Get a Free Quote

Our representative will contact you soon.
Email
Mobile/WhatsApp
Name
Company Name
Message
0/1000

Blog

 >  Blog

News

The Step That Still Needed A Person

Time : 2026-11-20

The process sheet for a magnetic particle inspection cell fits on one line: clamp, spray the magnetic particle fluid, magnetise, inspect under ultraviolet light, demagnetise. Four of those five steps had been mechanised for years. The fifth was a person standing in front of a lamp.

That fifth step is where the takt time, the consistency and the traceability of the whole cell were decided, and it is the step that magnetic particle inspection automation has been slowest to touch, because reading a fluorescent indication is not the same problem as finding a dimensional error.

1. Four Steps Are Machine Work; One Is Judgement

Clamping, spraying, magnetising and demagnetising are all actions with a defined start and end. They can be sequenced, timed and repeated to a second, and they behave the same way on the thousandth part as on the first.

Inspection is different in kind. The magnetised part carries a pattern of fluorescent particles that reveals where the field leaked, and somebody has to decide which of those accumulations are cracks, which are geometry, which are harmless indications from a section change, and which are artefacts of the spray itself.

That decision is the cell's real output, and while it is made by eye it carries three costs that are not visible in a cycle time study.

2. Why The Manual Step Sets The Pace

The first cost is attention. Reading fluorescent indications means looking at a dark surface under ultraviolet light, repeatedly, for a whole shift, while deciding on each part. It is dense repetitive work, and the decline across a shift is not visible in any record because the record only contains the verdict, not the confidence behind it.

The second cost is the working environment. A dark booth, ultraviolet light, magnetic particle mist, and a posture held over a part for hours. That combination is hard to staff, and it is hard to staff consistently, which is a production problem disguised as a recruitment problem.

The third cost is traceability. A manual verdict produces a signature and, at best, a note. It does not produce an image of the indication that was judged, so a dispute six months later is an argument about memory rather than a review of evidence.

3. Turning A Judgement Into An Image

The shift from manual to automated magnetic particle inspection is not a change in the magnetic process. Magnetisation, spray and demagnetisation stay as they were. What changes is that the reading step is replaced by an imaging step and a decision made by software.

The imaging hardware is a camera and an ultraviolet source, and the source matters more than the camera. A fluorescent indication is only bright relative to its surroundings, so the contrast the algorithm has to work with is set by how cleanly the surface is lit at the excitation wavelength and how little visible light reaches the booth. A source that produces visible spill washes out the indication before the camera sees it.

The decision is then a classification of what the image contains: a linear accumulation along a stress direction, an accumulation at a geometry change, a diffuse patch from a spray droplet, or nothing. Each of those is learnable, and none of them is a threshold that can be set once by hand.

4. What Makes It Harder Than Dimensional Vision

A dimensional station measures a feature that is designed to be there. A magnetic particle station judges a pattern that appears where the part is discontinuous, which means the pattern depends on the defect and on the magnetisation and on the spray, all at once.

Three consequences follow. The indications are low in contrast and small in area, so resolution matters more than frame rate. The same physical crack produces different images depending on magnetisation direction, which means a two-direction process needs the images from both directions judged together. And the particle concentration in the fluid changes over a shift, which changes the brightness of every indication and therefore any threshold set in absolute terms.

That last one is the reason a fixed threshold performs poorly over a full shift, and the reason fluid concentration is worth monitoring as part of the inspection rather than as part of maintenance.

5. What The Numbers Look Like

On railway turnout components, an automated magnetic particle inspection cell running the full five-step process reports a miss rate of 0.1 percent and a false call rate of 1 percent, at a cycle of 140 seconds including spraying and loading.

On crankshafts the criterion is finer: cracks of 0.1 mm by 2 mm, with the resolution to identify an indication of 0.1 square millimetres, over main journals, rod journals, oil hole openings, transition areas and non-machined surfaces. That coverage is worth noting, because the non-machined surface is exactly where a person scanning a rotating part tends to look least.

These are not universal figures, and presenting them as such would be dishonest. They describe what a cell achieves against a defined criterion on a defined part, and the same technology on a different geometry needs its own validation.

The two-direction requirement is worth stating explicitly, because it is where automated magnetic particle inspection differs from a camera replacing an eye. A single image of a part magnetised one way contains only the indications that the field produced; the crack that lies parallel to that field is not in the image at all. Judging the part therefore means judging two images together, from two magnetisation states, which is a fusion decision rather than a detection decision.

That fusion is also where the automation earns more than it costs. A person looking at a part after each magnetisation has to hold the first image in memory while looking at the second, which is exactly the task human attention is worst at over a shift. A system that keeps both images and evaluates them jointly does it the same way on the last part as on the first.

6. Why The Model Cannot Be Finished At Delivery

A rule-based inspection is complete when it is installed. A learned inspection is not, because the population of indications it will meet is larger than the sample it was trained on.

The practical answer is a platform that keeps learning from the line it is on: every part inspected produces an image, every image is a potential sample, and every judgement an inspector confirms becomes a label. Over weeks the model meets the indications this particular process produces, which is a different distribution from any general set.

This changes the acceptance criteria for such a station. Its accuracy on day one is a starting point rather than a specification, and the number that matters is how quickly it converges and whether anyone is reviewing the confirmations that drive the convergence.

7. Where The Person Still Belongs

Automating this step does not remove the person; it moves them.

The judgement that remains is the one the station cannot make: whether a defect class it flagged for the first time matters, whether a criterion should change because the foundry changed, and what to do with a part that sits on the boundary. Those are decisions about the specification rather than about the part, and they belong to somebody who can change the specification.

The role that disappears is the one that was unsustainable anyway: hours of fluorescent reading in a dark booth, with the whole cell's output waiting on it.

8. What This Changes Upstream And Downstream

Automating the reading step changes the requirements on both sides of it, and both are cleaning requirements.

Upstream, the surface has to be clean in a way an imaging step can verify. An oil film or a flash rust layer reduces particle mobility exactly where the indication should form, and the algorithm cannot recover information the physics never produced. A wash and dry stage ahead of the cell is not housekeeping; it is part of the measurement.

Downstream, the part leaves carrying particles in threads, blind holes and anywhere the carrier pooled, and it leaves magnetised unless it is demagnetised. An automated cell makes both of those steps more visible, because it produces a record for every part and the record shows what was and was not done.

The summary is that magnetic particle inspection was never short of a magnetising machine. It was short of a way to read what the magnetising produced, consistently, at rate, with evidence. Replacing that one step is what turns a five-step process from four machines and a person into a process.

Get a Free Quote

Our representative will contact you soon.
Email
Mobile/WhatsApp
Name
Company Name
Message
0/1000