MAS suite · Visual Inspection

The model proposes. The inspector decides. Both are recorded.

IBM Maximo Visual Inspection inside real inspection workflows: a written defect dictionary, a training set labelled with your inspectors, precision and recall per defect class, and every confirmation or override written to the work record in Manage.

Inspector working at plant equipment with a model-assisted view alongside, representing IBM Maximo Visual Inspection

One disagreement, five beats

Whether inspectors keep using this is settled the first time the model is wrong

A composite weld inspection. Details are illustrative; what each beat writes to Manage is not.

  1. On the round

    The inspector photographs a weld region against the asset the work order names. The model returns a crack indication at moderate confidence, region drawn on the image.

    Written Image against the work order, with model version, class returned and confidence value.

  2. Ten seconds

    She is standing in front of it. The surface is coating breakdown over sound metal, not the class returned, and she overrides. The model gets no second attempt at persuading her.

    Written Override with the reason taken from the agreed defect dictionary, plus the remedial task she judged necessary.

  3. Before she moves on

    She confirms the finding in her own words, under the competence regime that already governs the work. The model output is retained beside the image, not replaced by the correction.

    Written Finding, image, model version, model output and override, all on one work record.

  4. Friday

    The inspection lead works the week of disagreements with us, confidence value beside outcome, sorting each into model error, dictionary ambiguity or inspector calibration.

    Written Dictionary revisions and retraining candidates, with version history kept.

  5. After shadow

    The model screens routine volume on the two classes where measured agreement supports it, and stays advisory on the safety-critical class.

    Written The scope decision per class, with the measurement it was granted on.

The inspector decides on every class, every time, including the classes the model screens. Scope is granted one class at a time and can be withdrawn at a Friday review.

What a disagreement is evidence of

Three reasons an override happens, and only one of them is the model

Overrides are the most useful data this capability produces. Each kind has a different fix, and a wrong diagnosis is how a programme retrains a model to match an inconsistency.

Ref DiagnosisWhat it meansThe fix
O1 Model errorThe model was wrong on an image the dictionary handles clearly. The only one of the three that is a model problem.Retrain, or narrow the class the model is allowed to screen.
O2 Dictionary ambiguityThe boundary between two classes is genuinely unclear, so the inspectors have been applying it differently as well.Revise the written dictionary with the inspectors, then relabel the affected examples.
O3 Inspector calibrationAn inspector is applying an older threshold, learned from someone who has since left. Worth knowing before an auditor finds it.Calibration session against the dictionary. No model change at all.

Cite O1 to O3 in a scoping note and we will know what you mean.

Effort, in figures

What the training data actually costs

300–800
Labelled examples per defect class

Range across the workflows we have scoped, borderline examples included. Below roughly 150 we exclude the class rather than train something unmeasurable.

4–6 weeks
Dictionary and labelling, before any model build

Phase one on every engagement: classes written with boundaries and examples, historical images labelled with the inspectors judged against them.

6 weeks
Minimum shadow period with no authority

The model runs alongside the round and changes nothing. The inspection lead reviews disagreements weekly before scope is granted.

< 1 min
Inspector time to clear a flagged image

Measured on the Mobile inspection step. It decides how many false positives a round can absorb, which is why we publish the rate per class.

From no authority to screening

Four gates a model passes before it screens anything

Gates one and two happen before the model exists. That sequencing is why a model here does not arrive as a surprise to the people it affects.

  1. Gate 1 Written defect dictionary

    Signed by Inspection lead

    Passes

    Classes with boundaries and worked examples, agreed with the inspectors judged against them, disagreements resolved on the record.

    Held back

    Definitions that live only in people. No ground truth to train or measure against, so labelling cannot start.

  2. Gate 2 Training set from your own images

    Signed by Inspection lead with MaxIron

    Passes

    Historical images tied to a work order and an asset, labelled with your inspectors against the dictionary.

    Held back

    Folders named by date and device. Capture discipline becomes phase one and the model waits.

  3. Gate 3 Validation against inspector calls

    Signed by Asset or reliability lead

    Passes

    Precision and recall per class against historical decisions, with the weak classes named rather than averaged away.

    Held back

    A single accuracy figure across all classes. It hides exactly the class you care about.

  4. Gate 4 Shadow, then scope per class

    Signed by Competent person under the existing regime

    Passes

    Screening authority one class at a time, on measured agreement, withdrawn the same way if agreement falls.

    Held back

    Adjudication of a safety-critical call. That stays with the competent person, and we do not bid for it.

Scope and boundaries

Where a vision model stops being the right tool

It does not replace the inspector

The person in front of the asset decides on every class. We would not deliver it otherwise in a safety-critical regime.

It is not a general condition survey

Open-ended estate surveys are a poor fit. A defined asset population and a defined class list are where this earns its keep.

It goes stale if nobody watches it

Cameras, coatings, lighting and the inspector population change. Drift is instrumented in the run-state and retraining scheduled, so it is detected rather than discovered.

The model is not the deliverable

MAS ships the model lifecycle. We add the dictionary, the labelling, the Maximo Mobile workflow, the Manage integration and the run-state around it.

IBM Maximo Visual Inspection, frequently asked questions

What does IBM Maximo Visual Inspection do?
It applies image and video models to inspection tasks: defect detection on equipment, condition classification and anomaly spotting on assets that are already photographed on a round. The classification, the image and the inspector decision land on the work record in Manage.
How many images does a defect class need?
Between roughly 300 and 800 labelled examples per class before validation means anything, with the awkward and borderline examples included. Below about 150 we exclude the class rather than train a model that cannot be measured.
Who reviews a flagged image?
The inspector on the round, in the Mobile inspection, who confirms or overrides in seconds. Disagreements are then reviewed weekly as a set by the inspection lead. Safety-critical classes stay with the competent person named in the existing regime.
How are false positives handled?
On a screening class we tune for recall and publish the false-positive rate per class, because a flagged image costs an inspector under a minute while a missed defect does not. A class whose false-positive rate the round cannot absorb stays advisory.
Does MaxIron build proprietary computer-vision models?
No. We use the model lifecycle that ships with MAS, and add the defect dictionary, labelling with your inspectors, workflow design in Maximo Mobile, Manage integration and managed run-state on our cloud.
We inspect runways, rolling stock or lineside infrastructure. Is there a page for that?
Yes. The regulated regimes and field constraints are covered on Visual Inspection for transport, rail and airports. This page stays on the capability and its prerequisites.

Bring a hundred of your own inspection images.

Include the borderline ones your inspectors argue about, and whatever defect definitions exist today. We will say whether there is ground truth to train against, which classes are candidates, and what the dictionary work involves.

Bring this to the first call

  • One inspection workflow, and how often it runs
  • A hundred historical images, with the findings recorded against them
  • Whatever written defect definitions exist, however incomplete
  • The two or three classes where inspectors disagree most often
  • What a missed defect costs, and who signs the inspection today