Key takeaways
- The usual blocker is not sensor data. It is failure history: CMMS work orders that do not record what failed, why, or when.
- Pick a small number of critical assets with known failure modes before buying sensors or platforms.
- Validate models against held-out failures and agree a false alarm budget with maintenance before go-live.
- Every alert needs a named owner, a response time and a feedback loop into the CMMS.
- Model retraining or threshold changes are changes to how the plant is maintained. Put them through MOC.
Predictive maintenance is usually the first industrial AI use case a plant tries, and often the first one that quietly stalls. The pilot produces a nice dashboard. A few alerts fire. Maintenance does not trust them, or trusts them and gets burned by a false alarm on a Friday night. Six months later the sensors are still installed and nobody is looking at the screen.
The reasons are rarely about the algorithm. They are about data and governance: incomplete failure history, sensors on the wrong assets, models nobody validated, and alerts nobody owns. This page covers what you actually need and how to run a program that survives contact with the maintenance team.
Does this apply to you?
It does if you have critical assets whose unplanned failure stops a line, creates a safety or quality risk, or costs real money in repairs and lost output. It applies even more if a vendor has offered a predictive maintenance pilot, because the governance questions below are the ones to ask before you sign.
The data you need, and the gap you probably have
A predictive model needs three kinds of data. Most plants are strong on one, partial on another and weak on the third.
| Data | Typical source | Usual state |
|---|---|---|
| Condition signals (vibration, temperature, motor current, pressure, oil) | Historian, PLC, retrofit sensors | Often available or easy to add |
| Operating context (speed, load, product, state, ambient) | PLC, MES, historian | Partly available, often not aligned |
| Failure history (what failed, how, when, why) | CMMS work orders | Usually the gap |
Why CMMS history is the gap
Work orders were designed to schedule labor and track cost. They were not designed to train models. Common problems:
- Free-text descriptions ("bearing noisy, replaced") with no structured failure mode.
- The completion date recorded, not the date the fault started.
- Corrective work logged against the wrong asset or a parent asset.
- Run-to-failure events fixed by operators and never logged.
- Inconsistent codes across shifts and sites.
Without this, a model can find anomalies but cannot learn which anomalies precede which failures. That is the difference between "this looks unusual" and "this bearing has a few weeks left."
Plant-floor reality: The fastest improvement in many programs is not a new sensor. It is a shorter, mandatory failure-code list on corrective work orders, enforced by the maintenance supervisor, for the dozen assets in scope.
Fixing failure history
- Pick a failure-mode list per asset class. Keep it short. ISO 17359 gives general guidance on structuring condition monitoring programs around failure modes.
- Make failure mode, failure date and cause required fields for in-scope assets.
- Back-fill the last two or three years for in-scope assets by interviewing technicians. It is tedious and it is worth it.
- Record near-misses and "found during PM" conditions, not just breakdowns.
Choose assets before sensors
Start with a criticality review, not a sensor catalog:
- Which assets stop the line or create a safety or quality risk when they fail?
- Which of those fail often enough to learn from, but not so often that a simple PM fix is better?
- Which failure modes give a detectable warning before failure (for example, bearing wear shows in vibration; some electronic failures give no warning)?
A short list of five to fifteen assets with clear failure modes beats sensors on everything.
Sensor retrofit: practical points
- Match sensor to failure mode. Vibration for rotating equipment, thermal for electrical connections, current signature for motors, pressure and flow for hydraulics.
- Sample rate matters. Vibration analysis often needs high-frequency data processed at the edge, with only features (RMS, peaks, spectra bands) sent upstream. See edge vs cloud AI.
- Wireless sensors are network devices. They have gateways, firmware and credentials. Put them in your asset inventory and zone model.
- OEM warranties. Some OEMs object to drilling or mounting on their equipment. Ask first.
- Calibration and mounting. A loosely mounted sensor produces drift that looks exactly like a developing fault. See industrial data quality.
Model validation
Validation is where pilots most often skip steps. A minimum standard:
- Hold out time, not just rows. Train on earlier periods and test on later ones. Random splits leak future information.
- Count failures, not data points. If you have six real failures, you have a very small test set. Say so.
- Measure lead time. An alert two hours before failure is not useful if a repair needs a two-day parts lead time.
- Measure false alarms per asset per month. This is what maintenance will judge you on.
- Test across operating modes. Changeovers, start-ups and different products often trigger false alarms.
- Document it. Record data period, features, thresholds, validation results and known limits in the AI use case register.
The false alarm budget
Agree, before go-live, how many false alarms maintenance will tolerate per asset or per area per month. Then tune thresholds to that budget, not to the highest possible detection rate. A model that catches every failure and cries wolf twice a shift will be ignored within weeks, and an ignored model is worse than none.
Governance: who acts on an alert
An alert without an owner is noise. For each in-scope asset, define:
| Question | Example answer |
|---|---|
| Who receives the alert? | Area maintenance planner, not a shared inbox |
| What is the expected response time? | Review within one shift |
| What actions are allowed? | Inspect, schedule, or dismiss with reason |
| Can the alert stop or change the process? | No. Advisory only unless separately approved |
| Where is the outcome recorded? | CMMS work order linked to the alert ID |
| Who reviews performance? | Reliability engineer, monthly |
The link between alert and work order is the feedback loop. It turns every alert into new labeled history and lets you measure real precision over time.
Watch out: If a predictive maintenance output can trigger an automatic shutdown, change a setpoint or defer a safety-related inspection, it is no longer advisory. Treat it as a higher-risk use case and run a full AI risk assessment.
MOC when models change
Retraining a model, changing a threshold or adding a feature changes how the plant is maintained. Put these through management of change:
- Record what changed and why.
- Re-run validation on the held-out set.
- Tell the people who receive alerts.
- Keep the previous version so you can roll back.
Vendor-hosted models count too. Your contract should say when the vendor retrains, how you are notified and whether you can pin a version. Cover this in AI vendor due diligence.
ROI realism
Build the business case per asset, from your own numbers:
- Cost of an hour of unplanned downtime on that line.
- How often the asset fails per year, from (cleaned) CMMS history.
- Share of failures with a detectable early warning.
- Repair cost of planned vs unplanned intervention.
- Program cost: sensors, connectivity, platform, and the people time to review alerts.
If an asset fails once every three years and the downtime is cheap, a better PM routine may beat any model. Be wary of headline savings figures that are not tied to assets and failure modes like yours. Track realized savings from linked work orders, not projected savings from the dashboard.
Common mistakes
- Buying a platform before checking failure history.
- Sensors on everything, alerts owned by nobody.
- Validating on random splits and reporting high accuracy.
- No false alarm budget, so thresholds are tuned for the demo.
- Retraining silently, especially by a vendor.
- Measuring success by number of alerts rather than avoided failures.
Monday-morning checklist
- List your ten most critical assets and their top failure modes.
- Pull three years of CMMS corrective work orders for them and check how many record failure mode and date.
- Make failure mode a required field for those assets from today.
- Register the use case in the AI use case inventory with a business owner.
- Agree alert ownership, response time and a false alarm budget with the maintenance manager.
- Only then scope sensors and platforms.
Effort: cleaning failure history for a dozen assets is typically days to a few weeks of reliability and technician time. The ongoing cost is the alert review discipline, roughly a standing slot in the weekly maintenance meeting.
What good looks like
A short list of critical assets with clear failure modes. Clean, coded failure history. Validated models with known lead times and false alarm rates. Every alert routed to a named person, closed with a reason and linked to a work order. Model changes tracked through MOC. And a monthly review that shows avoided failures in plain numbers, not just a trend line.
Sources and further reading
- ISO 17359:2018 Condition monitoring and diagnostics of machines: General guidelines (ISO)
- NIST AI Risk Management Framework
- NIST SP 800-82 Rev. 3, Guide to Operational Technology (OT) Security
This guide is general information for manufacturers, not legal advice. Confirm regulatory obligations with qualified counsel.
FAQ
Frequently asked questions
What data do I need for predictive maintenance?
You need condition data from the asset (vibration, temperature, current, pressure or similar), operating context (load, speed, product, state), and failure history with dates, failure modes and causes. Most plants have the first two in a historian or can retrofit them. The third is usually incomplete, because CMMS work orders were written for scheduling and cost, not for analysis.
How do I validate a predictive maintenance model?
Test it on failures it did not see in training, ideally from a later time period. Measure how many real failures it caught, how far ahead, and how many false alarms it raised per asset per month. Agree those targets with maintenance before deployment, then monitor them in production and re-check after any process or equipment change.
Is condition-based maintenance the same as predictive maintenance?
They overlap. Condition-based maintenance acts when a measured condition crosses a threshold, such as vibration above a limit. Predictive maintenance tries to estimate when a failure will occur, often with a model trained on history. Many successful programs start with well-governed condition-based rules and add prediction only where history supports it.
What ROI should we expect from predictive maintenance?
It depends on the cost of unplanned downtime on the chosen assets, how often they fail, and how much of the failure is detectable early. Build the case asset by asset from your own downtime and repair costs. Be skeptical of headline percentages from vendors that are not tied to assets like yours.
Related guides
Data Foundations
Industrial data quality for analytics and AI: what breaks and who fixes it
Industrial AI fails most often on data quality problems that are specific to plants: clocks that disagree, sensors that drift, historians that discard detail, missing production context…
Guide · 6 min read
Data Foundations
Edge vs cloud AI in manufacturing: a governance and security decision guide
Run AI at the edge when it needs fast, deterministic responses, high-volume raw data, or must keep working without the internet.
Guide · 6 min read
AI Governance
AI risk assessment for manufacturers: a practical tiering scheme
Tier each industrial AI use case by what it can do to the process, people and product: does it advise, act in closed loop or touch safety; does it use worker data; does it release product…
Guide · 7 min read
AI Governance
How to build an AI use-case inventory in a manufacturing plant
An AI use-case inventory is a register of every AI system you build, buy or have embedded in equipment, with an owner, a risk tier and enough detail to act on.
Guide · 7 min read