Reading a Trend Against Reading a Value

Why this matters

Two techs can hold the same instrument, take the same reading, and be doing two completely different jobs. One is asking "is this number acceptable," which is a comparison against something published. The other is asking "is this number moving," which is a comparison against the machine's own past. Those two questions place opposite demands on the instrument, and a shop that does not know which demand it is under buys the wrong meter, trusts the wrong digit, and either replaces good parts or misses a part that is on its way out.

The useful way to hold it: a trend forgives an instrument that is wrong in a consistent way, and a single value does not. A single value forgives a record with holes in it, and a trend does not. Pick the mode that matches the property your instrument and your records are weakest in.

The two modes fail on opposite properties

Accuracy, resolution and repeatability are three separate properties, and a sibling article works through the separation in detail. For this purpose the shorthand is enough: accuracy is how close the reading is to the true value, repeatability is how consistently the instrument returns the same reading on the same quantity, and resolution is the smallest difference the display can express.

A single value is a claim about where the quantity actually sits, so it lives or dies on accuracy. A trend is a claim about a difference, and a difference cancels any error that is the same in both readings. That is the whole mechanism, and everything below follows from it.

What a trend forgives, and what it will not

It forgives a stable offset. An instrument reading consistently high by a fixed amount produces readings that are all wrong by that amount, and differences between them that are exactly right. This is why a trend can be run on an instrument you know is off, provided you know it is off in a stable way and not in a wandering way.

It forgives an unknown absolute reference, for the same reason. You can trend a quantity whose true value you could not state, as long as you keep taking it the same way.

It does not forgive poor repeatability. Scatter does not cancel. An instrument that returns a spread on the same quantity puts that spread into every difference, and the change you are trying to detect has to be larger than the scatter before it is visible at all. This is the property people ignore when they buy on accuracy specifications alone.

It does not forgive a changed instrument, a changed point, or changed conditions. Swap the instrument between readings and the two offsets no longer cancel; the difference now contains the difference between two instruments plus whatever the machine did. Move the port and you are trending two quantities. Take one reading at a different operating state and you have manufactured an event.

It does not forgive a gap. A trend is a shape, and a missing interval is a piece of the shape you get to guess at. Two points four months apart cannot distinguish a steady climb from a step that happened in week two and has been flat since, and those two carry different actions.

What a single value forgives, and what it will not

It forgives a gappy history. A value read today against a published limit is complete on its own. This is the mode to use on equipment you have never seen before and will probably not see again.

It forgives a changed instrument, since nothing is being differenced. Any instrument accurate enough for the margin you are judging will do.

It does not forgive a bias. The whole comparison is against a true value, so an offset goes straight into the decision.

It does not forgive an unstated condition. A published limit was established under conditions, and a reading taken outside them is not comparable, whatever the instrument's accuracy. That is a separate failure from instrument error and it is not fixed by a better meter.

Picking the mode from your weaker property

Read the table as a shopping list and a filing list at once. If the shop's instruments are known-good but the record keeping is patchy, single-value work is what those instruments can honestly support. If the records are disciplined but the instruments are aging, trending is where they still earn their keep. Chasing accuracy on an instrument you use for trending, while letting the record fill with different instruments and different ports, is spending on the property you did not need.

Correcting one side means checking the other

Once you know an instrument's offset, the correction rule has a trap in it. A published limit is a true value, so an indicated reading gets its known offset removed before it is compared against one. But a comparison against your own history taken on that same biased instrument needs no correction on either side, because both carry the same offset and it cancels.

Correcting the reading and then comparing it against an uncorrected in-house baseline is the mistake, and it is subtle enough that it survives review, because every number in it is individually right. State in the record which side of the comparison has been corrected, in the same line as the number, or leave both uncorrected and say so.

Worked example: a gauge that reads high, and a limit that does not

A gauge on a service port is checked against the shop's reference and found to read 2 units high, repeatably. The gauge stays in service. Readings are taken from it at the same port, on the same installed gauge, at the same operating state, without breaking into the line at any point, on three visits: week 1 indicated 12, week 5 indicated 14, week 9 indicated 17. The manufacturer publishes a maximum of 18, stated as a true value.

The trend, uncorrected. From 12 to 17 over eight weeks is a rise of 5. The instrument's offset appears in both ends and cancels, so the corrected series - 10, 12, 15 - shows the identical rise of 5 over the identical eight weeks. Any conclusion about the rate is the same either way, which is the property the whole article rests on.

The rate is not one number. The first four weeks moved 2, which is 0.5 per week. The next four moved 3, which is 0.75 per week. Reporting a single average of 0.625 per week is true and it understates what is happening now. It is accelerating, and that belongs in the note.

The value, corrected. The comparison to the published 18 is a true-value comparison, so the indicated 17 becomes a corrected 15, and the headroom is 3, not 1. Projected at the recent rate of 0.75 per week, that is about 4 weeks. Projected at the eight-week average of 0.625, it is about 4.8 weeks. Both land inside the same service month, and both are a long way from what the uncorrected number implies: 1 unit of headroom at 0.75 per week reads as a bit over a week, and would have driven an emergency return that was not warranted.

Where a fabricated trend would have come from. Suppose the week 5 reading had actually been taken with a different, unbiased gauge on a different port, and logged as 14 without a note. Then that 14 is a true 14, and the series read as true values is 10, 14, 15: a jump of 4 in four weeks, then a near-flat 1 in the next four. A tech reading that would report an event around week 5 followed by stabilization, would go looking for what changed at week 5, and would find nothing, because nothing did. The instrument was fine, the reading was fine, and the trend was fiction, created entirely by swapping the instrument and the point.

The failure mode to catch. The fabricated version is more dangerous than the biased version, because the biased version was wrong by a known constant and the fabricated one is wrong in a way that looks like information. The defense is a field on the record, not a better gauge: which instrument, which port. A trend assembled from readings that do not carry those fields cannot be defended and should be read as a set of single values instead.

How to verify you are reading the mode you think you are

Take the series you are about to act on and try to break it in two ways.

Ask what would change if every reading were off by the same amount. If the conclusion is unchanged, you are genuinely reading a trend and the instrument's absolute accuracy is not on the critical path. If the conclusion moves, you are reading a value dressed up as a series, and the accuracy question is live.

Ask what would change if one reading were deleted. If the shape survives, you have a trend. If deleting one point turns a climb into a plateau, you have two points and an opinion. Two points establish a difference, not a trend, and calling it a trend in front of a customer is a promise about the future that the data does not carry.

References

• Manufacturer documentation for the instrument's accuracy, repeatability and resolution on the function and range in use, and for the conditions under which any published limit was established
• Trade-standard practice for condition monitoring and periodic maintenance readings
• See related: Accuracy, Resolution and Repeatability Are Three Different Things; What a Baseline Is Worth and When to Take It
• See related: Is It Getting Worse, the Same or Better: Reading the Trend