Adaptive Control of Network Centric Dynamic Systems Jagannathan (Jag) Sarangapani Professor of Department of Electrical and Computer Engineering & Computer Science University of Missouri-Rolla Rolla, Missouri 65409. Tel: 573-341-6775; Email:sarangap@umr.edu & Site Director, NSF I/UCRC on Intelligent Maintenance Systems 1 Outline • Overview of Research at my Lab/Center • Energy Aware Protocols – Adaptive Congestion Control – Routing Protocol • Conclusions 2 Network of Autonomous Robots Requirement: Continuous Communication among 10 robots/unattended sensors OBJECTIVE UMR Mote • Cooperative decision making and control of multiple unattended formation of robots (mobile adhoc networks) with continuous RF communication (UMR mote) BACKGROUND SENSOR “ARRAY” Wireless Point-to-point Mission Head Quarters Cluster Heads Sensor nodes • Network communication among a formation of robots is necessary in order to perform any task. This communication must be present at all times. This communication must be reliable even when the robots are in motion and when energy is limited. This requires a novel distributed hybrid system architecture, reliable communication hardware and energy efficient protocols for communication that guarantees quality of service requirements. APPROACH COST AND SCHEDULE • Milestone 1: Develop distributed embedded hybrid architecture MILESTONES Qtr 1 Qtr 2 Qtr 3 Qtr 4 • Milestone 2: Develop wide band communication-based UMR Mote Milestone 1 • Milestone 3: Integrate sensors and RF hardware on the robot Milestone 2 • Milestone 4: Implement the energy efficient communication protocols using energy delay metric Milestone 3 • Milestone 5: Implement control strategy and demonstrate DELIVERABLES Milestone 4 • UMR RF Mote Milestone 5 • Energy Efficient Protocols • Demonstration and Reports Total Cost $250K ($K) 75 100 50 25 POINT OF CONTACT Dr. J. Sarangapani, Electrical and Computer Eng Dept., UMR Email: sarangap@umr.edu; Phone: 573-341-6775; URL: www.umr.edu/~sarangap 3 3 Manipulation of Microscale/Nano Scale Objects using Micro/Nano Robots • Manipulation of a micro-sphere – Pickup a micro-sphere from a planar substrate • Sequence of operations – Lower a probe – Pickup the micro-sphere – Retract the probe • • Novel intelligent controllers designed outperform available ones AFM is used as a feedback for manipulation and drift compensation of nano particles Fig. 1. Image sequences taken at 256 sec intervals without drift compensation (first row) and with drift compensation (second row). The scanned area is 512×512nm2. 4 Prognostics on a Chip Wireless UMR Mote Service History Database Distributed Sensors Wireless Degradation Multivariable Analysis with Learning Trending Confidence Remaining Useful Life Confidence Severity UMR Mote Wireless Reliability Information 5 Prognostic Agent Methodology 6 Monitoring Using Wireless Sensor Networks Sensor Node Base Station Cluster Head Self Organizing Network Detect Damage Progression Diagnosis/Prognosis Emission Control 7 Auto-ID Solutions for Network-Centric Manufacturing Environments Objective: Develop concepts, models, and prototypes for effective integration of Auto-ID technology in a Network-Centric Manufacturing Environment*, aimed at reducing delays and eliminating non-value added production activities and responding rapidly to unexpected events on the shop floor. Process Modeling Auto-ID Solutions Development Research Highlight: SYSTEMS Process • Develop techniques for effective use of real-time data provided by Auto-ID technology. All aircraft assembled at Boeing • Demonstrate integration of AutoID technology with the aircraft manufacturing practice. Re-Engineer Re-Engineer Application Potential: 1) Inventory control of time and temperature sensitive materials 2) Receiving & shipping operations 3) Aircraft assembly line flow Map System Simulation (before) Manufacturing R&D Assembly Lab. Analysis: amount of waste no of man-hours Embedded Systems and Networking Laboratory (EE) Integrated Systems Facility (Emgt) Auto-ID Solutions Synthesis Synthesis System Simulation (after) Tag Simulation Analysis: amount of waste no of man-hours Simulation Prototypes Simulation with hardware-in-the-loop Prototypes Hardware ROADMAP * Network-centric Manufacturing Environment incorporates a dynamic network of self-organizing, autonomous units that operate, collaborate, cooperate, and compete upon basic principles of decentralization, participation, and coordination in order to accomplish the goals set at system level. 8 Networking Infrastructure for Adaptive Inventory Management Objective: To demonstrate integration of decision making in a multi-technology environment. Conclusions: (1)Modularity of the architecture facilitates expansion of the application domain and experimentation with complex models, and (2) the infrastructure allows for investigating different networking topologies, protocols, and alternative hardware. 9 Neural Emission Control • Lean operation and with high EGR levels in certain SI Engine can reduce emissions (HC, CO & NOx) by as much as 30% and also it improves fuel efficiency by as much as 5 ~ 10% (Inoue, 1993). Heywood, 1988 Cyclic dispersion without control Objective: Minimize the cyclic dispersion at lean engine operation by applying the NN controller 10 Nonlinear Discrete-time System x1 k 1 f1 x1 k , x2 k g1 x1 k , x2 k x2 k d1 k x2 k 1 f 2 x1 k , x2 k g 2 x1 k , x2 k u k d 2 k y k 1 f3 x1 k , x2 k Back Stepping Nonlinear NonStrict 11 Comparison of Experimental and Simulation Data Experimental data (Sutton, 2000) Simulation result 12 NN Output Feedback Lean Emission Control Neural Network (NN) Controller VT (.) 1 (.) 2 (.) 3 x1 WT x2 xn Inputs Engine (.) y1 (.) y2 (.) L Hidden Layer (.) ym outputs Control Inputs Measurements NN Observer Heat release return map without control, = 0.705 Heat release return map with output control, = 0.705 12 9 Heat release Fiexed point 8 10 Heat release Fiexed point 8 Heat Release (i+1) Heat Release (i+1) 7 6 4 6 5 4 3 2 2 1 0 0 2 4 6 Heat Release (i) 8 10 12 0 0 1 2 3 4 5 Heat Release (i) 6 7 8 9 13 Contributions • Nonstrict Feedback Nonlinear Discrete-time Systems introduced and optimal controllers designed • Noncausal control problem is overcome • Separation Principle, Certainty Equivalence, Persistency of Excitation Condition, and Linear in the Unknown Parameters are relaxed • Fuel efficiency improvement by 5 to 10%, NOx by 98% and uHC by 30% was noted 14 NN Control Book in 2006 15 Outline • Overview of Research at ESNL • Energy Aware Protocols – Congestion Control – Routing • Conclusions 16 Performance Requirements • Congestion causes – Reduced channel capacity – Energy wastage • QoS suffers – Throughput, network efficiency – Delay , jitter – Fairness – Energy-efficiency • Proposed scheme consists of – Congestion prediction and control mechanisms to prevents buffer overflows – Fairness mechanism 17 Congestion in Wireless Sensor Networks (WSN) • Congestion can be a result of: – Channel fading – Traffic exceeding channel capacity • Note: Typical WSN has only one “sink” 18 Ways of Alleviating Congestion • Energy aware congestion control • Adaptive Back off Interval scheme • Adaptive energy delay routing * Cross Layer Design is needed 19 Background • Sensor versus ad hoc – Processing, memory communication and energy constrains in WSN in contrast to ad hoc networks • Previous works – End-to-end protocols (e.g. TCP) • Drawback of large feedback latency (round-trip) – CODA, Fusion • Small processing overhead • Congestion message is broadcasted to throttle traffic, with sources reducing introduced traffic • Use fixed thresholds to initiate flow control • Dropping packets when the congestion occurs 20 Objectives of Congestion Control • Detection of the congestion and an onset of one – Channel estimation from Distributed Power Control (DPC - our previous work) predicts severe fading and temporarily suspend outgoing and incoming flows – Buffer occupancy used to control incoming/outgoing flows • Quality of Service (QoS) – Weighted Fairness expressed as: W f (t1 , t 2 ) f W m (t1 , t 2 ) m 0 (1) where W – throughput, φ – weight • Congestion control using backpressure – a set of three proposed schemes (described later) 21 Proposed Scheme • Rate-based congestion control with weights changed adaptively – Rate selection to prevent buffer overflow – Rate allocation to flows according to weights (using fair scheduling) – Selection of back-off interval to achieve the selected rate (predictive, distributed, mathematically guaranteed) • (OPTIONAL) Adaptive weights allocated to each packet to improve weighted fairness – Adaptive re-calculation of weights for each packet – Fair scheduling and rate allocation to flows based on adopted weights 22 Buffer occupancy and Error Dynamics • Consider buffer occupancy at a particular node qi k +1 = Satp qi k + T ui k fi ui 1 (k ) d (k ) (2) – where T is the measurement interval, qi(k) is the buffer occupancy of node ‘i’ at time instant k, ui(k) is a regulated (incoming) traffic rate, and fi(k) is an outgoing traffic rate. • Consider the desired buffer occupancy at node i to be qid. Then, buffer occupancy error defined as ei(k)=qi(k)-qid can be expressed as ebi ( k 1) qi ( k ) T u i ( k ) f i (ui 1 ( k )) d ( k ) qid (3) 23 Rate Selection • Define the traffic rate input, ui(k) as ui (k ) T 1 qid qi (k ) T f i (k ) k v ei (k ) where kv is a gain parameter. • Unknown outgoing traffic is estimated using adaptive scheme fˆi (ui 1 (k )) ˆ(k ) f i (k 1) (4) (5) where the parameter vector is updated • Selected incoming rate is divided fairly (φj) among incoming flows 24 Theorem: Ideal Case Given the incoming rate selection scheme above with variable i estimated accurately (no estimation error), and if the incoming traffic is updated as (4), then the mean estimation error of the variable i along with the mean error in queue utilization converges to zero asymptotically, if the parameter i is updated as ˆi (k 1) ˆi (k ) ui (k ) e fi (k 1) provided: (a) ui k 2 < 1 and (b) K fvmax < 1 δ , where δ = 1 1 ui k 2 , K fvmax is the maximum singular value of K fv , is the adaptation gain, and e fi (k ) f i (k ) fˆi ( k ) is the error between the estimated value and the actual one. 25 Simulation Results for Rate-based Buffer Control qi estimated fout 15 Queue utilization and estimation of the outgoing flow. fout 10 10 5 0 0 5 10 15 Time [iteration] 20 Queue utilization error and outgoing traffic estimation error. 25 Buffer occupancy error [packets] Throughput estimation error [packets/iteration] Queue level [packets] Throughput [packets / iteration] 20 ebi ef 5 0 -5 -10 0 5 10 15 Time [iteration] 20 25 26 Adaptive Back-off Selection Algorithm • Goal Select back-off interval BOi at i-th transmitting node such that the actual throughput meets the desired outgoing rate fi(k). • Consider inverse of the back-off interval, and call it a virtual rate VRi VRi = 1 / BOi (6) where VRi is the virtual rate at i-th node, and BOi is the corresponding back-off interval. • NOTE: the virtual rate is not equal to the actual rate; instead, the virtual rate is proportional to the actual rate. 27 Outgoing Rate Selection • The actual rate of an i-th node is a fraction of the channel bandwidth B(t) defined as Bt VRi t Bt VRi t Ri t = = TVRi t VRl t (7) lSi where TVRi is the sum of all virtual rates for all neighbor nodes. • Differentiating and transforming into discreet time domain the outgoing rate is equal where Ri k +1 = Ri k αi k + βi k vi k (8) αi k = 1 TVRi k +1 TVRi k βi k = Ri k VRi k vi k = VRi k +1 = 1 BOi k +1 (9) 28 Closed-loop Control of Backoff Selection • Now consider feedback equation for closed-loop controller vi (k ) i (k ) 1 f (k ) R (k ) (k ) K e k (10) i ij i v Ri where âi(k) is estimate of ai(k), and ei(k)=qi(k)-qid is the throughput error • Parameters updates taken as aˆi k +1 = αˆi k + Ri k eRi k 1 • The closed loop throughput error system with estimation error, ε(k), as ei k 1 K v ei k i (k ) Ri (k ) (k ) (11) (12) 29 Convergence and Stability • Theorem (General case): – Given the back-off selection scheme above with an interval updated as (10), and with uncertainties estimated by (11), with ε(k) as the error in estimation which is considered bounded above εk ε N , with εN a known constant. – Then the mean error in throughput and the estimated parameters are bounded provided and σ R k 2 < 1 hold K <1 δ vmax – where δ = 1 1 σ Ri k ,K i 2 is the maximum singular values of Kv, and σ is the adaptation gain. vmax 30 Simulation Results Proposed scheme w/o weight adap - 600 Proposed scheme DPC IEEE802.11 500 400 CODA 300 200 1.4 100 1.2 0 1 1 2 3 4 5 flow ID Performance for unbalanced tree topology. Weight * Delay Throughput / Weight [kbps] 700 Proposed scheme w/o weight adapProposed scheme DPC CODA 0.8 0.6 0.4 0.2 Weighted delay with equal flow weights (const=0.2). 0 1 2 3 Flow ID 4 5 31 Impact on Routing Protocol • Hardware limitation explored – Memory limits queue size – Processing limits number of connections that can be handled (also for information aggregation) • Localized congestion – For example due to electro–magnetic interferences – Can be mitigated by selecting a alternative route around the congested area 32 OEDSR • Optimal Energy Delay Sub-Network Routing (OEDSR) protocol – Maximizes Link Cost Factor (LCF) to perform optimal routing based on network parameters – LCF is given by Ei LCFi Di xi – Where Ei is the energy in the next node, Di is the delay, and xi is the distance from the next node to the base station • Cluster heads (CH) and relay nodes (RN) are used to route data from a data source to the base station (BS) • Implemented on UMR hardware 33 OEDSR: Optimal Relay Node CH 1 1 1 2 3 4 5 6 7 3 1 6 CH 1 CH 2 5 6 7 6 7 12 4 5 1 2 3 4 11 2 7 15 2 BS 8 9 3 10 13 14 CH 2 CH 3 6 7 7 “HELLO” packet “RESPONSE” packet “HELLO_CH” packet Node ID Distance from BS Node ID End-to-End Delay Energy Available Distance from BS List of all nodes in Range “RELAY_SELECT” packet 34 OEDSR: Optimal Relay Node CH 1 1 1 2 3 4 5 6 7 3 1 6 CH 1 CH 2 5 6 7 6 7 7 15 2 9 CH 2 CH 3 10 13 14 7 7 Link _ cos t _ factor BS 8 3 6 7 12 4 5 1 2 3 4 11 2 4 12 15 8 Ern Delay Dist Ern - energy available in the given node Delay - average end-to-end delay between any two CHs Dist - distance from the node to the BS 35 Modification to OEDSR • Link Preference Factor (LPF) Ei LPFi i Di xi – Load factor i is for balancing load by distributing traffic between several nodes i Fi 1 f i – where, Fi is the maximum designed capacity of a node i, and fi is the current load at node i (measured in number of flows routed at the node) 36 Experimental Results for Routing Protocol • Simple topology used to verify claim of balancing load between two relay nodes versus sending the traffic through one node only BS Sources 3 sources Relay nodes Total Throughput Average per source Average per relay node Through 1 relay 6.88 kbps 2.29 kbps 6.88 kbps (1 relay) Through 2 relays 8.01 kbps 2.68 kbps 2.43 kbps (2 flows) 5.69 kbps (1 flow) 37 Demonstration <Play video> 38 Contributions • Novel adaptive schemes based on control theory developed • Analytically Guaranteed • Many of these are demonstrated on Mote Hardware • Deployed in various industrial environments 39 New Book to appear in 2007 • Wireless Ad hoc and Sensor Networks: Protocols, Performance, and Control • CRC Press, 471 pages 40 Conclusions • Analytical and simulation results show that the proposed scheme – Increases throughput – Guarantee desired QoS and weighted fairness – Also during congestion and fading channels. • The proposed scheme mitigates congestion using a hop by hop mechanism for throttling packet flow rate • The convergence analysis is demonstrated by using a Lyapunov-based analysis • Experimental results show that Congestion-aware routing protocol improves performance of the resource constrained network 41 Questions? 42