Advanced Control and State Estimation

Introduction

Modern mechatronic products—robotic manipulators, autonomous vehicles, precision stages, drones, and smart manufacturing cells—must regulate motion and interaction despite disturbances, uncertain parameters, nonlinear mechanics, sensor noise, computation limits, and actuator saturation. Classical single-loop control remains useful, but advanced systems are commonly represented in state space and designed using observers, optimal control, robust methods, adaptive control, nonlinear techniques, or model-predictive control (MPC).

This chapter develops a design-oriented framework. The central idea is that control quality depends not only on a controller equation but also on the model, sensing architecture, sampling, constraints, uncertainty description, and verification process.

Learning Objectives

By the end of this chapter, a student should be able to:

• formulate discrete-time state-space models for mechatronic systems;
• assess controllability, observability, stability, and sampling effects;
• design state-feedback and observer-based controllers;
• explain LQR, robust, adaptive, nonlinear, and MPC methods;
• select and implement estimators such as the Kalman filter;
• analyze constraints, uncertainty, saturation, noise, and computational limitations;
• evaluate a complete controller architecture in an applied engineering case.

1. State-Space Foundations

A continuous-time linear time-invariant (LTI) model is

Here, x(t) is the state vector, u(t) the control input, y(t) the measured output, and w(t) and v(t) process and measurement disturbances. Matrices A, B, C, and D have compatible dimensions. The model assumes linear dynamics around a fixed operating point; nonlinear systems may be locally approximated by linearization.

With zero-order-hold input over a sampling interval T_s, the exact discrete model is

where

and similarly E_d = ∫₀^T_se^AτE dτ. The state is not necessarily a directly measurable physical quantity; it may include position, velocity, electrical current, temperature, or flexible-mode amplitudes.

Stability, controllability, and observability

For a discrete LTI system, internal asymptotic stability requires every eigenvalue of A_d to lie strictly inside the unit circle: [[MATH(|λ_i(A_d)| < 1)]]. Continuous-time stability instead requires negative real parts.

The controllability matrix is

An n-state system is completely controllable if rank(𝒞)=n. Controllability means that an input can move the state between appropriate points in finite time. It does not guarantee low control effort or good numerical conditioning.

The observability matrix is

The system is observable if rank(𝒪)=n. Observability concerns reconstructing the state from measured outputs. Poorly conditioned controllability or observability can make a mathematically feasible design highly sensitive to noise and finite precision.

Worked example: sampled mass–spring–damper

For a mass–spring–damper system,

choose x=[q\ q̇]ᵀ. Then

For m=1 kg, c=2 N·s/m, and k=10 N/m,

The units are consistent: acceleration is in m/s², force in newtons, position in metres, and velocity in m/s. A sampling period should be selected with respect to the dominant closed-loop bandwidth, sensor dynamics, and processor capacity—not merely the plant’s natural frequency.

2. State Feedback and Separation of Estimation

A state-feedback controller has the form

where K is the feedback gain, r_k is a reference, and N is a reference precompensator when required. The closed-loop state matrix is A_d−B_dK. Pole placement selects K so that this matrix has desired eigenvalues, provided the pair [[MATH(A_d,B_d)]] is controllable.

If all states are not measured, an observer estimates them:

The estimation error e_k=x_k−x̂_k evolves as

Thus, observer poles can be placed if [[MATH(A_d,C)]] is observable. In the ideal LTI case, the separation principle allows controller and observer gains to be designed independently. In practice, saturation, nonlinearities, noise, unmodelled dynamics, and numerical limitations weaken this ideal separation.

An observer that is excessively fast may amplify measurement noise and excite neglected high-frequency dynamics. A slow observer may produce unacceptable transient errors. Pole placement is therefore a design trade-off, not a command to make observer poles arbitrarily fast.

3. Optimal Control

Linear-quadratic regulation (LQR) selects a feedback law by minimizing

where Q≥0 weights state deviation and R>0 weights actuator effort. The units of Q and R should be chosen so that both terms represent comparable engineering cost. Cross-weight terms can be included, but the standard form is often adequate.

For the discrete system x_k+1=A_dx_k+B_du_k, the stabilizing solution P of the discrete algebraic Riccati equation is

and

so that u_k=−Kx_k. Existence of a stabilizing solution requires appropriate stabilizability and detectability conditions; positive semidefinite Q and positive definite R are standard assumptions.

LQR gives a systematic effort–performance compromise, but the unconstrained solution does not explicitly enforce position limits, current limits, or collision boundaries. Saturation handling and anti-windup or constrained optimization must be added.

Kalman filtering

For a stochastic linear model,

assume zero-mean white disturbances with covariances

where W≥0 and V>0. The prediction step is

The measurement update is

The Kalman gain balances model prediction against sensor information. Increasing W makes the filter trust measurements more; increasing V makes it trust the model more. Covariances must be experimentally or analytically justified rather than adjusted solely to obtain visually smooth signals.

For nonlinear systems, the extended Kalman filter (EKF) linearizes the dynamics and measurement equations about the current estimate. The unscented Kalman filter propagates sigma points and can better capture some nonlinear transformations, but neither method guarantees global convergence.

4. Robust and Adaptive Control

Robust control addresses performance degradation caused by uncertainty. A nominal model may be written as

where G₀ is nominal and ΔG represents modelling error. Uncertainty descriptions include additive, multiplicative, parametric, and structured uncertainty. Robust stability means stability for every uncertainty within the specified set; robust performance additionally requires acceptable tracking and disturbance rejection.

H-infinity design frames disturbance attenuation as a worst-case optimization. A generalized plant is shaped using weighting functions to express tracking, control effort, and noise priorities. The resulting controller is often more conservative than an LQR design, but it can provide explicit guarantees under the selected uncertainty model. A guarantee is only meaningful if the uncertainty bounds are credible.

Adaptive control updates controller parameters online. In model-reference adaptive control, the plant is driven toward a reference model. In self-tuning control, parameters are estimated and a controller is redesigned from the estimate. A generic parameter estimator may use a regression model

where φ_k is a regressor, θ an unknown parameter vector, and ε_k an error term. Persistent excitation is often needed to identify parameters reliably. Adaptation can become unsafe when measurements are poor, actuators saturate, or the true plant is outside the assumed model class. Projection, parameter bounds, dead zones, normalization, and supervisory logic are practical safeguards.

5. Nonlinear Control

Many mechatronic systems obey

where f and g are nonlinear vector fields. Local linear control may work near one equilibrium but fail over a wide operating range.

Feedback linearization attempts to cancel known nonlinearities and impose a desired linear error model. It requires a sufficiently accurate model and well-defined input authority; singularities can make the transformation invalid.

Sliding-mode control defines a surface, commonly

for tracking error e and λ>0. A discontinuous term drives trajectories toward s=0, providing robustness to certain matched uncertainties. However, ideal switching produces chattering. Boundary layers, higher-order algorithms, or filtered implementations reduce chattering but alter the exact robustness properties.

Lyapunov analysis provides a stability framework. If a positive-definite function V(x) satisfies

for a positive-definite function α, stability and, under suitable conditions, asymptotic convergence can be established. For digital controllers, the corresponding difference condition is V(x_k+1)−V(x_k)<0 outside the desired equilibrium region.

6. Model-Predictive Control

MPC repeatedly solves a finite-horizon optimization. For a linear model, a typical quadratic problem is

subject to

where N is the prediction horizon, 𝒳 and 𝒰 are state and input constraint sets, and P is a terminal weight. Only the first computed input is applied; the horizon then shifts and the optimization is repeated.

MPC is valuable when constraints are central: actuator current, velocity, temperature, workspace, and energy may all be included. Feasibility is a major issue. If disturbances or modelling errors make the optimization infeasible, a controller must specify a recovery policy, such as softened constraints, a safe fallback mode, or constraint tightening.

Stability may be supported by terminal costs, terminal invariant sets, or suitable receding-horizon constructions. Nonlinear MPC uses nonlinear prediction models and generally requires more computation. Explicit MPC precomputes a piecewise-affine policy for some constrained linear problems, trading memory for online speed.

7. Implementation Effects and Verification

A practical digital controller includes sampling, computation delay, quantization, communication delay, zero-order hold, sensor filtering, and actuator dynamics. These effects modify the effective closed-loop system. Sampling too slowly can produce aliasing and phase loss; sampling too quickly can overload computation and expose high-frequency noise.

Important implementation practices include:

1. Use anti-aliasing filters before sampling.
2. Time-stamp measurements and account for variable latency.
3. Use numerically stable matrix factorizations rather than explicit matrix inverses where possible.
4. Include actuator saturation in simulation and test anti-windup or constrained control.
5. Monitor estimator innovation, residuals, covariance consistency, and sensor faults.
6. Test nominal, worst-case, fault, startup, and recovery scenarios.
7. Apply hardware-in-the-loop testing before deployment.

A Kalman innovation is ν_k=y_k−Cx̂_k|k−1, with theoretical covariance S_k. Persistent innovation bias may indicate sensor bias, incorrect dynamics, or an unmodelled disturbance. It should not be dismissed as random noise.

Safety-critical systems require traceable requirements, independent verification, fault containment, access control, secure software updates, and defined human override. An AI-based estimator or learned controller should be treated as a component with a validation envelope, not as an unquestionable authority. Data-driven adaptation can create cybersecurity and accountability risks if its behavior changes after deployment.

8. Integrated Case Study: Precision Linear-Stage Controller

The following is an illustrative engineering case. A motor-driven linear stage must follow a position trajectory with low steady-state error. Its load varies, the position encoder is accurate but velocity is not directly measured, and motor current is limited.

Background and problem

A nominal second-order model is selected with position and velocity as states. Friction varies with temperature, and the drive introduces a one-sample computational delay. A proportional position loop produces overshoot when the payload changes, while numerical differentiation of encoder position creates noisy velocity estimates.

Analysis

The design team first identifies the plant over the expected operating range and obtains discrete matrices using the actual sampling period. Controllability and observability ranks are checked. Current saturation is represented explicitly. Closed-loop tests include payload extremes and encoder quantization.

Approach and solution

An observer estimates velocity from position and input. The controller uses discrete LQR with position error and estimated velocity in the state penalty. The input weighting is increased until peak current remains within the drive limit. Because LQR alone does not enforce constraints, a supervisory layer limits commands and activates a safe deceleration profile near travel limits.

The team compares this design with MPC. MPC improves handling of travel and current constraints but requires a solver that meets the worst-case computation deadline. The selected implementation uses LQR for the fast inner loop and a slower trajectory generator that enforces position, velocity, and acceleration limits.

Outcome

Testing shows lower overshoot and less noise than numerical differentiation. At payload extremes, performance degrades but remains stable. Innovation monitoring detects a deliberately introduced encoder bias, causing the system to enter a restricted-speed mode rather than continue normal operation.

Learning points

The case demonstrates that state estimation is part of control design, not an optional software module. It also shows why nominal performance, constraint compliance, fault handling, timing, and validation must be evaluated together. A theoretically elegant controller can be unsuitable if its computational or safety assumptions are violated.

Challenges and Design Trade-offs

No method dominates in every application. Pole placement is transparent but does not directly optimize effort. LQR is systematic but unconstrained. Robust control handles uncertainty but depends on uncertainty descriptions and may be conservative. Adaptive control responds to changing parameters but can be fragile during poor excitation or faults. Nonlinear control can achieve broad-range performance but requires model knowledge and careful stability analysis. MPC handles constraints naturally but demands computation, feasibility management, and a sufficiently useful prediction model.

A defensible design process therefore proceeds from requirements to modelling, sensing, controllability and observability analysis, controller selection, simulation, real-time testing, fault analysis, and field validation. Performance metrics should include tracking error, settling time, overshoot, control energy, constraint violations, estimator residuals, worst-case computation time, and safety response.

Chapter Summary

Advanced mechatronic control uses state-space models to coordinate multiple variables, incorporate estimation, and analyze stability. Discrete state feedback places or optimizes closed-loop dynamics, while observers reconstruct unavailable states. LQR formalizes the trade-off between state regulation and actuator effort. Kalman filtering combines model prediction and noisy measurements according to uncertainty covariances. Robust control addresses bounded uncertainty; adaptive control updates parameters; nonlinear methods exploit system structure; and MPC solves constrained finite-horizon problems repeatedly. Successful implementation must account for sampling, delay, saturation, numerical conditioning, sensor faults, uncertainty, computational deadlines, cybersecurity, and safety governance. The most reliable architecture is selected by matching method assumptions to the physical system and validating the complete closed loop.

Key Terms

State-space model: A representation using state variables, inputs, outputs, and their dynamics.

Discrete-time model: A model describing system evolution at sampled instants.

Controllability: Ability to move the state through input action.

Observability: Ability to reconstruct internal state from outputs.

State feedback: Control using measured or estimated state variables.

Observer: An algorithm that estimates unmeasured states.

Separation principle: Ideal result allowing controller and observer design to be separated for suitable LTI systems.

LQR: Linear-quadratic regulator minimizing a quadratic state-and-input cost.

Riccati equation: Matrix equation whose stabilizing solution produces LQR gains.

Kalman filter: Minimum-variance linear estimator under stated stochastic assumptions.

Innovation: Difference between a measurement and its predicted value.

Robust control: Control designed to maintain specified properties under uncertainty.

Adaptive control: Control that updates parameters or laws online.

Feedback linearization: Nonlinear control method that cancels model nonlinearities when valid.

Sliding mode: Robust variable-structure control based on attraction to a switching surface.

Lyapunov function: Scalar function used to analyze stability.

Model-predictive control: Receding-horizon optimization subject to a predictive model and constraints.

Constraint tightening: Conservative restriction used to preserve feasibility under uncertainty.

Anti-windup: Method limiting integrator accumulation during actuator saturation.

Hardware-in-the-loop: Real-time testing combining actual hardware with simulated plant components.

Review Questions

Short Answer Questions

1. What is the difference between controllability and observability?
2. Why can an excessively fast observer amplify measurement noise?
3. What roles do Q and R play in LQR?
4. What is the purpose of the innovation in a Kalman filter?
5. Why does MPC require a feasibility or fallback strategy?

Long Answer Questions

1. Derive the observer error dynamics and explain the effect of observer pole placement.
2. Compare pole placement, LQR, robust control, and MPC for a motor-positioning system.
3. Explain the assumptions behind the Kalman filter and discuss how incorrect covariance estimates affect behavior.
4. Describe how sampling, delay, quantization, and saturation alter an otherwise stable controller.
5. Explain how Lyapunov reasoning supports nonlinear controller design.

Analytical / Application Questions

1. A two-state stage model is controllable but its velocity sensor is unavailable. Design a conceptual observer-based feedback architecture and identify the tests required before hardware deployment.
2. A robot actuator repeatedly reaches its current limit during trajectory tracking. Decide whether LQR, adaptive control, or MPC is most appropriate, justify the choice, and describe safeguards.
3. A Kalman-filter innovation develops a persistent nonzero mean after a sensor replacement. Formulate at least three plausible causes and propose a diagnostic sequence.

Multiple Choice Questions with Answers

Question 1

For a discrete-time LTI system, asymptotic stability requires:

A. All eigenvalues of the state matrix to have positive real parts

B. All eigenvalues of the state matrix to lie strictly inside the unit circle

C. The controllability matrix to be diagonal

D. The output matrix to be invertible

Correct Answer: B. All eigenvalues of the state matrix to lie strictly inside the unit circle

Question 2

A system with n states is completely controllable when:

A. rank(𝒞)=n

B. rank(C)=n in every case

C. All inputs are bounded

D. Its output has zero noise

Correct Answer: A. rank(𝒞)=n

Question 3

In observer design, the estimation error for the standard observer evolves according to:

A. (A+B K)e_k

B. (A−LC)e_k

C. (C−A L)e_k

D. (A+C)e_k

Correct Answer: B. (A−LC)e_k

Question 4

In the standard LQR cost, increasing R generally places greater penalty on:

A. State estimation error

B. Measurement noise

C. Control effort

D. Sampling frequency

Correct Answer: C. Control effort

Question 5

The Kalman filter innovation is:

A. The actuator command minus its limit

B. The measured output minus the predicted output

C. The state estimate minus the control input

D. The process covariance minus the measurement covariance

Correct Answer: B. The measured output minus the predicted output

Question 6

A principal advantage of MPC over unconstrained LQR is that MPC:

A. Never requires a model

B. Can explicitly incorporate input and state constraints

C. Eliminates all measurement noise

D. Always needs less computation

Correct Answer: B. Can explicitly incorporate input and state constraints

Question 7

Persistent excitation in adaptive control is primarily associated with:

A. Reliable parameter identification

B. Removing actuator saturation

C. Guaranteeing zero sensor noise

D. Increasing sampling jitter

Correct Answer: A. Reliable parameter identification

Question 8

A major practical disadvantage of ideal sliding-mode control is:

A. Lack of any disturbance rejection

B. Chattering caused by discontinuous switching

C. Inability to use feedback

D. Requirement for an unconstrained actuator

Correct Answer: B. Chattering caused by discontinuous switching

Question 9

For the continuous-time system matrix A, exact zero-order-hold discretization uses:

A. A_d=A/T_s

B. A_d=e^{AT_s}

C. A_d=A+B

D. A_d=C A

Correct Answer: B. A_d=e^{AT_s}

Question 10

A persistent bias in Kalman-filter innovations may indicate:

A. A perfectly matched model

B. An unmodelled disturbance or sensor bias

C. Guaranteed closed-loop optimality

D. That measurement covariance must be zero

Correct Answer: B. An unmodelled disturbance or sensor bias