Thickness grid statistics and projection

A single thickness reading is a sample, not a measurement of a component. Surface roughness, probe placement, couplant thickness, local pitting and the operator's hand all move the number, and the spread between readings at one location is usually larger than the metal lost between two inspections. Taking a small grid — four quadrant points, or a 3 × 3 pattern over a corroded patch — and treating the set statistically is what turns readings into data.

The minimum is what most codes want, because pressure retention is governed by the thinnest section, and API 510 and 570 both work from minimum measured thickness at a condition monitoring location. The mean is more stable and is what a corrosion rate should be trended on where the loss is genuinely general, while the standard deviation and the range are the honest statement of how repeatable the location is. A grid whose spread is several tenths of a millimetre cannot resolve a 0.05 mm/yr corrosion rate over a five-year interval, no matter how many decimal places the gauge displays.

Projecting forward is a straight subtraction: multiply the corrosion rate by the years ahead and take it off the wall. Doing it from the minimum gives the conservative answer used for scheduling; doing it from the mean gives the general-thinning answer; and the mean less two standard deviations is a useful indicator of where the thinnest point of the grid is likely to sit if the pattern continues.

Keep the grid pattern and its reference marks identical between inspections. Readings taken from a re-drawn grid are a different sample, and a change of a few tenths between them is more likely to be relocation than corrosion.

n = number of readings entered
mean = Σ t / n
s = √( ( Σt² − n·mean² ) / (n − 1) )   (sample standard deviation)
projected minimum = t_min − CR × years
projected mean    = t_mean − CR × years
statistical lower bound = mean − 2 s