Remote Inspection:
How Computer Vision can support Industrial Maintenance

The customer wanted to investigate whether part of the visual inspection process could be performed remotely.

Relevant image material was already available. The idea was to use this information more systematically, allowing maintenance personnel to identify specific visual features without having to examine every location in person.

The objective was deliberately limited. The system wasn't intended to take over maintenance decisions or replace experienced technicians.

Instead, computer vision would help analyze the available images and provide information that could support subsequent inspections.

Whether this approach would be useful depended on two questions: could the relevant features be identified reliably, and would the information make the existing maintenance process easier?

Together with the customer, we developed a computer vision system designed to identify relevant features in image material.

Unlike conventional image processing, which typically relies on explicitly defined rules, machine learning-based computer vision can learn to recognize visual characteristics from examples.

This can be useful when the objects or features being inspected don't always look exactly the same.

Images may differ in perspective, lighting, resolution or background. Equipment may also show signs of wear, contamination or other changes that affect its appearance.

Such variations make image-based inspection more challenging than recognizing an object under controlled laboratory conditions.

For the development of the system, the important task was to establish whether the relevant features could be identified with sufficient reliability in the customer's available image material.

The images also needed to support the actual inspection objective. Recognizing something in a picture is only useful if it provides information that maintenance personnel can act on.

A computer vision model can achieve promising results during development and still be of limited use in daily operations.

For example, an inspection system might correctly identify a relevant feature in most images but struggle with particular camera angles or poor lighting conditions.

These limitations matter when maintenance personnel use the results to decide whether further investigation is necessary.

The system therefore needs to be considered in the context of the inspection process.

What information should be presented to the maintenance team? How can users recognize when an image doesn't allow a reliable assessment? And which findings still require an inspection on site?

These questions are particularly relevant when remote inspection is intended to reduce the need to access difficult or potentially hazardous locations.

The goal is to make better use of available information without creating a false sense of certainty.

Remote visual inspections have a practical limitation: an image only shows what was captured at a particular moment and from a particular perspective.

Some conditions may be clearly visible, while others require closer examination, additional measurements or physical access to the equipment.

Computer vision doesn't eliminate these limitations.

It can, however, help maintenance teams examine image material more consistently and focus attention on features relevant to the inspection.

Where the image quality and application allow it, this may reduce the need for certain routine on-site checks or help determine where a closer inspection is warranted.

The value lies in how the information contributes to maintenance decisions.

A system that identifies visual features but leaves users with additional manual work hasn't necessarily improved the process.

That is why the practical use of the results matters as much as the recognition task itself.

The project highlighted an important distinction between developing a computer vision model and developing a useful inspection system.

The technical task was to identify specific features in existing image material. The broader challenge was to make that capability relevant to the customer's maintenance activities.

Several considerations follow from this.

The suitability of the available images has to be assessed against the intended inspection task. Detection results need to be presented in a way that maintenance personnel can interpret. And limitations must remain visible, particularly when image quality or viewing conditions affect the reliability of the assessment.

Working with the customer is essential to understanding these requirements. The people responsible for maintenance know which observations matter, what requires further investigation and where remote assessment can reasonably support their work.

This knowledge cannot be derived from image data alone.

Computer vision can be particularly interesting for inspection tasks that involve recurring visual checks, difficult access conditions or large amounts of image material.

Potential applications include inspecting visible components, identifying changes in equipment condition or supporting the review of images captured during scheduled inspections.

Not every inspection task is suitable for automation.

The feasibility depends on whether the relevant characteristics are visible, whether appropriate image material is available and whether the results can support a meaningful maintenance decision.

At craftworks, we approach these questions together with our customers before deciding how much automation is appropriate.

In this project, the starting point was an existing maintenance challenge and the opportunity to use available images more effectively.

The resulting computer vision system was developed to support that specific task.

Could computer vision make your maintenance inspections easier?

craftworks develops computer vision and industrial AI solutions tailored to existing equipment, data and operational processes. If you're exploring remote inspection or automated image analysis, we're happy to discuss your use case.