AIAI in Energy & UtilitiesEnergy & Utilities

Read how Duke Energy cut turbine failures using sensor AI

Duke Energy built one of the most operationally consequential AI deployments in U.S. generation by wiring sensor data into machine learning models that flag failures weeks before they happen. The mechanics of what they did, and what transfers to your assets, are worth examining closely.

Neo NeumannAI Practice LeadSeptember 30, 2026

Listen to the podcast

4 min

Chapters

Key takeaways

Pick your three most expensive failure modes and pull historian data on the last ten occurrences to look for patterns in the weeks before.
Label past failures in your data historian first, since without labelled history the model has nothing to learn from.
Set false positive thresholds by asset value: tolerate more false alarms on a turbine, skip the effort on a cheap pump.
Check the sensor sampling rate on your most critical asset; if it is minutes rather than seconds, no model can see the early signal.
Treat vendor deployment and benchmark claims as marketing until an operator or your own pilot confirms them.

Read the full transcript

Host:Welcome back to Leaders Insights. Read how Duke Energy cut turbine failures using sensor AI, and why it matters this week. Duke Energy just went public with the numbers, and they're the kind that make plant managers sit up. Weeks of warning before a turbine dies. What did they actually build?

Expert:They wired vibration and temperature sensors on their generation fleet into machine learning models — software that learns patterns from data instead of following fixed rules — and trained those models to spot the signature of a failure before it happens. The headline they're pushing is early detection two to three weeks out.

Host:Two to three weeks. Let's pull that claim apart, because "we caught it early" is what every vendor says right before the demo breaks.

Expert:Fair, and here's the part that's real. The value isn't the prediction, it's the lead time against the maintenance calendar. A turbine that fails unplanned costs you replacement power on the spot market — sometimes at peak prices — plus emergency crews. Push that same repair into a scheduled outage and you're paying a fraction. The model earns its keep in the gap between "surprise" and "Tuesday."

Host:So the artefact we're dissecting is that sensor-to-model pipeline. Start at the front. What's actually feeding it?

Expert:High-frequency sensor streams — think thousands of readings per second on bearing vibration, oil temperature, rotor speed. That's the strong part. Turbines are heavily instrumented already, so Duke wasn't installing hardware from scratch, they were finally using signals they'd been throwing away for years.

Host:Throwing away being the polite word for "logging to a database nobody opened."

Expert:Exactly that. Most utilities have a data historian — an archive of decades of sensor readings — sitting cold. Duke's real move was labelling it: going back through past failures and tagging what the sensors looked like in the days before each one. Without that labelled history, the model has nothing to learn from. That's the unglamorous 80% of the work.

Host:Where does it get shaky?

Expert:The failure modes it can't see. A model trained on bearing wear and thermal drift will catch bearing wear and thermal drift. Hit it with a failure type that never appeared in the training data — a manufacturing defect, a one-off — and it stays silent right up to the bang. Predictive maintenance is only as smart as your history is complete, and no fleet's history is complete.

Host:How do they keep it from crying wolf? Because a model that flags a failure every third Tuesday gets ignored by week two.

Expert:That's the failure that kills these projects. If false alarms are high, crews stop trusting the alert and you're back to run-to-failure. The tuning question is where you set the threshold — do you tolerate more false positives to never miss a real one, or the reverse? For a turbine where an unplanned trip costs millions, you lean toward more false alarms. For a cheap pump, you don't bother.

Host:Let's talk tooling, because I've seen the sources. Some of these platform figures come from vendors.

Expert:Right, and I'll flag it. A lot of the "ease of deployment" numbers floating around cite platforms like Hugging Face — that's a company that hosts machine learning models and sells the tooling to run them. Useful, but they're selling the shovels, so treat their "deploy in days" claims as marketing until an operator confirms it. Same with OpenAI's benchmark figures on model performance — they're an AI lab with product to move. Cross-check against your own pilot, not their slide.

Host:So a listener running a mid-size plant. What actually transfers, and what's Duke-only because they've got a research budget you don't?

Expert:What transfers: pick your three most expensive failure modes, pull the historian data on the last ten times they happened, and see if there's a visible pattern in the weeks before. You can do that with a data scientist and a spreadsheet before you spend a cent on a platform.

Host:And what doesn't?

Expert:The fleet-wide scale. Duke can pool patterns across dozens of identical turbines. One plant with three units doesn't have the volume, so you lean harder on physics-based rules and less on pure learning.

Host:Give me the one thing to do Monday.

Expert:Open your data historian and check one thing — what's your sensor sampling rate on your most critical asset. If it's minutes, not seconds, you're already blind to the early signal, and no model fixes that. Fix the sampling first.

Host:Sources for today's episode: TechCrunch AI, The Decoder, Ars Technica AI, Hugging Face (vendor — AI platform), OpenAI (vendor — AI lab). That's a wrap. Fresh AI briefings drop daily at mba-training.com.

Duke Energy operates roughly 50,000 megawatts of generation capacity across six states, under FERC oversight at the transmission level and multiple state commissions at the distribution level. By 2023, the company was facing a familiar but expensive problem: unplanned outages on combustion turbines at natural gas peakers were costing far more than scheduled maintenance windows, both in direct repair costs and in NERC reliability exposure. A single forced outage on a large generating unit can trigger compliance scrutiny under NERC FAC and TOP standards, and the associated replacement power costs on PJM or MISO markets can run into the hundreds of thousands of dollars per day. The business case for doing something different was not theoretical.

How Duke Energy built its sensor-based fault detection program

Duke's approach started with the instrumentation that was already in place. Modern gas turbines carry hundreds of sensors monitoring vibration, bearing temperatures, exhaust profiles, fuel flow ratios, and inlet conditions. The data existed; what was missing was a model layer that could distinguish a meaningful deviation from normal thermal cycling noise. Duke partnered with SparkCognition, an industrial AI vendor, to deploy machine learning models trained on historical sensor telemetry correlated with past failure events. The company's own engineers were central to the labeling process, identifying which prior anomalies had preceded actual failures and which had been false positives that wasted crew time.

The architecture was not a single monolithic model. Different asset classes, gas turbines, hydro generators, and high-voltage transformers, each required separate feature engineering because the physics of degradation differ. A transformer approaching insulation failure shows a different chemical signature in dissolved gas analysis than a turbine bearing about to fail shows in vibration frequency. Duke's team built asset-specific models and ran them against incoming SCADA and historian data in near real time, with alert thresholds calibrated to maintenance scheduling windows, giving dispatchers enough lead time to order parts and schedule outages during off-peak demand periods rather than emergency windows.

One decision that made the program work in practice: the output was surfaced to human engineers as a risk score with contributing factors, not as a binary "fail or not" flag. This matters for regulatory and liability reasons as much as operational ones. If a utility takes a unit offline based on an AI recommendation and that action affects grid reliability commitments, there needs to be a defensible audit trail.Understanding what NERC reliability standards actually demand from generators is not optional background for anyone deploying predictive maintenance in this sector; it shapes what the AI system must document, not just what it must predict.

What did Duke Energy's predictive maintenance program actually deliver?

Duke has publicly attributed meaningful reductions in unplanned outages to this program, though the company has been careful about specific percentage claims in earnings communications, which is worth noting. Reported outcomes include detecting a compressor blade degradation issue on a combined-cycle unit approximately six weeks before it would have caused a forced outage, allowing a planned replacement during a scheduled spring maintenance window. The avoided cost in that single event, counting replacement power procurement, crew overtime, and parts expediting, was estimated internally at over $2 million. Duke has not published a consolidated program-wide ROI figure in SEC filings, so portfolio-level numbers circulating in vendor marketing materials should be treated with appropriate skepticism.

What is independently verifiable: Duke's 2024 and 2025 annual reports showed continued capital allocation toward grid modernization and predictive analytics, and the company expanded the program beyond gas generation into substation transformer monitoring across its Carolinas territory. The transformer application is particularly significant given that large power transformers carry lead times of 12 to 18 months from order to installation; catching a failure in the incipient stage is the only scenario where the utility can realistically avoid a prolonged outage.

What transfers from Duke's case to your assets, and where it breaks down

The core lesson is sequencing: Duke spent significant time on data quality and historical labeling before building models, not after. Utilities with fragmented historian systems, mixed SCADA vintages, or gaps in failure event documentation will struggle to replicate this approach without adata readiness investment that most AI vendors understate in their sales process. A model trained on incomplete or mislabeled failure histories will generate false positives that erode operator trust, and once maintenance crews start ignoring alerts, the program is functionally dead regardless of technical accuracy.

Context differences matter. Duke operates under vertically integrated utility structures in the Carolinas, where capital expenditure on O&M technology can be argued into rate cases before the NCUC and PSCSC. A merchant generator operating in a competitive wholesale market like ERCOT does not have that regulatory cost recovery path; the ROI calculation is purely operational and the risk tolerance for false positives is different because there is no captive ratepayer to share the cost. Municipal utilities and cooperatives face a third constraint: their SCADA infrastructure is often older, and the sensor density on distribution assets may be too low to support the same model architecture Duke used on its large generation fleet.

For transmission asset owners specifically, FERC Order 881 requirements on ambient-adjusted line ratings have pushed more utilities toward real-time thermal monitoring of lines and conductors since 2025. That instrumentation, already being deployed for compliance, creates a secondary data stream that predictive maintenance models can use with relatively low incremental cost.

The program also required Duke to resolve an organizational question before a technical one: who owns an AI-generated maintenance recommendation when it conflicts with a senior engineer's judgment? Duke formalized this in its asset management governance structure. Without that clarity, the AI output becomes a political object rather than an operational input, and it will lose every time there is disagreement.

The single most transferable point from Duke's experience is that the value is not in the model; it is in the lead time the model creates. Six weeks of advance warning converts an emergency repair into a planned outage, and that conversion is where almost all the financial return concentrates.

The full course on this sector:AI in Energy & Utilities.

Frequently asked questions

How much sensor data does a utility actually need before training a predictive maintenance model?

Predictive maintenance AI for generation assets typically requires at least two to three years of continuous sensor telemetry correlated with documented failure events to produce reliable fault signatures. Without historical failure labels, models cannot distinguish meaningful degradation from normal operating variation, which is why Duke Energy's engineers spent significant time on the labeling process before any model training began.

Can predictive maintenance AI be used on distribution assets like poles and switches, not just large generators?

Yes, but the data infrastructure requirements are harder to meet at the distribution level. Large generation assets like gas turbines carry hundreds of sensors by design, while distribution equipment is often minimally instrumented, making it difficult to build the same feature-rich models Duke used. Utilities are increasingly adding smart sensors to distribution assets as part of grid modernization capital programs, which will change this calculus over the next several years.

Does an AI-generated maintenance recommendation create any NERC or FERC compliance exposure for the utility?

It can, which is why audit trail design matters as much as model accuracy. If a utility takes a generating unit offline or derates it based on an AI recommendation, that action must be defensible to NERC reliability coordinators and, in some cases, to state commissions reviewing O&M practices in rate proceedings. Duke addressed this by presenting AI outputs as risk scores with supporting evidence rather than automated commands, preserving human authorization at each step.

How do merchant generators justify predictive maintenance AI without the rate case cost recovery that regulated utilities use?

Merchant generators must build the ROI case entirely on avoided forced outage costs and capacity market performance. In competitive markets like PJM or MISO, a forced outage during a high-demand period means buying replacement energy at real-time prices that can be multiples of the generator's marginal cost, so the avoided-outage value can be large. The challenge is that the business case is harder to present to a CFO before the first avoided failure occurs, which makes piloting on the highest-risk assets first the standard approach.

Go deeper

The lessons that take this article further, free to read.

Sources

Finished reading?

Validate your read to earn XP and feed your radar.

AI articles Back to Blog