Control Systems 2 Lecture 6: Uncertainty and robustness in SISO systems Roy Smith 2016-4-4 6.1 Robustness analysis Nominal stability (NS) Is the closed-loop system stable when the plant is known exactly? Robust stability (RS) Is the closed-loop system stable when there is uncertainty in our knowledge of the plant? Nominal performance (NP) Does the closed-loop system meet the performance specifications when the plant is known exactly? Robust performance (RP) Does the closed-loop system meet the performance specifications when there is uncertainty in our knowledge of the plant? 2016-4-4 6.2 Uncertain models Model sets G(s) ∈ { Gnom (s) + ∆ | k∆k ≤ γ } k∆k Gnom (s) Gnom (s) = Nominal plant ∆ = unknown, but bounded, perturbation (i/o operator) Typically, ∆ is stable, causal and satisfies, k∆kH∞ ≤ γ. 2016-4-4 6.3 Uncertain models Multiplicative perturbation model d Gd (s) y + Wm (ω) ∆ + y(s) = G(s)u(s) + Gd (s)d(s), Gnom (s) where u G(s) ∈ G, G = { (1 + Wm (ω)∆)Gnom (s) | ∆ is causal, stable; k∆kH∞ ≤ 1 } . 2016-4-4 6.4 Uncertain models Multiplicative perturbation model d Gd (s) y y + = = Wm (ω) ∆ + Gnom (s) u (I + Wm (ω)∆)Gnom u + Gd d Gnom u | {z } + Wm (ω)∆Gnom u | {z } nominal response + Gd d |{z} perturbation response disturbance response 2016-4-4 6.5 Example: uncertain model (first order plus delay) Ke−λs G= 1 + τs Nominal case Magnitude 10 Gnom (jω) 0.1 1 10 0.1 1 10 log ω (rad/sec) 1 K = 10 τ = 1.0 0.1 λ = 0.5 0 log ω (rad/sec) −90 Gnom (jω) −180 −270 2016-4-4 Phase (deg.) 6.6 Example: uncertain model (first order plus delay) imag Ke−λs G= 1 + τs 4 2 Nominal case real K = 10 −4 τ = 1.0 −2 2 4 6 8 10 −2 λ = 0.5 −4 Gnom (jω) −6 −8 2016-4-4 6.7 Example: uncertain model (first order plus delay) Random examples Ke−λs G= 1 + τs Magnitude 10 Gnom (jω) 0.1 1 10 log ω (rad/sec) 1 G(jω) Perturbed case ±15% uncertainty K ∈ [8.5, 11.5] τ ∈ [0.85, 1.15] λ ∈ [0.425, 0.575] also includes boundary cases. 2016-4-4 0.1 0.1 1 10 0 log ω (rad/sec) −90 Gnom (jω) −180 −270 G(jω) Phase (deg.) 6.8 Example: uncertain model (first order plus delay) Absolute error (random examples) Error as a function of frequency: |Gnom (jω) − G(jω)| Magnitude 10 Gnom (jω) 0.1 1 10 log ω (rad/sec) 1 |Gnom (jω) − G(jω)| 0.1 2016-4-4 6.9 Example: uncertain model (first order plus delay) Relative error (random examples) Error as a function of frequency: |Gnom (jω) − G( jω)| |Gnom (jω)| Magnitude 10 Gnom (jω) 0.1 1 10 log ω (rad/sec) 1 |G(jω) − Gnom (jω)| |Gnom (jω)| 0.1 2016-4-4 6.10 Example: uncertain model (first order plus delay) Modeling problem: choose Gnom and Wm Wm (ω) y ∆ + Gnom (s) u Upper bound on relative uncertainty: choosing Wm (s) G(s) = (1 + Wm (ω)∆)Gnom (s) =⇒ so G(jω) − Gnom (jω) = Wm (ω)∆ Gnom (jω) G(jω) − Gnom (jω) ≤ |Wm (ω)| =⇒ |∆| ≤ 1. Gnom (jω) 2016-4-4 6.11 Example: uncertain model (first order plus delay) Upper bound on relative uncertainty: choosing Wm (s) G(jω) − Gnom (jω) ≤ |Wm (ω)|. Gnom (jω) (See Laughlin et al. for the Wm (s) formula) Magnitude 10 Gnom (jω) 0.1 1 10 log ω (rad/sec) 1 Wm (jω) |G(jω) − Gnom (jω)| |Gnom (jω)| 0.1 2016-4-4 6.12 Example: uncertain model (first order plus delay) imag Random examples 4 G= −λs Ke 1 + τs 2 real Perturbations: ±15% −4 −2 2 4 6 8 10 −2 on K, λ, τ . −4 Gnom (jω) −6 −8 2016-4-4 6.13 Example: uncertain model (first order plus delay) imag Random examples 4 G= −λs Ke 1 + τs |Gnom (jω)Wm (jω)| 2 real Perturbations: ±15% on K, λ, τ . −4 −2 2 4 6 8 10 −2 −4 Gnom (jω) −6 −8 2016-4-4 6.14 Example: uncertain model (first order plus delay) imag Random examples 4 G= −λs Ke 1 + τs |Gnom (jω)Wm (jω)| 2 real −4 Perturbations: ±15% −2 2 4 6 8 10 −2 on K, λ, τ . −4 Gnom (jω) −6 −8 perturbation regions 2016-4-4 6.15 Example: uncertain model (first order plus delay) G(s) = ( G(s) ∈ ) Ke γ = 10, τ = 1, λ = 0.5, (all ± 15%) 1 + τs −λs (1 + Wm (s)∆)Gnom (s) where here, Wm (s) = ∆kH∞ ≤ 1 , 1.15(1 + τ s) 0.15λs e −1 (1 + 0.85τ s) (up to its maximum, and then constant). Magnitude 10 Gnom (jω) 0.1 1 10 log ω (rad/sec) 1 Wm (jω) 0.1 2016-4-4 6.16 Design example: Loopshaping design Gnom (s) = Ke−λs , 1 + τs K = 10, τ = 1, λ = 0.5 j 0.15(1 + τ s) K(s) = s and 0.15Ke−λs Lnom (s) = s imag 2 real −2 2 4 6 8 −1 real 10 −2 −4 Gnom (jω) −6 Lnom (jω) Lnom (jω) −8 −j imag −10 2016-4-4 6.17 Nominal stability Nyquist criterion in more detail 0.15Ke−λs Lnom (s) = s imag -1 r −→ 0 real R −→ ∞ real Lnom (jω) imag 2016-4-4 6.18 Nominal stability example: Loopshaping design Ke−λs Gnom (s) = , 1 + τs 0.15(1 + τ s) K(s) = , s 0.15Ke−λs Lnom (s) = s Magnitude 10 Gnom (jω) 0.1 1 10 log ω (rad/sec) 1 Lnom (jω) K(jω) 0.1 0.1 1 10 0 K(jω) −90 Gnom (jω) −180 −270 log ω (rad/sec) Lnom (jω) Phase (deg.) 2016-4-4 6.19 Nominal performance |Wp (jω)Snom (jω)| < 1 for all ω. imag Lnom (jω) |Wp (jω)| −1 real |1 + Lnom (jω)| 2016-4-4 6.20 Nominal performance Nominal weighted sensitivity: Nominal performance ⇐⇒ In the loopshaping example: 1 1 + Gnom (s)K(s) Snom (s) = |Snom (jω)| < Wp( s) = 1 |Wp (jω)| s + 2.5 . 5s Magnitude 10 Wp −1 (jω) 0.1 1 log ω (rad/sec) 10 1 Snom (jω) 0.1 0.01 2016-4-4 6.21 Robust stability Closed-loop multiplicative perturbation model Wm (ω) y + ∆ Gnom (s) K(s) + r − Robust stability Is this closed-loop stable for all stable, causal ∆, with k∆kH∞ ≤ 1? 2016-4-4 6.22 Robust stability Wm (ω) y ∆ + Gnom (s) K(s) r + − Robust stability S(s) = 1 1 = 1 + G(s)K(s) 1 + (1 + Wm (s)∆)Gnom (s)K(s) is stable for all ∆, k∆kH∞ ≤ 1. 2016-4-4 6.23 Robust stability Wm (ω) y ∆ + Gnom (s) K(s) r + − −Wm (ω)Tnom (s) Snom (s)Wm (ω) y 2016-4-4 + v ∆ z + Tnom (s) r 6.24 Robust stability: an equivalent test −Wm (ω)Tnom (s) v Snom (s)Wm (ω) y z ∆ + + Tnom (s) r ∆ z v Wm (s)Tnom (s) 2016-4-4 6.25 Small gain theorem u1 + e1 y2 y1 M1 (s) M2 (s) e2 + u2 Given e1 , e2 with ke1 k < ∞ and ke2 k < ∞, define, u1 = e1 − M2 (s)e2 and u2 = e2 − M1 (s)e1 . Suppose there exists γ1 > 0 and γ2 > 0 such that, kM1 (s)e1 k ≤ γ1 ke1 k and kM2 (s)e2 k ≤ γ2 ke2 k. If γ1 γ2 < 1 then, ky1 k ≤ 2016-4-4 γ1 (ku1 k + γ2 ku2 k). 1 − γ1 γ2 6.26 Robust stability: an equivalent test ∆ z v Wm (s)Tnom (s) By applying the small-gain theorem: If k∆kkWm (s)Tnom (s)k < 1 then the perturbed closed-loop system is stable. Or: If kWm (s)Tnom (s)kH∞ < 1 then the perturbed closed-loop system is stable (RS). 2016-4-4 6.27 Robust stability Robust stability ⇐⇒ 1 1 + G(s)K(s) is stable for all G(s) ∈ G. ⇐⇒ 1 1 + (1 + Wm (s)∆)Gnom (s)K(s) is stable for all k∆kH∞ ≤ 1 2016-4-4 ⇐⇒ kWm (s)Tnom (s)kH∞ < 1 ⇐⇒ |Wm (jω)Tnom (jω)| < 1 ⇐⇒ |Tnom (jω)| < 1 |Wm (jω)| for all ω for all ω. 6.28 Robust stability |1 + Gnom (jω)K(jω)| > |Gnom (jω)Wm (jω)K(jω)| ⇐⇒ |Tnom (jω)| < 1 |Wm (jω)| for all ω. imag Lnom (jω) −1 real |1 + Lnom (jω)| |Lnom (jω)Wm (jω)| 2016-4-4 6.29 Robust stability: Loopshaping design example Robust stability ⇐⇒ |Tnom (jω)| < 1 |Wm (jω)| for all ω. Magnitude 10 Wm −1 (jω) 0.1 1 10 log ω (rad/sec) 1 Tnom (jω) 0.1 0.01 2016-4-4 6.30 Example: uncertain model (first order plus delay) Nyquist criteria: perturbed loopshape regions imag −2 −1 real |Lnom (jω)Wm (jω)| loopshape perturbation regions −2 Lnom (jω) 2016-4-4 6.31 Loopshaping approximations Nominal performance and robust stability If Lnom (jω) 1 (high freq.) Magnitude 100 T (jω) < 1/|Wm (jω)| is approximately Lnom (jω) |Lnom (jω)| < |Wm−1 (jω)| 10 Wp (jω) 0.1 1 log ω (rad/sec) 10 1 Wm−1 (jω) 0.1 If Lnom (jω) 1 (low freq.) S(jω) < 1/|Wp (jω)| 0.01 is approximately |Lnom (jω)| > |Wp (jω)| 2016-4-4 6.32 Robust performance Robust performance ⇐⇒ ⇐⇒ W (s) p 1 + G(s)K(s) H∞ < 1, for all G(s) ∈ G. W (s) p 1 + (1 + Wm (s)∆)Gnom (s)K(s) < 1, H∞ for all k∆kH∞ ≤ 1 ⇐⇒ |Wp (jω)| < 1, |1 + (1 + Wm (jω)∆)Gnom (jω)K(jω)| for all ω and all |∆| ≤ 1 which is equivalent to, |Wp (jω)Snom (jω)| + |Wm (jω)Tnom (jω)| < 1 for all ω. 2016-4-4 6.33 Robust performance |Wp (jω)Snom (jω)| + |Wm (jω)Tnom (jω)| < 1 for all ω. imag Lnom (jω) |Wp (jω)| −1 real |1 + Lnom (jω)| |Lnom (jω)Wm (jω)| 2016-4-4 6.34 Robust performance example |Wp (jω)Snom (jω)| + |Wm (jω)Tnom (jω)| < 1 for all ω. 1.25 1.0 0.75 Robust performance 0.5 Nominal performance 0.25 Robust stability 0 0.1 1 10 log ω (rad/sec) 2016-4-4 6.35 Notes and references Skogestad & Postlethwaite (2nd Ed.) Pertubation models: sections 7.1, 7.2, 7.3, 7.4 SISO robust stability: section 7.5 SISO robust performance: section 7.6 Perturbation bound, Wm (s), for the RS example “Smith predictor design for robust performance,”DL Laughlin, DE Rivera, M Morari, Int. J. Control, v. 46, no. 2, pp. 477–504, 1987. 2016-4-4 6.36