Skip to main content
Not all details of the prototype are public reductions were necessary. We also withdrew any conclusions that apply to the specific product and manufacturing sites keeping the presentation generally around anomaly detection.

​
Summary

Having a comprehensive solution for controlling the quality of newly introduced manufactured products is essential for quality control programs. This prototype system allows the acquisition, quantification and visualization of the quality of the seams made during the manufacturing process. The system uses an industrial-grade camera to produce an image of the seam’s quality using backlighting of the manufactured product. The image is stored or processed by an embedded compute node that runs a series of algorithms to detect anomalies in the seam quality. This documentation outlines the development of algorithms and the necessary automation required for deploying such algorithms in the factory.

​
Goals

This effort has the following goals:
  1. To develop one or more datasets suitable for model development.
  2. To develop one or more models for detecting seam quality anomalies and alerting the field personnel when consecutive anomalies happen.
    • The detection of a seam quality anomaly condition must be done ideally per product, in real-time, and with performance requirements determined jointly as part of acceptance testing.
    • The visual / audio anomaly events will be published in a stream processing platform suitable for real time reporting and analytics.
  3. To design the production environment, possibly spanning hyperedge, edge and cloud environments, that will allow the real-time operation of the solution, its horizontal scaling across factory sites as well as support quality monitoring processes by offering introspection into the various deployed pipelines and corresponding analytic facilities.

​
Results

The main result after trying supervised, unsupervised and self-supervised models shows an unsupervised pretrained model using as training input both nominal and abnormal images has achieved a test AU-ROC of approximately 0.97 out of 1.0 (theoretical maximum) when paired with UMAP dimensionality reduction and a kNN algorithm that uses majority voting for the determination of the anomalous images.

Software Architecture

System architecture and component overview

Data Pipeline

Data collection, processing, and the Seamagine dataset

Unsupervised Models

Anomaly detection using pretrained CNN embeddings

Inference Pipeline

Real-time inference and edge deployment
Edit this page on GitHub or file an issue.
Connect these docs to Claude, VSCode, and more via MCP for real-time answers.