Faster, more consistent inspections, with a defensible audit trail

IBM Maximo Visual Inspection applies image and video models to inspection tasks: defect detection on equipment, condition classification, and anomaly spotting on assets that are routinely photographed or filmed. The output feeds into the inspection workflow and the work record in Manage, with an audit trail that survives a change of inspector and a regulatory review.

When Visual Inspection earns its licence

Three conditions usually have to be true. The inspection is frequent enough that consistency between inspectors varies materially. The cost of a missed defect is high — safety-critical, regulatory, or operationally expensive. The inspection target is well-defined enough that a model has a fair chance of learning it. Where those conditions hold — runway inspections, rolling-stock checks, large-population equipment surveys, safety-critical asset reviews — Visual Inspection is one of the few AI-led MAS components with a clear payback. Where they do not, it produces a model that the inspectors quietly stop trusting.

What we deliver

• Workflow scoping with the inspectors who actually do the inspection, not a process map written in isolation
• Image and video labelling against historical defect data, with inspector calibration
• Model build and validation using the MAS Visual Inspection lifecycle
• Manage integration so classification becomes part of the work record, not a parallel system
• Mobile and field workflow design so inspectors are faster, not slower, with the model in the loop
• Audit and evidence trail so a regulator or an internal auditor can reproduce the basis of every decision
• Managed run-state on our cloud, with model retraining and operational support owned by the same team that runs Manage

Where the model earns trust

A vision model that the inspectors do not trust is worse than no model at all. The few we have seen earn lasting trust have four practical disciplines underneath them, and they tend to decide whether the rollout sticks long after the technology selection.

• Labelling discipline. The training set is built by, or with, the inspectors who actually do the inspection. Labels are calibrated against a written defect dictionary, not "what looked wrong to whoever was free that afternoon." Disagreements between labellers are resolved deliberately, and the resolution is recorded.
• Evaluation against ground truth. The model is measured against the inspectors’ own historical decisions, broken down by defect class and by inspector. We are explicit about precision and recall on the classes that matter, and equally explicit about the classes where the model is not yet good enough to operate alone.
• Drift monitoring. The condition of the asset estate, the camera hardware, the lighting and the inspector population all change over time. The model goes stale. We instrument the run-state so drift is detected, not discovered. Retraining is a scheduled operation, not a fire.
• Audit-grade evidence. Every classification carries the model version, the input image and the inspector’s confirm or override, attached to the work record in Manage. A regulator or an internal auditor can reproduce the basis of every decision months later. This is the discipline that keeps Visual Inspection inside the inspection regime, rather than parallel to it.

How this fits with the rest of the programme

Visual Inspection complements Health on safety-critical asset classes, where the inspection output drives the health score. It often surfaces alongside Maximo Mobile rollouts, because the practical workflow lives on the inspector's device. And it sits on the same managed run-state we operate for the rest of MAS — see managed Maximo and MAS hosting.