Control Systems Interview Questions

Open-loop systems operate without feedback - the output has no influence on the control action (e.g., a toaster timer). Closed-loop systems use feedback to compare actual output with desired output and adjust the control action accordingly (e.g., thermostat-controlled heating). Closed-loop systems are more accurate, can compensate for disturbances, but are more complex. Open-loop systems are simpler and cheaper but cannot correct for errors or disturbances.

A transfer function G(s) is the ratio of the Laplace transform of output to input, assuming zero initial conditions. It completely characterizes a linear time-invariant (LTI) system's input-output behavior in the s-domain. The transfer function is independent of input signal and describes the system's inherent dynamics. Poles (denominator roots) determine stability and natural response modes; zeros (numerator roots) affect the response shape.

Feedback provides: improved accuracy (error between desired and actual is corrected), reduced sensitivity to parameter variations (changes in plant don't affect output as much), disturbance rejection (compensates for external disturbances), and ability to modify system dynamics (change response speed, damping). Trade-offs include increased complexity, potential instability if poorly designed, and additional cost of sensors and controllers. Feedback is fundamental to most industrial control systems.

A PID controller uses three terms: Proportional (P) provides output proportional to current error (faster response but steady-state error), Integral (I) sums past errors to eliminate steady-state error (adds lag), and Derivative (D) predicts future error from rate of change (improves stability and reduces overshoot). The controller output is: u(t) = Kp*e(t) + Ki*integral(e) + Kd*de/dt. PID is the most widely used industrial controller due to its simplicity and effectiveness.

A system is stable if its output remains bounded for any bounded input (BIBO stability). For linear systems, this means all poles must have negative real parts (lie in left half of s-plane). An unstable system's output grows unbounded, potentially causing damage. Marginally stable systems (poles on imaginary axis) produce sustained oscillations. Stability is the first requirement in control system design - an unstable closed-loop system is dangerous and useless.

A first-order system has a transfer function with one pole: G(s) = K/(tau*s + 1), where K is gain and tau is time constant. Step response rises exponentially toward final value: y(t) = K(1 - e^(-t/tau)). The time constant tau determines response speed - after one tau, response reaches 63.2% of final value; after 5 tau, it is 99% complete. First-order systems have no overshoot. Examples include RC circuits and thermal systems.

A second-order system has transfer function: G(s) = omega_n^2/(s^2 + 2*zeta*omega_n*s + omega_n^2), where omega_n is natural frequency and zeta is damping ratio. Response depends on zeta: underdamped (zeta<1) shows overshoot and oscillation, critically damped (zeta=1) is fastest without overshoot, overdamped (zeta>1) is slow without overshoot. Typical design targets zeta = 0.5-0.8 for good balance of speed and overshoot. Many physical systems are second-order or approximated as such.

Steady-state error is the difference between desired and actual output after transients settle. System type (number of integrators in forward path) determines steady-state error for different inputs: Type 0 has constant error to step input, Type 1 has zero error to step but constant error to ramp, Type 2 tracks step and ramp with zero error. Integral action in controllers increases system type. Position, velocity, and acceleration error constants (Kp, Kv, Ka) quantify steady-state errors.

Block diagrams represent systems using: blocks (transfer function), summing junctions (add/subtract signals), and pickoff points (signal branching). Reduction rules include: series blocks multiply, parallel blocks add, feedback loop TF = G/(1+GH), moving blocks past summing junction or pickoff requires compensation. Block diagram reduction simplifies complex systems to single transfer function. Mason's gain formula provides direct solution for complex diagrams without iterative reduction.

Time-domain specifications describe transient response to step input: rise time (time to go from 10% to 90% of final value), peak time (time to first peak), overshoot (percentage exceeding final value), settling time (time to stay within 2% or 5% of final value), and steady-state error. These specifications are used to evaluate system performance and guide controller design. Trade-offs exist - reducing rise time often increases overshoot. Design targets depend on application requirements.

Negative feedback subtracts the feedback signal from the reference, driving error toward zero. This creates a self-correcting system that maintains stability and tracks the reference. Positive feedback adds the feedback signal, amplifying any deviation from the desired value, leading to instability or saturation. Negative feedback reduces gain but improves linearity, stability, and bandwidth. Positive feedback is used intentionally only in oscillators and comparators with hysteresis.

Increasing proportional gain (Kp) reduces rise time and steady-state error but increases overshoot and may cause instability if too high. Proportional control alone cannot eliminate steady-state error in Type 0 systems - some error is always needed to generate the control effort. The closed-loop response with P control is: steady-state error = 1/(1 + Kp*G(0)). Very high Kp makes the system respond quickly but oscillate excessively; too low makes it sluggish.

Integral action accumulates past errors and adjusts control output until error becomes zero, eliminating steady-state error. As long as any error exists, the integral term keeps changing, driving the system toward zero error. However, integral action adds phase lag, potentially causing instability if gain is too high. It can also cause integral windup when actuators saturate. Anti-windup schemes limit integral accumulation during saturation. Integral action is essential for accurate setpoint tracking.

Derivative action responds to the rate of change of error, providing anticipatory control. When error is changing rapidly, derivative action applies strong corrective effort before the error becomes large. This reduces overshoot and improves stability. However, derivative action amplifies noise (since noise has high frequency content), so filtering or limited derivative gain is necessary. Pure derivative-only control is not used because it cannot eliminate steady-state error. Derivative is often set to zero (PI control) in noisy systems.

Frequency response describes how a system responds to sinusoidal inputs at different frequencies - specifically the gain (amplitude ratio) and phase shift as functions of frequency. It is obtained by substituting s = j*omega in the transfer function. Frequency response is important because: it characterizes system bandwidth, it predicts stability margins, it guides controller design, and it can be measured experimentally. Bode plots, Nyquist plots, and Nichols charts are common frequency response representations.

Routh-Hurwitz determines stability by examining the characteristic polynomial coefficients without solving for roots. A Routh array is constructed with first two rows from polynomial coefficients and subsequent rows calculated from determinants. The system is stable if all elements in the first column are positive. The number of sign changes equals the number of right-half-plane poles. Special cases (zero in first column, row of zeros) require special handling. It also determines the range of gain for stability.

Root locus shows how closed-loop pole locations change as a parameter (usually gain K) varies from 0 to infinity. Locus starts at open-loop poles (K=0) and ends at open-loop zeros (K=infinity). Rules determine locus shape: number of branches, real-axis segments, asymptotes, breakaway/break-in points, and angle of departure/arrival. Designers use root locus to select gain for desired pole location (damping, speed) and assess stability. MATLAB and similar tools plot root locus automatically.

Bode plots display magnitude (in dB) and phase (in degrees) versus frequency (log scale). Construction uses asymptotic approximations: each pole contributes -20 dB/decade slope above its corner frequency and -90 degrees phase, while zeros contribute +20 dB/decade and +90 degrees. Complex poles show resonant peaks. Gain margin is read at phase crossover frequency (-180 degrees); phase margin at gain crossover frequency (0 dB). Design ensures adequate margins (typically GM > 6 dB, PM > 45 degrees).

The Nyquist criterion determines closed-loop stability by examining the open-loop frequency response plot in the complex plane. The Nyquist plot is G(j*omega)*H(j*omega) for omega from -infinity to +infinity. The system is stable if the number of counterclockwise encirclements of the critical point (-1,0) equals the number of right-half-plane open-loop poles. For most systems, zero encirclements means stability. Nyquist handles open-loop unstable systems and systems with time delay, which root locus cannot address directly.

PID tuning methods include: Ziegler-Nichols (closed-loop - find ultimate gain and period, or open-loop - measure delay and time constant from step response), Cohen-Coon (based on process reaction curve), IMC (Internal Model Control - based on process model), relay autotuning (inject relay, measure oscillation), and optimization-based methods (ITAE, ISE criteria). Each method has trade-offs: Ziegler-Nichols gives aggressive tuning (may need detuning), model-based methods require accurate identification. Modern systems use autotuning algorithms.

Lead compensator (zero before pole) adds positive phase, increasing phase margin and stability, speeding up response. Used when more phase margin is needed. Transfer function: Gc = (s+z)/(s+p), where z < p. Lag compensator (pole before zero) increases low-frequency gain without affecting high-frequency behavior, reducing steady-state error. Transfer function: Gc = (s+z)/(s+p), where z > p. Lead-lag combines both effects. These compensators are simpler alternatives to PID when full three-mode control is not needed.

State-space represents systems using first-order differential equations: dx/dt = Ax + Bu (state equation), y = Cx + Du (output equation). State variables capture internal system dynamics. Advantages over transfer functions include: handles multiple inputs/outputs (MIMO) naturally, reveals all system modes (including those canceled in TF), facilitates modern control design (optimal control, observers), and directly simulates on computers. State matrix A eigenvalues equal transfer function poles. Controllability and observability are key state-space properties.

Controllability means the system can be driven from any initial state to any final state in finite time using appropriate input. The controllability matrix Qc = [B AB A^2B ... A^(n-1)B] must be full rank. Observability means all states can be determined from output measurements over time. The observability matrix Qo = [C CA CA^2 ... CA^(n-1)]^T must be full rank. These properties are essential for controller and observer design - a state feedback controller can only place poles of controllable modes.

Pole placement (pole assignment) designs state feedback gain K such that closed-loop poles are at desired locations. With u = -Kx + r, the closed-loop system has poles at eigenvalues of (A - BK). Designer chooses poles based on desired transient response (damping, speed). Ackermann's formula or direct coefficient matching computes K. The system must be controllable. All states must be measured or estimated (using an observer). Pole placement gives complete freedom in pole location unlike classical compensators.

A state observer estimates unmeasured states from output measurements. Full-order Luenberger observer: dx_hat/dt = Ax_hat + Bu + L(y - Cx_hat). Observer gain L is chosen to make (A - LC) have desired eigenvalues, typically 2-5 times faster than controller poles. System must be observable. Reduced-order observers estimate only unmeasured states. Separation principle allows independent design of controller and observer gains. Kalman filter is optimal observer for systems with noise, minimizing estimation error covariance.

Gain margin (GM) is how much gain can increase before instability, measured at frequency where phase is -180 degrees. Phase margin (PM) is how much phase can decrease before instability, measured at frequency where gain is 0 dB (unity). Both indicate robustness - how much plant variation the system can tolerate. Typical requirements: GM >= 6 dB (factor of 2), PM >= 45 degrees. Low margins indicate a system close to instability, sensitive to parameter changes. Phase margin relates to damping ratio in second-order approximation.

Cascade control uses two controllers: an inner (secondary) loop controlling a faster variable, and an outer (primary) loop controlling the main variable through the inner loop setpoint. Benefits include: faster disturbance rejection (inner loop catches disturbances before they affect main variable), improved primary control (inner loop linearizes/speeds up inner dynamics), and reduced effect of inner-loop nonlinearities. Common in temperature control (flow is inner loop) and motor control (current is inner loop). Inner loop must be faster than outer loop.

Feedforward control measures disturbances and compensates before they affect output, rather than waiting for output deviation like feedback. It requires knowledge of disturbance effect on output. Perfect feedforward cancels disturbance completely, but model errors leave residual effect. Feedforward is combined with feedback: feedforward handles measured disturbances quickly, feedback corrects model errors and unmeasured disturbances. Common in process control where major disturbances (load changes) are measurable.

System identification determines mathematical models from input-output data. Methods include: step response (measure delay, time constant, gain from step test), frequency response (sweep frequency, measure gain and phase), correlation methods (use random input, compute impulse response), and parametric methods (fit model parameters to minimize prediction error - ARX, ARMAX). Modern tools use least squares, maximum likelihood, and subspace methods. Model validation compares model output to measured data for fresh input sequence.

Integral windup occurs when actuator saturates but integral keeps accumulating error, causing large overshoot when error reverses. Prevention methods include: back-calculation (reduce integral when output saturates), conditional integration (stop integrating during saturation), integrator clamping (limit integral term directly), and tracking mode (modify integral based on actual vs commanded output). Most industrial PID controllers include anti-windup logic. The choice of method affects performance during saturation and recovery.

Digital controllers sample signals at fixed intervals and compute control output discretely. Continuous designs are discretized using: bilinear transformation (Tustin method, preserves frequency response), forward/backward difference (simpler but may cause instability), or zero-order hold equivalent (matches step response). Z-transform replaces Laplace transform: z = e^(sT) where T is sampling period. Sampling rate should be 10-20 times system bandwidth. Digital implementation includes quantization effects, computational delay, and anti-alias filtering.

The sensitivity function S(s) = 1/(1 + G(s)H(s)) describes how the closed-loop transfer function changes with plant variations. Low |S| means the system is insensitive to plant changes and rejects disturbances well. Complementary sensitivity T(s) = G(s)H(s)/(1 + G(s)H(s)) is the closed-loop transfer function; S + T = 1. Design trade-off: reducing sensitivity at low frequencies (good tracking/rejection) increases it at high frequencies (noise amplification, robustness issues). Robust control designs with sensitivity bounds.

Time delays (dead time) add phase lag without affecting magnitude: G_delay(s) = e^(-Ls). This reduces phase margin and limits achievable bandwidth. Handling approaches include: Pade approximation (rational approximation of delay for analysis), Smith predictor (model-based compensation that effectively moves delay outside the loop), delay reduction (increase sampling rate, reduce transport lag), and conservative tuning (accept slower response). Significant delay relative to time constants severely limits achievable performance.

Pole locations in s-plane determine response: real part sigma affects decay rate (more negative = faster decay), imaginary part omega_d determines oscillation frequency. For dominant complex poles: damping ratio zeta = -sigma/|pole|, natural frequency omega_n = |pole|, settling time approx 4/|sigma|, overshoot depends on zeta. Poles further left give faster response but require more control effort. Zeros can cause undershoot or slow response. Design often targets specific s-plane regions for desired transient response.

IMC uses an explicit model of the plant in the controller structure. Controller C = Q/(1 - Q*Gm) where Q is IMC controller and Gm is plant model. When model is perfect (Gm = G), closed-loop TF equals Q - designer directly specifies desired response. Robustness is achieved by low-pass filtering Q: Q = Gm^(-1) * f(s) where f(s) is filter. Model mismatch causes error, but IMC structure provides disturbance rejection. IMC can be converted to equivalent PID for implementation. Foundation for model predictive control.

MPC uses a dynamic model to predict future system behavior and optimizes control actions over a receding horizon while satisfying constraints. At each step, it solves an optimization problem (typically quadratic programming for linear MPC) minimizing predicted error and control effort, applies first control action, then repeats. Advantages over PID: handles MIMO systems naturally, explicitly handles constraints on inputs and outputs, accommodates preview information, and manages complex interactions. Challenges include computational load and need for accurate models. Widely used in process industries and autonomous vehicles.

LQR finds optimal state feedback gain K minimizing quadratic cost: J = integral(x'Qx + u'Ru)dt, where Q penalizes state deviation and R penalizes control effort. The optimal gain K = R^(-1)*B'*P where P is the solution of the algebraic Riccati equation. Q/R trade-off: larger Q/R ratio gives faster response but more control effort. LQR guarantees stability and good margins (GM >= 6 dB, PM >= 60 degrees) for single input. Extensions include LQG (adds Kalman filter for noisy measurements) and LQI (includes integral action).

H-infinity control designs controllers minimizing the H-infinity norm (maximum gain over frequency) of a closed-loop transfer function, typically from disturbance to error. This provides guaranteed worst-case performance bounds despite model uncertainty. Design specifies weighting functions shaping sensitivity at different frequencies. Robust stability is achieved by keeping complementary sensitivity bounded. H-infinity outperforms classical methods when model uncertainty is significant. Solved using Riccati equations or LMI optimization. Common in aerospace, where uncertainty bounds must be guaranteed.

Nonlinear control techniques include: Lyapunov-based design (prove stability by finding Lyapunov function, design control to make its derivative negative), feedback linearization (cancel nonlinearities using measured states, apply linear control), sliding mode control (force trajectory onto sliding surface, robust to uncertainties), backstepping (recursive design for cascaded systems), and gain scheduling (interpolate between linear controllers designed at operating points). Model predictive control handles nonlinear models directly. Analysis uses describing functions, phase plane, and center manifold theory.

Adaptive control adjusts controller parameters in real-time to handle unknown or time-varying plant parameters. Two main architectures: Model Reference Adaptive Control (MRAC) - adjusts parameters to make plant behave like a reference model, uses Lyapunov/MIT rule for adaptation; Self-Tuning Regulator (STR) - estimates plant parameters online, designs controller based on estimates. Challenges include: persistent excitation requirement, handling unmodeled dynamics, transient performance, and stability proofs. Used when plant varies (wear, environmental changes) or initial uncertainty is large.

MIMO systems have interactions between control loops - one input affects multiple outputs. Challenges include: interaction analysis (Relative Gain Array quantifies coupling), loop pairing (match inputs to outputs), decoupling (design compensators to minimize interaction), directionality (different gains in different input/output directions - singular value decomposition), and right-half-plane zeros (limit achievable bandwidth in specific directions). Classical SISO methods fail; use frequency-domain methods (SVD, Nyquist) or state-space (LQR, pole placement). Coordinated controller design is essential.

The Kalman filter optimally estimates system states from noisy measurements using a predict-correct cycle. Prediction: project state and covariance forward using system model. Correction: update prediction using new measurement, weighted by Kalman gain K (balances model uncertainty vs measurement noise). K = P*H'*(H*P*H' + R)^(-1). The filter minimizes mean-squared estimation error. Requires known process (Q) and measurement (R) noise covariances. Extensions include Extended KF (linearizes nonlinear systems) and Unscented KF (better nonlinear approximation). Ubiquitous in navigation, tracking, and sensor fusion.

Robust stability analysis determines stability despite model uncertainty. Unstructured uncertainty: represented as bounded perturbation, stability checked using Small Gain Theorem (|T|*|Delta| < 1 for stability). Structured uncertainty: use mu (structured singular value) analysis for tight bounds. Parametric uncertainty: Kharitonov's theorem checks interval polynomials, vertex-based analysis for polytopic uncertainty. Robustness margins quantify maximum tolerable uncertainty. Design iteration shapes closed-loop to achieve robust stability and performance despite worst-case uncertainty combinations.

Digital implementation considerations include: sampling rate selection (10-20x bandwidth minimum, faster for stability-critical), anti-aliasing filter design (attenuate frequencies above Nyquist), quantization effects (ADC/DAC resolution, coefficient word length), computational latency (affects phase margin), fixed vs floating point (precision vs speed tradeoffs), numerical stability of algorithm (avoid difference of large numbers), overflow protection, and bumpless transfer between manual and automatic modes. Real-time operating system requirements and deterministic execution are critical for control applications.

Passivity-based control exploits energy concepts for stability - a passive system cannot generate energy (output power <= stored energy change). Interconnection of passive systems is passive and stable. Design shapes energy function (storage function) so the closed-loop has minimum at desired equilibrium. Port-Hamiltonian framework provides systematic design method. Benefits include: inherent stability robustness, physical interpretation, handles nonlinearities naturally, and applies to networked systems. Used in robotics, power electronics, and mechanical systems where energy structure is clear.

Distributed control involves multiple controllers coordinating without centralized computation. Challenges include: limited communication (bandwidth, delays, packet loss), local information only (each agent has partial view), consensus achievement (agreement on common value), formation control (maintain spatial relationships), and scalability (local control laws independent of system size). Approaches include consensus protocols, graph-theoretic analysis, distributed optimization, and multi-agent coordination. Applications span power grids, robot swarms, traffic networks, and large-scale process plants.

Fault-tolerant control maintains acceptable performance despite component failures. Key elements: fault detection and isolation (FDI) - detect when/where failure occurred using residuals, observers, or parameter estimation; fault accommodation - reconfigure controller for degraded plant (virtual actuator/sensor concepts, control reallocation, adaptive control). Redundancy levels include hardware redundancy (multiple actuators/sensors), analytical redundancy (estimate from model), and functional redundancy (use alternative control paths). Critical systems (aircraft, nuclear) require formal verification of fault tolerance.

Hybrid systems combine continuous dynamics with discrete events (switching, logic, resets). Examples include bouncing ball, thermostat, gear shifting, and supervisory control. Modeling uses hybrid automata, switched systems, or piecewise affine systems. Stability analysis uses multiple Lyapunov functions, dwell-time conditions, or common Lyapunov functions. Control design includes switched controller synthesis, optimal switching, and MPC for hybrid systems. Challenges include Zeno behavior (infinite switches in finite time), stability under arbitrary switching, and computational complexity of hybrid MPC.

ILC improves performance over repeated operations of the same task by learning from previous attempts. Control signal for trial k+1 uses error from trial k: u_{k+1}(t) = u_k(t) + L*e_k(t). The learning operator L is designed for convergence and robustness. ILC is effective when: task repeats identically, initial conditions reset, and error can be measured. Applications include robotic pick-and-place, batch processes, CNC machining, and printing. ILC complements feedback - feedback handles disturbances, ILC handles repetitive errors.

Reinforcement learning (RL) learns control policies by interacting with the environment and receiving rewards, without requiring an explicit model. Policy optimization directly learns control law; value-based methods learn value function then derive policy. Deep RL uses neural networks for function approximation in high-dimensional spaces. Applications include robotics manipulation, autonomous driving, and HVAC optimization. Challenges include sample efficiency (many interactions needed), safety during learning (exploration may cause damage), stability guarantees, and sim-to-real transfer. Combines with model-based methods for improved efficiency.