industrial//maintenance//condition monitoring

Condition monitoring is the practice of watching a machine's health through measurements that change as it degrades (vibration, temperature, motor current, oil debris) and intervening when an indicator crosses a limit, and it is the maintenance strategy that turns *replace every 4,000 hours* into *replace when this bearing starts to show it*. When the intervention is triggered this way it is called **condition-based maintenance**. It needs one thing above all: an indicator that warns early enough to plan, a long P-F interval.

Condition monitoring is the practice of watching a machine's health through measurements that change as it degrades (vibration, temperature, motor current, oil debris) and intervening when an indicator crosses a limit, and it is the maintenance strategy that turns replace every 4,000 hours into replace when this bearing starts to show it. When the intervention is triggered this way it is called condition-based maintenance. It needs one thing above all: an indicator that warns early enough to plan, a long P-F interval.

The techniques are mature industrial tools, each with its own failure modes in view. Vibration analysis is the workhorse for rotating machinery: imbalance, misalignment, looseness and bearing defects each leave a signature in the spectrum. Thermography sees hot electrical connections and overloaded bearings; oil analysis counts wear particles and contamination in gearboxes and engines; motor current analysis reads mechanical faults through the drive's own current; ultrasound hears leaks and early lubrication problems. Most plants combine a few, chosen by which failure modes matter and how early each technique sees them.

Compute features at the edge and ship the numbers.

One vibration channel sampled at 25.6 kHz with 16 bits produces 51 kB/s, about 4.4 GB a day, and a hundred channels 440 GB, which nobody pays to transmit or store. An edge gateway computes the RMS, peak, crest factor and band energies, sends a few dozen numbers every few minutes to the historian, and keeps a raw capture only when something changes.

With limits set per machine, those edge features already catch a large share of problems across a fleet of motors, pumps and fans; continuous raw streaming is kept for the few critical machines where an engineer will actually look at the waveforms.

The limits are usually two, trended: an alert that schedules an inspection and an alarm that schedules the stop, set from severity standards for the overall level and from each machine's own baseline for its spectrum.

Historian data are a poor substitute for raw captures. Compression discards the transients a detector wanted, interpolation invents values in the gaps, and the timestamp is when the value was stored; before modelling, compare a short period of raw data with what the historian kept.

Condition monitoring says now; projecting the trend to say when is prognostics, and acting on that projection with costs is predictive maintenance.