Predictive maintenance for manufacturing comes down to one decision: buy a platform that arrives with its own sensors and pre-trained models, or build a model on the machine data your plant already produces. This guide covers what predictive maintenance actually predicts, what each route asks of you, how both sit alongside the CMMS you already run, and a rule for choosing between them.
The prize is well documented. Deloitte’s Industry 4.0 research on predictive technologies for asset maintenance reports maintenance cost reductions in the range of 25% to 30% and equipment downtime reductions of up to 50%. Treat the upper figures as a ceiling reached by mature programmes on well-instrumented assets, rather than a first-year expectation.
What predictive maintenance for manufacturing actually predicts
Predictive maintenance estimates how much useful life a specific asset has left, and flags the failure mode that is developing. That is a different job from preventive maintenance, which replaces parts on a calendar whether they need it or not, and from condition monitoring, which reports a current reading without forecasting anything.
Most programmes start with rotating equipment, because it fails in ways that instruments can see early. Vibration analysis picks up bearing wear, imbalance and misalignment weeks before a failure. Motor current signature analysis reads electrical draw to infer mechanical trouble. Thermal and acoustic data add coverage. Plants that already buy condition monitoring services on a route-based schedule are collecting much of this data already, just at monthly intervals rather than continuously.
Two things determine whether prediction is possible at all. The asset needs a failure mode that develops over a measurable window, and you need enough recorded failures for a model to learn the pattern. An asset that fails instantly and without warning is a candidate for redundancy, not for predictive maintenance AI.
Route one: off-the-shelf predictive maintenance software
What the platforms do well
The vendor market has specialised rather than converged, which makes the choice easier than it looks.
- Augury focuses on vibration and acoustic analysis for rotating equipment, with models trained on a large base of machine hours.
- Samotics uses electrical signature analysis read from the motor control cabinet, so nothing has to be mounted on the asset itself. That suits submerged, hazardous or otherwise hard to reach equipment.
- Tractian ships its own sensors alongside condition-monitoring models and a built-in CMMS, aimed at maintenance technicians rather than data teams.
- Uptake sells asset performance management for heavy industry such as mining, energy and rail, and suits organisations that already run a historian.
- Senseye, acquired by Siemens in June 2022, sits at the enterprise end and now ships with a generative maintenance assistant.
- MachineMetrics is worth naming precisely: it concentrates on CNC machine monitoring, utilisation and OEE rather than full failure prediction.
What you get from any of these is speed. Sensors, models and dashboards arrive together, and a competent vendor can have a pilot cell reporting inside a quarter.
What you supply, and what you rent
You supply access, asset lists and the maintenance team’s time. You rent the models. Two consequences follow, and both matter more at year three than at month three.
The first is data location. Most predictive maintenance software processes machine data in the vendor’s cloud. For many plants that is acceptable. For defence work, contract manufacturing under customer NDAs, or sites with an air-gapped OT network, it is the point where procurement stops.
The second is portability. The model that learns your assets belongs to the vendor. If you change platforms, learning restarts. Per-asset pricing also means the economics work against you exactly as coverage expands across the plant.
Route two: a model built on your own data
What data you need before you start
A custom build needs three inputs, and the third is the one that stalls projects.
Sensor history. Continuous readings at a sample rate high enough to see the failure developing. Vibration typically needs high-frequency capture; temperature, pressure and current are less demanding.
Asset context. What the machine is, what it runs, and under what load. A model that cannot distinguish a heavy production run from an idle shift will read normal variation as a fault.
Labelled failures. Work orders that record what actually broke and when. This is where most plants discover their maintenance records describe the repair rather than the cause. Six to twelve months of clean failure history is a realistic starting point for a first model, and building that record is often the first phase of the work.
What you own at the end
Predictive maintenance machine learning built on your own data produces an asset you keep: the trained model, the pipeline that feeds it, and the labelled dataset underneath. It runs where you choose, including entirely inside your own network on a private, on-premise deployment, so machine data never leaves the site.
It also covers assets no vendor has a product for. Custom machinery, older equipment and process lines with unusual failure modes are common in plants that have grown by acquisition, and they are exactly what off-the-shelf catalogues skip.
The cost is time. There is no pre-trained model waiting, so the first useful alert arrives later than it would from a platform.
The two routes side by side
| Option | Best fit | What you supply | What you own at the end | Time to first prediction |
|---|---|---|---|---|
| Off-the-shelf platform | Standard rotating equipment, motors and pumps. Cloud processing acceptable. Small maintenance data team. | Asset list, site access, technician time, a per-asset subscription | A subscription. Sensors may be leased; models stay with the vendor | Weeks to a quarter |
| Model built on your data | Custom or legacy machinery, unusual failure modes, data that has to stay on site, coverage expanding across many assets | Sensor history, asset context, labelled failure records, a data owner | The model, the pipeline and the labelled dataset, running where you choose | One to two quarters, longer where failure history needs building |
Where your CMMS fits
This is the most common confusion in the buying process, so it is worth stating plainly. A CMMS schedules and records maintenance work. It does not predict failures.
UpKeep, Limble, MaintainX, eMaint and Hippo all handle preventive scheduling on time or meter triggers, and eMaint adds condition-based triggers on higher tiers. Fiix, acquired by Rockwell Automation in 2021, integrates closely with Allen-Bradley environments. All of them are the right destination for a prediction, and none of them is the source of one.
The practical pattern is that CMMS software becomes the work-order layer and the prediction becomes an input to it. Whichever route you take, the alert should raise a work order in the system your technicians already open every morning, rather than in a second dashboard nobody watches.
The same applies to machine downtime tracking. Downtime tracking software tells you what already stopped and why. That record is genuinely valuable, and it is also the labelled history a predictive model learns from, which is why plants with disciplined downtime coding reach a working model faster.
The decision rule
Three questions settle it in most plants.
Can your machine data leave the site? If the answer is no, on any material share of assets, the build route is the only one that clears procurement. Everything else is secondary to this.
Is your equipment standard or unusual? Fleets of common motors, pumps and gearboxes are what platform catalogues are built around, and buying is usually faster and cheaper. Custom machinery, process lines and older assets are where predictive maintenance solutions from a catalogue tend to run out of coverage.
How many assets will this eventually cover? Per-asset subscriptions are efficient for a pilot and get steadily less so across a large estate. Run the arithmetic at your three-year asset count, not at the pilot count.
A short answer for most mid-market plants: pilot a platform on standard rotating equipment to prove the value quickly, and build where the data cannot move or the assets are your own design. Those two paths coexist comfortably, and the second is easier to justify once the first has demonstrated a number.
Getting either route into production is an engineering problem more than a modelling one. Our machine learning consulting and MLOps consulting engagements cover the sizing, the data pipeline, the retraining schedule and the handover, and the wider set of plant use cases sits on our manufacturing AI solutions page.
Working out which route fits your plant?
Bring your asset list, your sensor coverage and your maintenance records. We will come back with the assets worth starting on and a realistic date for the first useful alert. See manufacturing AI solutions or book an AI strategy session.
