INDUSTRY & QUALITY
Automated defect classification: decide at the pace of the line
Automated defect classification with TIMi identifies and categorizes, during production, the type of defect detected on a part or product, to guide the decision quickly (scrap, rework, accept), and shows the measures behind each classification.
Defect categories
Each part gets a defect category or a conformity status.
Inline speed
At AGC, defects are classified in under 12 seconds.
Explained per part
Each call shows the measures behind it.
HOW IT WORKS
How automated defect classification works.
Three steps, in one engine, on a standard server with no cluster.
1. Prepare the data
Anatella gathers the measures coming from sensors and inline inspection systems, and the history of defects already classified manually, then structures them into variables.
2. Build the model
Modeler matches the measures of a part under inspection with the profiles of known defects, to produce a classification.
3. Classify and explain
Each inspected part gets a defect category or a conformity status, with the measures that motivate the classification.
Anatella
ModelerGETTING IT RIGHT
Choices that shape a classification model.
A classification is only useful if it arrives in time and can be checked.
Keep up with the line
AGC’s engine classifies laser-scanned defects into 12 categories in under 12 seconds, just in time for cutting. The latency requirement has to be defined with the production constraints.
Start from labeled history
A limited number of cases already classified can be enough if the available measures are discriminating enough. Modeler is designed to extract a usable signal from limited histories.
Keep a human check
A model trained on known categories will have more difficulty with a completely new defect. It can flag an out-of-norm measure as an anomaly to check manually. Human control stays on ambiguous cases and new types of defect.
EXPLAINABILITY
A class says what. The measures say why.
A classification alone says a part is defective and of which type, without detailing how the conclusion was reached. A white-box model shows the measures that motivated the classification for this part, so quality teams can quickly check the decision and adjust the process.
For quality teams
A classification can be audited measure by measure, which helps to trust an automated control running at high pace.
For continuous improvement
Understanding the recurring factors of a defect type helps to trace it back to its cause upstream of the quality control itself.
For model owners
Modeler’s built-in Drift Report shows when a model loses accuracy, so you update it when needed instead of on a fixed calendar.
DATA
What data a classification model uses.
Anatella connects to more than 80 sources, so this data can be gathered without prior custom development.
Inline measures
Dimensional, visual, thermal, depending on the context.
Classified defect history
Defects already labeled manually, as a learning base.
Production parameters
At the time of inspection.
Conformity thresholds
Already defined by the quality teams.
PROOF
Defect classification in production.
AGC is our reference for real-time defect classification.
Defects classified into 12 categories in under 12 seconds
More than 97% accuracy on laser-scanned defects, so each sheet is routed to the right customer and waste is reduced.
FAQ
Common questions.
Can it work at the pace of a production line?
Yes. At AGC, defects are classified in under 12 seconds, just in time for cutting. The latency depends on the deployment architecture and has to be validated against the production constraints.
How many historical defects do we need?
A limited number of already classified cases can be enough, provided the available measures are discriminating enough. Modeler is designed to extract a usable signal from limited histories.
Can it detect a new type of defect it has never seen?
A model trained on known categories will have more difficulty classifying a completely new defect. It can still flag an out-of-norm measure as an anomaly to check manually, next to the classification.
Does it replace human quality control?
It supports it, to absorb volume and pace. Human control is kept on ambiguous cases and on new types of defects detected as anomalies.
SEE FOR YOURSELF
Test it on your own data.
Run TIMi on your own data with a free trial, or talk to us about your case.
