AURA Technical White Paper
This is the AURA Model Technical White Paper. Aimed at technical reviewers (data scientists, equipment engineers, AI adoption evaluators), it explains how industrial AI models are built and how they are measured — through architecture, data, and scoring rules.
Download
| Language | Pages | File |
|---|---|---|
| Korean | 20 pages | PlantPulse AURA Technical White Paper (PDF) |
| English | 20 pages | PlantPulse AURA Technical White Paper (PDF) — same content as the Korean edition |
Revision 2026.09.
What's inside
| Chapter | Contents |
|---|---|
| 01 | Strengths — the difference architecture makes. A single GPU · no need to rebuild the index · distinguishing 0 from "not measured" · scoring without a judge model |
| 02 | Configuration — the three models in the language track and a separate lineage for viewing waveforms |
| 03 | Training and data — a single SFT stage, adapters, five branches that make answers come from rules |
| 04 | Recipes — teaching "failure mechanisms," not equipment. What is folded and what is kept in fine-tuning and quantization |
| 05 | Failure mechanism map — 11 failure types × 18 equipment types, with signal-wise variation range and degradation period |
| 06 | How to distinguish — a different order means a different failure. Tables of leading/lagging signals and distinction rules |
| 07 | How to select data — asking first "does the answer come from outside us?" |
| 08 | Data acquired — what is covered (by industry) |
| 09 | Simulator — why physics is fixed first and labels are derived |
| 10 | Saying "I don't know" — withholding judgment is also a correct answer |
| 11 | Source of answers — 30 calculation series · regulatory clauses · procedure documents · tool schemas |
| 12 | Training weighting — what and how much you feed in shapes the resulting capability |
| 13 | Identity and scope — through training, not prompts |
| 14 | Retrieval — for industrial RAG, retrieval quality is the ceiling on answer quality |
| 15 | Time series track — waveforms handled by small models, one per field |
| 16 | Evaluation — five benchmarks, all machine-scored without a judge model. Including how reviewers can verify for themselves |
| 17 | Field diagnostics — the composition of 158 questions and why false alarms are counted separately |
| 18 | On-premises deployment and the next release |
What changed in this edition
Expanded from 9 pages to 20 pages. Here is the gist of the eleven newly added chapters.
- Failure mechanism map (05·06) — signal-wise variation range and degradation period across 11 failure types × 18 equipment types, and "even the same two signals mean a different failure if the order differs." The rule for distinguishing four cases when temperature rises, and what the only mode in which power "drops" is
- How to select data (07) — "does the answer come from outside us?" · "lab or field?" · "a form with fewer built-in assumptions?" — and data with no causal link between sensor and label is not used, regardless of size
- Data acquired (08) — which public measured data fills which gap in coverage, by industry
- Simulator (09) — why labels are "derived" from physics (doing the reverse breaks causality), and numerical confirmation that causality was actually established
- Saying "I don't know" (10) — the criteria for treating withheld judgment as a correct answer, and why a third of the data are non-definitive answers
- Source of answers (11) — the 30 calculation series and seven holdouts, regulatory clauses, procedure documents fixed in advance, and tool schemas
- Training weighting (12) — if one branch dominates, the rest get diluted. Includes the measured finding that long prompts suppress tool calls
- Identity and scope (13) — how to make it hold up even without a prompt, or when overwritten by someone else's prompt
- Retrieval (14) — asymmetric encoding, why dimensionality is fixed by contract, and four ways evaluation leakage is prevented
- Time series track (15) — why million-parameter-scale models are a strength (on the Edge without a GPU, 10,000 models for 10,000 pieces of equipment) and four ways industrial signals differ from general time series
Why this edition has no performance figures
The trained model has just been rolled onto the product path. Figures will be included after being remeasured on that path — if we pin down numbers measured on the old path into the white paper, those numbers keep circulating as the reference even after later correction.
What is measured and how (question composition · scoring rules · data selection criteria) is already fixed, so reviewers can remeasure in their own environment using the same method. And gate criteria are set "before" measurement — deciding criteria after seeing results is not a criterion.
If we may recommend one reading order
You may read Chapter 16 (Evaluation) first. It covers what each of the five axes measures and how, how the scoring rules are mechanized, and how reviewers can verify for themselves — a way to directly check whether the trained version is actually being served, and whether training data has leaked into the evaluation.
See also
AURA Model Overview · Architecture · Data Lake Technical White Paper — the data infrastructure the model sits on