Retrofit or replace? Adding AI vision to a 20-year-old line

The question on a 20-year-old line is almost never whether AI can see the defect. If the image shows it, a modern model can usually be trained to find it. The real question is whether the line around the camera can support a reliable inspection: a camera you can get raw images from, a mount that doesn’t shake, a way to track each part to the reject gate, a PLC with room for new logic, guarding that can be changed safely, a validation path you can afford, and a downtime window long enough to do the work.

Score those seven things honestly before you buy anything. When most of them check out, a retrofit is faster and cheaper than a new station. When two or three fail badly, replacing the inspection station, or that section of the line, is usually the better use of the money.

Three options, not two

“Retrofit or replace” hides a middle option, and it is often the right one:

• Software retrofit. Keep the existing cameras, lights and brackets; route images to edge compute running a new model; keep or rework the reject logic. Least disruption, but you inherit every limitation of the existing optics.
• Station retrofit. Keep the line and its controls; replace cameras, lighting and mounts at the inspection point. Many retrofits land here once someone looks closely at the existing images.
• Replacement. A new inspection station or machine section, designed, installed and validated as a new system.

The seven questions

1. Camera interface and age

Start by pulling raw images from the existing camera, not screenshots from its HMI. Plenty of older smart cameras run closed firmware that exposes only a pass/fail bit and a compressed preview. If you can’t get full-resolution frames out, the camera is a sensor for its own software and nothing else.

Then look at the interface. GigE Vision and USB3 Vision cameras are usually easy to bring into a new pipeline. Analog and FireWire cameras need capture hardware that is increasingly hard to source, and standard-definition interlaced video limits the smallest defect you can find. Check rolling versus global shutter too, because a rolling-shutter sensor skews moving parts. Finally, check spares: a camera the manufacturer no longer supports is a single point of failure on your line.

2. Mounting rigidity

A model can’t separate vibration from a defect if the part moves several pixels between frames. A worked example with hypothetical numbers: at 0.1 mm per pixel, a bracket that vibrates 0.3 mm at line speed shifts the image by three pixels, enough to smear a fine crack or move an edge across a tolerance band. Old lines often have cameras clamped to guard rails, shimmed after a jam, and nudged by maintenance more than once.

The test is simple. Mount a static target where the part would be, run the line at full speed, and measure how far the target moves in the image. If it moves, budget for a rigid stand bolted to the machine frame or the floor, with dowel-pinned mounts so a camera removed for cleaning goes back in exactly the same place.

3. Encoder availability

To reject the right part, the system has to know where each part is between the camera and the reject gate. On a line with variable speed, time-based tracking (“fire the air blast 400 ms after the trigger”) misfires every time the line slows. An encoder on the conveyor turns that into distance-based tracking, which survives speed changes.

Check whether an encoder already exists and whether its signal can be shared: a splitter for differential signals, matching signal levels (5 V TTL or RS-422 versus 24 V HTL), and a spare counter input. If there is no encoder, check that a shaft or idler roller is accessible for mounting one. Our article on encoders, handshakes and reject gates covers the details.

4. PLC capacity and spare I/O

The PLC on a 20-year-old line may belong to a discontinued family, with little free memory, a scan time already near its limit, and a program whose comments left with its author. Before designing any integration, answer five questions:

• Is the source program available, and does it match what is running?
• Is there spare digital I/O near the reject gate?
• Is there room in memory and scan time for tracking logic?
• Which protocols does it speak: EtherNet/IP, PROFINET, Modbus TCP, or only a serial or legacy fieldbus link?
• Who is allowed to change it, and how?

If the PLC can’t take new logic, the vision system can drive the reject device directly through isolated digital outputs, with a hardwired interlock to the PLC for line stops and alarms. That is a legitimate design, but it has to appear in the machine’s electrical drawings and in the safety review.

5. Guarding and safety

New cameras and lights often have to sit inside guarded areas, and a reject mechanism (an air blast, a pusher or a diverter) is a new moving hazard. Any change to guarding, interlocks or moving parts should go through your site’s machine risk assessment process; ISO 12100 sets out the general method. Plan for lens cleaning without bypassing guards, and for strobes that won’t dazzle operators. These items rarely kill a retrofit, but discovered late they routinely add weeks.

6. Validation burden

On regulated lines, the paperwork can outweigh the hardware. In pharmaceutical and medical-device manufacturing, adding or changing an inspection on a validated line triggers change control and requalification, and the software around the model needs validating too. Automotive suppliers may owe their customer a formal change notification and, depending on the customer’s requirements, a new PPAP submission. Food plants may need to update their food safety plan if the inspection is a control point.

Estimate this effort early and compare honestly. Modifying a validated system can take as much documentation as qualifying a new one.

7. Downtime windows for installation

A well-planned retrofit needs very little line time. Cameras and lights go in during a short planned stop. The model then runs in shadow mode for a few weeks, logging decisions without rejecting anything. One more window wires the reject output and runs the site acceptance test. A replacement station usually needs a longer, single block of downtime for mechanical installation, commissioning and acceptance. If your next long shutdown is eight months away, that alone may decide the question.

Scoring the line

Factor	Retrofit-friendly	Points toward replacement
Camera interface	GigE Vision or USB3 Vision; raw frames available	Closed smart camera, analog or obsolete interface, no spares
Image content	Defect clearly visible in existing images	Wrong angle, lighting or resolution for the defect
Mounting	Rigid frame; repeatable mounts possible	Vibration you can’t fix without rebuilding the station
Part tracking	Encoder present or easy to add	No encoder access, erratic spacing, too little distance to the reject point
PLC	Spare I/O, source available, scan-time headroom	Discontinued, full, undocumented, or due for migration
Guarding	Space inside the existing guards	Guarding redesign needed anyway
Validation	Change control is routine	Modifying costs as much as qualifying new
Downtime	Short stops available; shadow mode possible	Only one long shutdown a year

Before the site survey: a checklist

When we scope a retrofit, we ask for these before anyone travels:

• Photos of the inspection point, camera, lights, mounts and reject device
• Camera make, model, interface and firmware, plus 50–100 raw full-resolution images that include known defects
• Line speed range, part pitch, and how parts are presented (fixtured, random orientation, touching)
• The distance from the camera to the reject point, and the type of reject mechanism
• Whether an encoder exists, and its type and signal level
• PLC model, firmware, program access and spare I/O count
• Electrical and guarding drawings for the station
• The validation status of the line and the change-control process that applies
• The planned shutdown calendar for the next 12 months
• Who signs off: quality, production, maintenance, controls and safety

When replacement wins

Replacement is the better choice when:

• the existing camera can’t deliver raw images, or its interface is obsolete with no spares;
• the defect can’t be seen from the existing camera position, and the station has no room for the lighting geometry it needs;
• the mount can’t be made rigid without rebuilding the station anyway;
• parts can’t be tracked reliably to a reject point;
• the PLC is already due for migration, so integrating with it now is throwaway work;
• revalidating a modified station costs about as much as validating a new one.

A worked example

Consider a hypothetical cap-and-label line installed two decades ago. A monochrome GigE camera running rule-based software checks for cap presence. Quality now wants cocked caps and torn labels caught as well. The camera exports raw frames, and the images show the cap clearly, but glare washes out the label. There is an encoder on the main conveyor and two spare outputs on the PLC near the reject gate.

That line scores well on most of the seven factors. The likely plan: keep the camera, add a polarized light for the label, move everything onto a rigid stand, run the new model in shadow mode for a few weeks, then wire the reject through the spare outputs. The same reasoning sits behind our weld-nut retrofit example engagement, where the plant’s existing cameras stayed in place.

Score first, then decide

Retrofits fail when the AI gets all the attention and the bracket, the encoder and the PLC get none. Score the seven factors, look at real raw images, and decide with both columns of the table in front of you. Our vision retrofit work starts with exactly this survey. When the answer is replacement, a new station falls under AI visual inspection systems, and the survey work carries straight over.