Industrial dataset creation
The asset that outlives every model you will train on it — and the one most projects never build properly.
Models are replaced every year or two. The dataset outlives all of them
Models are replaced every year or two. A well-built dataset is still useful in five years, and it is the part of an AI project that nobody can sell you off the shelf, because it only exists if someone measured your machines under your conditions and wrote down what the labels mean.
The failure mode is always the same and always expensive: recordings made during one good week, labels applied by one person with a definition in their head, no metadata about load or variant, no version history. Six months later nobody can say why two nominally identical files disagree, and the dataset quietly becomes unusable.
We build datasets the way a laboratory builds a reference measurement. The campaign covers the process window deliberately. The label specification is written with your quality engineers and tested by having two people apply it independently. Metadata, versions and splits are fixed so a result from last year can be reproduced.
What is included
Measurement planning across the process window; synchronised multi-modal acquisition; label specification written with your quality engineers; annotation QA and inter-annotator checks; metadata and versioning; documentation and handover. Optionally an internal benchmark with fixed splits and baselines.
A measurement plan that covers the process window
Variants, loads, shifts, operators, seasons — the variation that will exist in production, sampled on purpose rather than by luck.
A written label specification
Agreed with your quality engineers, with the boundary cases named and illustrated.
Annotation QA
Inter-annotator agreement measured, disagreements adjudicated, and the specification tightened where they cluster.
Versioning, metadata and handover
Fixed splits, documented provenance, and a package your team can train on without asking us anything.
Four stages, each with an output you can check
Label specification
Written first, before a single recording, because it determines what you need to measure.
Campaign
Synchronised multi-modal acquisition across the planned process window, with metadata captured at the time rather than reconstructed later.
Annotation and QA
Independent double-annotation on a sample, agreement measured, specification revised.
Benchmark and handover
Optional internal benchmark with fixed splits and baselines, so future work has something honest to beat.
Sectors that buy this most often
What clients ask before they start
How much data is enough?
Usually less than people fear, and it is almost never the binding constraint. Representativeness and label quality dominate volume for industrial sensing problems.
Can you work with data we have already collected?
Yes, and the first step is usually an audit of it: what is actually in the metadata, how consistent the labels are, and what is missing. That audit alone often changes the project plan.
Do you keep a copy?
Only for the duration of the project and only where you have agreed to it. We do not train across clients and we do not reuse your data.
Ask about Industrial dataset creation
Tell us the asset or process and what goes wrong. We answer every enquiry within 48 hours on working days, and the first call is free.
- NDAs signed before the first call if you prefer
- Anonymised or synthetic samples are fine to start
- Your data is never used to train models for anyone else
Services clients combine with this one
Ready to find out if it works on your machines?
Start with a free 30 to 45 minute discovery call. Bring the problem, the recordings you already have and your questions. We will tell you honestly whether sensing-based AI is the right tool and what the next step would cost in days, not months.