AI systems
Concept studyVision inspection that had to run on the edge
Visual inspection for a high-mix production line, designed around the physical constraints of the cell: latency, vibration, and no continuous network.
- Engagement
- Project-based delivery
- Duration
- Six months, feasibility through field deployment
- Year
- 2024
- Client
- Not applicable
The challenge
What made this difficult.
Manual inspection of finished assemblies was the last quality gate, and it was both slow and inconsistently calibrated between shifts. A cloud-hosted model was not viable: the network in the cell was unreliable, the latency budget was a few hundred milliseconds, and the cameras moved with the line. The system had to make its decision where the part was.
Constraints
Non-negotiables we designed around.
Hard latency budget
Decisions had to return within the takt time of the line, which ruled out round-tripping to a remote service per part.
Vibration and moving optics
Camera mounting and exposure had to be robust to a line that was never perfectly still, without over-constraining mechanical tolerances.
No continuous connectivity
The cell could lose the network for extended periods. Degraded operation had to be defined in advance, not improvised.
High mix, low volume per variant
Retraining per SKU was not viable, so the model needed to generalise across variation rather than memorise a catalogue.
Approach
How we would build it.
Established a baseline before optimising anything
The first deliverable was a measurement of current inspection performance — including inter-rater agreement between human inspectors. Without that, any later improvement claim would have been unfalsifiable.
Sized the model to the accelerator, not the other way round
Target hardware and its memory bandwidth set the model budget, and the architecture was selected to fit. This avoided the common failure of prototyping a model that cannot be deployed at the required rate.
Deployed the decision on the edge, kept the loop closed
Inference runs locally. Edge nodes hold enough state to continue evaluating when the network is down, buffer results, and reconcile on reconnect — so connectivity loss degrades reporting rather than inspection.
Treated the vision stack as part of the mechanical design
Lighting, mount rigidity, and trigger timing were specified alongside the model. Much of the achievable accuracy came from the optics, not the network.
Bounded the model's authority
The system flags; it does not reject. Confident and uncertain outcomes are separated, and uncertain cases route to a human with the underlying evidence attached.
Architecture
How the pieces fit together.
Edge inspection architecture. The decision is made at the cell; the network improves the system but is not required for it to inspect.
Capture
- Triggered cameras
- Controlled lighting
- Rigid mounts
- Frame integrity checks
Preprocess
- Geometric correction
- Normalisation
- Region of interest selection
- Quality gating
Inference
- Quantised model
- Edge accelerator
- Deterministic runtime
- Sub-budget timing checks
Decide
- Confidence scoring
- Uncertain-case routing
- Evidence capture
- Traceable output
Reconcile
- Offline buffering
- Store-and-forward
- Drift monitoring
- Retraining feedback
Stack
What it would run on.
Model
- Compact CNN
- Quantisation for edge inference
- Deterministic runtime
- Per-variant generalisation
Hardware
- Edge accelerator module
- Triggered industrial cameras
- Controlled lighting rig
Software
- C++ inference service
- Time-series storage
- Buffer and reconciliation layer
- Operator console
Expected outcomes
What success would look like.
Baseline
Measured first
Human inter-rater agreement was measured before any model work, so later comparisons had a defensible reference point.
Latency
Within takt
Model budget was derived from target hardware and the line's takt time, rather than tuned on a workstation and hoped to fit.
Offline behaviour
Defined
Connectivity loss degrades reporting and improvement signals, not the inspection itself — specified in advance, not improvised.
Authority
Flag, not reject
Uncertain outcomes route to a human with evidence attached, which kept the system's adoption compatible with existing practice.
What we learned
The conclusions we would carry forward.
Measuring the existing human baseline first made every subsequent claim falsifiable.
For edge deployment, the target hardware is an architecture input, not a deployment detail.
Much of the achievable accuracy came from lighting and mounting — the vision stack belongs in the mechanical design conversation.
Related