How machine vision systems detect defects in composite airframes

Composite airframes offer weight and performance advantages, but their layered construction makes inspection a demanding task. A faint wrinkle in a newly laid ply may indicate a manufacturing problem, while a barely visible surface mark on an aircraft in service may warrant a closer assessment. For quality control and safety teams, the challenge is not simply finding an unusual feature; it is deciding what that feature means, documenting it, and routing it to the right inspection or disposition process.

Machine vision systems support those decisions by combining cameras, controlled illumination, positioning equipment, and image analysis. They can make surface inspection more repeatable and provide records for review. They do not, however, see every defect in a composite laminate. Understanding that distinction is essential when building an inspection process that supports airworthiness rather than creating a false sense of certainty.

What machine vision can see on a composite airframe

A vision system detects differences in the light returned from a surface and interprets those differences against defined inspection criteria. Depending on the setup, an image may reveal changes in texture, shape, color, or reflectivity. The system then identifies features that could correspond to a defect, measures them where possible, and presents the result for acceptance or review.

Typical targets include:

• Fiber misalignment and wrinkles: Visible deviations in fiber direction or surface contour, particularly during layup when material is still accessible.
• Gaps, overlaps, and misplaced material: Irregularities between adjacent tapes or plies that may be observable before later layers conceal them.
• Surface cracks and scratches: Linear indications on accessible surfaces, subject to image resolution, lighting, and surface condition.
• Impact-related indications: Dents, coating damage, or other visible marks that can prompt further investigation.
• Foreign material and surface contamination: Objects or residue that contrast sufficiently with the surrounding material under the chosen imaging conditions.

Detecting an indication is not the same as establishing its structural significance. A visible mark may be cosmetic, or it may be associated with damage extending beneath the surface. Conversely, a laminate can contain subsurface damage without an obvious external sign. The applicable engineering criteria and inspection instructions determine what happens next.

From light to an inspection decision

Imagine a production team inspecting a curved composite skin after a layup operation. An operator can scan the area by eye, but fatigue, reflections, and variations in working distance can make subtle features difficult to assess consistently. A machine vision station addresses those variables by controlling how images are captured and analyzed.

1. Capture an image suited to the defect

The camera, lens, and illumination must be selected together. Resolution and field of view need to be sufficient for the smallest indication the process is expected to detect, while focus and working distance must remain suitable across the inspected area. A system covering a large fuselage section in one image may lack the detail needed for a fine surface crack; several overlapping views may be more appropriate.

Lighting is equally consequential. Low-angle illumination can make shallow surface relief more apparent through shadows, while diffuse lighting can reduce distracting reflections. Multiple lighting directions may be needed on glossy or curved surfaces. No single arrangement reliably reveals every defect type, so the intended application should drive the imaging design.

2. Locate and analyze relevant features

After acquisition, software may correct image distortion, align views to a reference, and isolate the inspection area. Traditional image-processing methods can measure edges, texture, contrast, and geometry against defined thresholds. For example, analysis of visible fiber patterns may flag a local change in orientation relative to an expected direction.

Some systems use machine learning to classify or highlight patterns in images. These models learn from examples rather than relying solely on manually defined rules, but their performance depends on the quality and representativeness of the training and validation data. Changes in material, lighting, tooling, or surface finish can alter the appearance of both acceptable features and defects.

3. Record, review, and disposition

A useful output is more than a red box on a screen. Reviewers need the image, location, inspection context, and the reason an area was flagged. Measurements should be presented with the resolution and accuracy limits of the system in mind. An authorized person or approved process can then compare the indication with applicable acceptance criteria and decide whether to accept, investigate, repair, or escalate it.

For a large part, the location must be mapped to a usable part coordinate or reference feature. Without that link, an inspector may spend valuable time finding the same indication again, especially after the component has moved between stations.

Where the approach fits in manufacturing and maintenance

Vision inspection is most useful when it is placed at a point where the relevant feature is visible and a decision can still change the outcome.

During material placement, cameras can monitor exposed plies or automated tape and fiber placement paths for visible gaps, overlaps, wrinkles, and placement deviations. Inspecting between layers matters: once another ply covers the area, an optical camera may no longer have access to the feature. The inspection plan should account for machine motion, material variation, and the time available before the next operation.

After curing and finishing, a vision system can examine accessible surfaces for cracks, scratches, contour changes, or other defined indications. Coatings, sealants, fasteners, and complex geometry can obscure features or generate misleading contrast. Surface preparation and the choice of viewing angles therefore become part of the inspection method, not incidental details.

In service, imaging can support the documentation and comparison of visible damage on accessible airframe areas. Repeatable views can help teams track an indication over time, provided the images are tied to the same location and captured under sufficiently controlled conditions. A changed appearance alone does not establish damage growth; measurements and follow-up must follow the applicable maintenance instructions.

Why optical inspection needs other NDI methods

Conventional visible-light cameras primarily assess what is exposed at the surface. Delamination, porosity, disbonds, and other internal conditions may not produce a detectable visual signature. Barely visible impact damage is a particular concern because limited external evidence does not necessarily describe the full extent of subsurface damage.

Non-destructive inspection methods such as ultrasonic testing, thermography, radiography, or shearography may be appropriate for particular materials, geometries, and suspected damage mechanisms. Each has its own access requirements and limitations. For example, an ultrasonic method may help characterize certain internal indications, while thermography can reveal some subsurface anomalies through their thermal response. The right choice depends on the approved inspection task, not on a general assumption that one technique covers every composite defect.

Vision can still play an important role alongside these methods: it can document the surface, identify areas needing closer examination, and associate results with part locations. It should not replace a required non-destructive inspection merely because an image appears clear.

How to judge whether a vision solution is ready for use

For a quality manager, the most persuasive demonstration is not a collection of striking defect images. It is evidence that the proposed inspection works on the actual parts, defect types, and operating conditions it is intended to cover.

• Define the inspection target. Specify the material, surface state, accessible area, defect type, and minimum relevant feature size. Separate detection requirements from measurement requirements.
• Challenge the system with representative samples. Include acceptable variation as well as relevant indications, across curved geometry, lighting changes, and realistic surface finishes. Use independently assessed references where possible.
• Examine both misses and false calls. Missed indications can create safety risk; excessive false calls can overload reviewers and encourage routine dismissal of alerts. Report performance for the defined test conditions rather than relying on a single headline accuracy figure.
• Check repeatability. Reinspect representative areas after repositioning the part or system, and assess the effects of different shifts, environmental conditions, and equipment settings.
• Plan for exceptions. Define what happens when an area is obscured, the image is out of focus, coverage is incomplete, or the software cannot produce a reliable result.
• Maintain traceability. Link images and results to the part, location, inspection procedure, equipment configuration, software version, time, and reviewer decision as required by the organization’s quality system.

Acceptance thresholds deserve particular attention. Tightening a threshold may reveal more subtle indications, but it can also increase the number of benign features sent for review. The appropriate balance depends on the consequence of a miss, the reliability of subsequent inspection, and the approved criteria for the part.

Keeping the inspection trustworthy over time

A system validated on installation does not remain trustworthy automatically. Camera movement, lighting degradation, lens contamination, software updates, new material batches, and changes in part finish can all affect results. Establish checks for image quality and coverage, control configuration changes, and define when revalidation is needed. Machine-learning models require the same discipline: new training data or a revised model can change behavior and should be assessed before operational release.

Human review also needs clear boundaries. Inspectors should know which indications the system is qualified to flag, where it cannot see, and how to handle ambiguous results. If an alert is overridden, the reason should be recorded in accordance with the applicable process. This makes the record useful for investigation and improvement rather than merely proving that an image was taken.

Aerospace inspection requirements are governed by the applicable design, production, maintenance, and regulatory framework. FAA or EASA oversight and relevant standards may be involved depending on the operation and jurisdiction, but installing a vision system does not by itself establish compliance or authorization to change an inspection task. Any substitution or modification must follow the applicable approved procedures and organizational controls.

Used within those boundaries, machine vision systems give quality and safety teams a more consistent way to examine and document visible composite features. Their greatest value comes from a well-defined inspection target, evidence of performance on representative parts, and a clear route from an image indication to the next engineering decision.