Predictive Maintenance: Turn Data Into Action for Maintenance Leaders
Predictive maintenance is the right strategy for high-criticality assets whose degradation can be measured before they fail. The U.S. Department of Energy reports reductions of up to 30% in maintenance costs, and ISO 17359 gives you the standard procedure for monitoring. The path runs from use case to proof of concept to deployment.
TL;DR:
- Use predictive monitoring when failures are costly, degradation is measurable, and warnings arrive early; schedule medium risk assets and run cheap parts to failure.
- DOE benchmarks reach 75% fewer breakdowns and 45% less downtime, but these are best case ceilings against reactive or fixed schedule maintenance, not guarantees.
- Start with detection if you lack failure records; prognosis needs usage, load, and failure history, while prescriptive decisions also require cost, parts, and staffing data.
- Set pilot thresholds before installing sensors, then confirm warnings allow repairs in time, false alarms stay low, and several alerts lead to completed work orders.
- Assign each alert an owner, response playbook, stocked parts plan, and shutdown authority; otherwise, a valid prediction can still fail to prevent downtime.
Table of Contents
- What is predictive maintenance and what does it return?
- Which technologies and predictive approaches fit your assets?
- How do you choose the right approach for your ambition and data?
- How do you run a pilot and scale it to deployment?
- What data and infrastructure do you need?
- How do you close the data-action gap?
- How do you size the business case and measure ROI?
- What are the best practices and pitfalls?
- Our perspective: prediction is the easy part
- Turn your pilot plan into a working predictive system
- FAQ
- Sources
What is predictive maintenance and what does it return?
Predictive maintenance (PdM) is condition-based, data-driven forecasting: you measure how an asset is degrading and schedule work when the evidence says it is needed. Reactive maintenance waits for the breakdown. Preventive maintenance acts on a calendar or a usage counter, whether or not the part needs attention. Predictive maintenance acts on the asset’s actual condition.
The U.S. Department of Energy reports that well-run programs can cut overall maintenance costs by up to 30%, eliminate breakdowns by up to 75%, and reduce downtime by up to 45% compared with reactive or fixed-schedule approaches. Treat these as ceilings from the best cases, not promises for your plant.
The same guidance shows where to spend. Assign predictive monitoring to high-criticality assets, preventive schedules to medium-criticality assets, and run-to-failure to cheap, easily replaced parts. A good PdM candidate meets three tests:
- High failure impact: a breakdown stops a line, endangers people, or triggers expensive emergency repairs.
- Detectable degradation: the fault announces itself through vibration, heat, current draw or contamination well before the failure.
- Enough lead time: the warning arrives early enough for you to order parts and book a repair window.
Typical outcomes are concrete. A bearing flagged by rising vibration gets replaced during a planned stop instead of seizing mid-shift. A motor drawing abnormal power is rewound before it burns out. A gearbox with contaminated oil gets flushed before the gears wear. The ISO 17359 guidance on condition monitoring asks you to run an equipment audit and a criticality analysis first, and that order protects your budget.
Which technologies and predictive approaches fit your assets?
Condition monitoring rests on a short list of measurable parameters. ISO 17359 names vibration, temperature, flow, contamination, power and speed as the core set. Sensors and IoT gateways capture them, and an analytics layer turns the readings into a health signal.
Three approach classes cover most programs, plus a hybrid:
- Experience and knowledge-based: rules and thresholds drawn from technician expertise and reliability statistics. Cheap to start, but limited to failure modes you already understand.
- Data analytics and machine learning: models learn patterns from historical and live data. They need enough labeled history, and they struggle when failures are rare.
- Physics and model-based: equations of wear, heat and stress describe how the asset should behave. They work with little failure data but demand engineering effort for every asset type.
- Hybrid physics-guided models: machine learning constrained by domain knowledge, which keeps predictions physically plausible.
The University of Twente categorizes PdM approaches into experience-based, reliability statistics, condition extrapolation, stressor-based, data analytics and physics-based families, and recommends selecting by suitability and feasibility. Combining case-specific information with generic degradation models tends to be the most cost-effective route.
The physics-guided option has measurable support. The Scientific Reports PG-XPM framework folds thermomechanical knowledge into feature engineering and monotonic constraints.
Physics-guided constraints cut the false negative rate from 0.360 to 0.153, a 57.5% relative reduction against a well-configured XGBoost baseline on the AI4I 2020 dataset (Scientific Reports). Fewer missed failures is the metric that matters most to a maintenance leader.
The trade-off is engineering time: someone who understands the asset’s physics has to encode it. Pure machine learning scales faster across similar assets but leans harder on clean, labeled data.
How do you choose the right approach for your ambition and data?
Start from what you want to decide, then check what data you have. The MDPI decision framework links ambition levels to required data types and offers a way to pick feasible methods when ambition and data do not match. Three ambition levels cover most needs:
- Detection: is something abnormal happening now? Supports inspection decisions.
- Prognosis: how long until failure? Supports scheduling and parts ordering.
- Prescriptive: what should we do, and when? Supports automated work orders and trade-off decisions.
Data requirements climb with ambition. Detection needs a healthy baseline and condition-monitoring signals. Prognosis adds usage and load history plus enough failure records to model degradation. Prescriptive work also needs cost, parts and resource data. Quick quality checks before you commit: confirm timestamps align across sources, sensor gaps are rare, and failure records say what failed and why.
Use this checklist:
- Pick the ambition level the business decision actually requires.
- List the data that level needs and compare it with what you collect today.
- If data and ambition match, choose the method class that fits and move to a proof of concept.
- If they do not match, collect more data, lower the ambition, or start with fault detection.
- Revisit the ambition level once the first results arrive and data coverage improves.
The University of Twente work adds a structured way to score each candidate approach on suitability and feasibility, which keeps the choice defensible when stakeholders ask why.
How do you run a pilot and scale it to deployment?
A Frontiers case study frames PdM delivery in three phases: use case analysis, proof of concept, and deployment. It also shows how transferable solution patterns let a model built in one factory move to another with far less rework.
Use case analysis. Build domain understanding with the people who run the asset. Run a failure mode and effects analysis (FMEA), rank assets by criticality, and pick a pilot where failures are costly and signals are measurable. Skip the asset that fails once a decade.
Proof of concept. Establish a sensor and data baseline, choose a model class from your decision checklist, and decide where inference runs, at the edge for low latency or in the cloud for heavier training. The PoC should answer one question: can we warn early enough to act?
Deployment and scale. Integrate with your CMMS or EAM so alerts become work orders automatically. Monitor for model drift, document the response workflow, and package the solution so it transfers to similar assets. Our field service automation guide covers the pilot-to-scale pattern for routing alerts to crews.
Require these results before scaling:
- False negatives: missed failures at or below a rate agreed with operations before the pilot starts.
- Lead time: warnings arrive early enough to order parts and schedule the repair.
- Latency: data-to-alert delay short enough for the failure mode in question.
- False alarms: low enough that technicians still trust and act on alerts.
- Workflow proof: at least several alerts carried through to completed work orders.
Pro Tip: Write the PoC success thresholds down before the first sensor goes on. Teams that define “good enough” afterward tend to declare success on whatever the data happened to show.
What data and infrastructure do you need?
The minimum viable architecture is simpler than vendors suggest: sensors, a way to move and store readings, a place to train and run models, and a connection to your maintenance system. ISO 17359 specifies measurement techniques, data acquisition rates, monitoring intervals and records to keep, with annexes giving example parameters and mapping faults to measured parameters. Use those annexes to set sampling rates per asset rather than guessing. As a rule of thumb, fast mechanical signals such as vibration need high-frequency capture, while temperature, power and flow often tolerate slower sampling.
Run this data quality checklist:
- Baseline: collect readings from the asset in known healthy condition, across normal load ranges.
- Time alignment: synchronize clocks across sensors and systems so events line up.
- Outliers: define rules for spikes, dropouts and stuck sensors before they reach the model.
- Labels: supervised models need records of what failed, when and why, so clean up work-order history early.
- Infrastructure: preprocess at the edge, train in the cloud, set retention and security policies, and connect to your CMMS or EAM.
Limited data is the norm. A Chalmers deployment thesis at Parker Hannifin documents limited sensor equipment, small or poor-quality historical datasets and heavy data cleansing. Practical mitigations include synthetic augmentation, transfer learning from similar assets, and starting with detection-based methods that need only a healthy baseline. Our guide to step-by-step digital transformation for manufacturers walks through connecting these data sources.
How do you close the data-action gap?
The most common reason PdM programs disappoint is not the model. Research on maintenance data maturity from the University of Sunderland describes a data-action gap: predictions exist, but the workflows, spare parts or trained staff needed to act on them do not. An alert nobody can act on is just a notification.
Typical failure modes:
- No workflow: the alert lands in a dashboard nobody owns.
- Missing spares: the part the alert calls for has a long lead time and is not stocked.
- Untrained staff: technicians do not trust or understand the signal.
- No authority: operators cannot take a machine down on a prediction.
The fix is governance. Write response playbooks for each alert type, set a spares strategy tied to predicted failure modes, and assign roles with KPIs. Track time-to-action, the percentage of alerts resolved within the target window, and mean time to repair (MTTR) after predictive alerts. Our partners at Hawkeye Systems build equipment and inventory tracking for field operations, the kind of tooling that links an alert to a crew and a part.
Pro Tip: Run a tabletop drill with a fake alert. Walk it from detection to completed repair and note every point where someone says “I’m not sure who does that.”
How do you size the business case and measure ROI?
Keep the business case simple enough to defend in one meeting. Cost items:
- Sensors and installation.
- Connectivity and gateways.
- Compute and software, edge and cloud.
- Integration with CMMS, EAM and data systems.
- Staffing and training.
Benefit buckets are reduced unplanned downtime, avoided repair costs, and extended asset life. Start from your own plant baseline: annual downtime hours for the pilot assets, cost per hour of lost production, and current repair spend. Then use the DOE benchmark as a sanity check.
The Department of Energy cites maintenance cost reductions of up to 30% and downtime reductions of up to 45% as upper-bound benchmarks. If your model assumes results above these ceilings, revisit the assumptions.
Target a fraction of the benchmark in year one and track unplanned downtime, repair cost per event, alert-to-action time and false alarm rate. Early evidence from a single critical asset family is usually enough to justify scaling.
What are the best practices and pitfalls?
Use this prioritized checklist:
- Do start with critical assets chosen by a criticality analysis.
- Do run focused proofs of concept with written success thresholds.
- Do document response workflows and rehearse them.
- Do measure impact against a baseline you recorded before launch.
- Do design monitoring programs with ISO 17359 as the reference procedure.
- Don’t sensorize low-value assets because sensors are cheap.
- Don’t deploy models without governance for retraining and drift.
- Don’t ignore spare-parts readiness.
Our perspective: prediction is the easy part
The industry talks about algorithms, but the programs that last are the ones that treat PdM as an operating change. In our experience, the model is rarely what blocks a pilot. Unclear ownership and messy source data are. Our case study on catching a failing server before it goes down shows the pattern: early warning only pays off when someone is set up to respond. If you want to test an idea on one critical asset, we are glad to talk it through.
— Matija
Turn your pilot plan into a working predictive system
We build the pieces this roadmap calls for. Our proof-of-concept development, from 5,000 EUR one-off, tests your use case against real data before you commit to scale. Once the PoC passes your thresholds, our AI automation services wire models, alerts and work-order workflows into your existing systems.
For rollout, our Agile Engagement model runs from 5,000 to 40,000 EUR per month, and a Fixed-Price Project suits a defined scope. Book a discovery call and bring your most critical asset: we will help you pick the pilot and define the success metrics.
FAQ
What are predictive maintenance techniques?
Predictive maintenance techniques fall into experience-based methods, reliability statistics, condition extrapolation, stressor-based methods, data analytics and physics-based models, as the University of Twente framework categorizes them. Most programs combine condition monitoring of vibration, temperature, power and similar signals with analytics or machine learning.
What is the 80/20 rule in maintenance?
A small share of assets causes most of the downtime and cost. In practice this supports criticality analysis: focus predictive monitoring on the few assets where failure hurts most, as the Department of Energy guidance recommends, and keep simpler strategies elsewhere.
What are the four types of maintenance strategies?
The four common strategies are reactive (run to failure), preventive (time or usage-based), predictive (condition-based forecasting) and prescriptive, which adds recommended actions to the prediction. Most plants mix them by asset criticality.
What is TBM and CBM in TPM?
TBM is time-based maintenance, work done on a fixed schedule. CBM is condition-based maintenance, work triggered by measured asset condition. Both sit inside total productive maintenance (TPM), and predictive maintenance extends CBM by forecasting how long until a fault matters.
Sources
- Physics-guided explainable predictive maintenance framework for industrial IoT systems | Scientific Reports
- 17359
- Frontiers | Implementation and Transfer of Predictive Analytics for Smart Maintenance: A Case Study
