What Is System Identification? Learning How a Real Machine Behaves.

What Is System Identification?

System identification is the process of developing a mathematical model of a system using measured information about its inputs and outputs.

Instead of relying exclusively on theoretical equations or manufacturer specifications, engineers observe how the physical system responds to known changes.

For example, a robotic actuator may receive controlled motor commands while sensors record its resulting movement.

The measurements can be used to estimate a model of the actuator's dynamic behavior.

System identification is useful for control design, simulation, and understanding equipment whose physical properties are not completely known.

Key Takeaways

• System identification builds mathematical models from observed system behavior.
• It uses measured inputs and outputs to estimate relationships and model parameters.
• A model must be validated using appropriate data before relying on its predictions.

How Does System Identification Work?

Imagine a robot joint whose response to motor commands is not fully understood.

An engineer commands a series of small, controlled movements.

The system records information such as motor input, joint position, and velocity.

The measurements reveal how the joint accelerates, slows down, and responds to changes.

Software then estimates a mathematical model that reproduces relevant aspects of this behavior.

For example, the model may contain parameters associated with inertia, damping, or other dynamic effects.

The engineer compares the model's predictions with measured responses.

If the model performs poorly, the identification procedure may need different data, a different model structure, or additional measurements.

What Are Input and Output Data?

An input is an action or signal applied to the system.

An output is the measured response.

For example, a motor-control experiment might use commanded electrical current as its input.

Joint velocity may be one measured output.

The relationship between input and output contains information about the system's dynamics.

However, the data must be collected carefully.

Noise, delays, unmeasured disturbances, and inappropriate excitation can make estimation difficult.

White-Box vs. Black-Box vs. Grey-Box Models

White-box model: Built primarily from known physical equations and parameters.

Black-box model: Estimates relationships from data without requiring a detailed physical interpretation of every internal element.

Grey-box model: Combines known physical structure with parameters estimated from measurements.

For example, an engineer might know the basic equation describing a motor's movement but not its exact friction coefficient.

A grey-box approach can use the known equation and estimate the missing parameter.

Different modeling approaches provide different balances of flexibility, interpretability, and accuracy.

System Identification vs. State Estimation

System identification estimates properties or dynamics of the system itself.

State estimation determines the system's current condition.

For example, system identification might estimate a robot joint's inertia or damping.

State estimation might calculate its current position and velocity.

The two processes can support one another.

A better system model can improve state estimation and control.

Why Is System Identification Important for Simulation?

A physics simulator requires assumptions about how motors, joints, and other components behave.

If those assumptions are inaccurate, simulated movements may differ from physical movements.

System identification provides a way to use actual measurements to improve the model.

For example, engineers may compare simulated joint movements with physical test results and adjust model parameters.

This can help reduce some aspects of the sim-to-real gap.

Frequently Asked Questions

Does system identification require artificial intelligence?

No. Traditional statistical estimation and control-theory techniques are widely used, although machine learning methods can also be applied.

Can a model be identified without testing the machine?

A model can be built from specifications or theory, but identifying it from real input-output behavior requires relevant measurement data.

The Bottom Line

System identification converts observed machine behavior into mathematical models. It is valuable for robotics because accurate models support better simulation, control, and understanding of physical equipment.