Event-Triggering of Large-Scale Systems without Zeno Behavior Claudio De Persis, Rudolf Sailer, Fabian Wirth University of Würzburg 09.07.2012 Outline 1 Motivation 2 Preliminaries (Standing Assumptions and Introduction to Small-Gain Conditions) 3 Practical Stabilization 4 Numerical Example 5 Outlook and Summary Fabian Wirth (University of Würzburg) MTNS12 09.07.2012 2 / 26 Outline 1 Motivation 2 Preliminaries (Standing Assumptions and Introduction to Small-Gain Conditions) 3 Practical Stabilization 4 Numerical Example 5 Outlook and Summary Fabian Wirth (University of Würzburg) MTNS12 09.07.2012 3 / 26 Motivation Stabilization of a large-scale system over digital channels Every subsystem communicates over a digital channel. Communication is a limited resource. Aim: reduce the amount of communication. Fabian Wirth (University of Würzburg) MTNS12 09.07.2012 4 / 26 Motivation Approach Lower the amount of superfluous data transmissions Fabian Wirth (University of Würzburg) MTNS12 09.07.2012 4 / 26 State Periodic Sampling Fabian Wirth (University of Würzburg) t MTNS12 09.07.2012 5 / 26 State Event-Based Sampling (δ-Sampling) Fabian Wirth (University of Würzburg) t MTNS12 09.07.2012 6 / 26 Drawback of Event-Triggering: Zeno Solutions Zeno Event-triggering may lead to Zeno solutions Example: Bouncing ball (with damping) Infinite number of discrete events in finite time How can we deal with the Zeno phenomenon? Fabian Wirth (University of Würzburg) MTNS12 09.07.2012 7 / 26 Outline 1 Motivation 2 Preliminaries (Standing Assumptions and Introduction to Small-Gain Conditions) 3 Practical Stabilization 4 Numerical Example 5 Outlook and Summary Fabian Wirth (University of Würzburg) MTNS12 09.07.2012 8 / 26 Systems under Consideration: x state, x̂ information available to the controller Single system ẋi = fi (x, ui ) x̂˙ = 0 ui = ki (x + e) e = x̂ − x xi , ei ∈ Rni x = (x1T , . . . , xnT )T i = 1, . . . , n e = (e1T , . . . , enT )T Overall system f1 (x, k1 (x + e)) .. ẋ = f (x, k(x + e)) = f (x, k(x̂)) = . fn (x, kn (x + e)) Fabian Wirth (University of Würzburg) MTNS12 09.07.2012 9 / 26 Event-Triggering Given continuous triggering functions Ti , i = 1, . . . , n, the condition Ti (xi , ei ) ≥ 0 implicitly defines event times (triggering times) tk by t1 := inf{t > 0 : ∃i s.t. Ti (xi (t), ei (t)) ≥ 0} tk+1 := inf{t > tk : ∃i s.t. Ti (xi (t), ei (t)) ≥ 0} The system i for which Ti ≥ 0 holds, transmits its state to all controllers kj at tk . Hence the ith error will be set to zero. Ti (xi , ei ) ≥ 0 ⇒ ei (tk+ ) = 0 Fabian Wirth (University of Würzburg) MTNS12 09.07.2012 10 / 26 Standing Assumption 1 For each subsystem there exists a proper, smooth function Vi : Rni → R+ and γij , ηij ∈ K∞ ∪ {0}, αi positive definite, ci > 0 such that Assumption 1 (ISpS-Lyapunov Function) Vi (xi ) ≥ max{γij (Vj (xj )), ηij (||ej ||), ci } ⇒ ∇Vi (xi )fi (x, ki (x + e)) ≤ −αi (||xi ||). j K∞ functions are ΓHsL continuous zero at zero strictly increasing unbounded Fabian Wirth (University of Würzburg) s MTNS12 09.07.2012 11 / 26 Some Remarks on the Gains γ, η ISpS-Lyapunov function Vi (xi ) ≥ max{γij (Vj (xj )), ηij (||ej ||), ci } ⇒ ∇Vi (xi )fi (x, ki (x + e)) ≤ −αi (||xi ||) j γi1 γ1i Σ1 η11 ηi1 η1i γ1i γi1 k1 Fabian Wirth (University of Würzburg) γni γin Σi ηii ki MTNS12 Σn ηni ηin γin γni ηnn kn 09.07.2012 12 / 26 Standing Assumption 2 Define max{γ11 (σ1 (r )), . . . , γ1n (σn (r )), ϕ11 (r ), . . . , ϕ1n (r )} .. Γ(σ(r ), ϕ(r )) := . max{γn1 (σ1 (r )), . . . , γnn (σn (r )), ϕn1 (r ), . . . , ϕnn (r )} Assumption 2: A small-gain condition holds n and ϕ ∈ Kn×n such that There exists σ ∈ K∞ ∞ Γ(σ(r ), ϕ(r )) < σ(r ) , Fabian Wirth (University of Würzburg) MTNS12 ∀r > 0 . 09.07.2012 13 / 26 Outline 1 Motivation 2 Preliminaries (Standing Assumptions and Introduction to Small-Gain Conditions) 3 Practical Stabilization 4 Numerical Example 5 Outlook and Summary Fabian Wirth (University of Würzburg) MTNS12 09.07.2012 14 / 26 Existence of a Lyapunov Function for the Overall System Theorem 1 Let there be an ISpS-Lyapunov function Vi for each subsystem. Let the small gain condition hold. Define V (x) = max σi−1 (Vi (xi )) , i η̂j = max ϕ−1 ij ◦ ηij . i Then there exits a positive definite α : R+ → R+ such that if max{σi−1 (Vi (xi )), ci } ≥ η̂i (kei k) ∀i , then V (x) ≥ ĉ ⇒ hp, f (x, k(x + e))i ≤ −α(||x||), ∀p ∈ ∂V (x) , with ĉ = maxi {ci , σi−1 (ci )}. Fabian Wirth (University of Würzburg) MTNS12 09.07.2012 15 / 26 Intuition Recall max{σi−1 (Vi (xi )), ci } ≥ η̂i (kei k) , η̂j = max ϕ−1 ij ◦ ηij . i Intuition The small-gain condition ensures that the interconnected system is stable if e ≡ 0. Scaling σi comes from the small-gain condition ηij describes the effect of ej to system i. ϕij is a damping of ηij such that the small gain condition holds despite e. Overall: If the error is small compared to the state, it cannot destroy stability. Fabian Wirth (University of Würzburg) MTNS12 09.07.2012 16 / 26 What Does the Triggering Condition Look Like? Corollary 1 Consider again ẋ = f (x, k(x̂)) . Define the triggering condition of the ith subsystem as Ti (xi , ei ) = η̂i (kei k) − max{σi−1 (Vi (xi )), ci } ≥ 0 , then the equilibrium is practically asymptotically stable. Important: The triggering condition only uses local information. Fabian Wirth (University of Würzburg) MTNS12 09.07.2012 17 / 26 What Does the Triggering Condition Look Like? Corollary 1 Consider again ẋ = f (x, k(x̂)) . Define the triggering condition of the ith subsystem as Ti (xi , ei ) = η̂i (kei k) − max{σi−1 (Vi (xi )), ci } ≥ 0 , then the equilibrium is practically asymptotically stable. Important: The triggering condition only uses local information. Fabian Wirth (University of Würzburg) MTNS12 09.07.2012 17 / 26 Practical Stabilization Prevents Zeno Behavior Lemma 1 Setting ci = 0 in Corollary 1 can lead to Zeno solutions. If the triggering condition from Theorem 1 exhibits Zeno behavior, then a subsystem triggering infinitely often has to converge to zero in finite time. Vi (xi ) Vi (xi ) ci kei k kei k η̂i (kei k) ≤ max{σi−1 (Vi (xi )), ci } η̂i (kei k) ≤ σi−1 (Vi (xi )) Fabian Wirth (University of Würzburg) MTNS12 09.07.2012 18 / 26 Outline 1 Motivation 2 Preliminaries (Standing Assumptions and Introduction to Small-Gain Conditions) 3 Practical Stabilization 4 Numerical Example 5 Outlook and Summary Fabian Wirth (University of Würzburg) MTNS12 09.07.2012 19 / 26 Example ẋ1 =x1 x2 + x12 u1 ẋ2 =x12 + u2 u1 = −(x1 + e1 ) , u2 = −k(x2 + e2 ) , k > 0 . 1 Vi (xi ) = xi2 , 2 1 σ1 ◦ η̂1 (r ) = √ r 2 , 8ν 4 σ2 ◦ η̂2 (r ) = ν 2k 2 2 r 4 4.5 4 3 3.5 2 3 1 2.5 2 0 1.5 −1 1 −2 0.5 −3 0 0.05 0.1 0.15 0.2 0.25 0.3 Fabian Wirth (University of Würzburg) 0.35 0.4 0.45 0.5 0 0 MTNS12 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 09.07.2012 0.45 0.5 20 / 26 Effect of different ci 4 4.5 4 3 3.5 17 events 2 c2 = 0.16 3 1 2.5 2 0 1.5 −1 1 −2 0.5 −3 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.35 0.4 0.45 0.5 4.5 4 4 3 3.5 6 events 2 c2 = 1.86 3 1 2.5 2 0 1.5 −1 1 −2 −3 0.5 0 0.05 0.1 0.15 0.2 0.25 0.3 Fabian Wirth (University of Würzburg) 0.35 0.4 0.45 0.5 MTNS12 0 0 0.05 0.1 0.15 0.2 0.25 0.3 09.07.2012 0.5 21 / 26 Larger ci do not necessarily lead to fewer events 350 # of events 300 250 200 150 100 c=1.86 50 0 0 1 Fabian Wirth (University of Würzburg) 2 3 4 MTNS12 5 6 7 c2 8 09.07.2012 22 / 26 Outline 1 Motivation 2 Preliminaries (Standing Assumptions and Introduction to Small-Gain Conditions) 3 Practical Stabilization 4 Numerical Example 5 Outlook and Summary Fabian Wirth (University of Würzburg) MTNS12 09.07.2012 23 / 26 Some Remarks Possible extensions Different comparison functions for the gains (e.g. + instead of max) Relation between Assumption 2 and known small-gain theorems External disturbances x̂˙ 6= 0 Quantization ci → 0 sufficiently slowly will result in asymptotic stability Open problems Delay and packet loss Collision avoidance Choosing ci in an “optimal” manner Fabian Wirth (University of Würzburg) MTNS12 09.07.2012 24 / 26 Summary Summary ISS Lyapunov functions + small gain + event based sampling stability Event based sampling can lead to fewer transmissions The systems decide how often they have to transmit to ensure stability This decision is based only on local information Hybrid systems can exhibit Zeno behavior Practical stability can prevent the occurrence of Zeno behavior Choosing ci that leads to a minimal number of events is not trivial Fabian Wirth (University of Würzburg) MTNS12 09.07.2012 25 / 26 Thank you for your attention! Fabian Wirth (University of Würzburg) MTNS12 09.07.2012 26 / 26