Blog

  • Vibration baselines drift with the season, and nobody tells you

    A vibration baseline established in February does not hold in August. Ambient temperature changes bearing clearances and mounting stiffness enough to move the spectrum, and a threshold set against the winter baseline will alarm through the summer on perfectly healthy equipment.

    We now fit the baseline as a function of ambient temperature rather than as a constant. It removed most of our seasonal false positives and, unexpectedly, made genuine bearing faults easier to see because the residual is quieter.

  • Wartungspilot v1.9: fault attribution across coupled machines

    On a shared frame the vibration signature of one machine appears in the sensors of its neighbours, so a single developing fault produces alarms on three assets and a maintenance team that does not know which one to open.

    v1.9 models the coupling explicitly and attributes spectral energy back to a source asset. It is not perfect on tightly coupled pairs, but it has cut the number of unnecessary teardowns substantially.

  • Why we do not stream raw sensor data to the cloud

    Raw high-rate vibration data from a plant is large, and shipping all of it is expensive. That argument is well known. The stronger argument is that plant operators will not agree to it, and a product that requires them to is a product that does not get deployed.

    Feature extraction happens on the edge gateway and only the derived spectra and health indicators leave the site. It constrains what we can do retrospectively, and that constraint has been worth accepting.

  • Notes on commissioning: the first two weeks are all data quality

    Documentation says a sensor is mounted radially on the drive end. Often it is mounted axially, or on the housing rather than the bearing cap, or on a machine that was replaced two years ago and never updated in the asset register.

    We now spend the first fortnight of every commissioning validating mounting and asset mapping before enabling any alerting. Skipping it does not save time; it moves the same work to a point where it is mixed up with real faults.