TUTORIAL on Networked Control Systems with Delay Cicsyn2010 2nd International Conference on Computational Intelligence, Communication Systems and Networks Liverpool, UK, July 29th, 2010 Vasilis Tsoulkas: Center for Security Studies, Athens, Greece & Dept. of Mathematics, University of Athens Research Group: - Pantelous Athanasios., University of Liverpool, - Dritsas Leonidas., - Halikias George, Hellenic Airforce Academy City University, London, UK. 1 Contents 1. 2. 3. 4. 5. 6. 7. 8. Introduction – General Features NCS Modeling - the issue of network induced delays Discretization of NCS dynamics Decomposing the Uncertain Delay (nominal and uncertain parts) and the NCS dynamics Robust Stability Analysis based on the augmented closed-loop vector (“ξ”) Design of a Simple Output Tracking Controller Investigation of Robust Tracking Performance via Simulation - Numerical Examples for Networked Stable and Unstable systems Conclusions & Topics for further study 2 Schematics of Networked Control Systems Networked control systems (NCSs) are spatially distributed systems for which the communication between sensors, actuators, and controllers is supported by a shared communication network. Hespanha et al.: “Survey of Recent Results in Networked Control Systems” (Proceedings of the IEEE, Vol. 95, No. 1, January 2007) 3 Motivation & Some Benefits Easy and low cost installation, wiring, maintenance, configuration Distributed Controllers and Plant with low cost distributed sensors and actuators are all coupled over the same Real Time communications network The distributed nature of elements offers great flexibility of architectures. Applicable in a wide variety of fields such as: Remote surgery, mobile sensor networks, UAV’s, Space teleoperations and Robotics. 4 Distributed Networked System 5 Control networks are indicated by solid lines, and diagnostics networks are indicated by dashed lines. 6 1. Introduction Feedback control systems wherein the control loops are closed through a real-time network are called Networked Control Systems (NCSs) Defining feature of NCS: Information (reference input, plant output, control input, etc.) is exchanged using a network among control system components (sensors, controllers, actuators, 7 7 etc.). 1. Introduction Νetwork Induced Delays Information flow in the control loop is delayed due to – buffering, – access contention (the time a node waits until it gets access to the network), – computation delay (assume absorbed into {τca (k)} ) – propagation (“transmission”) delays. – Network-induced delays in NCS appear in the information flow between (“k” denotes the dependence on the kth sampling period). A). The sensor and the controller {τsc (k)}, (controller receives “outdated” information about process behavior) B). The controller and the actuator {τca (k)}, (control action cannot be applied “on time” and the controller does not know the exact instance the calculated control signal will be received by the actuator) 8 1. Introduction Νetwork Induced Delays When a static linear time invariant controller is employed, can lump the delays τsc (k), τca (k), into τk= τsc (k)+τca (k). Network-induced delays in NCS between the sensor and the controller {τsc (k)}, and between the controller and the actuator {τca (k)}, (“k” denotes the dependence on the kth sampling period). 9 1.Introduction Tracking Control Design for NCS The Usual Approach for NCS Analysis & Design: – design a controller ignoring the network, then – analyze stability, performance and robustness with respect to the effects of network-delays and scheduling policy…(usually via the selection of an appropriate scheduling protocol). The issue of Tracking Control over Networks has not been adequately met – very limited published work on NCS Tracking !!! – the majority of NCS publications concerns regulation , (“design a controller which brings the output/state to “0” ) – many results on tracking for Time Delayed Systems (TDS) but cannot be applied “as is” to NCS due to the “Networkcentric” nature of NCS e.g. special nature of delays in NCS the fundamental issue of “Packet Loss/Drops” Scheduling, Quality of Service, Middleware 10 1.Introduction Tracking Control Design for NCS Concerning NCS Robust Tracking Performance… – only preliminary results - no strict mathematical proofs – yet…useful “lessons learned” through extensive simulations on S.I.S.O systems – we investigate both constant unknown or time-varying uncertain delays with known bounds – we do not take into account the network delays in the tracking controller design process… – “a posteriori” analysis of stability, performance and conservatism of results – we do not take into account “packet drops” – Analysis & Synthesis in the continuous time domain – No need to assume knowledge of the P.D.Fs (not a stochastic approach) 11 2. NCS Modeling NCS with network-induced delays in the actuation and sensor path Assumptions made … • the dynamics of the NCS under investigation is a combination of a continuous–time LTI plant with a discrete–time controller. • Time Invariant controller can lump τsc (k), τca (k), into τk= τsc (k)+τca (k). • Single source of uncertainty and performance degradation the lumped transmission delay τk. 12 • No plant uncertainties or nonlinearities - No packet drops 2. NCS Modeling - Assumptions In Practice… the dynamics of the NCS under investigation is a combination of a continuous–time uncertain/nonlinear plant with a discrete–time (“sampled-data”) controller. The sampler is time-driven, whereas both controller and actuator are event-driven, (=they update their outputs as soon as they receive a new sample). Some packets are lost or intentionally dropped (contain obsolete/useless info) 13 The delays τksc , τkca , τk < h τksc = τsc (k) is the delay experienced by a state or output sample x(kh), y(kh), sampled at time instance “kh” and presented –after a delay τksc to the event–driven remote controller for control computation purposes. τkca =τ ca (k) is the delay experienced by the control–action, computed immediately after its reception at time instance kh+ τksc until it is transmitted via the network to the Z.O.H (and finally presented to the event–driven actuator). The computation delay is absorbed into τ kca 14 The delays τksc , τkca , τk < h τ k : Total delay within the kth sampling period, i.e. the time from the instant when the sampling node samples sensor data from the plant to the instant when actuators exert a control action –whose computation was based on this sample– to the plant. τk= τksc + τkca (since a static time invariant control law is employed) Known Bounds: 0 ≤ τ min < τk < τ max ≤ h 15 NCS Timing Diagram (τk < h) for “short” (τk < h) + bounded delay 0 ≤ τ min < τk < τ max ≤h 16 2. NCS Modeling Difficulties in case of Discrete “Sampled Data” Controller û(t) is the “most recent” control action presented to the event–driven actuator at the time instance t within a sampling period [kh, kh + h) & can take two values ûk or ûk-1 û(t) experiences a “jump” at the uncertain or unknown time instance kh+ τ k , changing from ûk-1 into ûk (uncertain actuation instance) Very Complicated Dynamics Impulse Delayed Systems, Asynchronous Dynamical Systems, Hybrid Systems, etc even for the regulation case (r=0) 17 2. NCS Modeling - the issue of network-induced delays NCS Timing Diagram form Zhang & Branicky paper (IEEE Control Systems Magazine, Febr.2001). Possible misconceptions if symbols are not adequately clarified… Authors clarify that the confusing symbol “u(kh)” denotes the actuation that takes place at kh+ τk and its value is u(kh) = -Kx(kh) Hence (unless τk is constant) it is not possible to treat the ensuing NCS in a standard “sampled data” or “time-delayed” setting. Instead a “hybrid” setup should rather be used, as for example the one presented in P. Naghshtabrizi and J. P. Hespanha, “Stability of network control systems with variable sampling and delays” in Proc. of the 44th Annual Allerton Conf. on Communication, Control, and Computing, 2006. 18 CONTENTS 1. 2. 3. 4. 5. 6. 7. 8. Introduction – General Features NCS Modeling - the issue of network-induced delays Discretization of NCS dynamics Decomposing the Uncertain Delay (nominal and uncertain parts) Robust Stability Analysis based on the closed-loop augmented vector (“ξ”) Design of A Simple Output Tracking Controller Investigation of Robust Tracking Performance via Simulation - Numerical Examples for Networked Stable and Unstable systems Conclusions & Topics for further study 19 3. Descretization of NCS state equation with “small delay” x(t ) Ac x(t ) Bcuˆ(t ) and y(t ) Cc x(t ) τk < h xk = x(kh) xk+1 = Φ xk + Γ0(τk) ûk + Γ1 (τk) ûk-1 (Σ1) notation xk, xk-1,… denotes the values x(kh), x(kh-h), … of the periodically sampled discrete–time signal coming out of the sampler. The same notation for yk, yk-1,… We keep the “hat” notation for ûk , ûk-1 as a reminder of the asynchronous, (“jump”) nature of these signals. Õn … is an n-column zero vector, In is the n x n identity matrix, 0n is the n x n zero matrix. MT is the transpose of a matrix. M > 0 (< 0) means that M is positive (negative) definite. 20 3. Discretization of state equation dynamics of NCS (Comments) xk+1 = Φ xk + Γ0(τk) ûk + Γ1 (τk) ûk-1 (Σ1) Exact Discretization between equidistant sampling instances finite dimensional dynamics… The uncertain time varying delay τk can still take any (out of infinite) values within the allowable interval the uncertainty of τk generates an uncertainty in the actuation instance System matrices (Γ0(τk), Γ1(τk)) are uncertain Presence of a delayed input term ûk-1 21 Exact Discretization despite the “jump’’ nature of û(t) xk = x(kh), Φ = exp(Ach) State Equation: x(t ) Ac x(t ) Bcuˆ (t ), y (t ) Cc x(t ) (1) leads to kh k x(kh h) exp( Ac h) x(kh) kh h exp( Ac (kh h - s )) Bcuˆk -1 Define the three matrices ,0 , 1 exp( Ac h), 1 ( ) k λ exp( Ac (kh h - s )) Bc ds, 0 ( ) λ kh h exp( Ac (kh h - s )) Bc ds kh kh x(kh h) x(kh) 0 (t )uˆk 1 (t )uˆk -1 , where: 0 ( ) k exp( Ac (kh h - s )) Bcuˆk ds kh k kh kh k k h k exp( Ac ) Bc d , 1 ( ) k h exp( A ) d c c h - k 0 we have used the following: kh h - s ( so d -ds since h const.) and changing the variable of integration into d -ds, the new limits of integration are (h - k ) and 0 so we get the simplified expression for Γ0 : 0 ( k ) kh h kh exp( Ac (kh h - s)) Bc ds 0 h- k exp( Ac )c d (- ) 0 exp( Ac )cd h- k 22 Exact Discretization despite the “jump’’ nature of û(t) xk = x(kh), Φ = exp(Ach) • Similarly from the definition of Γ1, using the same change of variables as previously k ( kh h - s d -ds ) the new limits of integration are h and (h - ) so we get : 1 ( ) k kh h - k exp( Ac (kh h - s)) Bc ds kh exp( Ac )c d (- ) h exp( Ac )c d h - k h Moreover assuming there is no uncertainty on output matrix we have : y (kh) Cc x(kh) or yk Cc xk A usefull identity for calculating integrals of matrix exponential functions ( such as 0 ( k )). Xn x n exp( 0m x n To Compute 0 ( k ) h - (X t) Yn x n e t ) 0m x m 0m x n t e Ydr ( Xr ) 0 Im e( Ac ) c d we can use above identity as: 0 0T 0 ( ) [ I n 0 ] e Identity I. 1 where 0 01 x n the zero row vector with n columns k T Ac Bc ( (h - k ) ) 0 0 23 Exact Discretization despite the “jump’’ nature of û(t) xk = x(kh), Φ = exp(Ach) • Notice that : (3.A). (3.B). From 2nd equation : 1 ( k ) h h- e( Ac )c d 0 h- k k h e( Ac )c d e( Ac )c d 0 h = - 0 ( k ) e( Ac )c d 0 24 Contents 1. 2. 3. 4. 5. 6. 7. 8. Introduction – General Features NCS Modeling - the issue of network-induced delays Descretization of NCS dynamics equation Decomposing the Uncertain Delay (nominal and uncertain parts) and the NCS dynamics Robust Stability Analysis based on the closed-loop augmented vector (“ξ”) Design of A Simple Output Tracking Controller Investigation of Robust Tracking Performance via Simulation - Numerical Examples for Networked Stable and Unstable systems Conclusions & Topics for further study 25 4. Decomposing the uncertain delay of the system (into nominal & uncertain part) Examples: τo = τ min τo = τ max τo = τ avg τo is chosen as constant and known («semi-arbitrary») Use of “Min Max” techniques for selection of τo The nominally delayed system, Stability Analysis and Controller Synthesis depend on the (user’s) choice of τo 26 4. Decomposing the uncertain delay of the system k (into nominal & uncertain part)0 ( ) (4.C). so the nominal part of 0 ( ) is and the uncertain part of is 0 ( ) ΔΓ 0 ( ) h 0 e( Ac ) c d and h k h e( Ac ) c d 27 4. Decomposing the uncertain delay of the system (into nominal & uncertain part) ( k ) 1 4.D h 0 The Uncertain part is: 1 ( k ) e( Ac ) c d and h k the Nominal Part is : Γ1 ( 0 ) h e( Ac ) c d h 0 28 4. Decomposing the uncertain delay of the system (into nominal & uncertain part) h The nominal part can be calculated from: 1 ( 0 ) 0 ( 0 ) e( Ac )c d 0 Moreover from previous ΔΓ0 ( k ) and 1 ( k ) we observe the following relation between the two uncertain matrices: 1 ( k ) 0 ( k ), k [ min , max ] 4.E 29 Contents 1. 2. 3. 4. 5. 6. 7. 8. Introduction NCS Modeling - the issue of network-induced delays Descretization of NCS dynamics equation Decomposing the Uncertain Delay (nominal and uncertain parts) Robust Stability Analysis based on the augmented closed-loop vector (“ξ”) Design of Simple Output Tracking Controller Investigate Robust Tracking Performance via Simulation Conclusions & Topics for further study 30 5. Robust Stability Analysis based on the closed-loop vector augmented (“ξ”) - Closing the loop THE AUGMENTED CLOSED–LOOP STATE VECTOR (“ACLSV”) “ξ” xk+1 = Φ xk + Γ0(τk) ûk + Γ1 (τk) ûk-1 ûk = -Ksf xk Static State Feedback (SSF) ûk-1 = -Ksfxk-1 Closed Loop Dynamics xk+1= [Φ- Γ0(τk) Ksf] xk + [- Γ1 (τk) Ksf ] xk-1 only periodically sampled state vector values {xk+1, xk, xk-1 } are present 31 5. Robust Stability Analysis based on the closed-loop vector (“ξ”) Define the augmented “sampled data” closed-loop state vector Assuming there is no uncertainty on output matrix it holds: yk Ck xk Cc [ I n 0n ] k [Cc 0] k C o k 0n is the zero square matrix (n x n). 32 5. Robust Stability Analysis based on the closed-loop vector (“ξ”) The Closed loop Matrix A sf Asf ( k , K sf , ) R 2 n x 2 n is a function of : the uncertain delay τ k the preselected gain K sf and the predetrmined nominal delay το The above matrix relation is manageable and Robust Control Methods now can be used. Using 4C, 4D, 4E : Asf can brake down to a nominal time invariant part A osf and an uncertain part sf ( k ) as : 0 ( ) sf -1 ( ) sf -ΔΓ0 ( ) sf -ΔΓ1 ( ) sf Asf + In 0n 0n 0n Asfo sf 33 Contents 1. Introduction 2. NCS Modeling - the issue of network-induced delays 3. Discretization of NCS dynamics 4. Decomposing the Uncertain Delay (nominal and uncertain parts) and NCS dynamics 5. Robust Stability Analysis based on the augmented closed-loop vector (“ξ”) 6. Design of Simple Output Tracking Controller 7. Investigate Robust Tracking Performance via Simulation 8. Conclusions & Topics for further study 34 6. Design of Simple (Set Point) Tracking Controllers SPCT = Set Point Tracking Controller(s) The reference signal to be tracked by the output is (piecewise) constant (a “set point”) –Assumption:both the plant and the controller under investigation are continuous–time LTI systems –Since the controller is time invariant, can lump the delays τsc (k), τca (k), into τk= τsc (k)+τca (k). – A “naïve” tracking controller consists of two parts: Feedback & Feedforward u(t) = -Kx(t)+Fr The feedback part (-Kx(t)) assures closed-loop stability The feedforward part (Fr) assures that the static gain is “1” (Stable Transfer Function from r to y) 35 6. Design of Simple (Set Point) Tracking Controller •Suffers from three drawbacks (“naïve”): • the plant must not contain integrators (system matrix A is nonsingular) • cannot handle disturbances and/or model uncertainties (it is NOT Robust) • Number of inputs ≥ Number of outputs (“overactuation”) 36 Contents 1. Introduction 2. NCS Modeling - the issue of network-induced delays 3. Descretization of NCS dynamics equation 4. Decomposing the Uncertain Delay (nominal and uncertain parts) 5. Robust Stability Analysis based on the closed-loop vector (“ξ”) 6. Design of Simple Output Tracking Controller 7. Investigate Robust Tracking Performance via Simulation 8. Conclusions & Topics for further study 37 7. Robustness of Tracking Performance Numerical Example 1: a networked stable & minimum phase system • A “benevolent” stable & minimum phase (=zeros in LHP) system with infinite Gain Margin and… • “Lightly Damped” = stable poles close to the Imaginary axis “damping ratio” is small damped oscillative open-loop behaviour (typical in aerospace and “flexible space structure” applications) •SPTC was designed via LQR with R=1, Q=1000*I2 u(t) = -30.63 x1(t) - 30.63 x2(t) + 31.63*r gives “perfect tracking” in the absence of delays 38 7. Robustness of Tracking Performance Nmerical Example 1: a networked stable & minimum phase system with constant delay The Networked Version with constant delay τk τsc =τca = 0.0131 s τk= τsc +τca=0.0262s •Assuming that τk ≤ h this delay corresponds (for the discrete time control case) to a sampling frequency of 38Hz a relatively “slow sampling”… • “slow sampling” is typical for NCS (fast sampling increases # of packets increases network traffic increases chances for collisions packet loss/drops) 39 7. Robustness of Tracking Performance Numerical Example 1: a networked stable & minimum phase system with constant delay The Networked Version with constant delay τk τsc =τca = 0.0131 s τk= τsc +τca=0.0262s • 7th order Pade Approximation used in simulations for the constant time-delay •Reference Signal(s) r are (piecewise) constant: • combination of step functions or • square pulse with period slower than the system’s time constants • Simulation needs time for Instability to occur (see next Figs) 40 7. Robustness of Tracking Performance Numerical Example 1: a networked stable & minimum phase system with constant delay 41 7. Robustness of Tracking Performance Numerical Example 1: a networked stable & minimum phase system with constant delay 42 7. Robustness of Tracking Performance Numerical Example 1: a networked stable & minimum phase system with uncertain (time-varying) delay • The Networked Version with uncertain time-varying delay τk varying between τmin = 0 and τmax = 0.0312s < h corresponding to a sampling frequency of 32 Hz • Implementation used in simulations: τk = τo + δ τunc , |δ| < 1 with τo = τavg = (τmax + τ min )/2 = 0.0156 s being the “mean value” (a constant “nominal” delay) and |δ|<1 being a random variable of uniform distribution. 43 7. Robustness of Tracking Performance Numerical Example 1: a networked stable & minimum phase system with uncertain (time-varying) delay 0=τmin ≤ τk ≤ τmax = 0.0312s An instance of the actual uncertainly varying delay used in simulations τk = 0.0156 + 0.0156* δ |δ| < 1 τk = τo + δ τunc , |δ| < 1 τo = τavg = (τmax + τ min )/2 44 7. Robustness of Tracking Performance Numerical Example 1: a networked stable & minimum phase system with uncertain (time-varying) delay τk = 0.0156 + 0.0156 *δ |δ| < 1 45 7. Robustness of Tracking Performance Numerical Example 1: a networked stable & minimum phase system with uncertain delay 46 7. Robustness of Tracking Performance Numerical Example 2: a networked unstable system •SPTC was designed via LQR with R=1, Q=100*I2 u(t) = -9.05 x1(t) -10.78 x2(t) + 10.05 *r gives “perfect tracking” in the absence of delays •The “Q” matrix was selected “small” in order to avoid high feedback gains…and yet 47 7. Robustness of Tracking Performance Numerical Example 2: a networked unstable system with constant delay τk= τsc +τca=0.0155s 48 7. Robustness of tracking performance…….some comments Many more simulation results with different 2nd order “benchmark” S.I.S.O systems from the literature (not shown) But….we can deduce useful conclusions (despite the lack of a mathematically rigorous approach) Clearly a more sophisticated approach is needed for the design of tracking controllers for NCS We cannot “pretend” that the “delays are not there” - must take them into account in the design phase. We can not compromise stability (avoid large gains) - rule of thump for Time-Delayed -Systems (mid ’50s result !!!) 49 CONCLUSIONS AND FUTURE WORK 1. The constant delay case (contrary to intuition) is as detrimental to tracking performance as the varying delay case. 2. The feedback gain must be kept “small”. –If an LQR design is employed : extensive trial-and-error simulations with various “Q” matrices must be carried out for the entire delay range to ensure (at least) stability. –Tracking for the case of unstable plants and/or lightly damped plants is not trivial. 3. For Unstable plants it is always difficult to enforce tracking (with or without delays). 4. When implementing the tracking controllers in discretetime special attention is needed due to (1) the interplay between sampling period and delay and (2) the “asynchronous / jump” nature of the control signal 50 Last Minute Thoughts – Dynamical Systems with Time Delays Consider the time delay systems: x(t ) Ax(t ) Ad x(t d ) (S1 ) x(t ) Ax(t ) Ad x(t d (t )) (S2 ) x(t ) (t ), t [-d , 0] x(t ) (t ), t [-d M ,0] state x(t) R n and (t) is the continuous I.C. Asumptions :1. d (t ) is a Cont. function satisfying 0 d(t) d M with d M a constant positive scalar. It is the upper bound of d(t) 2. d(t) is a differ. function satisfying 0 d(t) d M and d(t) 1. d M is given previously and is the u.b. of d(t). 3. d(t) is a differ. function satisfying 0 d m d (t ) d M and d(t) . d m , d M are given pos. scalars representing l. u. bounds of interval time varying delay d(t) and ρ as given previously. 51 Theorem 1.( S1 ) is asymptotically stable if there exist matrices P > 0 and Q > 0 A T P PA Q PAd such that : 0 T Ad P -Q where the Lyapunov - Krasovskii functional was used: t V(t) = xT (t ) Px(t ) xT ( s )Qx( s )ds t d Theorem 2. Under assumption 2, (S2 ) is As. Stable if there exist matrices P > 0 A T P PA Q PAd and Q > 0 such that : 0 T Ad P -(1- )Q the Lyapunov - Krasovskii functional of the form t V(t) = xT (t ) Px(t ) t d (t ) xT ( s)Qx( s)ds is used. 52 CONCLUSIONS AND FUTURE WORK Generalize achieved results for – MIMO NCS plants with multiple delays, Parametric Uncertainties & Actuator constraints The use of Robust Control Methodologies (H∞ or “Guaranteed Cost”) for the design of Feedback Gain The employment of Integral Action (apart from feedback and feedforward terms) in the tracking control Algorithm(s). Investigate Specific Applications: Aerospace & Robotics (Teleoperation) NCS’s indeed constitute a very interesting and rich field of control systems both in theoretical results as well as in future applications. 53 THANK YOU 54