top of page

Irregular Sensor Samples: When a Simple Average Answers the Wrong Question

Nov 3
4 min read

A dashboard's average load disagrees with an independent energy calculation. Both calculations may be arithmetically correct while answering different questions. Before adjusting the meter, check whether each observation represents one sample or a known duration.


Original conceptual illustration of differently sized measurement columns on a time rail beneath a balancing beam

Original conceptual illustration; not a live ProSight Software screen, specific product or customer installation.


Key takeaways


  • A mean of samples gives each retained observation equal weight.

  • Time weighting needs a justified interval model.

  • Unknown gaps must remain visible in the result.


1. Decide what average you need


The arithmetic mean of samples adds their values and divides by their count. That is useful when the question concerns those observations. A time-weighted mean instead gives values weights representing duration, where the measurement model justifies that interpretation.


This distinction matters when reporting intervals change, an event causes extra samples, or one sensor sends more messages than another. A quieter operating state can be under-represented in a sample mean if a changing state produces many more records. More messages do not necessarily mean more elapsed time.


2. Use a known-duration example


Suppose an isolated test load is known to remain at 2 kW for 50 minutes and then at 10 kW for 10 minutes. Taking the arithmetic mean of those two operating values gives 6 kW. Weighting by their known durations gives (2 × 50 + 10 × 10) / 60, or approximately 3.33 kW.


Over that complete hour, the hypothetical energy is approximately 3.33 kWh. Six kilowatts would incorrectly imply 6 kWh if treated as the hour's average power. The weighted calculation follows the general weighted-mean relationship: sum of value times weight, divided by sum of weights.


These are deliberately controlled, piecewise-constant conditions. Two instantaneous field readings taken 50 minutes apart do not prove that the first reading held throughout the intervening period. The assumption must come from the measurement and operating model, not from a convenient spreadsheet formula.


3. State how values represent intervals


A previous-value hold, straight-line interpolation and a device-reported interval average are different models. Identify which one applies before assigning time weights. For a held value, specify when its validity expires; an old reading should not represent an indefinitely long outage.


Clip intervals at the requested report boundaries. Decide how the first interval obtains its starting value and whether the final sample has any justified duration within the window. A timestamp alone is insufficient to settle those questions.


Keep instantaneous power distinct from interval energy and cumulative energy. Averaging energy totals as though they were power samples produces a different quantity. Confirm units and what the device actually measures before selecting an aggregation.


Known load of two kilowatts for fifty minutes and ten kilowatts for ten minutes has a time-weighted average of 3.33 kilowatts, compared with six kilowatts for the unweighted mean of the two values

Original worked-example diagram. Values are illustrative and assumptions are explained in the text; this is not a product specification or field result.


4. Make incomplete coverage explicit


Exclude invalid intervals according to a documented rule and report their duration. If only 40 minutes of an hour are valid, a mean over those 40 minutes is not automatically the mean of the full hour. Label the denominator and the coverage alongside the result.


Inserting zeros during an outage can bias the result downward; holding the last good reading can hide a real change. Neither is a neutral repair. If estimation is justified, preserve both the observed coverage and the estimated portion so readers can assess the result.


Do not describe duplicate messages as additional evidence of operating duration. First reconcile duplicates and out-of-order samples, then construct the intervals used in the calculation.


5. Build an independent acceptance test


Test a constant signal, a controlled two-state signal and a deliberate missing interval. Change the message frequency without changing the physical test signal. A duration-based result should behave according to its declared model rather than merely follow the number of retained rows.


Include an interval crossing the reporting boundary and an invalid-quality case. Save the raw observations, the assigned weights, excluded durations and independently calculated result. This is stronger evidence than matching two charts that may use the same aggregation behind the scenes.


  1. Identify whether the source is instantaneous, interval or cumulative data.

  2. Choose and document the interval representation.

  3. Assign only justified durations within the report window.

  4. Track invalid and unknown coverage separately.

  5. Compare against a controlled independent test calculation.


6. Document the result for its readers


An informative report names the quantity, units, window, aggregation method, valid duration and treatment of gaps. Keep enough detail in exports or commissioning records to reproduce the answer. Round the final presentation appropriately while retaining suitable calculation precision.


Confirm which aggregation and quality-handling options are available in your actual ProSight Software deployment. The method here is an engineering reporting requirement, not a claim that every deployment exposes a particular calculation control. For consequential decisions, the responsible application team should approve the measurement model and acceptance evidence.


Related guide and project resources




Technical references



Need help with your application?


Send a representative dataset, equipment details and the calculation or reporting rule you need to verify. Confirm the supported functions and integration requirements for your actual deployment.










Comments


prosight-software-logo (1).png
bottom of page