Half a part in frame: trigger and encoder faults that look like the model
In short
Put acquisition ahead of inference in the fault tree. Count triggers, images and parts over 100 pieces: more triggers than parts is a debounce fault, fewer is a sensor that cannot see your parts, equal counts with bad pictures means exposure, strobe or encoder. Then do the blur arithmetic — belt speed times exposure, over millimetres per pixel — because that number ends most retraining arguments.
Key takeaways
- Blur in pixels equals belt speed times exposure, divided by millimetres per pixel. Compute it before arguing.
- Cutting exposure to freeze motion costs light in exact proportion, which is why a strobe exists.
- Trigger latency times belt speed is a fixed offset; jitter in that latency is what actually clips parts.
- An encoder fault has a signature no other fault has: distortion along travel only, never across it.
- Acquisition parameters are per part number. Left out of the recipe record, every changeover re-rolls the dice.
If the pictures are wrong, nothing downstream can be right, and a surprising number of inspection projects spend weeks on a training set to fix a photoelectric sensor mounted 30 millimetres too far along the belt. Streaks across a surface, a part cut off at the frame edge, the same piece imaged twice, a piece never imaged at all — all four are acquisition faults, all four have physical causes, and three of them are settled by arithmetic rather than by opinion.
Start with a count, because it costs ten minutes and it partitions the problem before anybody opens a configuration screen.
Ten minutes: count triggers against parts
Run 100 parts and record three numbers: triggers fired, images acquired, and parts actually produced as counted at the machine. Three numbers, one shift-supervisor-sized task, and the pattern between them tells you which half of the system to open. Note the third row in particular — triggers matching parts while images fall short is a throughput ceiling rather than a triggering fault, and it belongs to it keeps up on the bench and falls behind at line speed.
| Pattern | Fault class | Where to look |
|---|---|---|
| Triggers greater than parts | Multiple triggering | Sensor sees more than one edge per part; debounce or lockout is missing or too short |
| Triggers fewer than parts | Missed triggering | Sensor type wrong for the material, gap between parts too small, or part presentation varies |
| Triggers equal parts, images fewer | Acquisition dropping frames | Buffer, interface or host cannot keep up; this is a throughput fault, not a trigger fault |
| All three equal, images unusable | Exposure, strobe or encoder | The rest of this page |
The blur arithmetic, done once, ends the exposure argument
That last sentence is the trap. Cutting exposure from 1 millisecond to 200 microseconds removes four fifths of the light reaching the sensor, so either the illumination gets five times brighter or the images get five times darker, and darker images lose exactly the low-contrast defects the shorter exposure was meant to preserve. This is why a strobe exists: it delivers a large pulse over tens of microseconds and freezes motion without asking a continuous lamp to be five times more powerful. It also introduces a new way to fail, because a strobe that fires outside the exposure window produces a dark frame with no obvious cause.
- Check the strobe fires inside the exposure window, not merely at the same time as the trigger. A scope on both signals settles in minutes what a week of image review will not.
- Check the strobe is not being asked to fire faster than its duty cycle allows. At high part rates some units skip pulses rather than overheat, which produces intermittently dark frames.
- Check whether the exposure was shortened by someone chasing blur after a line speed increase. This is the most common undocumented change on an inspection cell, and it degrades every low-contrast class at once — the same contrast loss described in the model that finds the scratch on day shift.
- Check that the smallest defect you must find still spans several pixels after any resolution change. If a class was marginal at feasibility it will be the first casualty, which is what the pre-spend gate in fifteen checks that tell you whether a camera can judge this part exists to catch.
Clipped parts are geometry: offset, latency and jitter
A part cut off at the edge of frame is almost never a cropping bug. Between the sensor detecting the part and the sensor array being exposed there is a chain of delays — sensor response, controller scan, cable, camera trigger handling — and during that time the part keeps moving. At 500 mm/s a total latency of 20 milliseconds carries the part 10 millimetres past where the trigger thought it was. That offset is constant and can be compensated with a trigger delay.
What cannot be compensated is jitter. If the latency varies by 5 milliseconds — a controller scanning on its own cycle rather than responding to an interrupt is the usual reason — the part lands 2.5 millimetres either side of nominal on every image. Set the field of view so that the part fits with a margin bigger than the jitter and you have a robust cell; set it tight to maximise resolution and you have a cell that clips one part in fifty and nobody can say why.
- Measure the offset. Image 20 parts, mark the leading edge position in pixels, and take the mean. That is your fixed offset, convertible to a trigger delay.
- Measure the jitter. Take the spread of those same 20 positions. That is what your margin has to absorb.
- Compare the jitter against the free space at the frame edge. If the spread is larger, the cell will clip parts regardless of the delay you set.
- Fix the source before the symptom. A trigger routed through a controller's scan cycle inherits that cycle's variability; a hardware trigger wired to the camera does not.
- Re-measure after any change to belt speed, part length or the sensor's mounting. All three move the numbers.
Encoder faults have a signature nothing else has
On a line-scan station the image is built one line at a time, clocked by an encoder on the drive. Every geometric property of the image along the travel direction therefore depends on the encoder being right, and encoder faults are unique in leaving distortion along travel only, never across it. A round hole photographs as an ellipse; a square part comes out as a rectangle; the cross-web dimension stays correct throughout. That asymmetry is the diagnosis.
There are two versions. The first is a wrong constant: an encoder giving 1000 pulses per revolution on a 100 mm diameter roller produces one pulse every 0.31 mm of belt travel, so a station configured for 0.1 mm per line is stretching every part by roughly three times unless the divider is set correctly. This is consistent and repeatable, and it shows up the moment you image a calibration artefact of known length. The second is slip: the encoder wheel loses grip on a dusty or loaded belt, so the stretch varies part to part and within a single part. Consistent distortion is arithmetic and is fixed in configuration; varying distortion is mechanical and is fixed with a hand on the encoder mount.
Distortion along the direction of travel and nowhere else is an encoder telling you it is not measuring the belt. No amount of training data will teach a model to un-stretch a part by an amount that changes every piece.
Duplicated and missing frames: debounce, pitch and what the sensor can see
A part with a slot, a hole or a stepped profile presents more than one edge to a photoelectric sensor, and each edge is a trigger. The result is two or three images of the same piece, which then arrive downstream as separate parts with separate verdicts, and a reject on any of them looks like a reject on the part. The fix is a one-shot lockout sized from the minimum part pitch: parts 300 mm apart at 500 mm/s arrive every 600 milliseconds, so a lockout around 400 milliseconds suppresses the extra edges while never suppressing a real part.
- Size the lockout from measured pitch, not from a default. Too long and it deletes genuine parts during a burst; too short and it does nothing.
- Match the sensor to the material. A diffuse-reflective sensor may not see a matt black or a translucent part at all, and swapping to through-beam or background suppression solves a triggering problem that looks like software.
- Watch changeovers. A shorter part at the same belt speed halves the pitch and can walk straight into a lockout tuned for the previous product.
- Keep the acquisition parameters in the recipe, not in an engineer's memory. Exposure, delay, lockout and encoder divider are per part number and belong with the part's other master data — the argument in master data before a plant goes live.
- Exclude bad frames from the training set explicitly. A labelling standard needs an 'unusable image' verdict, or blurred and clipped frames get labelled as parts and taught to the model, which is one of the adjudication rules in labelling defect images so two inspectors agree.
The decision tree: steel, wiring or code
- Counts unequal? Fix triggering first — sensor type, mounting, lockout. Nothing else can be assessed until one part produces one image.
- Smear across the whole image, in the direction of travel? Exposure against speed. Compute the blur, then decide between a shorter exposure with more light or a strobe.
- Smear on some frames only? Strobe synchronisation or duty cycle. Put a scope on the trigger and the strobe.
- Distortion along travel only? Encoder. Consistent means a divider or resolution setting; varying means slip, and that is a mechanical job.
- Parts consistently off-centre? Trigger offset, correctable with a delay. Parts randomly off-centre? Jitter, correctable only with margin or a hardware trigger.
- Everything measured and correct, images still poor? Now it is optics or presentation, and the model still has not been touched.
The organising principle is that acquisition is cheap to measure and expensive to assume. A morning with a tape measure, a scope and a calibration artefact answers questions that months of model work cannot, and it produces numbers a supplier and a plant team can argue about on the same terms. When we scope an inspection build, these measurements sit at the front of the plan rather than in a commissioning week at the end — the shape of work described under MVP and product builds. The rest of the cluster, from feasibility to acceptance, sits in visual inspection and defect detection, part of our manufacturing and industrial vision work.
Frequently asked questions
Short answers to the follow-ups this page tends to raise.
How do I know whether motion blur or the model is the problem?
Compute the blur in pixels: belt speed multiplied by exposure time, divided by millimetres per pixel. If the answer is larger than the smallest feature you need to detect, the defect is not in the image and no model can find it. A belt at 500 mm/s with a 1-millisecond exposure and 0.1 mm per pixel gives about 5 pixels of smear, which is enough to erase a fine scratch.
Why is the part cut off at the edge of the inspection image?
Because the part moves between the trigger firing and the sensor being exposed, and the field of view has no margin for the variation in that delay. The average delay is a fixed offset you can cancel with a trigger delay; the variation cannot be cancelled and has to be absorbed by leaving free space at the frame edge. Measure the leading-edge position on 20 parts to get both numbers.
What causes duplicate images of the same part?
A trigger sensor seeing more than one edge as the part passes — typical of parts with slots, holes or stepped profiles. Each edge fires a trigger and each trigger produces an image, so one part becomes two or three records with independent verdicts. Add a lockout after the first trigger, sized shorter than the minimum time between parts at full line speed.
How do I recognise an encoder fault on a line-scan station?
Look for distortion along the direction of travel with correct dimensions across it. Round holes appear elliptical and part length is wrong while width is right. If the error is the same on every part, the encoder resolution or divider is misconfigured; if it varies from part to part, the encoder wheel is slipping and the fix is mechanical.
- machine vision
- triggering
- encoders
- image acquisition
The work behind this page
Builds from our portfolio that this page draws on.
Open Vision PPE Monitoring
Boundary surveillance, PPE compliance monitoring, and intrusion detection via real-time video analytics. Runs fully on-premise — no cloud required.
Safety & ComplianceFactory OS
Production planning and task management for a tier-1 apparel manufacturer — replacing Excel with automated milestone planning, SOP gate enforcement, and real-time visibility.
ManufacturingRead next
- It keeps up on the bench and falls behind at line speedA cell that is fast on the bench and late on the line has either a latency problem or a throughput problem, and they need opposite fixes. Queue depth over ten minutes tells you which, and a per-part budget tells you where the time went.diagnostic
- The model finds the scratch on day shift and misses it after darkA detection rate that tracks the clock is an optics fault wearing a model's clothes. Re-scoring yesterday's archived images tonight proves the model is unchanged and points the investigation at the light instead.diagnostic
- The camera passed it and the customer found it: tracing an escape backwardsMost escape investigations start by retraining a model that never saw the surface in question. Working backwards through the retrieval chain — part to image to score to disposition — settles which of five very different faults you actually have.diagnostic
- False reject rate: what the line feels, not the accuracy on the slideA 1% false reject rate sounds like rounding. At 1,200 parts an hour it is 96 good parts a shift and over an hour of somebody re-checking them.definition
- Rejects double the morning a new coil or resin lot goes onA reject spike that starts at a lot changeover is a revalidation problem, not a retraining problem. The overlay takes an hour, and the 30-part retest tells you whether the material got worse or merely got different.diagnostic
- The anomaly score is a ranking; the threshold is a business decisionThe score orders parts by how far they sit from normal. It carries no units, no probability and no severity — which is why the threshold belongs to quality, not to engineering.definition
Working on something in this space?
Tell us where you are in a sentence or two. We'll tell you honestly whether we're the right team, and what a sensible first slice of the work looks like.
Start the conversation