Quality inspection that sees, localises and learns.
vKI standardises visual inspections, combines automatic detection with human review and turns images into usable quality information for decision-making, traceability and continuous improvement.
AI starts before the algorithm
Before a model can even be trained, the data foundation has to be right: camera, sensors, optics, lighting, part positioning and cycle are designed to produce reproducible images. For very small defects in particular, several cameras or sensors per part are sometimes required. The real barrier to entry for an industrial AI solution lies not only in model training, but in the know-how to generate stable, reproducible data from a real production process and to operate the entire chain reliably.
For the physical basis, we plan, build and install custom inspection tunnels and modular CamBoxes on site together with technology and integration partners: from a single camera station to a multi-cell inspection tunnel with continuous status and DMC display.
From camera image to inspection decision
Images are captured, analysed by the trained model and evaluated against defined thresholds. Detected features, defects or anomalies are visualised directly in the image and support the quality decision of the operator or inspector.
Two approaches, one goal: object detection and anomaly detection.
Depending on the inspection task, the appropriate method is used, often in combination.
Localise known defect patterns
Object detection models identify and localise defined features or defects in the image, including position, classification and confidence score. Suitable for completeness checks, variant control and the localisation of known defects.
Detect unknown deviations
The model learns the characteristic features of good parts (a dataset of good parts alone is often sufficient) and localises where, how large and in which area an irregularity deviates from the normal state.
Combined inspection systems
On a mounting plate, object detection first checks whether all bolts and components are present. Anomaly detection then identifies unknown damage, deformation or irregular surfaces. The result is a particularly robust inspection system.
Fast enough for your cycle time.
Inference times depend heavily on hardware, model complexity and image resolution.
Without acceleration, the large YOLO-NAS variant takes around 5 seconds on a CPU. All models are tested and optimised both for resource-efficient edge deployment and for high-performance hardware.
Reference cases with reliable figures
E-motor housings, fully automated 100% inspection: Up to 2,400 housings are inspected every day using 42 sensors. Defects from 0.2 mm are reliably detected in accordance with the defect catalogue, with an inspection time of no more than 36 seconds per part. Defective parts are automatically rejected, additionally checked by human review and packed at the same time.
Stator sleeve inspection at a production site in the Czech Republic: Inspection takes around one to five seconds per part, several thousand parts have already been inspected and manual inspection has been significantly reduced, with the human-in-the-loop principle applied throughout.
People stay in the process
vKI does not have to decide on every part fully automatically. Clear-cut results can be processed automatically, while conspicuous or uncertain cases are passed to an employee for review via vCHECK. This way, human experience is applied precisely where an automated assessment is not reliable enough.
When an employee confirms or corrects an AI result, this produces valuable training data for further developing the model in vKI Studio. Inspection and model maintenance thus become part of a continuous cycle.
What vKI changes in day-to-day operations
- Standardised, reproducible inspection results, regardless of how anyone is feeling on the day
- Relief for inspectors through automated pre-inspection
- Early localisation of defects instead of a simple OK/NOK result
- Seamless connection of inspection result, image and production context via vPRODIS
- Foundation for continuous model improvement via vKI Studio
Which inspection task is on your mind right now?
Together we look at your part, your defect patterns and your cycle time, and assess which method is the right fit.