Gains:
- Ability to understand visual inspection and NDT methods and use computer vision as a large surface pre-screening and attention guide.
- Ability to apply that the distinction between relevant/non-relevant and acceptance/rejection of an indication belongs to the certified examiner and the limit document.
- Ability to understand the risks of false negative and false positive and not end the examination with 'the system said clean'
Most of the care starts with the eyes. A crack, a corrosion spot, a loose fastener, a trace of a leak — most faults are first caught by visual inspection. In modern aviation, this eye is now accompanied by computer vision (artificial intelligence that detects objects, defects and patterns from images): drone body scanning, borescope image analysis, automatic defect marking. In this unit, we will cover how AI shares the burden of the eye in visual inspection and NDT (Non-Destructive Testing, methods for detecting internal/surface defects without damaging the part), but why the decision is still up to the authorized inspector.
Visual inspection and NDT basics
Visual inspection can be general (GVI) or detailed (DVI); It is done with the naked eye, a magnifying glass or a borescope. NDT methods find defects that cannot be seen with the naked eye:
- Eddy current (eddy current — finds surface/subsurface cracks electromagnetically),
- Ultrasonic (ultrasonic — measures internal defect/thickness with a sound wave),
- Dye penetrant (liquid penetrant — makes surface cracks visible with colored liquid),
- Magnetic particle (magnetic particle — ferromagnetic surface cracks),
- Radiography (radiography — internal structure with X-ray).
NDT requires certified personnel (e.g. NAS 410 / EN 4179 level 1/2/3). This is a mandate that AI can never take over. AI can extract a mark in the image; The certified examiner decides whether that mark is a real defect (relevant indication) or a false/non-relevant indication.
Where is computer vision useful?
Computer vision is powerful in repetitive and large-surface tasks where the human eye gets tired during visual inspection:
- Large surface pre-scanning: Scanning fuselage, wing overlay, lightning damage by drone/camera; marking suspicious areas.
- Borescope image support: Inner engine wing damage, highlighting the suspicious frame in combustion chamber control.
- Change detection: Comparing the previous and current image of the same area and marking new damage.
- Measurement support: Estimated scaling of damage size (exact measurement is instrumentation).
- Recording/tagging: Assistance in documenting the location of the indications found.
This is a layer of pre-screening and attention routing. "Look here," he says; It doesn't say "this is a crack and it's off limits". That ruling is based on SRM/AMM limits and examiner interpretation.
Caution: The two-way error of computer vision is critical. False negative (missing the real flaw) is a direct security risk; Just because the system says "clear" does not end the examination. False positive (marking a non-existent defect) creates unnecessary disassembly and cost. Therefore, computer vision complements, not replaces, human examination.
Why does the decision remain with the person?
A dark line in an image could be a crack, a scratch, an oil mark, a paint defect, or a light reflection. Models are wrong in certain conditions (illumination, angle, surface condition). Furthermore, whether an indication is acceptable (within the limit, in which zone, again) is an engineering judgment and depends on the approved limit documents. AI cannot legally enforce the limit nor does it carry a certificate. So the flow is always this: AI marks → examiner verifies → decides on limit document → authorized signs.
Tip: Use computer vision output as a "second pair of eyes", not as a "mono-eye". The best results are in the double-layer inspection, where the machine catches what the human might miss and what the machine incorrectly marks.
three mini cases
Case 1 — Drone scanning diverted attention. During a fuselage upper surface inspection, drone images were scanned with computer vision and two suspicious areas were marked. The examiner took a detailed look at these two areas: one was a trace of oil (non-relevant), the other was the beginning of an actual lightning damage. Hours of eye scanning on a large surface area have been reduced to a targeted examination; The real damage was caught early.
Case 2 — False positive hunted. In a borescope analysis, the system flagged a "crack" in a turbine blade. When the certified examiner magnified it and changed the angle, he saw that it was a surface deposit and not a crack. An unnecessary engine disassembly (very high cost + risk of new errors) was prevented. The system marked it, the human read it correctly.
Case 3 — Discipline versus false negative. The system was "clear" on a wing undersurface scan, he said. The inspector still procedurally inspected the critical joint area by hand and with a magnifying glass and found a subtle beginning of corrosion. "He said the system is clear" did not finish the examination; discipline caught a missed fault.
Four copyable templates
Role: Visual inspection pre-screening assistant.Task: Mark suspicious points that need attention in the image/area below and give a possible explanation (crack? scratch? corrosion? mark?) for each.Rules:- Make a definitive judgment of fault; Use “doubtful, should be confirmed” language. - State that the acceptance limit is defined in the SRM/AMM and that the examiner will decide. Image/context: [description or visual]
Role: Change detection assistant. Task: Compare the PREVIOUS and PRESENT state of the same area and mark the newly formed/growing spots. Rules: Point out that the lighting/angle difference may be misleading; precise measurement and decision.Input: [previous/current definition]
Role: NDT method selection consultant. Task: Explain which NDT method(s) would be appropriate for the following possible defect type and location and why. Rules: State that application of the method requires certified personnel and that the acceptance criteria are in the relevant specification. Situation: [defect type + material + location]
Role: Indication recording/labeling assistant. Task: Translate the following inspection finding into a structured record with location, type, size (estimated), photo reference. Rules: Retain "estimated"; Leave the accept/reject clause blank, it belongs to the examiner. Finding: [raw note]
Weak prompt / Strong prompt
Weak: "Are there any cracks in this photo, can the plane fly?"
Requires both diagnosis and airworthiness decision with a single output; Both are certified inspector and limit document jobs.
Güçlü: "Mark suspicious points in this area image and give a possible explanation for each (crack/scratch/corrosion/mark). Do not make a final judgment, say 'must be verified'; state that the acceptance limit is in the SRM and the decision belongs to the examiner. Point out the lighting/angle misleading."
This prompt puts AI in the pre-screening role and leaves the decision to the human.
Table: Role distribution in visual/NDT examination
step
Computer vision / AI
Human (examiner)
Large surface scanning
Suspicious area signs
Targeted detailed examination
Indication comment
Suggests possible explanation
Relevant/non-relevant decision
Size
Estimated scale
Calibrated measurement
Accept/reject
— (can't)
Limit document + provision
signature/registration
draft help
Certified signature
Common mistakes
- Ending the examination with "The system said clear". False negative is a security risk.
- Mistaking the sign for fault. The distinction between relevant/non-relevant is up to the examiner.
- Using the estimated size as exact measurement. The acceptance decision is based on calibrated measurement.
- Interpreting NDT without certification. Authorization and acceptance criteria depend on the specification.
- Ignoring the lighting/angle illusion. It produces false positive/negative.
In summary
Visual inspection and NDT are safety-critical flaw hunting, and computer vision is a powerful pre-screening layer here that shares the eye's burden: scanning large surfaces, marking change, directing attention. But the sign is not the verdict; Relevant/non-relevant distinction, size measurement and acceptance/rejection decision belongs to the certified inspector and limit documents. The best results are in the double-layer examination, where the machine and the human cover each other's blindness.
Application task
Choose an examination scenario (photo or description). Have the AI mark suspicious spots with the first template. Then decide for yourself for each sign: relevant, non-relevant, which SRM/AMM limit should be checked? List the points that the AI missed or incorrectly marked and write a “double-layer inspection” evaluation.
checklist
- [ ] I used the computer vision output as a preliminary screening, not as a verdict.
- [ ] Even though it said "the system is clean", I manually checked the critical areas.
- [ ] I made the relevant/non-relevant decision as the examiner.
- [ ] I based the size/acceptance decision on the calibrated measurement and limit document.
- [ ] I observed the certificate authority for the NDT comment.
- [ ] I took lighting/angle-related distortions into account.