How do I set historian logging rates and deadbands in SCADA without losing important data?

Fixed-interval logging stores the same number of samples whether the signal moves or not. At 1 second, 2,000 tags produce 2,000 × 86,400 = 172.8 million samples a day, and most of them repeat the previous value. Fixed slow logging has the opposite fault, as you found: a 60 second interval cannot show anything shorter than a minute.

The standard answer is exception-based logging. The historian looks at the value every scan but stores a sample only when it has changed by more than a deadband since the last stored value, or when a maximum time has passed. A quiet signal then costs almost nothing, and a fast event is recorded at full scan resolution.

Set the deadband from the instrument, not from one global number. There is no value in storing changes smaller than the measurement noise or the transmitter accuracy. For a 0 to 10 bar transmitter with 0.25 percent accuracy, anything below 0.025 bar is within the error band, so a deadband of 0.02 to 0.05 bar loses no real information.

Always pair the deadband with a maximum interval, for example 5 or 10 minutes, so there is a stored point even when nothing changes. Without it you cannot tell a steady signal from a dead one.

Many historians add a second stage, often a swinging door algorithm. Instead of comparing with the last stored value, it checks whether the new point could be reproduced within a tolerance by a straight line from the last stored point. Points on a steady ramp are dropped because the line between the end points recreates them; a change in slope forces a stored point. This keeps the shape of spikes and steps with far fewer samples than a plain deadband.

How the exception and compression settings should relate to each other differs between products, so follow the guidance in your historian's documentation.

For fault investigation, do not rely on the analog trend alone. Log the digital events (trips, valve commands, alarm transitions) on change, with the time stamp from the source where the PLC can provide it, because a SCADA poll every second can reorder events that happened 100 ms apart.

After changing the settings, verify them: overlay the compressed trend on a temporary 1 second log of the same tag for a day, and check that the peaks and the step times agree within the tolerance you chose.

The slow month-long trends are a separate issue from storage volume. A month of 1 second data is about 2.6 million points per pen, while the screen is perhaps 2,000 pixels wide, so drawing raw samples is wasted work.

Check whether your trend tool requests raw or aggregated data, and switch long time ranges to aggregates. Asking the historian for the minimum and maximum per interval, rather than an average, keeps short spikes visible even when zoomed out.