Logging Strategies: Triggers, Deadbands and Aggregation

Why not log everything?

Recording every sample at the highest rate is simple and, up to a point, cheap. But volume costs: storage, bandwidth, battery, upload fees, and above all the time to find the interesting minute in a terabyte. A logging strategy decides which samples are kept. The aim is to store less without losing the information you need, and to know what each strategy hides.

Strategy	Keeps	Hides
Periodic	A sample every interval	Anything faster than the interval
On change (deadband)	A value when it moves by more than a band	Changes smaller than the band; the time between records unless a heartbeat is added
Triggered (with pre-trigger)	A window of fast data around an event	Everything outside events, and events the trigger misses
Aggregated	Min, mean, max (and count) per interval	The shape inside the interval
Burst	Short fast records at set times	Everything between the bursts

Periodic logging

The default: one record every fixed interval. It is easy to analyze, since the samples are evenly spaced and every tool understands them. Choose the interval from the process (see Planning a Logging System) and filter before sampling (see Sampling, Aliasing and Filtering). Its weakness is a fixed price: a steady plant at 1 Hz stores the same as a violent one.

On change: deadbands

A deadband stores a value only when it differs from the last stored value by more than a set amount. A steady signal produces almost nothing; a moving one produces many records. Always add a heartbeat, a record at least every so often even if nothing changed, so that you can tell “no change” from “logger dead”.

• Compare with the last stored value, not the last sample, or a slow ramp never triggers.
• Use hysteresis for alarms (for example trip at 80, clear at 78), or a signal near the limit logs and alarms continuously.
• Log the time with every record: the spacing is irregular, so analysis must resample (see Data, Files and Parsing).
• Industrial historians use swinging-door compression, a refinement of the deadband that fits a straight line through the points within a tolerance.

Triggers and the pre-trigger buffer

To catch a rare, fast event (a fault, a spike, an impact) you cannot store everything at full rate. A trigger starts a fast recording when a condition occurs. But the interesting part is often what happened before the trigger, so the logger keeps a circular ring buffer of the most recent samples and, on a trigger, saves the buffer plus a post-trigger period.

Aggregation: min, mean, max

Instead of storing every sample, store the minimum, mean and maximum (and sometimes the count and standard deviation) over each interval. The mean gives the trend, and the min and max keep the extremes that a plain average would erase.

• Align intervals to the clock (on the minute, on the hour) so that data from several loggers lines up (see Time Stamps and Clocks).
• Weight a time-weighted mean when samples are irregular, or for on-change data.
• Averaging angles needs vectors, not numbers: the mean of 359° and 1° is 0°, not 180° (wind direction is the classic case).
• Counters and energy are summed or differenced, not averaged: store the totaliser value.
• Keep a count of valid samples in each interval, so that an average of three readings is not mistaken for one of six hundred.

Burst and adaptive logging

• Adaptive rate: log slowly in steady running and fast when a quantity is moving or an alarm is active, and record the rate with the data.
• Scheduled bursts catch slow trends in fast signals; event bursts catch the rare ones; use both.
• Edge processing: compute the FFT or the RMS on the logger and send the result, keeping raw data only when something looks wrong.

Comparing the strategies

Need	Use	Remember
Evenly spaced trends of a slow process	Periodic	Interval about τ ÷ 5
Mostly steady plant with occasional moves	Deadband + heartbeat	Band above 2 to 3 × noise; log the time
A rare fast fault	Trigger + ring buffer	Pre-trigger time; test the trigger
Long-term summary of a fast signal	Min / mean / max per interval	Keep the extremes and the count
Condition monitoring of machinery	Bursts + features	Fixed operating conditions for comparable bursts
Compliance record	Periodic at a defined interval, plus alarms	Do not reduce data that rules say to keep

Pitfalls

Problem	Result	Prevention
Deadband smaller than the noise	Almost no saving, noisy record	Filter or raise the band
No heartbeat	A flat stretch cannot be told from a dead logger	Always store a periodic record
Trigger never fires, or fires constantly	No data, or a full card in an hour	Test with a real event; add a re-arm time and a daily event cap
Averaging hides the extreme	A real excursion is missed in a report	Store min and max with the mean
Mixed strategies without a note	Later users assume even spacing	Record the strategy, band and interval in the metadata
Thinning before filtering	Aliasing (see Sampling, Aliasing and Filtering)	Low-pass or average, then keep every Nth

Related

• Data Logging: the overview of this guide.
• Storage, Formats and Data Rates: what each strategy saves in bytes.
• Data Quality, Analysis and Reporting: checking irregular and aggregated data.
• Embedded Systems: ring buffers and timers in firmware.
• Machine Learning: using features and events for prediction.

FAQ

What deadband should I use?

About the accuracy of the sensor, and above two to three times the noise. Anything smaller stores noise. Always add a heartbeat record.

How long should the pre-trigger be?

Long enough to show the cause: several times the time scale of the build-up to the event. Start with 10 to 30 % of the whole record and adjust after the first events.

Does averaging lose information?

Yes: the shape and the extremes inside the interval. Store the min and max with the mean, and keep raw data when you can.

Can I reduce data after the fact?

Yes, from the raw data, with filtering first. You cannot recover detail that was thinned or averaged away at the logger.

References

• OSIsoft / AVEVA PI documentation on exception reporting and swinging-door compression; Bristol, E. H., “Swinging door trending”, ISA Transactions, 1990.
• Smith, S. W., The Scientist and Engineer’s Guide to Digital Signal Processing (decimation).
• Vendor application notes on trigger and pre-trigger modes (National Instruments, Keysight, HOBO).

Last reviewed: October 2026. Event rates and file sizes are examples: use your own signals.