Time-series data comes in two fundamentally different shapes, and most of the mistakes people make with it trace back to not noticing which shape they are holding. A regular series has a sample at every fixed tick of a clock; an irregular series has samples whenever something happened, at whatever uneven moments those were. They look the same in a table of timestamps and values, but they behave completely differently under a calculation, and treating an irregular series as if it were regular is how a straightforward average quietly comes out wrong.
Regular vs Irregular Series in one line: A regular time series has samples at fixed, evenly spaced intervals - every second, every minute - so each sample represents the same span of time. An irregular time series has samples at uneven, unpredictable moments, typically because data is recorded only when a value changes or an event occurs. The difference matters because a plain average assumes every sample counts equally, which is true only for regular data; averaging an irregular series naively over-weights the closely spaced samples and under-weights the values that held for long stretches, which is why irregular data must be weighted by time rather than simply averaged.
A regular series is the product of periodic sampling: a clock ticks at a fixed rate and a value is recorded on every tick, whether or not anything changed. Scan-based historians and fixed-rate loggers produce this shape, and it has a comforting property - because every sample covers the same slice of time, the samples are directly comparable and most textbook statistics apply without adjustment. Count the samples and you know the duration; add them up and divide by the count and you have a mean that means what you expect. The cost of this tidiness is storage and bandwidth, because a value that sits perfectly still still generates a sample on every tick, so a regular series records a great deal of redundancy.
An irregular series is the product of event-driven or exception-based recording: a value is stored only when it changes meaningfully, or only when some event fires, so samples land at whatever uneven moments those changes happened. Report-by-exception telemetry, change-driven historian compression, alarm and event logs, and manually entered lab results all produce irregular series. The shape is efficient - a still signal produces almost no data, and a busy one produces a lot exactly when it matters - but it sacrifices the even spacing that made the regular series so easy to reason about. Two consecutive samples might be seconds apart during a transient and hours apart during a quiet stretch, and that unevenness carries real meaning: the long gaps are periods where the value simply held.
The two shapes are not just cosmetic variants; they encode different information. In a regular series the timestamps are almost noise, because you could reconstruct them from the fixed interval, so the values carry everything. In an irregular series the timestamps are as important as the values, because the spacing tells you how long each value persisted. That single fact - that in irregular data the gaps mean something - is the root of why the two behave so differently under any calculation that touches time.
Take a valve-position signal recorded on exception. Suppose it sat at 20 percent for an hour, then someone made three quick adjustments within a single minute - 40, 60, 55 - and it held at 55 for the rest of the hour. The stored samples are 20, 40, 60, and 55. A naive average of those four numbers is about 44 percent, which is nonsense: the signal spent almost the whole two hours near 20 and 55 and only flickered through 40 and 60 for seconds. The plain mean gave the three rapid-fire samples the same weight as the two long-held ones, so it drastically over-represents a brief flurry and under-represents the values that actually dominated the window. This is the trap of averaging irregular data as if it were regular.
The correct answer weights each value by how long it was in effect - the time-weighted average - which computes the area under the held-value curve and divides by the total span. Under that method the hour at 20 and the hour at 55 dominate, exactly as they should, and the brief excursions barely register. This is not an optional refinement for irregular data; it is the only average that respects what the series actually recorded. The naive mean is correct only when the samples are evenly spaced, because then equal weighting and time weighting coincide. The instant the spacing becomes uneven, the two diverge, and the plain mean becomes wrong.
The same reasoning extends to sums, integrals, and any statistic that should account for duration. An irregular series must first be made time-aware - either by weighting each value by its interval or by resampling it onto a regular grid with an appropriate hold-or-interpolate rule - before ordinary statistics can be applied. Skipping that step is the single most common way irregular data produces confidently wrong numbers, precisely because the arithmetic runs without error and the result looks reasonable. The failure is silent, which is what makes recognizing the series shape up front so valuable.
SCADA systems generate both shapes at once, often for the same physical measurement. A fast scan-based loop inside a controller sees a regular series; the historian that stores the same tag with change-driven compression turns it into an irregular one; a report-by-exception link from a remote site delivers it as irregular by design to save bandwidth. Field operations lean heavily toward irregular data because sending a value only when it changes is what makes battery-powered, cellular, and radio-linked remote sites practical. The consequence is that most of the interesting historian data in a distributed operation is irregular, which means the averaging and totaling pitfalls are the normal case, not an exotic one.
A cloud SCADA platform such as Merobix ingests irregular, exception-based streams from many remote wellpads, tanks, and stations and has to compute correct summaries from them - daily averages, totals, utilizations - despite the uneven spacing. Doing that faithfully means treating the stored data as what it is: time-stamped values that persisted until the next one, to be summarized by time-weighted methods rather than plain means, and to be resampled onto a regular grid with the right interpolation when a calculation genuinely needs even spacing. The platform's job is to hide none of this complexity in the wrong direction - it should never let an irregular series be averaged as if it were regular and hand back a distorted number.
For the field-operations team the practical upshot is that summaries over remote, exception-based data can be trusted to reflect duration rather than sample count. A daily average pressure on a remote separator should be weighted so a long steady stretch dominates a brief spike, and a utilization figure should reflect how long equipment actually ran, not how many times its status happened to be logged. Recognizing that field telemetry is overwhelmingly irregular, and handling it with time-aware math, is what keeps centralized reporting honest across a fleet where almost nothing is sampled on a neat, even clock.
Look at the spacing between consecutive timestamps. If every gap is the same fixed interval - one second, one minute - the series is regular and was produced by periodic sampling. If the gaps vary, with samples clustering during activity and thinning out during quiet periods, the series is irregular and was produced by exception-based or event-driven recording. In an irregular series the uneven spacing is meaningful, because the long gaps represent periods where the value held.
A plain average weights every stored sample equally, which is only valid when the samples are evenly spaced. In an irregular series some values held for long stretches while others were captured during a brief flurry of changes, so equal weighting over-represents the closely spaced samples and under-represents the long-held values. The correct method is a time-weighted average, which weights each value by how long it was in effect, so long steady periods dominate the result as they should.
Yes, by resampling it onto a fixed grid, deriving a value at each regular instant using an appropriate rule - linear interpolation for continuous analog signals, or holding the last value for state tags. This regularizes the data so ordinary evenly spaced statistics can be applied, but the result is only as good as the interpolation, and a coarse grid can average away short excursions the original irregular data captured. Regularizing is useful for combining tags or for tools that expect even spacing, but time-weighted methods can often summarize the irregular series directly without it.
Merobix reads your field devices into a cloud SCADA - the real thing behind these terms, live in days from any browser.