Dependable Wireless Control

Transcription

1 Dependable Wireless Control through Cyber-Physical Co-Design Chenyang Lu Cyber-Physical Systems Laboratory Department of Computer Science and Engineering

2 Wireless for Process Automa1on Emerson 5.9+ billion hours operating experience 26,200+ wireless field networks $ million by 2020 [Market and Market] Offshore Onshore Courtesy: Emerson Process Management Killer App of Sensor Networks! 2

3 WirelessHART Ø Reliability and predictability q Multi-channel TDMA MAC q One transmission per channel q Redundant routes Ø Centralized network manager q Collect topology information q Generate routes and schedule q Change when devices/links break Industrial wireless standard for process automation 3

4 The Control Challenge Most of today s industrial wireless networks are for monitoring. Sensor Actuator sensor data control command Controller Dependable control requires real-time control performance resilience to loss 4

5 Towards Dependable Wireless Control 1. Real-time wireless networks and analysis 2. Optimizing control performance over wireless 3. Resilient yet efficient wireless control under data loss. This cannot be accomplished by wireless or control design alone à Cyber-Physical Co-design of Wireless and Control 5

6 The Real-Time Problem Ø A feedback control loop incurs a flow F i q Route: sensor à à controller à à actuator q Generate packet every period P i q Multiple control loops share a network Ø Each flow must meet deadline D i ( P i ) q Stability and predictable control performance Ø Research problems q Real-time transmission scheduling à meet end-to-end deadlines q Fast schedulability analysis à adapt to wireless dynamics through admission control and rate adaptation 6

7 Delays in WirelessHART A transmission is delayed by Ø channel contention: all channels are assigned to other transmissions Ø transmission conflict over a same node q contributes significantly to latency! and 5 conflict 4 and 5 conflict 3 and 4 do not 7

8 Fast Delay Analysis Ø Compute upper bound of the delay for each flow q Sufficient condition for real-time guarantees Ø Channel contention à multiprocessor task scheduling q q q A channel à a processor Flow F i à a task with period P i, deadline D i, execution time C i Leverage existing response time analysis for multiprocessors Ø Account for delays due to transmission conflicts Ø Fast delay analyses q Fixed priority scheduling [RTAS 2011, TC, RTSS 2015] q Earliest Deadline First [IWQoS 2014] 8

9 Delay due to Conflict Ø Low-priority flow F l and highpriority flow F h, conflict à delay F l )*$%+*, F l delayed by 2 slots Ø Q(I,h): #transmissions of F h sharing nodes with F l q In the worst case, F h can delay F l by Q(l,h) slots Ø Conflicts contributes significantly to delays q Must be considered in algorithms and analysis! F l delayed by 2 slots F l delayed by 1 slot!"#$%&'"(!" # &!"#$%&'"(!" $ 9

10 Real-Time Wireless Networking Ø WirelessHART stack in TinyOS q Implementation on a testbed of 69 TelosB motes. q Network manager (scheduler + routing). Ø Conflict-aware real-time scheduling q Fixed priority assignment [ECRTS 2011] q Conflict-aware Least Laxity First [RTSS 2010] Ø Energy-efficient routing [IoTDI 2016] Ø Emergency communication [ICCPS 2015] WUSTL wireless sensor network testbed 10

11 Op1mize Control over Wireless Observation Ø Wireless resource is scarce and dynamic Ø Cannot afford separating wireless and control designs Cyber-Physical Co-Design Ø Holistic co-design of wireless and control Examples Ø Rate selection for wireless control [RTAS 2012, TECS] Ø Scheduling-control co-design [ICCPS 2013] 11

12 Rate Selec1on for Wireless Control Ø Optimize the sampling rates of control loops sharing a WirelessHART network. Ø Rate selection must balance control and network delay. q Low sampling rate à poor control performance q High sampling rate à long delay à poor control performance 12

13 Control Performance Index Ø Digital implementation of control loop i q q Periodic sampling at rate f i Performance deviates from continuous counterpart Ø Control cost of control loop i under rate f i [Seto RTSS 96] q Approximated as α i e β i f i with sensitivity coefficients α i, β i Ø Overall control cost of n loops: n i=1 α i e β i f i 13

14 The Rate Selec1on Problem Ø Constrained non-linear optimization Ø Determine sampling rates f = { f 1, f 2,, f n } minimize control cost n i=1 α i e β i f i subject to delay i 1/ f i f i min f i f i max Delay bound 14

15 Polynomial Time Delay Bounds Ø In terms of decision variables (rates), the delay bounds are Lagrange dual of objecdve q q q Non-linear Non-convex Non-differentiable Rate of control loop 6 15

16 Cyber-Physical Co-Design Ø Relax delay bound to simplify control optimization Ø Derive a convex and smooth, but less precise delay bound. Ø Rate selection becomes a convex optimization problem. Optimize control performance efficiently at run time! Control cost A. Saifullah, C. Wu, P. Tiwari, Y. Xu, Y. Fu, C. Lu and Y. Chen, Near Optimal Rate Selection for Wireless Control Systems, ACM Transactions on Embedded Computing Systems, 13(4s), Article 128, April

17 Resilient Control under Data Loss Ø Data loss causes instability and degrades control performance. Ø State of the Art q Control: control design to tolerate data loss. q Wireless: redundancy reduces loss at high resource cost. Ø Cyber-physical co-design q Incorporate resilient control design q Tailor wireless protocols for control design q Resilient wireless control at low resource cost 17

18 Handle Data Loss from Sensors Reference Model Predictive Control [ uk ( ), uk ( + 1),, uk ( + w)] Buffer uk ˆ( ) Actuators xk ˆ( ) Extended Kalman Filter yk ˆ( ) yk ( ) Sensors Plant Ø State Observer estimates system states based on a system model even if there is no new data from sensors B. Sinopoli et al., Kalman filtering with intermittent observations. IEEE Transactions on Automatic Control, 49(9): ,

19 Handle Data Loss from Controller Reference Model Predictive Control [ uk ( ), uk ( + 1),, uk ( + w)] Buffer uk ˆ( ) Actuators Ø Model Predictive Control 19 xk ˆ( ) Extended Kalman Filter yk ˆ( ) q Controller computes control inputs in the next w+1 sampling periods: u(k), u(k+2),... u(k+w). q Actuator applies u(k). Ø Buffered actuation yk ( ) q Actuator buffers previous control inputs u(k+1),... u(k+w). q Applies buffered control input if updated input is lost. Sensors Plant q A control horizon of w+1 à may tolerate w consecutive packet loss.

20 Exothermic Reac1on Plant Pump 1 a u 1 Reagent Tank 1 Plant: nonlinear chemical reaction Control input: u1 and u2 Objective: Maintain temperature in Tank 2 Tank 1 Reagent Tank 2 L 1 b Tank 2 L 2 Pump 2 Heater u 2 Wireless Cyber-Physical Simulator (WCPS) Integrate TOSSIM and Simulink Capture dynamics of both wireless networks and physical plants Holistic simulations of wireless control Open source: wcps.cse.wustl.edu 20

21 Resilience of State Observer Left: without state observer Without state observer under 60% sensor data loss Extended Kalman Filter (EKF) under 60% sensor data loss 21

22 Resilience of Buffered Actua1on Actuation buffer (size 8) under 60% controller data loss System is more sensitive to data loss to actuators than that from sensors. 22

23 Tradi1onal Rou1ng Ø Entire network uses uniform routing strategies Ø Source routing: single path routing q Efficient but unreliable Ø Graph routing: every node on the primary path has a back path. q More reliable at cost of capacity and energy Gateway 1 Gateway 2 Gateway 1 Gateway 2 5, 6 5, (6) 3, 4 3, 4 (5) 4 1, 2 1, 2 3 Primary Path Sensor (a) Source Routing Sensor (b) Graph Routing Backup Path 23

24 Asymmetric Rou1ng Ø Differentiated routing based on control needs Ø State observer handles data loss from sensors à Ø Source routing from sensors q State observer compensates for lower reliability q Cost less resource Ø Actuation is more sensitive to data loss à Ø Graph routing to actuators q Higher reliability q Higher resource cost needed by control Tailoring routing strategies for control design à Efficient and Resilient Wireless Control 24

25 Maximum Absolute Error -73dBm Noise -73dBm Noise Ø Source/Graph performs close to Graph routing under 3Hz Ø Source/Graph s resource efficiency allows 5Hz control q Further improve control performance B. Li, Y. Ma, T. Westenbroek, C. Wu, H. Gonzalez and C. Lu, Wireless Routing and Control: a Cyber- Physical Case Study, ACM/IEEE International Conference on Cyber-Physical Systems (ICCPS 16). 25

26 Towards Dependable Wireless Control Ø Real-time, predictable wireless networking q Protocols and delay analysis for bounded latency Ø Optimize control performance over wireless q Incorporate scheduling analysis in rate selection for wireless control Ø Resilient wireless control under data loss q Tailor routing strategies for control needs Ø Cyber-physical co-design helps overcome the dependability challenges! 26

27 Engineering Building Blocks Ø Industrial wireless networks have arrived q Industrial standards: WirelessHART, ISA100 q World-wide field deployments q Great opportunity for the sensor networks community! Ø WirelessHART implementation in TinyOS Ø WCPS: Wireless Cyber-Physical Simulator q Enable holistic simulations of wireless control systems 27

28 Future Direc1ons Ø Scaling up wireless control networks q From 100 nodes à 10,000 nodes q Dealing with dynamics locally q Hierarchical or decentralized architecture q Predictable protocols over low-power wide-area networks Ø Unified science and engineering of wireless control q Case studies à unified theory, architecture and methodology of cyber-physical co-design. q Bridge the gap between theory and systems 28

29 For More Informa1on Ø C. Lu, A. Saifullah, B. Li, M. Sha, H. Gonzalez, D. Gunatilaka, C. Wu, L. Nie and Y. Chen, Real-Time Wireless Sensor-Actuator Networks for Industrial Cyber-Physical Systems, Special Issue on Industrial Cyber-Physical Systems, Proceedings of the IEEE, accepted. Ø Research on Real-Time Industrial Wireless Sensor-Actuator Networks: Ø Wireless Cyber-Physical Simulator: 29