Anomaly detection, with a quality reference
Learning normal variation is the starting point. A production decision needs a defined quality reference.
Automotive inspection must recognize meaningful defects while accepting the variation that belongs to a conforming product. Surface finish, material batch, component configuration and part presentation can all change the image without changing the quality decision.
Learn normality. Establish conformity.
Anomaly detection can help when defect examples are scarce or the complete defect catalogue is difficult to enumerate. Edge Impulse’s visual anomaly-detection tutorial demonstrates a workflow trained on normal images, with anomalous examples used separately for testing.
In a production application, the model’s deviation score is an input to the inspection method. Product specifications, approved samples and reference measurements establish whether a finding is relevant and how it should be handled.
Match the method to the check.
An assembly may require several methods: component presence, variant recognition, local geometry, surface inspection and a separate functional test. A useful architecture gives each requirement a suitable method and connects the results to the same product identity.
Evaluate both kinds of error.
Missed defects and unnecessary rejects have different production consequences. Evaluate them separately by defect category, using representative normal variation, known defects and agreed boundary cases. Development data and acceptance references have different roles.
Keep the operating conditions visible.
Fixture wear, illumination changes, new materials and product revisions can affect an inspection. Configuration records and change review help determine when the original evidence needs to be revisited.
A 2026 Lincode article republished by the Edge AI and Vision Alliance highlights lighting, training-data gaps, operator training and plant integration as practical obstacles. This vendor perspective reinforces the importance of evaluating the complete inspection workflow.
Our engineering approach.
Reliova connects anomaly-detection methods with controlled imaging, complementary checks, station response and a reviewable acceptance process. The objective is a useful inspection system that production and quality teams can operate and maintain.