The full power of TDengine, now free forever for up to 5,000 tags.

Explore More

How far back should you pull sensor data after an equipment failure?

Juno Qiu

November 17, 2026 /

Start with the failure event, then use separate windows for the immediate sequence, the lead-up, and a comparable healthy baseline. The right duration and resolution depend on the failure mechanism you are testing and the data actually recorded. Pulling more months is not automatically better, and a high-resolution graph cannot reconstruct samples the historian never stored.

This is a window-selection method for a reliability engineer. It assumes the event and asset have been identified; it does not establish the cause of failure. If required detail was never collected, record the gap and adjust future instrumentation rather than interpolating an invented signal.

Use three windows for three questions

Window Question Starting approach
Event What changed immediately before and after the trip or fault? Begin around the best-supported event time; retain the finest available relevant timestamps and quality information.
Lead-up Did a condition change as the equipment approached failure? Expand far enough to include plausible operating cycles and the earliest observed change; compare at a resolution that retains the trend.
Baseline What did similar healthy operation look like? Select periods with comparable load, product, speed and operating state, avoiding known bad sensors or unrelated maintenance.

These are categories, not prescribed durations. A short electrical event and slowly developing wear need different views. You may need several baseline periods rather than one continuous window. Record why each boundary was chosen and expand it if the data do not answer the question.

Work backward from a hypothetical pump trip

Assume a pump tripped on Tuesday. A first pull covers the preceding 24 hours at one-minute averages and shows no obvious temperature trend. That does not prove temperature was stable at every instant or that the pump was healthy before Monday: the pull has both a time boundary and an aggregation boundary.

First, inspect the trip window at the finest trustworthy resolution available for relevant pressure, flow, motor status, and alarm events. Check whether the apparent sequence is credible given clock agreement, sampling, quality flags, and notification delay. If pressure was stored only once a minute, report that a shorter pressure transient could have been missed.

Next, expand the lead-up to encompass prior starts and operating cycles. If the equipment changes slowly, compare earlier weeks or months where available; if it operates only intermittently, count comparable run cycles as well as elapsed calendar time. A hypothetical increase across multiple comparable runs might justify further investigation. A flat trace from an unchanged or failed sensor does not prove the physical condition was flat.

Finally, choose a baseline from periods with similar duty. A high temperature at full load may not be comparable with a low temperature during idle operation. Record changes in sensor calibration, unit scaling, equipment configuration, and maintenance that could make an apparent long-term slope misleading. The example is illustrative and does not assert a universal bearing-deterioration timetable.

Choose resolution to preserve the decision

Ask how fast the phenomenon could change and what the collection chain can observe. A trend over months may be readable with coarser summaries, while a trip sequence may require individual stored events. An average can conceal a brief high or low value; inspect minimum and maximum over each bucket as well as sample counts and quality where available. Even extrema only summarize the points included in them. They cannot recover a transient that was never sampled or was filtered upstream.

When calculating a rate of change, include elapsed time, units, and the measurement method. Compare like operating states; an apparent slope caused by a load change is not a degradation rate. Sparse or irregular samples also constrain the inference. A rate can be useful evidence, but it is not always more informative than a single safety-critical threshold crossing.

TDengine’s time-series window documentation describes time and state windows for query aggregation. Those operations can organize stored data by time or state; they do not validate sensor quality, establish a healthy baseline, or prove a failure mechanism.

Save the window choice with the finding

For each view, retain the start and end, time zone, event-time basis, tags and asset identity, operating-state filter, sampling and aggregation, missing periods, and any changes to sensor or configuration. State whether the window was chosen before or after seeing the data. If an expanded search reveals a possible pattern, test it against other relevant periods rather than presenting an exploration as a preplanned confirmation.

An assistant can retrieve and summarize the selected windows, but it needs the same boundaries and limitations. The useful output is not simply “no anomaly found”; it is a traceable statement of what was inspected, at what resolution, and which potential behaviors the available data could not rule out.

Related Articles