Rewriting a neutron flux channel that had drifted for eleven years

The channel was in tolerance. It had been in tolerance at every single calibration for eleven years. That was the problem.

We were asked to look at a research reactor's ex-core neutron flux channel because the operators had lost confidence in it. Nothing had failed. No surveillance test had been missed. But the calibration factor had been adjusted at almost every outage, always by a small amount, always in the same direction, and nobody on the current staff could explain why.

The instrument had originally been commissioned in 2015. By the time we arrived the applied correction was 14% away from where it had started. Each individual step had been small enough to look like normal maintenance. Added together they were a trend that nobody had ever plotted, because each calibration record lived in the outage it belonged to.

The first thing we asked for was not a drawing. It was every calibration sheet since commissioning, in one place, in date order.

What the records actually showed

Eleven years of as-found and as-left readings, plotted together, produced a line that was almost perfectly straight. That immediately ruled out most of what people had assumed. Detector burn-up is not linear over that period. Neither is electronics ageing, which tends to be flat and then not. A straight line over eleven years suggests something environmental and slow.

It also told us something less comfortable. Because each recalibration had been performed against the previous reading rather than against an independent reference, the chain had been walking. Every adjustment was correct relative to the last one and wrong relative to physics. There was no point in the record we could treat as ground truth.

We could not calibrate our way out of this. There was no reading in eleven years of records that we could point at and call correct.Jen Barnes · Radiation Measurement Lead

Rebuilding from the geometry

The only reference not affected by the drift was the reactor itself. We modelled the expected flux at the detector position from core geometry and power history, using Monte Carlo transport, and predicted what the detector should read at a defined power level. That gave us a number derived from physics rather than from the instrument's own history.

The modelled value and the measured value disagreed by 13.6%. Close enough to the accumulated correction that the mechanism was clearly real, and close enough to confirm the model was sound.

Then we started removing candidates. The detector was pulled and tested against a calibrated source: within specification, burn-up consistent with its exposure history. The preamplifier was bench-tested across temperature: a small reference shift, entirely as its datasheet predicted, accounting for about 1.5% of the total. Real, but not the answer.

It was the cable penetration

The remaining candidate was the part nobody had touched: 22 metres of mineral-insulated cable and, in particular, the penetration where it passed through the biological shield.

Insulation resistance measured 40 megohms. When commissioned it had been specified at greater than 10 gigohms. For a high-impedance charge signal, that is not a marginal degradation — it is a parallel leakage path that scales directly with the signal you are trying to measure.

The cause was moisture ingress at a seal that had been thermally cycling for eleven years. It had not failed. It had simply become slightly worse every year, in a way that was invisible to any test anyone was performing, and that produced exactly the linear drift the calibration record described.

Insulation resistance had never been trended. It was measured at commissioning, recorded, and never looked at again.

What we changed

Replacing the penetration and the affected cable run brought the measured response to within 1.8% of the modelled value, with no correction factor applied at all. The channel now reads what the physics says it should.

The more useful change was to the programme around it:

• Insulation resistance is now trended continuously, not measured once at commissioning. A falling trend raises maintenance action years before it reaches the measurement.
• Calibration is referenced to the modelled flux profile, not to the previous reading. The chain can no longer walk.
• As-found readings are plotted across outages, on one chart, for the life of the installation. A drift that is invisible in any single record is obvious across eleven of them.
• The proof-test interval was re-derived from the safety integrity calculation rather than inherited from the previous supplier's convention.

The uncomfortable part

Every person who worked on that channel over eleven years did their job correctly. Every calibration was performed to procedure, documented and signed. The procedure was the problem: it defined correctness relative to the instrument's own past, and it had no mechanism that could ever detect a slow common-mode drift.

That is not unusual, and it is not limited to nuclear instrumentation. Any measurement chain calibrated against its own history will eventually tell you what it told you last time. The defence is an independent reference — and the discipline to plot the long line even when every individual point is inside the band.