A bearing does not fail suddenly. It fails slowly, for months, and then visibly, in one afternoon.
The failure everyone remembers is the afternoon. The failure that mattered was the four months before it, when the vibration signature was drifting and nobody was reading it.
Why threshold alarms miss almost everything
Most plants already have some vibration monitoring. It usually works like this: a sensor reports overall vibration velocity, SCADA compares it to a limit, and an alarm fires when the limit is crossed.
That catches catastrophic failure. It catches almost nothing else, for a specific reason.
Overall velocity is a single number summarising a complex signal. A developing bearing race defect might raise energy in one narrow frequency band by a large factor while barely moving the overall number. By the time overall velocity crosses a threshold, the defect is not developing. It is advanced.
Threshold alarms give you hours of warning. What maintenance planning actually needs is days — enough time to source the part, get it approved, and schedule the work into planned downtime rather than emergency downtime.
What the frequency spectrum shows that a single number can’t
Frequency analysis is decades old. Skilled vibration analysts have done it since long before machine learning existed. The honest question is what AI adds.
Three things.
It learns each machine’s own normal. Textbook fault frequencies assume textbook machines. A real pump, on a real foundation, with a real coupling, has its own baseline signature. A pattern recognition model trained on that specific machine’s history knows what normal looks like for that machine, rather than comparing it to a generic standard.
It watches continuously. A route-based analyst visits a machine monthly or quarterly. A model watches every reading. Slow degradation between visits is exactly the thing routes miss.
It scales past the analyst bottleneck. Most plants have one or two people who can genuinely read a spectrum, and several hundred rotating assets. The constraint is not knowledge. It is hours.
PlantWiz combines FFT, anomaly detection and a pattern recognition model at the edge, and outputs failure prediction 4–6 days ahead.
The output has to be in maintenance language
This is where a lot of condition monitoring projects quietly die. The system generates a spectrum, the maintenance engineer looks at a plot he was not trained to read, and the alert gets ignored until the system is switched off.
PlantWiz Automatic Diagnostics & Recommendations produces something a maintenance planner can act on directly: the fault type, the risk level, and the specific corrective action.
A real ADR output for a conveyor motor reads roughly like this — pattern of structural looseness observed at lower side amplitude, synchronous energy at 1× running speed slightly increased in the horizontal direction, structural looseness most likely diagnosed. Recommended action: monitor the overall vibration trend for significant increase, reinforce mounting and improve stiffness, check looseness of mounting bolts, failed grout, cracks or other weakness in the machinery mount, base plates and hold-down bolts in the upcoming maintenance cycle.
That is a work order. Not a plot.
Remaining Useful Life and what it changes about planning
Asset Health Score gives each machine a 0–100 rating with full history — what is quietly degrading, months before it stops.
Remaining Useful Life converts that into a planning input: days to warning, days to critical.
The consequence is that maintenance stops being calendar-driven and becomes evidence-driven. You are no longer replacing a bearing at 8,000 hours because the manual says 8,000 hours. You are replacing it when its own signature says it has weeks left. Some assets get serviced earlier than the calendar. Many get serviced considerably later — which is where a large share of the cost saving actually comes from.
What PlantWiz detects
Slow-progressing faults that fixed threshold alarms in SCADA and PLCs never catch:
Bearing defects (race, ball, cage) · imbalance · shaft misalignment · structural looseness · soft foot · gear mesh and tooth wear · belt and coupling faults · cavitation and flow instability · rotor bar and electrical faults · lubrication failure · resonance · thermal drift
Across assets: motors, pumps, fans and blowers, gearboxes, compressors, bearings, conveyors, chillers, cooling towers, air handling units and mills.
Results from Indian deployments
From first-year Indian manufacturing deployments: 2 of 3 typical failures prevented, 60–80% fewer emergency repairs, 25–30% better equipment uptime, and 9–12 months to ROI.
One specific case: an automotive press shop, where an unforeseen mechanical press failure stops the whole press line. Within weeks of deployment, downtime on a critical press-shop asset was reduced by 14 hours.
Across industry-leading manufacturers, benefits observed run up to 30% reduction in machine downtime.
Get the full AI/ML Predictive Maintenance & Asset Health flyer
Two pages covering the complete fault detection list, Asset Health Score, Automatic Diagnostics & Recommendations, Remaining Useful Life modelling, FFT and trend views, supported assets and industries, and first-year results from Indian deployments.
