The situation we are usually called into
A producer operates raw mills, kiln drives, fans and a large population of rotating equipment on which production depends absolutely. A condition-monitoring programme exists. Data is collected on routes, stored, and largely unused.
This is the most common failure in condition monitoring, and it is rarely described honestly: the organisation has data collection, not diagnosis.
Technicians gather readings competently. Nobody on site can interpret spectra reliably — distinguish an outer-race bearing defect from misalignment, separate looseness from imbalance, recognise the early signatures that appear long before a trend threshold trips. Alarms are set on overall levels, which move late. Faults are therefore detected close to failure, when the only remaining option is an unplanned outage. Trust declines, and with it the discipline of collecting the routes at all.
- Under a fixed-interval or overall-level regime, a fault is typically detected near the end of its degradation curve — the P-F interval is consumed before anyone acts.
- Roughly 10% or less of industrial equipment ever wears out on a predictable age pattern; most failures are random in onset and therefore only detectable by condition, not by calendar.
- Reactive-dominant operations are associated with materially higher defect rates and maintenance-related delays than planned regimes.
The real problem is competence, not instrumentation
Buying better analysers does not produce better diagnoses. Interpretation is a certified skill, and the certification matters precisely because internal judgement of who can read a spectrum is unreliable — particularly under staffing pressure, where the alternative to “he is competent” is “we cannot run the route.”
How we run the programme
Diagnose from spectra and phase
Rolling-element bearing defect frequencies, misalignment and its harmonic signature, imbalance, mechanical looseness, gear mesh, and electrical faults in motors — taught against real machine data, not idealised plots.
Acquire data that is diagnosable
Transducer selection and mounting, resolution, averaging, and the measurement locations that make a fault visible rather than obscured. Bad data cannot be rescued by good analysis.
Establish the P-F interval
Per failure mode, how much warning the chosen technique actually gives — and therefore what inspection frequency is defensible. Most sites pick a frequency by convention; this converts it into a derived quantity.
Set alarms on diagnostic parameters
Not on overall levels alone. Overall level is a lagging indicator of a fault the spectrum identified weeks earlier.
Report so that someone acts
A diagnosis, a confidence level, a recommended intervention and a timeframe — routed to a named person with authority to schedule it.
What makes it survive the next budget cycle
What the client is left with
Analysts who can state what is wrong with a machine, how confident they are, and how long the plant has before it matters — and a system that acts on what they say.