Find the cause in the data you already record.
Hypothesis-driven root cause analysis and predictive maintenance on operational and sensor data, with HGE, the Hypothesis Generation Engine.
The operational problem
Root cause is usually established after the fact, manually, by the people who already know the plant best. It is slow, it depends on who is on shift, and it rarely finds a cause nobody suspected beforehand. Off-the-shelf predictive maintenance models drift with the process and seldom explain why. The information needed is almost always already in the historian. It has just never been asked the right questions.
Method
HGE does not fit one model to your data. It runs a procedure over a large field of competing explanations.
- 1Generate. The engine proposes many falsifiable candidate causes for an event or a degradation, derived from structure in the data rather than from prior assumptions.
- 2Test. Each candidate is tested against your actual measurements. A cause that does not hold, falls. What stands is quantified.
- 3Rank. You receive a reproducible, ranked result, each cause traceable back to the raw data. The same question gives the same answer, every time.
Output
The result is a ranked table, not a needle on a dashboard. Table 1 shows the output format from a representative run.
| Rank | Candidate cause | Test against data | Score | Retained |
|---|---|---|---|---|
| 1 | Feed moisture on line B exceeds spec on night shift | moisture vs. yield loss | 0.91 | yes |
| 2 | Roaster bed temperature couples to intake air | lag correlation vs. throughput | 0.74 | yes |
| 3 | Acid plant conversion dips after clean-in-place | event-aligned comparison | 0.63 | yes |
| 4 | Conveyor vibration precedes trips by ~40 min | precursor test on trip log | 0.57 | yes |
| 5 | Reagent dosing drift vs. calibration date | drift vs. maintenance record | 0.19 | no |
Worked example
The hardest version of this problem is astronomy: finding a handful of real candidates in 1.8 billion observed objects, where a single wrong threshold drowns you in false positives. Run against ESA's Gaia DR3 catalogue, HGE surfaced and ranked 220,656 candidates, independently evaluable in an ESA context. Finding the one real cause of a process deviation among millions of tag values is the same class of problem, at smaller scale.
HGE reads from your historian and control system and sits beside them. It replaces nothing, and we commit to method and track record, never to estimated numbers before we have seen your data.