must fit around the conveyor, guards, and fixtures that already exist. That makes early site review more valuable than any single product feature.
Step 1: Audit the Line Before Choosing a System
A short audit prevents most surprises. It should cover the items below, each recorded in writing.
- Line speed and part spacing: the maximum parts per minute and the gap between parts, which sets the time available for image capture and processing.
- Part presentation: whether parts arrive in the same position and orientation, and whether they stop, slow, or move continuously at the inspection point.
- Physical space: room for the camera, lens working distance, lighting, and any shroud, plus access for cleaning and adjustment.
- Environment: vibration, dust, washdown, temperature, and ambient light from windows, dock doors, or nearby fixtures.
- Utilities and controls: available power, network connections, free PLC inputs and outputs, and the type of PLC in use.
The audit also confirms whether a smart camera will do the job or whether the line needs a controller-based system. Compact, all-in-one cameras fit tight spaces and stand-alone stations. Multi-camera lines and high-speed stations may need a controller that runs several cameras together.
Step 2: Choose the Inspection Point
Where the camera sits matters as much as what it sees. Three questions help place it.
The first is how early the defect can be seen. Catching a defect before further assembly or processing stops value from being added to a part that will fail. The second is how consistent the presentation is at that point. A position where parts are steady and well located gives more stable results than one where they are still settling. The third is whether the line can act on the result there. A pass/fail decision is only useful if a reject gate, diverter, or stop signal sits close enough downstream to remove the part.
If one camera cannot see everything that matters, a second inspection point or a second camera angle may be more reliable than forcing one view to do too much.
Step 3: Plan the Imaging
Most retrofit problems that appear later were decided at this stage. The image has to show the defect clearly and consistently, shift after shift.
Lighting and ambient light
Lighting controls contrast, which decides how reliably the system can see a defect. The method should suit the surface: backlighting for silhouettes, low-angle light for scratches and dents, diffuse light for shiny flat parts. Existing lines often have uncontrolled ambient light, so a simple shroud or enclosure around the station is common practice and prevents changes in sunlight or overhead lights from altering the image.
Field of view and mounting
The field of view should cover the full range of part positions with some margin, while keeping the smallest defect large enough in pixels to detect. Mounting should be rigid, because vibration and bumped brackets are frequent causes of drift. Cameras with software-controlled optical zoom, such as the ZoomTrax function on the VS Series, allow the field of view to be adjusted after installation without swapping lenses, which helps when the real line differs from the drawings.
Step 4: Plan the Integration
Integration links the vision system to the rest of the line, and it is the stage most likely to run over schedule. The following points should be agreed before installation.
- Trigger: what tells the camera to capture an image, such as a sensor signal or a PLC output, and how long after that signal the part is in position.
- Result output: how pass, fail, and any error state are sent back, through discrete I/O or an industrial network protocol supported by both the camera and the PLC.
- Reject action: which device removes a failed part, how far downstream it sits, and how the PLC tracks the failed part to that point.
- Fault handling: what the line does if the camera is not ready, loses its trigger, or returns an error, such as stopping, alerting, or passing parts to manual inspection.
- Data: what is logged, where it is stored, and who reviews it.
Interface support differs by model, so each point should be checked against the specifications of the system being considered. Writing the signal list and timing diagram first lets mechanical and controls teams work in parallel instead of waiting for each other.
Step 5: Collect Images and Train the AI
AI inspection learns from examples, so the quality of the examples decides the quality of the result. Images should come from the real line, not from a clean bench setup.
A good training set covers the normal range of the process: different shifts, material lots, operators, and lighting conditions. It includes good parts, clear defects, and borderline parts, since borderline cases are where decisions are hardest. Rare defects can be hard to collect, so teams often start with the defects that occur most often and add others as they appear.
A varied set of a few dozen images can be enough to begin with some AI tools, and the set grows as the system sees more of the process. It helps to store images from the first day, including good parts, because that archive becomes the training material for later improvements. The VS-G Series has built-in image storage in the controller for this purpose.
Step 6: Validate Before Going Live
Validation proves the system works on the actual line before it controls anything. A staged approach lowers risk.
Run in parallel
In shadow mode, the new system inspects parts and records results while the existing process stays in control. Its results are then compared with the current inspection and with known good and bad samples. This shows where the system agrees, where it differs, and why.
Measure what matters
The key results are the false reject rate, where good parts fail, the escape rate, where bad parts pass, and repeatability across shifts. Seeded tests, which run known defective parts through the line, check that real defects are caught. A system that rejects too many good parts will lose operator trust, so false rejects deserve the same attention as escapes.
Tune and re-test
When settings change, they should be re-tested against the stored images to confirm that earlier results did not shift. The VS-G Series Settings Verification function does this by replaying stored images against a proposed change before it goes live.
Confirm acceptance criteria
Pass/fail targets, test conditions, and sign-off roles should be written down before validation starts, so the decision to go live rests on agreed evidence.
Step 7: Go Live and Hand Off
Go-live is a transition, not a finish line. The system should move from monitoring to control in steps, for example by first enabling the reject action on a single shift, then on all shifts once results hold steady.
Handoff matters just as much. Operators need to know what the indicators mean and what to do when a part fails. Maintenance needs a routine for cleaning lenses, checking mounts, and confirming lighting output. Engineering needs controlled access to settings, with a record of what was changed, when, and why. Systems with stored images and verification tools make that record easier to keep, and the machine vision troubleshooting guide covers how to diagnose problems that appear after launch.
Teams that plan to add more stations should treat the first line as a template. Reusing the same platform, settings structure, and training process makes each later deployment faster. The buyer's guide to choosing an AI inspection system covers how to compare platforms, and the guide to AI inspection ROI and payback period covers how to measure results.
Common Pitfalls When Retrofitting AI Vision
Most retrofit problems repeat across projects, and most are avoidable.
- Choosing the system before the audit: a product picked first may not fit the space, speed, or controls.
- Uncontrolled ambient light: an inspection that works at night and fails at noon is usually a lighting problem.
- Training on clean sample parts: models trained on bench parts often struggle with real variation on the line.
- Late integration planning: unclear trigger timing or reject tracking is a common source of delays.
- No fault plan: the line needs a defined behavior for when the vision system is unavailable.
- No owner after launch: settings drift when nobody is responsible for them.
KEYENCE supports customers from selection through line operation, including on-site operating instructions and after-sales support. A demo or test on real parts, which can be requested through the demo and test page, is a practical first step. More machine vision systems and industry applications are available for teams planning a retrofit.
Frequently Asked Questions
Yes. Compact smart cameras and controller-based systems are both used on existing lines. The main requirements are room to mount the camera and lighting, a reliable trigger, and a way to act on the result, such as a reject gate or a stop signal to the PLC.
It depends on the application, the integration work, and how much validation is required. Mechanical mounting and image setup are often the quicker parts, while PLC integration and validation take more care. A written plan and signal list at the start shortens the schedule.
Usually only briefly. Much of the work, such as the audit, signal planning, and image collection, can be done while the line runs. Mounting and wiring typically need a short planned stop, and the system can run in monitor-only mode afterward without affecting production.
Through discrete inputs and outputs, an industrial network protocol, or both. A trigger signal starts the image capture, and the pass/fail result goes back to the PLC, which tracks the part to the reject device. Supported interfaces vary by model, so they should be confirmed in the specifications.
It varies by tool and application. Some AI tools can begin with a few dozen images, but a varied set from real production, including borderline parts, matters more than a large one.
Shadow mode runs the new inspection alongside the existing process without letting it control the line. It builds a record of how the system performs on real production before it makes any decisions.
- Vision System with Built-in AI VS series
Vision System with Built-in AI
VS series
- AI-Powered High-Performance Vision System VS-G series
AI-Powered High-Performance Vision System
VS-G series
