Digital Twin Multiphysics
Theory: combining model and sensor
Overview
Is a digital twin just running a detailed simulation alongside the real machine?
That alone isn't enough. Any model, however detailed, has errors in losses, cooling and material data, and drifts as it runs alongside the machine. Sensors are accurate but cover few places: on a motor you can measure the housing but not easily the winding interior where insulation degrades. The core of a digital twin is using the difference between model prediction and sensor reading to correct model states every step, so unmeasurable temperatures can be inferred through the model. In multiphysics terms, tracking loss (electrical), temperature (thermal) and life (material) together means pairing a small, fast model with a correction scheme such as a Kalman filter.
Two-node thermal model
T₁ is winding and T₂ housing temperature. C₁ = 500 J/K, C₂ = 5,000 J/K, R₁₂ = 0.05 K/W, R₂ₐ = 0.02 K/W, ambient 25 °C. The sensor measures housing temperature y every second with noise standard deviation 0.5 K. Load alternates 1,500 W for 600 s and 300 W for 600 s for three cycles (3,600 s). Actual loss is 20% above the model's.
Kalman filter correction
K is the Kalman gain. Adding loss error ΔP to the state lets the filter infer missing loss from the gap between predicted and measured housing temperature.
The estimator that went to the Moon
Kalman published his recursive method for estimating state from measurements in 1960. NASA's Ames Research Center took it up early for Apollo navigation, which had to keep estimating spacecraft position and velocity from noisy measurements on very limited computers. The same idea — combining model and measurement with little computation — now powers digital twins of factory machines, batteries and motors.
Worked examples
Example 1: winding temperature estimation error (20% loss model error)
| Method | Max error | Mean error, last cycle | Max error, last cycle |
|---|---|---|---|
| Model only | 21.0 K | 12.6 K | 21.0 K |
| Filter (2 temperatures) | 15.2 K | 9.0 K | 15.1 K |
| Filter + loss error | 9.3 K | 1.2 K | 7.9 K |
True peak winding temperature is 150.8 °C, with the housing at 60.8 °C. The final estimated loss error is 49 W (actual 60 W).
With the loss error estimated, mean error is 1.2 K, so why does a 7.9 K max remain even in the last cycle?
Because the model misrepresents the error's form. The real error is multiplicative — loss 20% high — so it's 60 W short at 300 W and 300 W short at 1,500 W. The filter estimates loss error as a single additive value, so right after a load step it still holds the old value and must wait for housing temperature to change before correcting. The housing has large heat capacity and responds slowly, so the winding estimate drifts in the meantime. Remedies: estimate the error as a ratio of loss, or model copper resistance rising with temperature to remove the error's cause. Digital-twin accuracy often depends less on the correction scheme than on deciding what the model represents correctly and what is left to estimation.
Modelling workflow
- Build a small model linking the location of interest to measurable points from detailed analysis.
- Add main error sources (loss, cooling) as estimated states.
- Set filter weights from sensor noise and model uncertainty.
- Validate against embedded winding thermocouples in tests.
- Integrate insulation life from estimated temperature for maintenance planning.
“A twin that spotted a clogged fan”
On one factory pump motor, the digital twin's estimated cooling thermal resistance crept up over several months. Temperature itself was still below its limit, but the estimated resistance trended clearly upward. Inspection found the cooling fan intake about half clogged with dust; after cleaning, the estimate returned to normal. Continuously estimating an unmeasurable state revealed the cause before any temperature alarm.
Common mistakes
Mistakes and fixes
| Mistake | Effect | Fix |
|---|---|---|
| Running the model alone | Errors accumulate | Correct with sensors |
| Using measured temperature as proxy | Dangerous temperature missed | Estimate via model |
| Correcting temperatures only | Bias remains | Add error causes as states |
| Overly detailed model | Can't run in real time | Use a ROM |
| No validation of estimates | Wrong decisions | Compare with test data |
I'd like to learn related topics.
Related pages include reduced-order models for coupled problems, model order reduction, data-driven multiphysics, motor thermal management and co-simulation coupling.
Related Topics
Experience the theory firsthand with the interactive simulator for this field
All Simulators