Table of Contents
Compressor and turbine health monitoring tracks a machine train's mechanical condition (vibration, bearings, seals, lube oil) and its aerothermal performance (power, efficiency, flow, pressures, temperatures) against a baseline of normal behaviour. The aim is to see fouling, wear and seal problems weeks before they force a trip or an unplanned outage.
Key takeaways
- Protection and prediction are separate jobs. An API 670 protection system shuts the machine down; condition monitoring finds degradation measured in hours, days or months.
- Compressor fouling is the main cause of gas turbine performance loss, with estimates of 70 to 85 percent of all losses, and fouling can normally be removed by cleaning.
- Performance must be compared at the same conditions: correct for inlet temperature, pressure and humidity, or the ambient swing hides the degradation.
- Field performance data carries an uncertainty of 2 to 3 percent or more, so trend each machine against its own baseline rather than against the datasheet alone.
- Machine learning helps most as a model of expected behaviour and as multivariate anomaly detection; remaining-life models need failure histories that few plants have.
What does health monitoring cover?
Online monitoring of rotating equipment is usually split into protection monitoring and prediction monitoring, and the choice depends on how critical the asset is. A large gas turbine-driven compressor train needs both.
The protection layer is defined by API Standard 670, whose sixth edition was released in July 2025. It sets minimum requirements for a machinery protection system measuring radial shaft vibration, casing vibration, axial position, speed, overspeed and critical temperatures, including shutdown voting logic (Emerson). It is built for failures that are fast, violent and dangerous.
Health monitoring sits beside it. It reads the same signals plus process and performance data, and looks for slow changes: a compressor losing efficiency, a seal leaking more each week, a bearing temperature creeping up. It advises people; it never trips the machine.
Which signals matter on a compressor train?
A typical train in gas processing or boosting is a gas turbine driving a centrifugal compressor. At Hassi R'Mel, for example, the gas boosting project includes 20 compression trains based on Frame 5 gas turbines and BCL compressors, across three stations whose commissioning Sonatrach scheduled between October 2026 and April 2027.
| Subsystem | Signals | What a change can reveal |
|---|---|---|
| Rotor and fluid-film bearings | Shaft displacement (X-Y proximity probes), axial position, bearing metal temperature, speed | Unbalance, misalignment, rubs, thrust bearing wear |
| Gas turbine aerothermal | Power or torque, exhaust gas temperature, compressor discharge pressure and temperature, inlet temperature and pressure, fuel flow, speeds | Compressor fouling, hot section degradation, inlet filter problems |
| Driven compressor | Suction and discharge pressure and temperature, flow, gas composition, antisurge valve position | Loss of head or efficiency, internal leakage, operation close to surge |
| Dry gas seals | Primary vent flow and pressure, seal gas supply pressure | Seal leakage, regulator faults, contamination |
| Lube oil system | Supply pressure and temperature, filter differential pressure, laboratory oil analysis | Wear, oil degradation, contamination |
| Inlet air system | Filter differential pressure | Filter loading, which also drives fouling |
On fluid-film bearings, shaft vibration is measured relative to the bearing with eddy current sensors, usually two per bearing mounted 90 degrees apart so the shaft orbit can be recorded.
Why does gas turbine performance degrade?
Rainer Kurz and Klaus Brun list the mechanisms: fouling, erosion, hot corrosion, oxidation and changes in seal geometry (Turbomachinery Magazine, 2018). They separate two kinds. Fouling can normally be removed by cleaning. Increased tip clearances and changed airfoil geometry typically lead to non-recoverable degradation, which washing does not reverse. They also note that a degraded compressor has a reduced surge or stall margin.
Fouling dominates. Meher-Homji, Bromley and Stalder explain that deposits reduce compressor airflow and isentropic efficiency, which lowers power output and moves the surge line towards the operating line (Meher-Homji et al., 2013). Their tutorial gives figures that matter for North African sites:
- Typical air loadings range from 0.01 to 0.1 ppm by weight in rural areas to 0.1 to 700 ppm in desert regions.
- In a simulation of 92 gas turbines, smaller machines lost a larger share of power for the same imposed fouling.
- The shape of the trend matters: in their example, power falls by 5 percent over 8,000 hours either way, but a linear decline costs 2.5 percent of annual output and an exponential one about 4.65 percent.
How do you detect fouling and performance loss?
The same tutorial lists three indicators: a drop in power output at a given ambient temperature and corrected speed, a drop in compressor mass flow on fixed-geometry engines, and a drop in compressor efficiency and pressure ratio. Mass flow is the most sensitive of the three. Data already in the turbine control system is often enough to start.
Raw values are misleading because ambient conditions also move them. A common approach, the authors note, is to trend a power capacity factor, actual output divided by the corrected new-and-clean test output, with corrections for inlet temperature, inlet pressure, humidity and pressure drops. On a compressor map, corrected speed and corrected flow absorb changes in ambient temperature and inlet filter loss. The same corrected trends show whether each wash recovered what it should.
This is where machine learning first earns its place. Instead of a single OEM correction curve, a regression model trained on clean, post-wash operating data predicts the expected exhaust temperature, discharge pressure or power for the current ambient conditions and load. The residual, measured minus expected, becomes the health indicator to trend.
How do you monitor the driven centrifugal compressor?
Brun, Omatick and Jacoby describe compressor monitoring that trends polytropic head, polytropic efficiency, shaft power, discharge pressure and pressure ratio over time, and compares them with the performance expected from the original design data (H2TECH, 2023). They warn about instrument gaps: valves, flowmeters and piping between the transmitters and the flanges distort both the measurement and the prediction.
Accuracy limits what you can see. Southwest Research Institute engineers report that performance uncertainty ranges from less than 0.5 percent in laboratory tests to 2 to 3 percent or more in the field, and that results from different operating conditions can only be compared at duplicated conditions or through non-dimensional parameters (Poerner et al., 2019). Brun and Kurz add that, as a baseline for condition monitoring, accurate field test data is preferable to factory test data, because the site test uses the actual gas and piping (Turbomachinery Magazine, 2025).
In practice, a slow efficiency loss smaller than the measurement uncertainty only becomes visible by trending one machine against its own baseline over weeks.
What about vibration, seals and lube oil?
Vibration. Monitoring systems can connect to the buffered outputs of the existing vibration monitor and add spectra, orbit, waterfall and Bode plots without new hardware (H2TECH, 2023). Overall levels, the running-speed (1x) component and shaft position, trended per operating speed, make useful model features.
Dry gas seals. The usual alarms are simple setpoints: high alarms, high-high trips. A gradually increasing vent flow can therefore go unnoticed, which is why the authors of a 2018 symposium paper point to trend alarms based on change over time (Turbomachinery Magazine, 2018). A rising primary vent flow does not always mean a leaking seal; a faulty secondary seal gas regulator produces the same symptom.
Lube oil. NASA's reliability-centred maintenance guide describes oil analysis as a way to determine machine wear condition, lubricant condition and contamination, and advises correlating changes in wear metals with vibration analysis (NASA, 2008). Laboratory results arrive as periodic samples, so join them to the sensor data by sampling time.
Where does machine learning help?
| Approach | Data needed | Good for | Caution |
|---|---|---|---|
| Corrected-parameter trending | OEM curves, ambient data | Fouling, wash effectiveness | Correction curves may not match an aged machine |
| Expected-behaviour regression | Months of healthy operation | Residuals on EGT, power, efficiency | Retrain after washes, overhauls and modifications |
| Multivariate anomaly detection | Normal operation per mode | Unusual combinations across many tags | Alerts need diagnosis by an engineer |
| Fault classification | Labelled examples of each fault | Naming known faults | Few plants have enough labelled faults |
| Remaining useful life | Run-to-failure histories | Planning interventions | Public datasets are simulated |
For anomaly detection, a PCA model with T² and SPE is a good first step; our guide to anomaly detection methods compares the options. Fault classifiers depend on clean failure labels, which labelling failures in time-series data covers.
Be careful with remaining-life claims. Two widely used public datasets are simulations: NASA's C-MAPSS turbofan degradation data and the UCI naval propulsion dataset, generated from a gas turbine simulator with compressor and turbine decay coefficients. They are good for practice, not proof that a method works on your train.
How do you start, step by step?
- Pick one critical train with a known pain point, such as frequent washes, seal trips or a capacity shortfall, and name an owner.
- Collect the reference data: OEM performance curves or compressor map, test results, instrument list and the correction method.
- Extract at least a year of history from the historian, with washes, trips, overhauls and seal replacements from maintenance records; see historian data for machine learning.
- Check instruments first: flow, temperature and gas composition errors look like degradation.
- Build corrected performance trends, then an expected-behaviour model on clean periods, then anomaly detection across mechanical and seal signals.
- Run in parallel for several weeks and review every alert with rotating equipment engineers.
- Connect accepted alerts to work orders; see connecting an AI agent to a CMMS.
How do you keep the models trustworthy?
A compressor train changes its own baseline. A wash, an overhaul, a re-wheel or a new gas composition makes the old "normal" obsolete, so record these events and retrain or re-baseline after each. Treat start-ups, load changes and recycle operation as separate modes. Monitor the models as you monitor the machine; our guide to model drift monitoring covers the method. For the wider picture, see AI predictive maintenance in oil and gas and, for downhole pumps, ESP failure prediction.
What we have seen at KABAS AI
Our oil and gas page lists predictive maintenance among its use cases, and DALILE connects production data, maintenance records and equipment histories in one system. Our data science service covers data preparation, predictive models that anticipate failures, and MLOps pipelines that monitor models in production. We have not published a compressor or turbine monitoring project, so this section is guidance, not a reported result.
Our practical advice:
- Start with physics. Corrected performance trends that engineers already trust make machine learning alerts easier to accept.
- Leave protection alone. The monitoring layer reads data; it never writes to the protection system or the control system.
- Keep the data inside your network. Equipment condition and throughput data are commercially sensitive, and the models run well on-premise. Before you start, our oil and gas data readiness checklist helps check what you have.
Sources
- Meher-Homji, Bromley and Stalder, Gas Turbine Performance Deterioration and Compressor Washing, 2nd Middle East Turbomachinery Symposium, Texas A&M Turbomachinery Laboratory, 2013
- Kurz and Brun, Causes and effects of gas turbine degradation, Turbomachinery Magazine, 2018
- Emerson, Rotating Machinery Protection and API Standard 670
- Emerson, Measurement Types: Online Machinery Monitoring
- Brun, Omatick and Jacoby, Advanced compressor control and monitoring systems, H2TECH, August 2023
- Poerner, Allison and Delgado, Field Performance Testing for Centrifugal Compressors, Chemical Engineering, 2019
- Brun and Kurz, Myth: Compressor Field Performance Tests are Superfluous, Turbomachinery Magazine, 2025
- Guidelines for dry gas seal troubleshooting, Turbomachinery Magazine, 2018
- NASA, Reliability-Centered Maintenance Guide for Facilities and Collateral Equipment, September 2008
- NASA Prognostics Center of Excellence, Data Set Repository (C-MAPSS turbofan degradation simulation)
- Coraddu et al., Condition Based Maintenance of Naval Propulsion Plants, UCI Machine Learning Repository
- Sonatrach, Chairman and CEO visit to the Hassi R'Mel Boosting project, 31 July 2025
- World Oil, Baker Hughes, SONATRACH partner to increase production from Algeria's largest natural gas field, 23 May 2024
Frequently asked questions
What is the difference between machinery protection and condition monitoring?
A protection system, usually built to API 670, shuts the machine down automatically when vibration, axial position, speed or temperature crosses a limit. Condition monitoring looks at trends over hours, days or months to find degradation early enough to plan the work. Both use many of the same sensors, but the protection system must stay independent and unchanged.
Can we use the vibration data from our existing protection system?
Usually yes. Monitoring systems can connect to the buffered outputs of the vibration monitor, so no new sensors are needed. The monitoring layer should only read this data and never change protection setpoints or trip logic, which remain under the control of the instrumentation and machinery teams.
How often should a gas turbine compressor be washed?
There is no single interval. Some operators wash on a fixed schedule, others when performance parameters show a given loss, and the right choice depends on the site's fouling rate and how well washing works there. Tracking corrected power or compressor efficiency before and after each wash shows whether the schedule is right.
Do we need failure history to start monitoring with machine learning?
No. The first models learn normal or expected behaviour from healthy operating data and flag deviations, so they need clean data more than failure records. Recorded failures, trips and inspection findings are needed to test those models and, later, to build fault classifiers or remaining-life estimates.
What data should we ask the OEM for?
Ask for the design performance curves or compressor map, the acceptance or performance test results, the correction method used for ambient conditions, and the instrument list with sensor locations. These define the expected behaviour that measured data is compared against.
Can compressor and turbine monitoring run without internet access?
Yes. Performance calculations, anomaly detection models and dashboards run on ordinary servers inside the plant network, reading from the historian. Keeping them on-premise also keeps production rates and equipment condition data inside the company.



