Solution · Process Anomaly Detection
Detect change before it becomes scrap.
Anomaly detection built on clean values per cut — starting with simple comparisons against normal behaviour and moving to trained models only when the data supports it.
The problem
What you are up against.
- 01
Models trained on averaged data learn the average, not the anomaly.
- 02
A rating nobody can trace back to the raw data is a rating nobody acts on.
- 03
Rolling out a detector to every machine without a trial period is how false alarms are born.
The approach
Simple values first
From the full-detail sensor data, Toleron calculates values for each cut — taken only while the tool is cutting. Limits turn them into a first detector that anyone can explain, tried on real past data before it goes live.
Models, checked and observed
Trained models are checked when they are uploaded and can only be connected to machines whose data fits them. A new model first only observes; it is allowed to raise alarms only after an observation period, and you can switch back at any time.
Traceable by design
Because a rating is just another curve, you can lay it over the raw signal it came from, record it when it crosses a limit, and let it start a full recording of the conspicuous cut. The evaluation runs on the EdgeBox or the AI box in your plant: the raw data stays at the machine, only results are passed on.
What we use
- Values per cut from sensors and spindle current
- Limits you define
- Ratings from trained models
- Automatic recordings of conspicuous cuts
What you get
- First detectors built from simple limits that anyone can explain
- Trained models that observe first and go live deliberately
- Automatic full recordings of conspicuous cuts
- Ratings traceable back to the raw data
Tested on
- LabValues calculated on the EdgeBox from the fast current signal — no gaps, no duplicates
- LabEvaluation on the AI box — every result stored exactly once
Tested on the lab machines in Tuttlingen — proof that it works, not customer references.
Go deeper
Related
AI Models & Assistant
AI models that run at the machine and are tested before they go live — plus an assistant that answers from your own documents.
Learn moreAnalytics & OEE
Turns machine data into OEE, downtime and your own evaluations — set up once, reused on every machine.
Learn moreCut Detection
Evaluates only the time the tool is actually cutting — not the minutes in between.
Learn moreNext step
Bring process anomaly detection to your machines.
We start with one machine and the signals it already has. Tell us the controller and the question, and we will tell you what the first weeks look like.