Research & insights

Notes from the laboratory.

What we measured, what broke, and what it changed in the product. Written by the engineers who did the work, with the figures from the deploy reports.

Featured · Grinding

A sharpness index for grinding wheels from acoustic emission and spindle current

The first production model on the platform combines acoustic emission and spindle current into one indicator of wheel condition, and advises when the wheel needs dressing.

26 August 20265 min read

Grinding

Gating a grinding cycle: why raw channels cannot gate and windows must

Compared against a level, a raw current waveform crosses it on every cycle. A reduced window crosses it where the cut begins and ends. That difference is the whole design.

24 August 20263 min read

CNC Monitoring

Connecting a Haas TL-1 over MTConnect in one day

The fleet's first Haas Automation control went live in a day — with robust MTConnect streaming and the same screens as every other machine.

21 August 20263 min read

Research

Why we validate on real machines: the laboratory fleet in Tuttlingen

Simulated data does not have drifting controller clocks, imperfect power or protocol quirks. Real machines do, and that is why we build on them.

20 August 20264 min read

Sensor Technology

Acoustic emission in grinding: why acquisition decides what you can see

Acoustic emission is the most direct process signal in grinding and the most demanding to acquire. We instrumented a tool grinder with a high-speed digitiser to capture it end to end.

18 August 20264 min read

Tool Monitoring

Per-cut features: what spindle current tells you about the tool, and what it doesn't

Current is cheap, robust and honest — and it is a proxy. Here is how to turn it into features you can trend, and where its limits are.

15 August 20266 min read

Manufacturing Data

n/c is not zero: why a monitoring system must refuse to invent numbers

An OEE of 0 % and an OEE that cannot be computed look identical on most dashboards. We made them different on purpose, and it costs us screenshots.

14 August 20266 min read

Industrial AI

Shadow before active: how a model earns the right to be believed

Every new binding on the platform computes silently first. Promotion is a server-enforced state with an observation period, and rollback is always permitted.

12 August 20265 min read

Industrial AI

Three axes of a signal type, and the wrong number they prevent

Wiring a kilohertz block stream into a per-window step is not an error at runtime. It is a plausible-looking wrong number. The type system refuses it at authoring time.

10 August 20266 min read

Manufacturing Data

Complete or clearly marked: why a recording must never have a silent gap

A recording is evidence. If it can have a hole nobody sees, every analysis built on it inherits the hole. Here is the rule our recording engine is built around.

7 August 20264 min read

Edge Computing

Evaluating triggers at the edge, on the data's own clock

A trigger that needs a server round-trip misses the spike it was armed for. We moved evaluation onto the EdgeBox and timestamped decisions from the acquisition clock.

6 August 20264 min read

CNC Monitoring

Discovering over ten thousand nodes on a Sinumerik 840D sl without interrupting polling

Nobody should hand-type node IDs from a manual. Here is how the EdgeBox walks a controller's full address space while live acquisition keeps running.

5 August 20265 min read

Manufacturing Data

Tiered time series that never average the peak away

A year-long chart should still be able to find a forty-millisecond transient. Automatic tier selection with true min/max envelopes is how.

26 July 20265 min read

Next step

Have a machine that raises a question?

Most of these notes started as a question on a real machine. Bring yours.

Talk to Toleron See the platform