Full-Order State Observer Design

In control engineering, you often need to know every internal state of a system—like the position, velocity, and acceleration of a robotic arm—to implement sophisticated feedback laws. However, directly measuring every state is frequently impractical or too costly. A full-order state observer solves this problem by acting as a software sensor, intelligently reconstructing all system states using only the available output measurements and a mathematical model of the system. Mastering observer design is crucial for implementing modern state-space control techniques, bridging the gap between theoretical control laws and real-world application.

The Luenberger Observer Structure

The foundation of state estimation is the Luenberger observer, named after its inventor, David G. Luenberger. It is a model-based predictor with a crucial corrective feedback loop. Consider a linear time-invariant system defined by: $$\dot{x}(t)=Ax(t)+Bu(t)$$ $$y(t)=Cx(t)$$ where $x$ is the state vector (to be estimated), $u$ is the known input, $y$ is the measured output, and $A$, $B$, and $C$ are system matrices.

A full-order observer generates an estimate, denoted $\hat{x}(t)$, for all states. Its dynamic equation is: $$\dot{\hat{x}}(t)=A\hat{x}(t)+Bu(t)+L(y(t)-C\hat{x}(t))$$

This structure is elegant and intuitive. The term $A\hat{x}(t)+Bu(t)$ is a model-based predictor. It simulates the system's behavior using the known model and input. Alone, this open-loop simulation would drift due to model inaccuracies and unknown disturbances. The term $L(y(t)-C\hat{x}(t))$ is the correction gain applied to the output error. Here, $L$ is the observer gain matrix. The quantity $(y-C\hat{x})$ is the innovation or residual—the difference between the actual measured output and the predicted output based on our state estimate. This error signal is fed back through the gain $L$ to correct the state estimate continuously.

Observer Pole Placement and Performance Tuning

The performance of the observer—how fast and how accurately it converges to the true state—is entirely dictated by the choice of the observer gain matrix $L$. To analyze this, we examine the estimation error dynamics. Define the estimation error as $e(t)=x(t)-\hat{x}(t)$. By subtracting the observer equation from the true system equation, we derive the error dynamics: $$\dot{e}(t)=(A-LC)e(t)$$

This is a homogeneous differential equation. The error $e(t)$ will decay to zero if and only if the matrix $(A-LC)$ is Hurwitz (i.e., all its eigenvalues have negative real parts). The eigenvalues of $(A-LC)$ are the observer poles.

Therefore, designing the observer reduces to selecting an $L$ that places these poles in desired locations in the complex left-half plane. This process is analogous to controller pole placement for the matrix $(A-BK)$. A fundamental requirement for being able to place the observer poles arbitrarily is that the pair $(A,C)$ must be observable. Observability is a property of the system that guarantees the outputs contain enough information to uniquely reconstruct all initial states.

The choice of pole locations involves a critical trade-off:

• Fast Poles: Placing poles far to the left in the s-plane makes the error decay very rapidly. This gives a fast, responsive observer.
• Noise Sensitivity: In practice, measurements $y(t)$ are corrupted by sensor noise. A very fast observer aggressively corrects based on $y(t)$, amplifying this high-frequency noise in the state estimate $\hat{x}(t)$. This leads to a noisy, chattering estimate.

A good design practice is to make the observer poles 2 to 5 times faster than the desired closed-loop controller poles. This ensures the estimates settle quickly relative to the system response, without being so fast that they become dominated by measurement noise.

Stability and the Separation Principle

A primary use of a state observer is to provide estimates for a state feedback controller of the form $u(t)=-K\hat{x}(t)$. A profound result in linear system theory justifies designing the controller and observer independently. This is known as the separation principle.

It states that for a linear system, you can:

1. Design the state feedback gain $K$ assuming full state access ($u=-Kx$), placing the closed-loop poles of $(A-BK)$.
2. Design the observer gain $L$ to place the observer poles of $(A-LC)$.
3. Combine them into a controller-observer system ($u=-K\hat{x}$), and the poles of the overall closed-loop system will be the union of the poles from step 1 and step 2.

Mathematically, the combined system using the estimated state has dynamics governed by the state $x$ and the error $e$. The system matrix for this combined state is block diagonal: $$\left[\begin{array}{c}\dot{x}\\ \dot{e}\end{array}\right]=\left[\begin{array}{cc}A-BK& BK\\ 0& A-LC\end{array}\right]\left[\begin{array}{c}x\\ e\end{array}\right]$$

The eigenvalues of a block triangular matrix are the eigenvalues of its diagonal blocks: $(A-BK)$ and $(A-LC)$. Therefore, the controller poles and observer poles do not affect each other's locations. This principle dramatically simplifies the design process and guarantees that if you design a stabilizing controller and a stabilizing observer independently, the combined system will be stable.

Common Pitfalls

1. Ignoring Observability: Attempting to design an observer for an unobservable system is impossible. The unobservable modes of the system will be unreachable by the correction term $L(y-C\hat{x})$, and their associated estimation errors will not decay. Correction: Always check the observability matrix rank before proceeding with observer design. If the system is not observable, you may need to add or relocate sensors, or you can only design an observer for the observable subsystem.

1. Making the Observer Too Fast: It is tempting to place observer poles extremely far to the left for instantaneous estimation. In simulation with perfect measurements, this works beautifully. In reality, it leads to an estimator that is highly sensitive to measurement noise, resulting in unusable, jittery state estimates. Correction: Respect the bandwidth of your sensors. Design observer poles to be sufficiently faster than the plant dynamics but within a frequency range where your sensor signals are still reliable. Consider using a Kalman filter framework if noise statistics are known.

1. Using an Inaccurate Model: The observer's predictor is entirely based on the matrices $A$, $B$, and $C$. Significant discrepancies between the model and the real plant (parameter errors, unmodeled nonlinearities) mean the open-loop prediction is wrong. The correction term must then work harder to compensate, leading to steady-state estimation errors or even instability. Correction: Invest in accurate system identification. For systems with known parameter variations, consider adaptive or robust observer techniques. Always validate observer performance with a high-fidelity simulation model before deployment.

1. Forgetting the Initial Condition: The observer differential equation requires an initial state estimate $\hat{x}(0)$. If this initial guess is wildly incorrect, the transient period of estimation error convergence can be large and potentially drive the controller to produce excessive control signals. Correction: Initialize the observer with the best available guess. If no information is available, initializing states to zero is common. The observer will converge from any initial condition as long as it is stable, but a good initial guess improves transient performance.

Summary

• A full-order state observer reconstructs all internal system states using available outputs and a system model, following the Luenberger structure: $\dot{\hat{x}}=A\hat{x}+Bu+L(y-C\hat{x})$.
• Observer performance is tuned via the gain matrix $L$, which places the observer poles (eigenvalues of $A-LC$). The design involves a trade-off between fast convergence and sensitivity to measurement noise.
• The system must be observable for arbitrary pole placement. Observability guarantees the outputs contain sufficient information to estimate all states.
• The separation principle is a cornerstone result, allowing the state feedback controller gain $K$ and the observer gain $L$ to be designed independently while preserving the overall closed-loop stability.