Prediction is earned. It starts with knowing what normal looks like.
Building maintenance intelligence from measured equipment behaviour: condition monitoring first, accumulated failure history second, and prediction once the data supports it.
What is usually being asked for
The request is normally some version of “tell us before it breaks”. It is a reasonable thing to want. Unplanned downtime is expensive, and scheduled maintenance replaces parts that still had life in them.
The difficulty is that genuine prediction of a specific failure, with enough warning to act, generally requires recorded examples of that failure developing. Most operations do not have them, because failures are rare and nobody was recording the relevant signals when they happened. A programme that promises prediction from data that does not exist will disappoint, and will make the next proposal harder to get approved.
The sequence that works
- Measure. Instrument representative equipment properly and collect continuously. See legacy machine retrofit.
- Establish normal. Build a baseline for each machine under its own operating conditions. Machines of the same model in the same plant often differ enough that a shared threshold is useless.
- Detect departures. Alert when behaviour moves away from that baseline. This alone catches a useful share of developing faults and delivers value in months rather than years.
- Record outcomes. When something is investigated or repaired, capture what was found against the signals that preceded it. This is the step most programmes skip, and it is the one that makes everything later possible.
- Predict, once earned. With accumulated labelled history, estimating remaining life or failure probability becomes a question that can be answered with evidence rather than optimism.
What gets measured
Vibration for bearing condition, imbalance, misalignment and looseness. Motor current and power for load, developing mechanical resistance and cycle timing. Temperature and its trend relative to load. Acoustic signature, including ultrasonic ranges for leaks and early wear. Cycle and utilisation data for how the machine is actually being used, which often explains more than any single sensor.
Alerts people will act on
A programme lives or dies on whether maintenance staff trust it. Too many false alarms and alerts get ignored within weeks, at which point the investment is dead regardless of how good the underlying analysis is.
That argues for conservative alerting at the start, thresholds tuned per machine using real collected data rather than defaults, alerts that reach people in the tools they already use, and a feedback route so an investigated alert improves the system rather than vanishing.
How we help
- Scoping. Which assets justify instrumentation, based on what their failures actually cost you.
- Sensing and installation design. What to measure and how to mount it, so the data means something. See sensor selection and signal conditioning.
- Collection infrastructure. Devices, gateways, connectivity, storage and retention sized for the data rate you are committing to.
- Baselining and alerting. Per-machine normals and departure detection, tuned against your own data.
- Integration. Into your maintenance system, so alerts become work orders rather than emails.
- Analysis as history accumulates. Revisiting the prediction question when the dataset can support it. See AI for industrial machinery.
What we will not claim
We will not quote a percentage reduction in downtime before seeing your equipment and your data. Figures of that kind, quoted in advance, are marketing rather than engineering. What we will do is agree measurable criteria for the pilot, and report honestly against them, including where the data did not support the original hope.
Related
AI for industrial machinery, industrial IoT and equipment monitoring, cloud and dashboards.
Which failures hurt most?
Tell us what the equipment is, what unplanned downtime costs, and whether you already record anything about machine condition or repair history.
What clients ask before starting
Can you guarantee a reduction in downtime?
No, and we would treat any figure quoted before seeing your equipment as marketing rather than engineering. What we do is agree measurable criteria for a pilot and report honestly against them, including where the data did not support the original hope.
How long before the system is useful?
Condition monitoring and departure alerting usually deliver value within months, because they need a baseline rather than failure history. Genuine prediction depends on accumulating recorded failures, which takes longer and cannot be rushed.
We have no maintenance history. Can we still start?
Yes. Start by measuring and by recording what is found whenever equipment is investigated or repaired. That second habit is the one most programmes skip, and it is what makes prediction possible later.