Observer Design for State Estimation

Accurate knowledge of a system's internal state is the lifeblood of advanced control and monitoring, but directly measuring every variable is often impractical or impossible. Observers, also known as state estimators, solve this problem by intelligently reconstructing the full internal state using only the available output measurements and known system inputs. Mastering observer design allows you to build robust, high-performance control systems without an army of expensive sensors, transforming limited data into a complete real-time picture of your system's dynamics.

The Need for State Estimation

In control engineering, we describe dynamic systems using a state-space representation, which consists of state and output equations. The state equation, $\dot{x}=Ax+Bu$, describes how the internal state vector $x$ evolves over time based on the current state and the input $u$. The output equation, $y=Cx+Du$, tells us what we can actually measure. The fundamental challenge arises when the matrix $C$ is not invertible and we cannot directly compute $x=C^{-1}y$. Many critical states—like internal temperatures, chemical concentrations, or rotational speeds—are not directly instrumented due to cost, reliability, or physical constraints. An observer acts as a software-based sensor, fusing the mathematical model of the system (the $A$, $B$, $C$, $D$ matrices) with the real, noisy streams of $y(t)$ and $u(t)$ to produce a reliable estimate $\hat{x}(t)$.

The Luenberger Observer Architecture

The Luenberger observer is the foundational algorithm for deterministic state estimation. Its genius lies in its simple, corrective feedback structure. It replicates the plant model inside the computer but adds a critical correction term. The observer's state equation is:

$$\dot{\hat{x}}=A\hat{x}+Bu+L(y-\hat{y})$$

Here, $\hat{x}$ is the estimated state vector and $\hat{y}=C\hat{x}+Du$ is the estimated output. The term $(y-\hat{y})$ is the output estimation error—the difference between what the real system outputs and what our estimator predicts. This error drives the correction. The observer gain matrix $L$ is the design parameter we tune. It determines how aggressively we correct the estimator based on this error. Think of it as a blending knob between trusting the open-loop model and trusting the sensor measurement.

Designing the Observer Gain (L)

The goal of designing $L$ is to make the estimation error $e=x-\hat{x}$ converge to zero as quickly and smoothly as desired. We can derive the dynamics of this error by subtracting the observer equation from the true state equation:

$$\dot{e}=\dot{x}-\dot{\hat{x}}=(Ax+Bu)-(A\hat{x}+Bu+L(y-\hat{y}))$$

Substituting $y=Cx$ and $\hat{y}=C\hat{x}$ (assuming $D=0$ for simplicity), this simplifies to: $$\dot{e}=Ae-L(Cx-C\hat{x})=(A-LC)e$$

This is a crucial result. The error dynamics are governed by the matrix $(A-LC)$. The eigenvalues of this matrix are the observer poles. Just as we place controller poles $(A-BK)$ to achieve desired closed-loop system response, we place observer poles $(A-LC)$ to achieve desired estimator response. By choosing $L$, we directly control the rate at which the estimate $\hat{x}$ converges to the true state $x$.

Observer Pole Placement and Separation Principle

A standard and vital design rule is to place the observer poles 2 to 10 times faster (i.e., further left in the complex s-plane for continuous systems) than the dominant controller poles. This ensures the estimation error decays significantly faster than the system's controlled response. Why is this necessary? In a state feedback control law like $u=-K\hat{x}$, the control action depends entirely on the estimated state. If the observer is slow, the controller will be acting on poor, lagging information, which can degrade performance or even cause instability.

This leads to the Separation Principle, a key concept in modern control theory. It states that you can design the state feedback gain $K$ (assuming full state access) and the observer gain $L$ independently. The combined system's eigenvalues will be the union of the controller poles $(A-BK)$ and the observer poles $(A-LC)$. This allows a modular design process: first design a good controller, then design a faster observer.

A Worked Design Example

Consider a simple DC motor where we can measure shaft angle (output $y$) but not angular velocity (state $x_2$). A simplified state-space model is: $$\dot{x}=\left[\begin{array}{cc}0& 1\\ 0& -1\end{array}\right]x+\left[\begin{array}{c}0\\ 1\end{array}\right]u,y=\left[\begin{array}{cc}1& 0\end{array}\right]x$$ We want an observer to estimate velocity. Suppose we design a controller with poles at $s=-2\pm j$. Following the rule, we target observer poles at, say, $s=-10,-10$ (much faster).

We compute $L=\left[\begin{array}{c}{l}_1\\ l_2\end{array}\right]$ such that $\det (sI-(A-LC))=(s+10{)}^2$. The characteristic equation is: $$\det \left(\left[\begin{array}{cc}s& -1\\ l_1& s+1+l_2\end{array}\right]\right)=s(s+1+l_2)+l_1=s^2+(1+l_2)s+l_1$$ We want this equal to $s^2+20s+100$. Equating coefficients gives: $$1+l_2=20\Rightarrow l_2=19,l_1=100$$ Thus, $L=[100,19{]}^T$. Our Luenberger observer is: $$\dot{\hat{x}}=\left[\begin{array}{cc}0& 1\\ 0& -1\end{array}\right]\hat{x}+\left[\begin{array}{c}0\\ 1\end{array}\right]u+\left[\begin{array}{c}100\\ 19\end{array}\right](y-\hat{y})$$ This estimator will rapidly converge to the true angle and velocity, providing the state needed for feedback control.

Common Pitfalls

1. Making the Observer Too Fast (Overly Aggressive Gain): Placing observer poles extremely far left makes the estimator respond violently to the smallest output error. This amplifies measurement noise, causing the estimate $\hat{x}$ to be jittery and unreliable. The observer essentially stops trusting the model and overreacts to sensor noise. Always balance convergence speed with noise sensitivity.

2. Ignoring the Separation Principle's Assumptions: The Separation Principle holds perfectly for linear, time-invariant systems. If you have significant nonlinearities, time delays, or model uncertainties, the independently designed controller and observer may interact poorly, leading to degraded performance or instability. In such cases, more robust techniques (like sliding mode observers) may be required.

3. Forgetting About Real-World Signal Issues: The textbook Luenberger observer assumes perfect, instantaneous knowledge of $y(t)$ and $u(t)$. In reality, sensors have delays, and analog-to-digital conversion introduces sampling and quantization. Designing a discrete-time observer (e.g., using a Kalman filter formulation) that explicitly accounts for sample time and synchronizes correctly with the control loop is essential for implementation.

4. Assuming a Perfect Plant Model: The observer's model ($A$, $B$, $C$) is never perfect. Mismatches between the model and the real plant mean the estimation error $e$ will not converge to zero but to a bounded error that depends on the size of the mismatch. Robust observer design aims to minimize the sensitivity of the estimate to these modeling inaccuracies.

Summary

• Observers are algorithms that reconstruct a system's internal, unmeasured states by combining a dynamic model with available output and input signals.
• The Luenberger observer uses output error feedback, scaled by a tunable observer gain matrix $L$, to drive the state estimate $\hat{x}$ toward the true state $x$.
• The speed of convergence is determined by the observer poles, the eigenvalues of $(A-LC)$, which should typically be placed faster than the controller poles to ensure accurate estimates are available for feedback.
• The Separation Principle allows for the independent design of the state feedback controller and the state observer, with the combined system's poles being the union of the two sets.
• Successful practical implementation requires careful consideration of measurement noise, model inaccuracies, and discrete-time sampling effects to avoid unstable or noisy estimates.