Separate a bad reading from a real event
An extreme value may be noise, a communication error or the process event you most need to understand. Keep an untouched copy of the input and document each filter. Review suspicious points against machine state and maintenance notes. Removing outliers automatically can remove the evidence of a developing fault.
Inspect time before value statistics
Check ordering, duplicate timestamps, missing intervals and clock resets. A mean calculated over irregular samples is not necessarily a time-weighted mean. Record the sampling method and distinguish event-based reporting from fixed intervals. If timestamps are unavailable, state that limitation before drawing conclusions about duration or frequency.
Verify units and expected ranges
Confirm the engineering unit and scaling against a known condition. Look for stuck values, impossible sign changes and abrupt steps after maintenance. Compare statistics within operating modes instead of mixing startup, idle and full-load periods. A wider distribution may reflect a product mix change rather than a sensor problem.
Publish a quality note with the result
Report the source period, number of raw and retained points, exclusion rules and known gaps. Keep the original values available for review. Use generated sensor data to exercise the analysis workflow, but do not treat synthetic examples as evidence of real machine behavior or predictive model performance.
Put it into practice
Use the related tool to check your assumptions, then bring the results to your project discussion.