The path
The distance between a sensor and a failure prediction
Prediction is the last step of a path, not the first purchase. Each stage below produces value on its own, and each is a prerequisite the next cannot skip.
A flow of four stages — instrument the failure mode, baseline normal behaviour, detect deviation, predict remaining life — with a branch back to inspection where sensing is not economic.
- 01
Instrument the failure mode
A vibration sensor on the bearing that fails, a temperature point on the winding that overheats, a current signature on the motor that jams — chosen from your failure history, not from a catalogue.
- 02
Baseline normal behaviour
Weeks of readings across real operating conditions, so "different" has a reference. Without a baseline, every reading is just a number.
- 03
Detect the deviation
Alerts on drift from the baseline, reviewed by the people who know the machine — this stage alone converts some surprises into scheduled work.
- 04
Predict remaining useful life
Only once enough degradation history exists to learn from. A model trained on data that contains no failures has nothing to predict with.
Back to inspection
Where a failure mode cannot be sensed economically, the honest answer is a better inspection routine, not a model pretending to see it. Saying so is part of the design.
The blind spot
Calendar maintenance protects against wear; most failures are not wear
Time-based maintenance rests on an assumption: that failure probability rises with age, so replacing parts on a schedule gets ahead of it. For some components that assumption holds. For many it does not — failures triggered by a lubrication fault, a misalignment after a repair, a process upset or an installation defect do not care what the calendar says, and a plant can be fully compliant with its preventive schedule while its worst actors keep failing between services.
Worse, scheduled overhauls of healthy equipment introduce their own failures — a machine that was running well is opened, disturbed and reassembled, and some fraction of early-life defects come from exactly that. The point of condition monitoring is to replace both errors with evidence: intervene when the machine says something is changing, and leave it alone when it says nothing is. Vibration reads the mechanical state of rotating parts, temperature reads friction and load where they turn into heat, and a motor’s current signature reads the electrical and mechanical load pattern from the switchboard without touching the machine at all. None of these is exotic; the discipline is matching the measurement to the failure mode it can actually see.
The deciding factor
What separates a working failure model from an expensive dashboard
Two plants can buy the same sensors and the same software; one gets predictions and the other gets charts. The difference is almost always in these conditions, none of which is a product feature.
The sensor sees the failure mode
A vibration sensor in the wrong axis, or a temperature point far from the hot spot, produces clean data about the wrong thing.
Without it: Confident readings that never move before a failure.
History that contains degradation
A model learns from examples. If monitoring started last month, the examples do not exist yet, and the honest phase is baselining and alerting.
Without it: A prediction layer with nothing to learn from.
Readings tied to the right asset
A signal is only evidence if it is attached to a specific machine, its runtime and its maintenance record — not to a bare tag number.
Without it: Data that cannot answer "which machine, doing what".
Outcomes written back from work orders
When a technician opens the machine, what they found is the label the data needs. If findings die in a closed work order, the model never learns which alerts were real.
Without it: Alerts that stay opinions forever.
Our role
Where we start when a plant has no condition data yet
- 01
The failure history first
We start from your downtime record, not from a sensor list — the handful of machines that cause most of the lost hours decide where instrumentation earns its cost.
- 02
The data path, end to end
We build the pipeline that carries the readings into a queryable store, tied to the asset and its work orders, because that plumbing is what the prediction stage will stand on.
- 03
Prediction when it is earned
We move to models only when the baselines, the alerting and the labelled history support them — and we say plainly which equipment is not there yet and what would get it there.
The maintenance strategy and the intervention decisions stay with your reliability team; we build the monitoring and the data foundations that give those decisions something better than the calendar to stand on.
Common questions
What buyers ask
Can we skip condition monitoring and go straight to predictive maintenance?+
No, and any proposal that says otherwise deserves scrutiny. A prediction model learns from history that contains degradation and failures, and that history is exactly what condition monitoring accumulates. A plant without monitoring has no baselines, no labelled examples and usually no link between sensor readings and work-order outcomes — so a model bought today would have nothing to learn from. The good news is that the intermediate stages pay for themselves: baselining and deviation alerts convert some failures into scheduled work long before any model exists.
Which equipment should be condition-monitored first?+
The machines your own downtime record names, filtered by consequence. Rank equipment by what its failure costs — lost production, safety exposure, repair lead time — and cross that against how often it actually fails and whether its dominant failure modes are sensible to instrument. A critical pump with a history of bearing failures is a textbook first candidate; a machine that fails rarely, cheaply or in ways no economic sensor can see is not, however prominent it looks. Starting with five well-chosen machines beats starting with fifty.
What do vibration, temperature and current signatures each actually reveal?+
Three different windows on the same machine. Vibration is the richest for rotating equipment: bearing wear, imbalance, misalignment and looseness each disturb the machine in a distinguishable pattern, often weeks before failure. Temperature is simpler but honest — friction, lubrication problems and overload all end up as heat, so a trending temperature is a machine asking for attention even when the cause is not yet clear. A motor’s current signature is read from the electrical supply, which makes it attractive where mounting sensors is hard: changes in the load pattern can betray mechanical problems downstream of the motor. Most programmes combine them, because each failure mode is loud in one window and silent in another.
In this cluster
Walk one asset through it with us
A working session that picks one bad actor from your downtime record and maps which of its failure modes a sensor could have seen coming.
Start the conversation
