"We have plenty of data. Years of it." — the line heard most often at the kickoff of a manufacturing AI project, and the one most often overturned a few weeks later. Factories certainly overflow with data. Yet the moment you try to build a model, it turns out there is almost no data you can actually train on. This training gap between sensor data and training data is the most common reason factory AI fails — and, strangely, a problem that rarely gets named.
The Six Things That Create the Gap
1. No labels — Supervised learning needs ground truth, but 99.9% of field data is "normal." Even when failures and defects occurred, it is rare for that timing to be recorded in a way that links to the time series. Maintenance logs are handwritten, and their times do not line up with the time series.
2. No context — The data itself will not tell you which line, which piece of equipment, or what signal TAG_00042 is. Data accumulated without an asset model is ciphertext — to the model and to people alike.
3. No quality indicators — Sensors drift and communications drop out. If you cannot tell whether a value in that interval is a "measured zero" or a "dropped-connection zero," the model learns noise as pattern.
4. Time does not line up — Every piece of equipment has a different collection interval, and clocks drift apart. Fail to align events (alarms, stoppages) with the time series, and you produce training data with the causality inverted.
5. Operational context is lost — The same temperature curve is normal under product A's recipe and abnormal under product B's. A time series not linked to work orders, recipes, and shift information is only half the data.
6. The distribution shifts — Data distributions drift with the seasons, the raw materials, and equipment aging. That is why a model trained on last year's data goes silent this year, and why one line's model fails on the line next to it.
The Gap Is Closed at Collection, Not at Modeling
The crux of the problem is that all six are decided not when you build the model, but when you collect the data.
- Store quality codes together with the values — Record collection quality on every point, and "dropped zeros" become distinguishable from "measured zeros."
- Collect on top of an asset model — Give tags meaning through the ISA-95 hierarchy and an ontology, and the data carries context from the moment it is born.
- Events and time series on one foundation — Alarms, work orders, and time series must accumulate on the same time axis for causality to be preserved.
- Collection that loses nothing on disconnection — Store-and-forward buffering shrinks the missing intervals themselves.
- Labels accumulate through operations — Instead of a separate labeling project, every cycle of the anomaly detection, operator acknowledgment (Ack), and action-record workflow adds one more verified label.
- Start with baseline learning — Deliver value first with an unsupervised approach that learns each tag's normal behavior without labels, and accumulate labels on top of it.
Conclusion — AI Ready Is a Design, Not a State
"AI Ready data" is not a state you reach by cleaning things up later; it is the outcome of a design that the collection architecture produces from the very start. This is why PlantPulse® makes quality-code storage, an ontology-based asset model, unified storage of events and time series, store-and-forward, and the Ack feedback loop part of the platform's fundamentals. A factory where sensor data is training data — that is the only way to eliminate the gap.
Topic reference: IIoT Sensor Data Training Gap — The Factory AI Problem Nobody Names (Medium · CreedTec)