True predictive maintenance waits on a hall-wide dataset
The lean board wants fewer surprise stops. A slide promises predictive maintenance for the whole line. What lands is a clamp on three motor shafts that hears vibration, current, or noise, and the rest of the value stream stays as blind as before.
Why the term is hot and the floor still waits
Predictive maintenance is one of the most sought-after labels in lean manufacturing and TPM programs. It sounds like the clean next step after calendar preventives: act before the breakdown, cut waiting and overtime, protect flow. Boards and Industry 4.0 checklists repeat the word because remaining life on a critical asset is easy to want.
Wanting the word is not the same as owning the data. Reviews of manufacturing PdM keep pointing to the same stall: models can work in a lab while production fails on joins across systems, missing failure labels, and weak asset identity. Software claims that skip that foundation sell hope. Lean teams that buy the hope without checking the record still pay for surprise stops.
What most real implementations actually are
On the floor, mature predictive work is still mostly component condition monitoring. Vibration on bearings and shafts. Motor current signature on the supply. Acoustic noise near rotating parts. Temperature riding along. That stack is real and useful on pumps, fans, conveyors, and spindles. It is also narrow by design: one machine envelope, a healthy baseline, an anomaly or remaining-life model for that asset.
A typical lean moment: reliability instruments three critical motors. One early warn saves a weekend rebuild. Sponsorship treats the predictive program as done. PLC cycle and fault bits, product and quality results, humidity and ambient swings, who ran the station, who closed the last preventive, and which breakdown codes repeated last quarter never entered the model. The slide said plant predictive maintenance. The deliverable was shaft-level PdM.
That gap is not an insult to motor monitoring. It is a naming problem. Twin RFPs that collapse the same gap are walked under predictive maintenance as the default twin promise. Lean ownership still needs the split in plain language: useful component PdM versus plant-true prediction.
What plant-true prediction would actually ingest
True predictive maintenance for a line is not a wall of motor dashboards. It learns from a spatially oriented dataset: place and time for machines, material, people, and environment in one frame. That record has to carry more than shaft sensors.
Line and control state - PLC and related control signals: modes, faults, cycle times, interlocks, and load that explain how hard the asset was working when the signature changed.
Product and quality - What SKU or recipe was running, scrap and dimensional results, rework hops. A vibration spike with a known heavy mix is not the same event as the same spike on an empty night run.
Environment - Temperature and humidity swings, and other building loads that shift baselines by season or by open dock doors. Summer and winter healthy states are not one threshold.
People - Who operated the station, who repaired it, who performed the preventive, and whether that person was actually at the asset. Desk-closed work orders without presence are not the same fuel as work proven on the hall.
Maintenance history - Preventive done on time or late, breakdown codes, spare changes, and restore times from the CMMS. Prediction without labeled outcomes is storytelling.
On that joined record, correlation and causality analytics ask which co-occurring places and events keep preceding a stop. A predictive twin can later estimate what may happen next across the hall. Live logistics, the stored spatial replica, and that analytics layer stay equal jobs. Motor PdM remains a useful slice inside its envelope. It is not the hall model.
Nobody has arrived. It is a journey
Hall-scale predictive maintenance does not exists as something plants can buy end to end today. The ambition is not wrong. The fuel is missing. Most sites are still starting operational digital twins and the spatial datasets underneath them. Without months and years of PLC, product, environment, people, and maintenance history sharing one place-and-time frame, vendors ship what they can: isolated vibration and current packages labeled as predictive maintenance for the factory.
Do not treat those packages as proof that plant-true prediction is already done. Keep motor pilots when reliability owns a critical shaft list. Fund the operational hall and honest CMMS joins as the foundation for lean flow and later prediction. Refuse the claim that the first shaft pilot finished the plant AI story. Why prediction waits on stored history is under predictive digital twins wait on stored history. TPM place and due work sit under TPM as lean flow insurance on the hall and preventive maintenance only works when CMMS meets the hall.
How lean teams stay honest about the claim
1. Separate component PdM from plant predictive maintenance in the charter - Vibration, current, and noise on shafts are one buy. Hall-wide prediction on a spatially oriented dataset is another.
2. List the data classes the model must see - PLC and process state, product and quality, environment, people presence and ownership, CMMS preventives and breakdowns. If a class is missing, the claim is not plant-true yet.
3. Demand place and time for every join - Closed preventive without presence at the asset, overload without mover path, or a false trip without ambient and door state is incomplete fuel.
4. Treat the operational record as the current lean win - Search, dwell, overdue work, and correlation on today's hall pay now. Plant-scale prediction is the destination after that record is dense enough to learn from.
twinzo's clear scope is that operational layer: production monitoring, internal logistics optimization, and the dataset that correlation and causality can read before anyone claims factory-wide predictive maintenance. Capability depth sits on the features overview.
Get in touch if you want to map which data classes your lean predictive claim already has, and which ones still make the plant-wide promise a journey rather than a finished product.