How are datalogger data analyzed?
Turn recorded measurements into defensible decisions
Start data logger analysis with a question, then check the records before calculating an answer. You might need to locate a pressure loss, investigate a temperature excursion or compare operating shifts. A data logger supplies the history; the analysis must establish what that history can actually tell you. The application determines which statistics and events matter.
Keep the original and identify each field
Retrieve the records using the supported software or interface. Export to CSV or another suitable table format when further processing is needed, and work on a copy of the original file. Retain channel names, units, device identification, timing settings and quality flags. Understanding how logging works helps distinguish a point reading from an interval average, maximum or accumulated total. Check whether an interval timestamp identifies its beginning or end.
Inspect a few imported rows before trusting a spreadsheet. Decimal separators, delimiters and date formats can change how a file is interpreted. Verify that numerical fields remain numerical and dates are read correctly. If the recording configuration changed during the deployment, mark the change rather than treating the entire file as one uniform series.
Check continuity and measurement quality
A missing reading is not a zero. Separate blank cells, error codes and valid zero measurements. Look for duplicate timestamps, unexpected gaps, flatlined channels and values outside the sensor's range. Power interruptions, full memory and wireless transfer problems can affect the available record differently. A transfer gap may still be recoverable from local storage. Consult maintenance records and assess backup power where continuity matters. Investigate suspicious points before excluding them: a spike may be the fault you were trying to capture.
Plot against measurement time
Use measurement timestamps on the horizontal axis and label each vertical axis with its quantity and unit. Temperature, relative humidity and pressure need clearly identified scales. Changing the axis range can make a small variation look dramatic. Display gaps explicitly: a line between two readings does not prove that intermediate values were measured.
Before comparing devices, check time zones and clock drift. For remotely collected records, arrival at a server may occur well after measurement. A live dashboard may also show a different level of detail from the stored historical data. Use the underlying records appropriate to the investigation.
A comfortable average can hide an excursion
Consider these four point measurements, taken one minute apart. For this illustrative example, investigate readings above 25 °C. That threshold is an example, not a specification for any particular product or process.
| Time | Temperature | Above 25 °C? |
|---|---|---|
| 10:00 | 20 °C | No |
| 10:01 | 20 °C | No |
| 10:02 | 32 °C | Yes |
| 10:03 | 20 °C | No |
The mean is 23 °C, median 20 °C, minimum 20 °C and maximum 32 °C. The mean sits below the threshold while one reading exceeds it. One out of four samples, or 25%, is above the limit. This does not establish that the temperature exceeded the limit for 25% of the elapsed time, or for exactly one minute. Conditions between point readings are unknown. Any duration estimate needs a stated interpolation or holding assumption and an explanation of its limitations.
Choose statistics that match the decision
Minimum and maximum describe observed extremes; the median describes a typical level; standard deviation measures spread around the mean. These quantities do not establish sensor accuracy. An arithmetic mean is useful for equally spaced valid samples. Irregular records require duration weighting and an assumption about behaviour between readings to estimate a time average. If only interval averages were saved, the original brief peaks cannot be reconstructed.
Totals need the correct quantity, elapsed time and units. A constant load of 2 kW for 30 minutes uses 1 kWh. For changing power, calculate energy over the relevant time intervals. Adding kW readings alone does not calculate kWh.
Separate smoothing from event investigation
A moving average can reveal a trend while reducing a short peak and delaying the apparent response. Preserve the unfiltered series and document the filter and window length. The measurement input matters too: Pt100 probes, thermocouples and their installation can respond differently. The RTD and thermocouple comparison helps explain why similar-looking temperature channels may not track an event identically.
Align logger timestamps with SCADA event records, machine states or measurements read through Modbus. A door opening, load change or shutdown can provide useful context. Two signals moving together do not, on their own, establish cause and effect.
Make the result reproducible
Report the question, date range, valid record count, gaps, threshold, calculation method and finding. Include the relevant plot and event times, and retain sensor details and configuration information. Acceptance criteria differ across industries. When estimating system costs, include analysis software, licensing and reporting effort. Vendor software or Excel may suit a small investigation; Python, R or a suitable data platform can support larger repeatable workflows.
Technical references: Onset — plotting, analysis and export · Campbell Scientific — record fields and timestamps · Campbell Scientific — missing records
