Pilot programme · 75 days · 1,800 hours
Seventeen failures. Every one called in advance.
Between 1 August and 15 October 2026 we ran a ten-valve fleet under continuous monitoring and let it degrade. Eighteen failures occurred. The CANARY prognostic engine warned on every one that had a physical precursor — a minimum of 9.1 days ahead, a median of 10.6 days, and in two cases more than three weeks. Below is the complete result, including the one failure nothing could have caught.
Open the recording
Inspect the pilot for yourself
The complete 75-day recording opens in your browser: all ten valves, every condition measurement at five-minute resolution, the WARNING and ALARM transitions, and one-second burst capture around each of the eighteen failures. Scrub the timeline, select a valve, read the signals the engine was reading.
Opens in a new tab · 6.8 MB · no sign-in
Best viewed on a desktop browser
The headline finding
The predictions were conservative, every single time
This matters more than the lead time itself. On all sixteen failures where the engine issued a time-to-failure estimate, the predicted failure date arrived before the actual one — never after. The engine erred early in 16 of 16 cases, by a median of 5.8 hours against lead times measured in days.
16 / 16
Predictions that fell earlier than the actual failure. Not one estimate overshot, which is the direction that would have left a plant exposed.
5.8 hours
The typical gap between predicted and actual failure, against warnings issued nine to twenty-two days out.
2.1%
Median timing error expressed against the warning lead. The worst case across the whole programme was 2.8%.
Why the direction of the error is the point. A prognostic system that predicts failure later than it happens is worse than useless — it tells a planner they have time they do not have. CANARY’s estimator is deliberately tuned to resolve uncertainty toward the earlier bound, so the date you schedule against sits before the date the valve actually reaches its limit. Across 75 days and eighteen failures, that bias held without exception.
Results by failure mode
Every mode the engine watches, tested
Seventeen failures across seven degradation modes, on valves from three manufacturers in new, used and remanufactured condition. Lead time is measured from the WARNING being raised to the valve reaching its failure limit.
| Failure mode | Events | Shortest | Median | Longest |
|---|---|---|---|---|
| Spool & sleeve wear | 3 | 9.8 | 11.2 | 19.4 |
| Silting (contamination) | 3 | 9.1 | 9.6 | 11.6 |
| Cavitation & aeration | 3 | 9.6 | 10.6 | 22.0 |
| Erosion & internal leakage | 2 | 11.9 | 11.9 | 11.9 |
| Solenoid coil degradation | 2 | 10.1 | 10.4 | 10.8 |
| Valve electronics | 2 | 10.4 | 10.4 | 10.4 |
| Wiring & supply | 2 | 9.9 | 10.9 | 11.9 |
| All modes | 17 | 9.1 | 10.6 | 22.0 |
No mode fell below nine days
The weakest result in the entire programme was 9.1 days on a silting failure. Every one of the seven monitored modes produced warnings with enough runway to reach a planned maintenance window.
Condition made no difference to detection
Eight failures occurred on used valves, seven on remanufactured units and three on new ones. All were detected. The engine learns each valve’s own baseline, so a tired valve is judged against itself rather than a factory specification.
What we got wrong
Two calls that did not match the strip-down
Publishing only the successes would be easy and worthless. Three of eighteen events did not go cleanly. Here is each one, and what the post-mortem found.
A sudden mechanical break
A stem coupling fractured through the threaded root. The engine gave 1.6 days’ notice and named the wrong mode.
Post-mortem. The fracture face showed beach marks over roughly 70% of the section — high-cycle fatigue propagating under a closed crack. Crack growth of that kind changes neither leakage nor friction, so hysteresis, null offset and silt lock all stayed flat. There was nothing for a condition monitor to see until the remaining ligament began to flex in the final hours.
What it means. This class of failure is not detectable from valve signals by any monitor. It needs periodic NDT of the coupling. We say so rather than counting it as a near miss.
Two faults degrading at once
The engine warned 9.8 days ahead and named valve electronics. The strip-down found spool and sleeve wear: scored lands and a centring spring 14% down on free length.
Post-mortem. The unit was also carrying an ageing position-sensor board. Both were degrading simultaneously, and the electronic indicators — ripple, standing current, handshake delay — crossed their limits earlier and harder than the mechanical ones. The board was still functional at removal.
What it means. The engine was not wrong about what it measured; it was wrong about which part would fail first. The valve was still pulled in time, which is what the warning is for.
The third case. An erosion failure on FCV-302 was detected 11.9 days ahead but carried no confidence grading, because pressure transmitters were fitted to that valve partway through the programme and the pressure-gain baseline was still being learned. It is counted as a detection and excluded from the confidence statistics.
The test fleet
Ten valves, three manufacturers, real service
The fleet was deliberately mixed. Valves from Bosch Rexroth, Eaton Vickers and Danfoss, in new, used and remanufactured condition, across duty cycles from fast anti-surge trim to slow batch ramping.
| Tag | Valve | Condition | Service |
|---|---|---|---|
| FCV-101 | Bosch Rexroth 4WRPEH 6 | New | Boiler feedwater control |
| FCV-102 | Bosch Rexroth 4WRPH 10 | Used | Steam turbine bypass |
| PCV-201 | Eaton Vickers KBSDG4V-3 | Remanufactured | Main header pressure |
| TCV-202 | Danfoss PVG 32 | Used | Cooling water temperature |
| LCV-301 | Bosch Rexroth 4WRLE 16 | Remanufactured | Condensate drum level |
| FCV-302 | Eaton Vickers KDG5V-7 | Used | Compressor anti-surge |
| PCV-401 | Danfoss PVEH 157B | New | Fuel gas letdown |
| TCV-402 | Eaton Vickers KFDG5V-8 | Remanufactured | Reactor jacket cooling |
| FCV-501 | Bosch Rexroth 4WRTE 10 | Used | Column reflux flow |
| LCV-502 | Danfoss PVG 100 | New | Surge drum level |
75 days
Continuous monitoring, 1 August to 15 October 2026. 1,800 hours per valve.
21,600
Five-minute condition records per valve, with one-second burst capture around every event.
2.05
Valve-years of monitored operation across the fleet.
How to read these numbers
What this does and does not establish
What it establishes
- The engine detects all seven monitored degradation modes on real valve signals, not only in simulation
- Lead times of 9 to 22 days are achievable across mixed manufacturers and valve conditions
- The time-to-failure estimator is biased conservative, and held that bias in 16 of 16 cases
- These results align with the figures our simulation programme produced, which is the correlation we were testing for
What it does not establish
- A false-alarm rate. Seventy-five days on ten valves is not long enough to characterise false positives per valve-year
- Performance across every valve family, duty cycle or plant environment
- Behaviour over a full seasonal temperature swing
- Anything about failures with no physical precursor, which remain undetectable
The programme continues. Field data collection is ongoing and these figures will be restated as the evidence base grows. If you would like the detailed methodology, the per-event records or the post-mortem reports, we will share them under NDA during evaluation — write to engineering@canarypvis.com.