AI · Real-time scrap analytics · pilot operation

AI watches scrap and alerts you in time

A traditional report tells you how much scrap you made yesterday. Instead, we analyze the data in real time: AI recognizes an anomaly - a sudden spike, a slow creep, or their combination - and alerts the right person before the impact grows. The goal is simple: shorten the reaction time and significantly reduce scrap.

Now in pilot operationRelease expected in summer 2026 · real-time detection · notifications · backtest on your own history

Tested on real production data

SSA
+ CUSUM · proven detection algorithms
98/100
trained model quality (test data)
summer 2026
expected release for customers
Problem

The summary number arrives too late

  • Scrap is measured, but it's only noticed once the damage is done.
  • The daily report shows the workstation had a bad day. The shift is already lost.
  • Slow growth is the most dangerous. For weeks nobody notices it, then comes a jump and the quarter is ruined.
  • Before the information reaches the right person, the problem keeps running.
  • Causes are argued over at the quality meeting. Nobody has hard data on exactly when it started.

The cost: a late reaction means whole shifts of scrap, the fix targets the symptom instead of the cause, and recurring incidents mean there's no timely alert. The key isn't more data, but timely information for the person who can intervene.

How it works

From data to a timely notification

AI continuously watches your scrap data and tells you itself when something is off. Five steps, and the last one makes the biggest difference.

1
Data into bins
From TPM&M / MES we continuously collect scrap into time bins at the workstation, workstation × product and station × product levels.
2
The model learns
A statistical model (SSA, CUSUM) is trained for each series. With no labeled data. It learns your normal behavior.
3
Real-time evaluation
Every new bin is compared with the learned behavior. AI computes the deviation, severity and statistical significance.
4
Notification
When the deviation is significant, an alert reaches the right person with context: where, when, how far above normal, how confident.
5
Intervene, less scrap
The foreman or process engineer reacts at once, not after the shift. The problem's duration shrinks and scrap drops.
Anomaly types

Three patterns the AI recognizes

Every dashboard sees sudden jumps. A slow creep it doesn't. AI watches both and their combination, and for drift it alerts hours to days earlier than it shows up in the daily report.

Sudden spike

Something just went wrong

A bad tool change, a shifted parameter, a different material batch. A sharp jump above normal.

An alert within minutes with info: where, when, how far above normal and how confident the detection is.
Slow creep

Slow drift

Mold wear, gradual calibration drift, seasonal influence. The eye misses it in reports.

An alert as soon as the trend is significant, typically hours to days BEFORE it appears in the daily report.
Combination

Several signals at once

For example a sudden jump on top of a slow creep. The most serious situation, to be handled first.

Higher severity and significance. The system tells it apart from normal fluctuation.
What you see

Four views of one module

The “Scrap” module has four tabs. From live incidents through history and backtest to a look into the model’s quality.

Active scrap

Live incidents right now, with severity, status and context. What you need at the morning meeting.

History

The scrap-rate and runtime over time, with incidents marked in the chosen range.

Backtest

Virtual incidents: what the model would have caught on your history, by type and duration. Export to CSV.

Detection status

Model quality (a score 0 to 100), training and a watch on whether detection is actually running.

TypeContextLevelDurationSeveritySignif.
Sudden Welding shopWorkstation2 h 15 min18.5%12
Slow creep AssemblyWorkstation7 h 30 min6.6%8
Combination Press shopWorkstation13 h 15 min28.2%14

An illustrative example. Each incident has a type, context, severity (how far above normal) and statistical significance.

Backtest

Try it on your own history

Before you put anything into operation, you let the model replay your history. You see virtual incidents: how many there'd be, of what type (sudden vs. slow) and how long they'd last. From that we tune the sensitivity so it neither floods you nor misses things. No black box.

What you would catch

An overview of virtual incidents on your history: count, the split into sudden and slow, average duration.

Sensitivity tuning

More alerts versus fewer false alarms. We set it for each series separately.

Export to CSV

You take the backtest results out and we go through them together, on your data.

Detection status

You see how the model is doing

Detection works unsupervised, so it has no pre-labeled “correct” answers. That’s why we openly show how good the model’s data is. Where data is scarce, we tell you the result is less reliable.

Quality score

0 to 100 for each series

The score blends data volume, day coverage, scrap density and history length. Rated from “Excellent” to “Insufficient”.

Training

Automatic and manual

The model is retrained regularly (monthly) and whenever you run it manually. Training is idempotent and can be repeated safely.

Run watch

It knows when data stops coming

If no new bins arrive, the system reports it (Detection not running). You won’t find out only when an incident is missing.

The point of the whole module

A notification where you can intervene

Detection on its own won't reduce scrap. Timely action will. That's why the notification is the core of the module: it goes to the person who'll do something about it, at a moment when it still matters.

To the right person

The foreman on shift, the workstation's process engineer. Not into a generic report nobody reads.

With context

Where, when, what type, how far above normal and how confident the detection. You immediately know what to address.

On a channel you watch

Mobile push (Android and iOS), SMS and email. The alert reaches you even when you're not at a screen.

Guardrails

What the AI does NOT do

Detection isn't a decision
The system flags an anomaly. Whether it's a real problem is judged by the foreman.
It doesn't replace your quality control
Your SPC thresholds (CL, UCL, LCL) stay. AI runs in parallel and catches what statistical control misses.
It's not magic
Detection is statistical (SSA, CUSUM), with clear parameters. Every alert has a severity and significance.
It doesn't claim the cause
It shows that something is off and where. The cause is interpreted by a process engineer, not the model.
It doesn't hide uncertainty
Where data is scarce, the model openly shows lower quality and a less reliable result.
It's not a tool against people
It detects a deviation in the process, not in an operator's performance. It's not a disciplinary tool.
Input data

What the module needs

Production data from the floor The number of pieces produced and scrap in real time, time-stamped. Source: iDomino TPM&M or MES.
Scrap context Workstation, station, product. The monitored series are built from this.
Master data The workstation → station hierarchy and products, for the correct split into series.
History for training Enough historical data to learn normal behavior (on the order of months).

Technical prerequisites: a deployed TPM&M platform with TPM Analytics and a MES with a continuous data feed (bins). Detection quality grows with the volume and length of history, so the model openly shows a quality score for each series. The module is now in pilot operation, with release expected in summer 2026.

FAQ

What we're asked most often

It's now running in pilot operation on real production data. We plan release for summer 2026. If you want to be among the first, get in touch and we'll arrange pilot testing.

The model learns without labeled data, but it needs enough history to recognize your normal behavior. Where data is scarce, it openly shows a lower quality score.

Yes. The model has built-in thresholds for a minimum number of non-zero bins, so it doesn't raise false alarms on series with negligible output.

Sensitivity is configurable per series. Before launch we run a backtest on your history and show how many alerts you'd get. We tune the balance together.

Complementarily. Your SPC thresholds stay. AI catches what statistical control misses: drift below the limit and combined signals.

Automatically at a regular interval (monthly) and whenever you run it manually. If data stops coming, the system warns that detection isn't running.

Want to be among the first?

We test the module on real production data. We'll show it to you live and arrange how to try it at your site before it ships.

Book a presentation