32 Predictive Maintenance: Models and Rollout
More
Depth: Practitioner
Progress saves as you read. Mark sections complete as you go; your progress appears in the sidebar.
- Podcast audio coming soon
- Character narration audio coming soon
32.1 From Condition Evidence to Planned Work
A vibration model can flag a bearing pattern, but some failures appear first as heat and every alert must still become a planned job. Predictive maintenance (PdM) uses condition evidence to support that decision. This chapter follows the maintenance team from thermal inspection and model selection through work orders, economic review, and staged rollout.
Keep the 100-motor manufacturing plant in view: its planning case has eight historical failures per year, and the team must decide whether warnings can prevent enough failures to justify the investment. Before counting savings, it needs evidence that a warning exposes the target fault, that technicians can confirm it, and that a repair fits the production window. A useful prediction must change maintenance action.
This is part 2 of 2. Predictive Maintenance: Signals and Vibration supplies the signal foundations. Here the problem is connecting those signals to a defensible decision. The approaches are to combine thermal and vibration evidence, select models within their operating limits, and expand only after alert quality and maintenance outcomes support the economics.
KEY CONCEPTS
- Thermal evidence
Compare heat with a healthy baseline and operating context.
- Model limits
Match labels and degradation histories to the question.
- Maintenance follow-through
Carry reviewed evidence into inspection, repair and feedback.
- Measured rollout
Use local costs and confirmed outcomes to gate expansion.
32.2 Learning Objectives
By the end of this chapter, you will be able to:
- combine thermal and vibration evidence for maintenance decisions
- select a machine-learning model with stated operating limits
- connect alerts to work orders, return on investment (ROI) evidence, and staged rollout
32.3 Thermal Imaging
Infrared cameras detect thermal anomalies, adding evidence when a defect first appears as heat. The first choice is how to collect a repeatable reading.
Look at Figure 1 for the tool an inspector actually carries on a route, and for what it cannot do.
Look first at the body in Figure 1. It is a pistol grip with a trigger, built to be aimed one-handed while the other hand holds a torch or a clipboard. The screen sits above the grip, so a reading is checked at the asset rather than back at a desk. The lens at the front sets the field of view, which decides how close an inspector must stand to a live panel. Nothing here records a route on its own. That is the limit the rest of this section builds on: a handheld sweep is a snapshot taken by a person, so repeatable trending needs a fixed sensor or a strict capture procedure.
Figure 2 answers what the camera alone cannot: how a hot reading turns into an action.
Read Figure 2 downward in three moves. The top band is sensing, and it already forces a choice: a handheld camera suits an inspection route, a fixed sensor suits repeatable trending, and a drone survey suits assets nobody can reach safely. The middle band is where a temperature becomes evidence, because a reading only means something next to a baseline and the rise above ambient. That comparison sorts a site into normal, warning and critical. The bottom band then aims the effort at bus bars, bearings or heat exchangers. The same three steps serve all three families, so the workflow is reusable even when the failure physics differ.
Electrical applications.
- Hot spots on connections indicate high resistance. Compare readings with a healthy baseline and the rise above ambient to interpret the heating.
- Overheated components indicate overload. Record load and environmental context because temperature also changes with operating conditions.
- Phase imbalance in motors. Phase-to-phase comparisons can flag electrical connection or load asymmetry for review.
- Can detect developing thermal anomalies before equipment fails, with lead time depending on the defect and inspection interval. A handheld sweep provides a snapshot; repeatable trending needs a fixed sensor or a strict capture procedure. Reviewed warnings should connect to inspection and planned maintenance.
Mechanical applications.
- Bearing overheating (friction). Compare temperature with a healthy baseline and operating context; thermal and vibration evidence can support the maintenance decision together.
- Belt misalignment (heat buildup). Review the rise above ambient and the temperature trend to assess whether the condition is stable or worsening.
- Lubrication issues (dry bearings). Tie the thermal warning to a recommended inspection so technicians can confirm the fault and record findings.
- Coupling problems. Keep the component and suspected fault mechanism connected to the evidence. Temperature changes with load and ambient conditions, so an unusual reading needs review before maintenance action.
Temperature thresholds. The following planning rubric distinguishes absolute temperature from rise above ambient. Review readings against the actual component limit and site conditions before action.
| Component | Normal | Warning | Critical |
|---|---|---|---|
| Motor bearings | <70°C | 70-85°C | >85°C |
| Electrical connections | <40°C rise | 40-70°C rise | >70°C rise |
| Gearbox oil | <80°C | 80-95°C | >95°C |
32.4 Machine Learning Models
Vibration and thermal rules handle many faults directly. When conditions vary by load, recipe, season, and maintenance history, machine learning (ML) can help, but only if the model choice matches the evidence you actually have. It can learn normal behavior and detect anomalies; its output remains evidence for review.
Figure 3 sorts model families by the labels you hold, not by the algorithm you like.
Begin at the question box in Figure 3, which asks what the maintenance team needs to know. The three branches below answer it in the order of evidence you can supply. Supervised learning comes first, and it needs labelled failures, which many sites do not have. Unsupervised learning needs only a record of normal running, so it suits a plant with clean history and few recorded faults. Forecasting sits last because it needs runs that were followed all the way to failure, the hardest data of all. Each branch also states its own question: will this asset fail soon, is this pattern odd now, how long is left.
Supervised learning requires labeled failure data to train classifiers. Random Forest and XGBoost support classification, while neural networks can learn complex patterns. The output answers a bounded question such as “Will this bearing fail in the next 30 days?” with a yes/no prediction and probability. Training needs dozens to hundreds of historical failure examples, consistent sensor data leading up to failures, and domain expertise to label failure modes. With labeled outcomes, supervised learning can support classification or regression.
Unsupervised learning learns normal operation without failure labels. Autoencoders use reconstruction error to indicate an anomaly; Isolation Forests detect outliers; one-class SVM is another option. Instead of a failure probability, the output is an anomaly score answering “Is this vibration signature abnormal?” This approach works without historical failures, can detect novel failure modes, and suits rare events. An unusual pattern still needs engineer review before action.
Time-series forecasting predicts remaining useful life (RUL) from degradation trends. The model families here include LSTM neural networks, Prophet for trend and seasonality, and Gaussian Process Regression. The output answers “How many hours or days until failure?” with an RUL estimate and confidence interval. Evaluate forecasts using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and the percentage within 10% or 20% tolerance.
RUL and time-series estimates apply only to the asset class, fault mode, and operating context represented in training data. They should include uncertainty; clean normal-running history does not establish an exact time to failure. LSTM Neural Networks explains the recurrent architecture used for time-series sensor data.
32.5 Alert to Work Order
An automotive smart-factory maintenance program usually succeeds or fails at the handoff between analytics and maintenance execution. The useful case-study pattern is not “AI predicted a fault” by itself; it is a closed loop from condition evidence to work order, inspection, repair, and model feedback.
Example scope:
- Critical conveyors, robots, compressors, pumps, spindles, and drives are ranked by failure impact.
- Vibration, current, temperature, cycle-count, and controller-state data are captured with asset id, speed/load context, and maintenance history.
- Edge gateways compute features near the machine while historians, OPC UA servers, MQTT brokers, or MES connectors, or computerized maintenance management system (CMMS) connectors move selected evidence upward.
- Maintenance planners review severity, confidence, spare-part lead time, and production windows before scheduling work.
What to measure:
- Alert precision and missed-fault rate by failure mode.
- Time from first warning to confirmed inspection.
- Planned work percentage versus emergency work percentage.
- Downtime avoided, repair hours, spare-part waste, and technician trust.
Figure 4 shows where an alert stops being an analytics output and becomes scheduled work.
Follow the numbered loop in Figure 4. The first two stages are the part most teams build first: capture the signal with its load context, then compare features against a baseline. The third stage decides whether the system is trusted, because an engineer can reject a bad alert before it costs a shift. Stages four and five turn a confirmed fault into a repair booked against spare-part lead time and a production window. The dashed return is why this is drawn as a loop rather than a line: technician outcomes go back into thresholds and models. The band underneath lists what each pass must leave behind.
The workflow must leave an auditable trail. Each alert should identify the asset, feature values, baseline comparison, severity, recommended inspection, technician disposition, replaced part, and return-to-service result.
Lesson learned: Success requires maintenance adoption, not just analytics. Technicians need enough evidence to challenge bad alerts, confirm real faults, and feed the outcome back into thresholds and models.
32.5.1 Carry the Same Evidence Across Plant Systems
Highly automated electronics plants often use the same building blocks discussed in this chapter: PLCs and Process Field Network (PROFINET) or industrial Ethernet at the machine layer, OPC UA or historian interfaces for operations data, radio-frequency identification (RFID) or traceability records for product context, and analytics that compare equipment behavior against known-good baselines.
Integration pattern:
- Production equipment emits process values, alarms, cycle counts, and quality results.
- Asset health features are tied to product, recipe, shift, maintenance event, and environmental context.
- Quality, maintenance, and operations teams review the same asset history rather than separate dashboards.
- Digital twin or simulation work is used for what-if planning, not as a replacement for measured condition data.
Business impact to verify locally:
- Fewer emergency repairs on critical bottleneck assets.
- Higher planned-maintenance ratio without excessive part replacement.
- Lower false-alarm burden for technicians.
- Better root-cause records for repeat failures.
Key Success Factor: The plant treats PdM as a maintenance decision system with measured outcomes, not as a standalone ML demo.
OPC UA Standard explains industrial interoperability for collecting this evidence. Digital Twins describes virtual equipment replicas that combine real-time sensor data with physics-based models for advanced failure prediction; what-if planning still depends on measured condition data.
32.6 ROI Calculation Framework
A reviewed alert creates value only when the resulting intervention avoids a cost. Build the business case from local costs, prevented failures, and the operating expense of keeping the system useful.
32.6.1 Price the Whole System
Investment costs (illustrative ranges; replace with local quotes):
- Sensors: $100-500 per motor (vibration, temperature)
- Gateways: $500-2,000 per zone
- Software: $50,000-500,000 (depending on scale)
- Integration: 2-5x hardware cost for brownfield
- Training: $1,000-5,000 per technician
Operating costs (illustrative ranges; replace with local contracts):
- Platform licensing: $10-50 per asset/month
- Connectivity: $5-20 per gateway/month
- Data storage: $0.02-0.05 per GB/month
- Analyst time: $50,000-100,000/year for dedicated resources
32.6.2 Tie Savings to Maintenance Outcomes
Direct savings:
- Reduced emergency repairs (labor + parts + expediting)
- Extended equipment life (deferred replacement)
- Lower spare parts inventory (order when needed)
- Reduced energy consumption (efficient equipment)
Indirect savings:
- Avoided production losses (unplanned downtime)
- Improved quality (equipment in specification)
- Reduced safety incidents (early warning of hazards)
- Better capital planning (known equipment condition)
32.6.3 Sample ROI Calculation
Scenario: 100-motor manufacturing plant using editable planning assumptions.
| Item | Value |
|---|---|
| Average motor replacement cost | $15,000 |
| Historical failures per year | 8 |
| Average downtime per failure | 12 hours |
| Downtime cost per hour | $5,000 |
| Annual failure cost | $600,000 |
With predictive maintenance:
The illustrative investment assumes a $300 sensor per motor, a $90,000 software platform and integration costing twice the sensor hardware. Select Load chapter case in the calculator to reproduce the table.
| Item | Value |
|---|---|
| Investment (sensors, software, integration) | $180,000 |
| Annual operating cost | $36,000 |
| Expected share of failures prevented | 85% |
| Prevented failures | 6.8 per year |
| Annual savings | $510,000 |
| Payback period (after annual operating cost) | 4.6 months |
Annual net savings equal (prevented failures × failure cost) − operating cost. Payback in years equals investment divided by annual net savings. For the table, $510,000 − $36,000 = $474,000 net per year, so $180,000 / $474,000 × 12 is about 4.6 months. Detection alone does not establish prevention: validated alerts must lead to completed interventions.
The separate chemical-plant example uses $200,000 upfront, $50,000/year operating cost, and five historical unplanned failures costing $80,000 each. Preventing four of those five gives $320,000 gross and $270,000 net annual savings. Payback is $200,000 / $270,000, about 0.74 years or nine months.
32.7 Implementation Roadmap
Economics provide a reason to pilot, not permission to cover the whole facility immediately. Inspect Figure 5 to see what each rollout stage must prove before more assets are added.
Prove one asset first. Expand only after measured value. Read Figure 5 from the 10–20-asset pilot to expansion and facility coverage. The pilot installs basic sensing, builds normal baselines, and must catch one previously unseen issue. Expansion to 50–100 assets adds anomaly models, CMMS work orders, and technician training; success is a measured drop in emergency work on those asset classes. Only then does the final stage add remaining-life forecasts, parts automation, and continuous model improvement. Coverage grows after alert quality and maintenance follow-through are stable, not simply when the model emits more warnings.
32.7.1 Phase 1: Pilot (Months 1-6)
- Select 10-20 critical assets. Rank them by failure impact and identify the fault mechanism and maintenance action that sensing should support.
- Deploy basic vibration and temperature sensors. Combine evidence where it helps expose the target fault under the asset’s operating conditions.
- Establish data collection infrastructure. Capture asset identity, speed and load context, and maintenance history alongside condition data.
- Create baseline normal operation profiles. These provide the reference needed to distinguish normal operating changes from potential faults.
- Success metric: Detect one previously undetected issue. Use technician findings to confirm alert quality, and expand only after maintenance follow-through and measured value support the decision.
32.7.2 Phase 2: Expansion (Months 7-18)
- Expand to 50-100 assets. Add coverage after the pilot demonstrates stable alert quality, maintenance follow-through, and measured value.
- Implement ML-based anomaly detection. Match the model to available evidence and retain operating context so normal changes are not confused with faults.
- Integrate with CMMS for work order generation. Reviewed alerts should carry asset identity, baseline comparison, severity, and recommended inspection into maintenance planning.
- Train maintenance technicians on new tools. Their findings help confirm faults and feed outcomes back into thresholds and models.
- Success metric: measured reduction in emergency work on pilot asset classes. Track maintenance outcomes to assess whether expanded monitoring supports planned repair.
32.7.3 Phase 3: Optimization (Months 19-36)
- Full facility coverage (all critical assets)
- Remaining useful life predictions
- Automated parts ordering
- Continuous model improvement
- Success metric: sustained planned-maintenance ratio and technician-confirmed alert quality
32.8 Establish the Baseline Before Enabling Alerts
Consider the failure pattern: A factory installs vibration sensors on 50 motors and immediately expects anomaly alerts. After 2 weeks, they get zero alerts and assume the system is broken — or worse, they tune sensitivity so high that false alarms overwhelm maintenance.
The reason is straightforward: Machine learning models need to learn “normal” before detecting “abnormal.” Each motor has a unique vibration signature based on its age, mounting, load, and environment. Without baseline data, the model has no reference.
The chapter’s food-processing example makes the consequence concrete: A food processing plant deployed predictive maintenance sensors on 30 pumps. They expected immediate failure predictions. Instead, they got alerts on pumps that had run the same way for 10 years. The “anomalies” were just normal operating characteristics the model hadn’t seen yet.
Build the baseline in a deliberate sequence:
- Run in learning mode for 4-8 weeks to establish baseline per motor
- Capture full operating envelope: startup, shutdown, light load, heavy load, seasonal variations
- Label known-good periods in training data; exclude startups and maintenance events from the steady-operation baseline while retaining their context separately.
- Tune thresholds after baseline — start conservative (only flag extreme deviations)
- Continuous retraining as equipment ages (bearing wear shifts baseline)
Timeline:
- Weeks 1-4: Passive data collection, no alerts enabled
- Weeks 5-8: Model training on baseline data, internal validation
- Week 9: Enable alerts at conservative thresholds (low sensitivity)
- Weeks 10-16: Adjust thresholds based on technician feedback
- Month 4+: Confidence in predictions, adjust sensitivity upward
Rushing to production without baseline data causes alert fatigue (“boy who cried wolf”) that destroys user trust. Technicians who ignore 10 false alarms will ignore the 11th real one. The 4-8 week investment in baseline data pays for itself by preventing trust erosion.
Vibration, temperature, current, and pressure change with speed, load, recipe, ambient condition, and recent maintenance. Store that context with each feature window. A high anomaly score still needs alert review, work-order creation, technician disposition, part condition, a return-to-service record, and a model or threshold update.
32.8.1 Planning Inputs To Localize
| Input | Why It Matters |
|---|---|
| Failure cost | Sets the value ceiling for prevented failures |
| Downtime hours and rate | Converts a technical failure into business impact |
| Sensor and installation cost | Determines whether the asset is worth instrumenting |
| False-positive burden | Controls technician trust and inspection workload |
| Spare-part lead time | Defines how early the warning must arrive |
32.8.2 Vibration Frequency Signatures
| Defect | Common Frequency Clue | Design Note |
|---|---|---|
| Imbalance | 1x shaft speed | Compare against speed/load baseline |
| Misalignment | Often strong 2x component | Confirm with axial/radial measurements |
| Bearing defects | BPFO/BPFI bands and harmonics | Requires bearing geometry and good mounting |
| Gear mesh | Tooth count x shaft speed | Sidebands and load context matter |
32.8.3 Temperature Review
| Check | Why It Matters |
|---|---|
| Rise above ambient | Separates equipment heating from room-temperature change |
| Phase-to-phase imbalance | Flags electrical connection or load asymmetry |
| Trend slope | Identifies whether the condition is stable or worsening |
| Component limit | Keeps decisions tied to the actual device rating |
The Vibration Analysis Sensors reference covers microelectromechanical systems (MEMS) accelerometer specifications for industrial predictive maintenance: 100-1000 Hz sampling and ±50g range. Fast Fourier transform (FFT) analysis transforms time-domain vibration data to frequency-domain bearing-defect harmonics. Data Storage and Databases explains time-series design, including InfluxDB and TimescaleDB, for high-frequency vibration data (100-1000 Hz) with millisecond timestamps. OPC UA extracts vibration, temperature and power data from PLCs for predictive models.
Summary
Predictive maintenance creates value when observable failure modes connect to closed maintenance workflows. Reactive, preventive, and predictive strategies trade emergency repair, planned replacement, condition evidence, and downtime risk differently. Vibration, thermal, acoustic, current, pressure, oil, and controller-state signals help only when they expose the target fault under the asset’s operating conditions.
Supervised models need labeled outcomes, anomaly models need disciplined baseline review, and RUL models need degradation histories for the relevant asset class and fault mode. Start with the asset, component, fault mechanism, consequence, detection method, and maintenance action. Then prove alert quality against technician findings and measure the workflow and economics before expanding. Cultural change, technician training, and organizational commitment matter alongside technology.
A model warning becomes maintenance value through confirmed evidence, a completed intervention, and a measured outcome. Expand coverage only after that loop works.

