Why is this asset failing, and what do I test next?
AssetBlue gives reliability engineers a defensible diagnosis from messy field evidence — photos, HMI screens, alarms, readings, logs, manuals, and voice notes — reasoning through failure physics and operating envelopes to the next best test, with the evidence shown.
Assets rarely fail cleanly. They degrade quietly first.
Energy creeps up. Throughput drops. Alarms repeat. Temperatures drift. Vibration rises. Efficiency falls. The expensive problems hide in the grey zone between healthy and failed, where an asset still runs but quietly degrades. Most industrial systems record what happened; few explain why. AssetBlue diagnoses soft degradation and efficiency loss before it becomes a hard failure.
Cavitation, impeller wear, suction restriction, air entrainment, fouling, recirculation, or operating off its best-efficiency point.
Scaling, soot deposition, excess air, combustion instability, blowdown losses, economizer fouling, and degraded heat transfer.
Misalignment, imbalance, bearing wear, cooling issues, harmonics, insulation degradation, and poor load matching.
Oil contamination, moisture ingress, insulation ageing, thermal stress, cooling degradation, and partial discharge.
Start without a data project. Deepen with every connection.
AssetBlue's asset graphs can begin from sparse field evidence (a photo, HMI screenshot, alarm list, voice note, or single reading) and improve as manuals, CMMS history, historian data, and prior cases are connected.
Its own reasoning brain
A pre-trained diagnostic model. It does not need your historian, your data lake, or a training project to start reasoning.
Point, shoot, and answer
Start from a photo, a few words, or a single reading. AssetBlue asks the questions it needs and reasons from there.
No new sensors
It works from the evidence you already collect. Nothing to mount, wire, or commission on the asset.
Runs from your HMI
Use it right at the panel where your engineers already stand. No extra console to roll out.
AssetBlue does not end with a report. Each investigation becomes part of the asset's memory: symptoms, evidence, failed hypotheses, confirmed mechanisms, corrective actions, recurrence patterns, and source-backed reasoning. The next engineer starts from accumulated plant intelligence, not from a blank form.
From a messy symptom to an audit-ready root cause: one inspectable causal chain.
A neurosymbolic causal chain fuses failure physics and your engineering knowledge graph with live field evidence, so you can see, and challenge, the model's reasoning at every node.
Symptom
Describe it in plain language, photo, voice, or sensor export.
Hypotheses
Competing causes, ranked by posterior confidence.
Evidence for & against
Each claim weighed, with the source attached.
Next best test
The one test that best discriminates the cause.
Human validation
No work order without a named engineer's sign-off.
Audit-ready RCA
Traceable report, work order, and asset memory.
Dashboards show what changed. AssetBlue reasons why it changed, with artifacts you already trust.
It turns symptoms into competing failure hypotheses, shows the evidence for and against each one, recommends the next best discriminating test, and converts validated findings into an audit-ready RCA.
Gate logic, condition by condition
The leading mechanism's AND/OR gate, with each condition marked confirmed, unconfirmed, or refuted from the evidence.
Cause-and-effect, every branch
The 6M view: candidate-cause families fanned off the spine, narrowed to the confirmed mechanical path.
Reading against the limits
Live sensor value plotted against the normal band and threshold lines, so a number always carries its context.
For & against
Every signal, weighed and cited.
Risk, scored
Likelihood × severity, with detection.
Cited, not implied
OEM manuals, standards, prior cases.
What it ruled out
And exactly why, kept on the record.
Built for the cases vanilla AI misses.
Industrial RCA rarely starts with complete evidence — a short incident note, partial readings, conflicting symptoms, and one urgent question: what is failing, why, and what should we test next? AssetBlue wraps the model in a structured diagnostic graph — components, symptoms, diagnostics, operating envelopes, maintenance actions, failure mechanisms, and causal families — and the lift grows with difficulty.
| Benchmark | Failure mechanism | Root cause | Reasoning | Overall |
|---|---|---|---|---|
| SYNTH-V7 · 40-case Sonnet | +6.5% | +9.7% | +5.9% | +7.3% |
| SYNTH-V7 · medium cases | — | — | — | +14.1% |
| SYNTH-V7 · hard cases | — | — | — | +40.7% |
| Qwen3.5-9B · full evidence | +17.9% | 0.0% | +8.6% | +9.0% |
Graph traversal lifts the bare Sonnet solver from 84.0 to 90.2 / 100. The lift rises with difficulty — easy +2.3%, medium +14.1%, hard +40.7% — where vanilla reasoning drifts to the wrong mechanism. The strongest contribution is diagnostic structure: the right failure-mechanism family, a stronger root-cause chain, and a more coherent evidence-to-conclusion path.
| Rare mechanism | Vanilla LLM | AssetBlue KB | Lift |
|---|---|---|---|
| Dealloying | 30 / 90 | 65 / 90 | +116.7% |
| Corrosion fatigue cracking | 48 / 90 | 75 / 90 | +56.3% |
| Brittle fracture | 20 / 90 | 42 / 90 | +110.0% |
| Dissimilar metal weld cracking | 60 / 90 | 77 / 90 | +28.3% |
RARE-V6 stresses rare mechanisms from sparse summaries. The same graph also runs on Qwen3.5-9B, small enough to deploy at the edge, lifting it to 80.7 / 100 so the reasoning holds on-prem.
Built for field conditions: high-contrast dark mode, rugged-tablet layouts, HMI screenshots, offline capture, and evidence review in low-light environments.