What problem does it solve?
Plants and power stations run on pumps, compressors, valves, turbines, motors and conveyors that wear out. Many are maintained on a fixed calendar or run until they fail. Calendar maintenance replaces parts that still had life in them; run to failure means an unplanned stop, emergency parts at premium prices, lost production and, in energy, safety and environmental risk.
The signals of an approaching failure are usually there: a bearing runs slightly warmer, vibration creeps up, a valve takes longer to close. But a large site has thousands of sensors, and the few experienced engineers who can read those patterns cannot watch all of them. Some companies also expect to lose that knowledge: Georgia-Pacific said many of its site experts were "retiring soon" (AWS). Predictive maintenance lets models watch every asset continuously and send the experts only the cases that need their judgment.
How does it work?
- Collect the signals. Temperature, vibration, pressure, flow, current and process data stream from the control systems and historians into one data platform.
- Learn normal behaviour. For each asset, a model learns how its readings relate to each other under normal operation, or is trained on labelled past failures where they exist.
- Detect and predict. The model flags deviations early and, where history allows, estimates the likely failure mode and the time left, weeks or months ahead.
- Triage centrally. Analysts in a central monitoring and diagnostics centre review early warnings, set aside false alerts and confirm real issues with the site. At the time of the AVEVA story, Duke Energy ran such a centre, with five analysts, for over 87% of its generating fleet (AVEVA). At most Shell assets, a remote engineer vets each alert before it goes to the asset engineers (Shell).
- Plan the work. Confirmed issues become prioritised work orders with the parts and the window for the repair, planned into the next outage instead of forcing an unplanned one.
- Feed back. What the technicians find on the equipment is recorded against the alert, so the models and thresholds improve.
- Audience
- Employee facing
- Autonomy
- Assist
- Adoption
- Early adopters
- Channels
- Internal tools, API and system to system
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Cost savings | Not pooled | at least USD 34 million | 1 | 1 vendor |
Value drivers: Risk and loss reduction, Lower cost to serve, Speed and cycle time, Employee productivity.
Indicative value
A process plant or power station with 150 critical rotating and process assets
USD 200,000 to USD 2.7 million
Unplanned downtime cost avoided per year
How this is calculated
Formula: downtimeHours * costPerHour * avoidedShare. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Unplanned downtime hours per year on critical assets downtimeHours, hours per year | 100 | 150 | Editorial assumption, replace with your own downtime records. |
| Cost of one hour of unplanned downtime costPerHour, USD per hour | 20,000 | 60,000 | Editorial assumption covering lost production, emergency repair and restart. Replace with your own figure. |
| Share of unplanned downtime avoided or converted into planned work avoidedShare, fraction of downtime hours | 0.1 | 0.3 | Editorial assumption. The evidence on this page reports one early catch at Duke Energy that AVEVA says avoided more than 34 million US dollars in cost for that single event (AVEVA), but no fleet wide downtime reduction, so the range stays conservative. |
What it leaves out: Counts avoided downtime only. It leaves out sensors, data platform and analyst costs, the cost of the planned repairs that replace breakdowns, savings from fewer unnecessary calendar maintenance tasks, and the safety and environmental value of avoided failures.
Who already uses it?
3 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Shell
United Kingdom · Energy and utilities · 2022
Shell runs a predictive maintenance programme built on the C3 AI platform. Machine learning models flag equipment degradation and likely failures early so operators can intervene before unplanned downtime, production interruptions or safety and environmental risks; C3 AI names control valves, pumps and compressors among the monitored equipment. In a March 2022 C3 AI press release, Shell's Dan Jeavons said that "Monitoring 10,000 pieces of critical equipment" with AI predictive maintenance was a target Shell had set for 2021 and achieved. The wording "more than 10,000", the asset scope and the technical figures in the release are C3 AI's. In its own TechXplorer Digest article, Shell describes the rollout asset by asset (a Dutch refinery in 2020, where the models flagged 65 control valves in need of repair that traditional methods would have missed, then Singapore, the USA and Canada in early 2021) and a process in which, for most assets, a remote engineer vets each alert before it reaches the asset engineers.
No outcome disclosed.
Georgia-Pacific
United States · Manufacturing · 2019
Georgia-Pacific, a pulp, paper and building products manufacturer, streams data from equipment at its North American facilities into an operations data lake on AWS and analyses it with an AWS based advanced analytics solution that includes Amazon SageMaker machine learning models. For selected assets the company can now predict equipment failure 60 to 90 days in advance, so it can plan equipment downtime instead of suffering unscheduled production stoppages. No outcome figure is published for the failure prediction; the same data platform also runs process optimization models for converting line speeds, which are outside this use case.
No outcome disclosed.
Duke Energy
United States · Energy and utilities · 2016
Duke Energy runs a central Monitoring and Diagnostics (M&D) centre that watches coal, gas, combined cycle and other generating units in several US states with predictive asset analytics software; the AVEVA page names PRiSM Predictive Asset Analytics among its tools. Early warning notifications of equipment problems let a small team of experienced analysts alert the plants before a failure. AVEVA reports that the centre covers over 87% of Duke's generating fleet with over 11,000 models, and that a single early catch in 2016 avoided more than 34 million US dollars in cost had the problem gone undetected.
- Cost savings: at least USD 34 million, a single early catch event in 2016, counterfactual avoided cost
"Savings of over $34 millions in a single early catch event in 2016."
Claimed by: vendor
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Historian or control system data per asset at a useful sampling rate
- An asset register with criticality, failure modes and maintenance history
- Past failures and work orders linked to the sensor data where possible
- Engineering knowledge of normal operating ranges per asset type
Systems to integrate
- Plant historian and control systems (SCADA, DCS)
- Data platform for streaming and model training
- Enterprise asset management or computerised maintenance management system for work orders
- Alerting to the monitoring centre and site teams
Complexity: High
The models are well understood; the hard parts are clean sensor data from many different control systems, too few recorded failures to learn from, and changing the maintenance process so that alerts actually become planned work.
- 1
Start with critical assets and known failure modes
Rank assets by the cost and risk of failure and pick one asset class with sensor coverage and a few documented failures, such as large pumps or compressors.
- 2
Get the data flowing and trusted
Connect the historian, fix tag names and units, and agree with the site which readings are reliable. Many first alerts turn out to be sensor faults, not machine faults.
- 3
Stand up a central monitoring team
Put a small group of experienced engineers between the models and the sites to triage alerts, so the sites only see confirmed issues.
- 4
Wire alerts into maintenance planning
Create work orders in the maintenance system from confirmed alerts, with a priority and a repair window, and record what the technician found.
- 5
Scale by asset class, then by site
Reuse models and templates across identical equipment, track avoided failures with a written case per catch, and widen coverage one asset class at a time. Shell set itself a target of 10,000 monitored pieces of critical equipment for 2021 and reported reaching it.
Guardrails
- Alerts advise; protection systems and trips stay in the certified control and safety systems
- A human engineer confirms every alert before a work order or shutdown is raised
- Model changes tested against past data before release, with version history
- Sensor health checks so that faulty instruments are not read as failing machines
KPIs to instrument
- Unplanned downtime hours on monitored assets, before and after
- Documented early catches and their estimated avoided cost
- Share of alerts confirmed as real issues
- Lead time between first alert and failure or repair
- Failures on monitored assets that the models missed
Human in the loop
Monitoring and diagnostics engineers triage every alert and decide whether it becomes work. Site maintenance planners choose when to repair, and reliability engineers review missed failures and false alarms each month to tune the models.
Common failure modes
- Alert fatigue
- Too many alerts with too little context and sites stop reacting. Triage centrally and report the confirmation rate.
- No link to the maintenance process
- Alerts land in a dashboard nobody plans from. Create work orders from confirmed alerts in the maintenance system.
- Too few failures to learn from
- Critical assets rarely fail, so supervised models lack examples. Use anomaly detection on normal behaviour and engineering rules alongside.
- Savings nobody believes
- Claimed savings that cannot be traced to a specific catch lose credibility. Write up each early catch with what would have happened.
What are the risks and rules?
EU AI Act
Depends on design
A system that advises engineers on the condition of equipment is usually minimal risk. Annex III point 2 lists AI systems intended as safety components in the management and operation of critical digital infrastructure, road traffic and the supply of water, gas, heating or electricity; if predictive maintenance acts on protection or control in a utility network, it can become high risk. Article 6(1) can also apply when the AI is a safety component of machinery or another product covered by Annex I legislation and that product must undergo a third party conformity assessment.
Guidance
- Annex III, high risk AI systems referred to in Article 6(2) (European Union, Europe). Point 2 covers safety components in the management and operation of critical infrastructure, including the supply of water, gas, heating and electricity.
- Article 6, classification rules for high risk AI systems (European Union, Europe). Explains when an AI system that is a safety component of a product under Annex I legislation, such as machinery, is high risk.
- AI Risk Management Framework (NIST, North America). Voluntary framework to map, measure and manage the risks of AI systems, useful for documenting model limits and monitoring.
Controls to put in place
- Asset and model inventory with an owner per model and the assets it covers
- Documented separation between advisory models and certified protection systems
- Written record of every confirmed catch and every missed failure
- Cybersecurity controls on data flows from operational technology, in line with NIS2 where it applies
Frequently asked questions
- What results do companies report from predictive maintenance?
- AVEVA reports that a single early catch by Duke Energy's monitoring and diagnostics centre in 2016 avoided more than 34 million US dollars in cost had the problem gone undetected; that is one event, not a yearly saving. AWS reports that Georgia-Pacific can predict failure of selected assets 60 to 90 days ahead, without an outcome figure. Shell writes that at one Dutch refinery its models flagged 65 control valves in need of repair that traditional methods would have missed. Published fleet wide downtime figures are rare, so measure your own avoided failures case by case.
- How many assets can one programme cover?
- In a March 2022 press release, C3 AI said Shell's predictive maintenance programme on its platform monitors more than 10,000 pieces of equipment; Shell, quoted in the release, called monitoring 10,000 pieces of critical equipment a target set for 2021 and achieved. Coverage grows asset class by asset class, reusing models across identical equipment.
- Is predictive maintenance high risk under the EU AI Act?
- Usually not while it advises engineers. It can become high risk if it acts as a safety component in the supply of water, gas, heating or electricity (Annex III point 2), or a safety component of machinery covered by Annex I legislation that is subject to third party conformity assessment (Article 6(1)).
How to cite this page
Blits.ai AI Use Case Library, "AI predictive maintenance for industrial and energy assets", last verified 27 September 2026, https://www.blits.ai/ai-use-cases/industrial-asset-predictive-maintenance. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published