Scaling Detection in a Condensing Steam Turbine Using Physics-Guided Machine Learning

Introduction

Scaling in condensing steam turbines is the accumulation of crusts on blading, diaphragms, and exhaust hoods. It begins when impurities carried with steam precipitate onto wetted surfaces as the vapor expands and cools below ambient pressure. Dissolved solids from boiler carry-over, impure attemperation spray, or volatile silica oversaturate condensate micro-films, crystalising into tenacious scale.

Once established, deposits roughen blade profiles, disturb boundary layers, and choke throats, forcing steam to deviate from its designed velocity triangles. The immediate consequence is a drop in efficiency and capacity, reflected by a rising heat-rate, condenser back-pressure, and wheel-chamber pressures. Asymmetrical build-up further unbalances the last-stage rotor, amplifying vibration over time and encouraging fatigue at blade roots. Left unchecked, corrosion-under-deposit sets in, fragments spall downstream, and maintenance shifts from on-load water rinse to prolonged chemical cleaning or blade replacement.

Conventional heat-balance trending or walk-down inspections register the damage only after a decline in profitability. Performance monitoring and modern machine learning approaches have their own challenges with sensor limitations, the lack of labeled data for classification models, and difficulty distinguishing between degradation within the turbine versus issues with driven equipment. This case study describes a hybrid approach that combines thermodynamic performance monitoring, supervised machine learning, and physics-guided machine learning to enable a scaling detection system that overcomes many of the inherent challenges.

Challenges

Threshold & Performance Monitoring Limitations

Traditional approaches to detection struggle to expose early scaling. Wheel-chamber pressure, for example, is an important indicator of flow restriction, yet a single transmitter must infer the behavior of an entire stage row; drift or impulse-line plugging can trigger false alarms, while modest scaling on the opposite side of the rotor goes unnoticed until efficiency is already lost. Flow-rate uncertainty compounds the problem: with flow rates either unavailable or susceptible to accuracy loss due to deposits or drift.

Mechanical-drive units add another blind spot; there is no generator to provide a clean kilowatt output signal. Without shaft-power feedback, classic performance indices such as isentropic efficiency or heat-rate must be estimated, not measured, widening confidence bands and masking the small changes that mark incipient fouling.

Even where a full thermodynamic model is available, it hinges on exhaust enthalpy, yet wetsteam quality is rarely instrumented. Assuming a constant dryness fraction forces analysts to attribute every additional kilojoule of drop to scaling when it may simply reflect wetter exhaust. Finally, condenser conditions resemble a moving target: a modest rise in backpressure caused by tube fouling or cooling-water temperature can mimic the same power loss created by blade deposits. Threshold rules therefore oscillate between nuisance trips and late warnings, proving that reliable scaling detection demands more context than conventional tactics can deliver.

‍

Machine Learning Challenges

Machine-learning ML) promises to detect turbine scaling without first-principles models, yet are exposed to some stubborn limitations.

Labeled data is scarce: Scaling evolves slowly, is removed during outages, and is rarely instrumented with direct verification such as borescope photos. Supervised classifiers therefore lack the fault tags needed to separate “clean” from “fouled” operating points, forcing practitioners toward weakly supervised or fully unsupervised techniques.

Interpretability remains poor: Popular multivariate anomaly detectors—autoencoders, isolation forests, variational recurrent networks—may flag a deviation in the highdimensional sensor space, but they seldom explain whether the culprit is crust in last-stage blades, condenser vacuum drift, or a drifting pressure transmitter. Operators confronted with opaque “red scores” become skeptical, tune the sensitivity down, and miss the early window for corrective washing.

Condensing service multiplies confounding factors: The thermodynamic signature of light scaling (a two-per-cent heat-rate rise, subtle wheel-chamber pressure increase) mirrors

seasonal swings in cooling-water temperature or a partial condenser tube blockage. An ML model must see thousands of operating hours across the full envelope of condenser conditions to learn the true baseline, a volume of data that many plants—and especially new builds—do not yet possess.

Mechanical-drive configurations obscure the target: Without generator output, shaft power is estimated from torque meters or inferred from driven-equipment conditions. ML algorithms therefore struggle to decide whether declining efficiency is rooted in turbine fouling or driven equipment issues, producing cross-equipment false positives.

Together these gaps underscore why ML alone rarely delivers reliable scaling alarms; additional physics context is needed to supply labels, improve explainability, and disentangle turbine health from its ever-changing boundary conditions.

‍

Equipment Overview

Asset:

A Condensing Steam Turbine driving a centrifugal compressor.

Instrumentation Gaps:

• No direct power/torque measurement
• No throttle steam flow rate
• No exhaust steam quality
• Variable exhaust steam conditions

Available Instrumentation:

• Steam Turbine: inlet pressure and temperature, exhaust pressure and temperature, wheel chamber pressure, overall radial vibrations, turbine speed, throttle steam valve differential pressure, and valve position feedback
• Centrifugal compressor: suction pressure and temperature, gas composition, discharge pressure and temperature, suction flowmeter differential pressure, compressor speed

To overcome these challenges, a hybrid solution was implemented, leveraging both physics-based modeling principles and machine learning models, each compensating for weaknesses inherent in the other when deployed alone.

Technical Solution

Solution Architecture

Multi-layer Modeling Approach:

The following modeling approach was utilized to account for the absence of important sensor data and the inherent challenges that come with condensing steam turbines:

1. Physics-guided machine learning for steam turbine power prediction
2. Supervised machine learning for wheel chamber pressure prediction
3. Supervised machine learning for radial vibration prediction
4. Physics-based performance models for the steam turbine, centrifugal compressor, steam throttling valve, and the compressor suction flowmete

‍

MODEL 1:

Physics-Guided Machine Learning Model for Steam Turbine Power

A crucial requirement for scaling detection is the ability to detect when the steam turbine output power is dropping at given inlet steam conditions (pressure, temperature, and flow). In this application, historical data from a time where the centrifugal compressor condition was known to be acceptable was used to perform thermodynamic calculations to estimated compressor gas power. This synthetic gas power was then utilized with historical steam turbine process data from the same timeframe in a supervised machine learning model for steam turbine power prediction.

Crucially, this predictor incorporated established centrifugal compressor performance curves as embedded physics constraints, ensuring that inferred power estimations remain physically plausible absent direct torque/power measurement instrumentation.

‍

MODEL 2:

Wheel Chamber Pressure Prediction Model

A supervised machine learning model for predicting wheel chamber pressure was created using historical data at healthy system conditions. The intent of this model was to provide an independent means of evaluating steam turbine condition. Combining information from multiple independent models helps increase the confidence in model findings and reduce false positives.

‍

MODEL 3:

Radial Vibration Prediction Model

Historical data was also used to create multiple models for the prediction of overall, unfiltered, peak-to-peak vibration amplitudes for the steam turbine radial vibration probes. A significant advantage of using ML models to predict overall vibrations is that the dependence of vibrations on parameters such as load, lube oil pressure, and temperature can be ‘learned’, with deviations providing very early indications of issues such as varnish or scaling. Once deviations have been detected early, high-speed vibration spectra can be utilized to correlate findings with the underlying physical phenomena. Evaluating the data in conjunction with data from physics models and other machine learning models provides further insights into potential underlying causes.

‍

MODELS 4-7:

Physics-based Performance Models

Robust physics-based performance models were configured for the following equipment:

• Steam Turbine
• Centrifugal Compressor
• Inlet Steam Throttling Valve
• Compressor suction venturi flowmeter

The models were baselined on historical data and used to compare actual performance to baseline performance. A key aspect of physics-based performance models is to have as many independent

methods of calculating a variable as possible. For example, estimating steam mass flow using control valve CV, chest pressure, and power balance gives three independent values. Matching values increases confidence in the calculated variable, whereas differences can be correlated to issues such as sensor drift.

‍

ORCHESTRATION & DIAGNOSTICS

Data flow orchestration and robust fault diagnostics are crucial elements in converting performance calculations and model predictions into actionable insights. The diagnostic engine continuously ingests all relevant real-time feeds while referencing historical baselines established during healthy operations periods:

1. Each asset-level model independently predicts its target variable(s);
2. Results feed into higher-order logic evaluating cross-component consistency, for example, confirming that detected wheel chamber pressure anomalies vary logically with inferred shifts in compressor/turbine/control valve performances;
3. Multiple independent “votes” against common failure hypotheses enable robust fault confirmation, even if some sensor signals degrade/fail entirely.

By orchestrating information across physics-guided predictors and independent machine learning anomaly detectors, with embedded domain expertise hardcoded throughout, the resulting platform maintains resilience against bad/missing sensors while minimizing nuisance alarms through multi-signal validation protocols.

Results

The following results demonstrate how this integrated approach reliably identified, and quantified, the development and remediation cycles associated with turbine scaling over an

extended observation period:

<insert image>

A gradual increase in wheel chamber pressure deviation was observed over months. Partial recovery followed cleaning activities; complete recovery occurred after blade replacement. The confidence in the deviations was further strengthened due to similar findings from the power prediction model that showed an increased steam rate, that is, more steam utilization for the same turbine output power.

A parallel trend was captured via estimated exhaust steam quality metrics generated by performance models configured specifically for condensing service environments. Once again, the increase in estimated exhaust steam quality over time, eventually reaching a value of 1, indicated scaling which when evaluated in conjunction with the wheel chamber pressure prediction and output power prediction provided high confidence in the outcome.

This unambiguous correlation between increasing wheel chamber deviation, confirmed by independent power prediction and physics models as well as physical inspection outcomes validates both sensitivity and specificity advantages provided by fusing physical modeling constraints into anomaly prediction workflows.

Summary

Direct evidence from this deployment confirms several unique benefits compared with legacy methods:

Causality via Performance Models: By comparing actual versus estimated outputs from rigorously constructed first-principles benchmarks, rather than relying solely on correlational statistics, it becomes possible not only to flag anomalies but also to identify root causes that are directly traceable back through asset subcomponents.

Historical Context through Machine Learning: Supervisory predictors leverage deep contextual awareness drawn from long-duration plant history, allowing rapid adaptation whenever “normal” operating envelopes shift due to either process adjustments or gradual wear-out phenomena.

Embedded Domain Knowledge = Larger Fault Universe: Physics guidance extends diagnosable event coverage beyond what purely black-box AI systems can achieve, offering targeted insight into rare/unusual failure modes otherwise invisible amid routine variability.

Tolerance to Incomplete/Missing Data: Cross-validating predictions among parallel input channels ensures that individual sensor drift or failure does not automatically trigger incorrect alarms; the system demonstrates strong tolerance even amid incomplete or missing field data.

Greater Fault Mode Awareness / Sensor Failure Tolerance: Multi-layer validation logic distinguishes between genuine process disturbances versus benign signal dropout/noise, limiting risk that spurious warnings distract operators during critical periods.

Lower False Alerts: Multi-model evaluation chains filter out transients attributable merely to isolated sensor noise/process fluctuations, ensuring that issued recommendations correspond almost exclusively with verified physical changes inside assets themselves.

Greater Awareness/Tolerance For Operational Variability: Unlike rigid threshold systems vulnerable whenever plant setpoints shift unexpectedly, physics-guided frameworks dynamically adjust their reference baselines, ensuring accurate diagnostics remain valid across evolving site conditions without manual retuning.

Conclusion

This case study demonstrates that converging rigorous physical modeling discipline with best-in-class supervised machine learning yields transformative gains over both legacy rule-based and “black box” anomaly-detection platforms. In particular:

Actionable Diagnostics Under Realistic Constraints: Even when direct measurement coverage is sparse/incomplete, as is typical throughout mature industrial facilities, the described hybrid approach consistently uncovers root-cause phenomena such as internal scaling well before late-stage outages occur.

Robustness To Data Gaps/Sensor Failures: Through continual cross-validation among redundant input sources/models, reliability engineers can trust flagged recommendations reflect genuine emerging threats, not simply artifacts produced by defective field instrumentation.

Efficient Adaptation To Complex Process Environments: By embedding deep domain knowledge within every algorithmic layer, diagnostic precision persists regardless of how frequently “normal” operations evolve over time, a must-have capability wherever plants face shifting market/process requirements year after year.

‍

‍

‍

‍

‍

‍

‍