1 On several composite quadratic Lyapunov functions for switched systems Tingshu Hu, Umass Lowell Liqiang Ma, Umass Lowell Zongli Lin, Univ. of Virginia
2 Outline Background on switched systems and sliding motion Approach of this work, main issues Use three Lyapunov functions to construct switching laws How to ensure stability in the presence of sliding motion? Main Results Switching law based on directional derivatives Directional derivatives along sliding surface Stabilization via min function Dual result for max function and convex hull function Conclusions
3 Switched systems and LDIs Given a family of linear systems A linear differential inclusion (LDI): Switching controlled by unknown force Expect the worst case Switching orchestrated by controller, Can be optimized for best performance This work considers the second type, the switched systems A switched system Two approaches: Analytical methods for lower-order sys [ Antsaklis, Michel, Hu, Xu, Zhai] Using Lyapunov functions for switching laws construction and stability analysis
4 Earlier constructive approaches Find P > 0 and [0,1] such that [Wicks, Peletics & Decarlo, 1998] A switching law can be constructed for quadratic stability Find P 1, P 2 >0, ≥ or ≤ such that [Wicks & Decarlo, 1997] A switching law can be constructed based on V(x) = max{x T P 1 x, x T P 2 x}, or V(x) = min{x T P 1 x, x T P 2 x}. Stability ensured only if no sliding motion occurs. Both methods involving BMIs, harder than LMIs but numerically possible More recent development based on BMIs: [Decarlo, Branicky, Pettesson, Lennartson, Zhai, etc].
5 Effect of sliding motion Assume the matrix condition is satisfied [Wicks & Decarlo, 1997] When sliding motion occurs, the system can be stable or unstable Sliding motion not unusual in switched systems; It may be a result of optimization Not realizable but can be approximated via hysteresis, delay, fast sampling, etc. Based on V min Based on V max
6 Approach of this work, main issues Approach: Use three types of Lyapunov functions to construct switching laws Functions composed from a family of quadratic functions; Lead to semi-definite matrix conditions, numerically possible; Two types of functions are not everywhere differentiable Main issues: How to dealing with non-differentiable Lyapunov functions? Use directional derivatives How to address stability in the presence of possible sliding motion? Exclude the existence of sliding motion Characterize directional derivatives along sliding direction
7 Three Lyapunov functions Given matrices : 3) The convex hull function: 2) The max function : 1) The min function: Level set = Only the convex hull function is everywhere differentiable V c and V max are convex conjugate pairs, they have been successfully applied to LDIs and saturated systems [Goebel, Hu, Lin, Teel, Zaccarian]. Let Level set =
8 Switching law based on directional derivatives Consider a function V ( x ). The one-sided directional derivative at x along the direction is, Then, For the family of linear systems Let the switching law be constructed as V(x) can be V max (x), V min (x ), or V c (x)
9 How sliding motion complicates the analysis? Along sliding direction: Sliding along the set of differentiable points is easy to deal with Sliding along the set of non-differentiable points is more complicated What really matters is
10 Directional derivatives along sliding surfaces A1xA1x A2xA2x A 1 x+(1- )A 2 x Different situations w.r.t V max and V min A2xA2x A1xA1x A 1 x+(1- )A 2 x Not necessary to require Not sufficient to require How can we ensure along sliding direction?
11 Some key points in this work A1xA1x A2xA2x A 1 x+(1- )A 2 x For V max, If there exists a s.t. at the non-differentiable point, then along the switching surface. For V min, no sliding motion exist in the set of non-differentiable points Note: A 1 x and A 2 x points away from switching surface Only need to consider the set of differentiable points A2xA2x A1xA1x A 1 x+(1- )A 2 x
12 Stabilization via min function The switching law: Proposition 1: There exist no sliding motion in the set of x where V min ( x ) is not differentiable. Matrix condition: Consider V min = min{x T P j x: j=1,2,…,J}. If there exist ij ≥0, ij ≥0, i=1 N ij =1, such that Then for every solution, Stability ensured by matrix condition even if sliding motion occurs.
13 Stabilization via min function Special case with two A i ’ s: The matrix inequalities [Wicks & Decarlo, 1997] ensures stabilization regardless of sliding motion. The number of matrices P j ’s (J) does not need to be equal to the number of A i ’s (N) : J≥N, or J<N. As J increases, the convergence rate increases. Example: System cannot be stabilized via quadratic Lyapunov functions: No P>0 and satisfy both neutrally stable With J=2, maximal ; With J=3, maximal ; With J=4, maximal ;
14 A dual result for V max and V c Recall V max and V c are conjugate functions Consider the dual switched systems: Proposition: Suppose N=2. Sys 1 is stable iff Sys 2 is stable. Sys 1: Sys 2: Remarks: ₋ Results also obtained for the case N >2. ₋ It is easier to obtain matrix conditions via Sys. 2 since V max is piecewise quadratic.
15 Example: A pair of dual systems Sys 1: Sys 2: Sliding motion occurs for both systems. They are stable with the same convergence rate w.r.t correspoding Lyapunov function.
16 Conclusions Switching laws constructed via three Lyapunov functions The min function The max function The convex hull function Sliding motion carefully considered by using the directional derivatives When min function is used, sliding motion does not exist in the set of non-differentiable points When max function is used, V max decreases along the sliding surface iff it decreases along a certain convex combination of A 1 x and A 2 x. A dual result obtained via max function and convex hull function Condition for stabilization characterized by BMIs.