What is a propulsion Digital Twin?
A propulsion Digital Twin is a dynamically updated digital representation of a specific propulsion asset that combines physical models, sensor observations, estimated states, historical evidence and predictive models to understand current condition and forecast future behaviour within a defined operating envelope.
How is it different from simulation?
A simulation runs a model under inputs someone chose, and answers what would happen. A Digital Twin is tied to one specific propulsion asset: that asset's telemetry continuously updates the model's states and parameters, so the twin answers what is happening to this asset and what is likely to happen next. A simulation is a component of a twin; it is not one by itself.
Why is pressure monitored in rocket propulsion?
Pressure is monitored because every function of a rocket engine is a pressure difference. Propellant flows only from higher pressure to lower, thrust comes from chamber pressure acting through the nozzle, and the margins that protect the engine, against cavitation, combustion instability and structural limits, are stated as pressures. Pressure is also measured quickly and accurately, which makes it the best window into the engine's condition.
What is chamber pressure?
Chamber pressure is the static pressure of the burning gas in the combustion chamber. With a choked nozzle it is proportional to the propellant mass flow and to how completely the propellants burn, which makes it the engine's primary indicator of thrust and performance and the reference that the other pressures in the engine are compared with.
What is model residual?
A model residual is the difference between what is observed and what a model expects for the same conditions: Residual = Observation − Model Expectation. A residual near zero means the system is behaving as the model says it should. A residual that grows, and stays outside its healthy scatter, is evidence that the system, a sensor or the model has changed.
What AI models are used in Digital Twins?
Digital Twins use regression to estimate expected values and correct sensors, classification to identify faults and operating states, unsupervised anomaly detection such as PCA, Isolation Forest and autoencoders where labelled failures are scarce, time-series models such as autoregressive, LSTM and Transformer models for temporal behaviour, and degradation or survival models for prognostics. Each is used for a specific job and validated for it.
What is a physics-informed Digital Twin?
A physics-informed Digital Twin combines physical models with learned models so that each covers the other's weakness. Physics supplies the structure and stays valid outside tested conditions; learning captures what the physics model leaves out. Typical forms are physics-derived features fed to a classifier, a learned correction added to a physics prediction, and learned models trained under physical constraints.
How are propulsion faults detected?
Propulsion faults are detected by comparing measurements with limits, with their own rate of change, and above all with what a model expects. Threshold checks catch gross exceedances, rate checks catch abrupt changes, model residuals and statistical monitoring catch small and gradual ones, and learned anomaly detectors catch unfamiliar patterns across many channels. Robust schemes require independent methods to agree before declaring a fault.
How do we distinguish sensor drift from real engine degradation?
Sensor drift and real engine degradation are distinguished by what else changes. A real change in the engine moves every measurement that physics ties to it: a second sensor, the thrust proxy, flow and pump pressure all shift together. A drifting sensor moves alone. Hardware redundancy, cross-sensor consistency and a model-based virtual sensor together decide which has happened.
What is state estimation?
State estimation is the calculation of the most probable state of a system from noisy measurements and an imperfect model, together with the uncertainty of that estimate. Kalman filters and their nonlinear variants predict the state with the model and correct it with each measurement, and can reach states and parameters that no sensor measures directly.
What is Digital Twin validation?
Digital Twin validation is the demonstration, against independent measurements of the physical system, that the twin's models are accurate enough for a stated use within a stated operating envelope. It is distinct from verification, which checks that the model was built correctly, and from calibration, which fits parameters to data. Agreement with the data used for fitting is not validation.
Can AI replace propulsion physics?
No. AI cannot replace propulsion physics. A learned model is reliable only inside the conditions it was trained on, and the conditions that matter most for safety are the ones for which little or no data exists. Physics remains mandatory for limits, protection and extrapolation. AI is valuable beside it: for screening, for ranking candidate causes and for making expensive physics fast.
What is FDIR?
FDIR is fault detection, isolation and recovery: detecting that behaviour is no longer nominal, isolating the fault to a component and a mechanism, and recovering by reconfiguring, reducing power or shutting down safely. A Digital Twin contributes mainly to detection and isolation, and informs recovery decisions that control logic and engineers take.
What is prognostics?
Prognostics is the estimation of how a system's condition will evolve: the rate of degradation, the margin that remains and the probability that a limit will be reached within a given time. A prognostic result is a distribution with a confidence interval, not a single date, and it is only as good as the degradation model and data behind it.
What is Digital Twin uncertainty?
Digital Twin uncertainty is the stated range within which the twin's estimates and predictions are expected to lie. It combines sensor and calibration uncertainty, model-form and parameter uncertainty, uncertainty in the environment and the error of any learned model. A twin that presents a prediction without it is claiming knowledge it does not have.
Glossary
- Telemetry
- Measurements sent from sensors and controllers for monitoring and recording, with their timestamps and quality flags.
- Reduced-order model
- A simplified model that keeps the dominant physics so that it runs fast enough to work beside live data.
- Net positive suction head (NPSH)
- The margin between the pressure at a pump's inlet and the liquid's vapour pressure, expressed as a head.
- Cavitation
- Local boiling of a liquid where its pressure falls to vapour pressure, typically at a pump inlet.
- Characteristic velocity (c*)
- Chamber pressure times throat area divided by mass flow: a measure of how well the chamber turns propellant into pressure.
- Mixture ratio
- Oxidiser mass flow divided by fuel mass flow.
- Redline
- A limit on a monitored quantity that, once exceeded and confirmed, triggers a protective action.
- Virtual sensor
- A value computed from other measurements through a model, standing in for a measurement that is missing or in doubt.
- Analytical redundancy
- Checking a measurement against a model-based estimate of the same quantity, instead of against a second sensor.
- IVHM / PHM
- Integrated vehicle health management, and prognostics and health management: the disciplines of monitoring condition and predicting its evolution.
- Out of distribution
- Conditions unlike those a learned model was trained on, where its output cannot be relied on.
- Operating envelope
- The range of conditions within which a model has been shown to be adequate for its use.