New inspection systems adapt in hours instead of weeks, leveraging synthetic training data and real-world validation to finally make automated inspection practical for short-run, custom production.

Configuration correctness checking — demonstration sample

For the past 7 years, our team at Spiral Technology has worked on quality control solutions across industrial manufacturing — from aerospace composites and wind turbine blades to precision machine components and non-destructive testing. We’ve seen firsthand how OEMs navigate the tradeoffs between accuracy, efficiency, and the organizational challenges of digitizing legacy processes.

In the first two articles of this series, we discussed the challenges of quality control in high-mix, low-volume production — where greater SKU variety and smaller batch sizes add complexity — and we explored how modern computer vision can be deployed on edge devices to address these issues. In this article, we take a closer look at how the technology actually works in practice.

AVIS Workflow Overview

AVIS, the Automated Visual Inspection System, uses a cascade of computer vision models that work in sequence to validate assemblies. The process begins with part identification, where individual components are detected and classified. These results are then compared against the expected configuration to confirm presence, position, and orientation. Next, surface regions are segmented — isolating areas that should be free of subcomponents — and analyzed for anomalies such as defects or foreign objects. If markings are present, optical character recognition (OCR) validates their accuracy. Finally, all results are combined in a comprehensive validation stage, producing a clear, auditable pass/fail decision.

Press enter or click to view image in full size

AVIS modules / inspection steps

AVIS Configuration

To keep pace with rapid production changeovers, training pipelines had to be accelerated dramatically. In manufacturing, lack of clean datasets — particularly for small batches and rare defects — is a well-known challenge.

Our approach begins with automated image generation for component identification. At first glance this might seem limiting, but the sheer volume of synthetic data that can be generated in essentially no time more than compensates. Importantly, a complete CAD model is not required. This means proprietary design data stays within the OEM organization, while inspection models are still trained effectively.

Press enter or click to view image in full size

Real vs synthetic part

Correct and incorrect assembly configurations can be derived from just a few minutes of video footage, or they can be synthesized in the same pipeline. By augmenting data this way, the system learns both what a correct assembly looks like and how to flag deviations in real time.

Learning incorrect configurations, demonstration sample

Press enter or click to view image in full size

Synthetic data generation for correct and incorrect configurations

Surface defect detection combines two strategies. For known defects — scratches, cracks, FOD, pinholes — supervised detection and segmentation models are used. For unknown or novel anomalies, autoencoders are applied. These models learn to reconstruct the “normal” appearance of a surface; when reconstruction fails, the residual differences highlight anomalies. The result can be visualized in heatmaps, allowing both automated detection and human interpretability.

Autoencoder heatmap

This layered approach ensures both the expected defects and the unexpected irregularities are captured, even in complex assemblies where variability is high.

Technical Parameters

Engineers evaluating such systems expect concrete specifications, not just promises. Performance metrics provide a clear picture of what’s feasible.

Accuracy: Positional tolerance within ±10 mm for component presence and alignment checks; reliable recognition of text characters down to 3 mm in height.
Scale: Effective for features as small as 5 mm and assemblies as large as 5 meters.
Setup Speed: New part families configured in under 24 hours; new variations ready in as little as an hour.
Hardware: Operates on smartphones and standard HD cameras; no need for dedicated scanners.
Connectivity: Runs without an internet connection; logs inspections locally and syncs when possible.
Compliance: Exports audit-ready records, supporting ISO 9001 and AS9100 frameworks.

These numbers reflect a balance between precision and practicality: good enough to catch what matters in production, fast enough to keep pace with HMLV changeovers.

Input Requirements

CAD model of the assembly to be inspected
Go-around video of the assembly if CAD is missing
Articulated acceptance criteria
Lighting of ~800 lux on the shop floor

Value Proposition

There is far more to the underlying technology than we can cover here, and many details are simplified for clarity. What matters most is the value proposition: the democratization of advanced computer vision for the modern factory.

For small and medium-sized manufacturers — often tier-2 or tier-3 suppliers to OEMs like Airbus, Boeing, or BAE Systems — launching an in-house AI program is unrealistic. Their assemblies are of medium complexity and their production runs too small to justify traditional, high-cost machine vision systems. AVIS addresses this gap as a plug-and-play solution: mobile, adaptable, and practical. By lowering the entry barrier, it makes inspection technology that was once reserved for large enterprises accessible to the broader manufacturing ecosystem.

Learn more about AVIS — portable computer vision for high-mix, low-volume manufacturing. Verify component configuration, markings, and surface condition in real time using cameras your teams already own.