OUTPUT FEEDBACK PREDICTOR COMPENSATING TIME-DELAYS IN DISTRIBUTED REAL-TIME SYSTEMS Lilantha Samaranayake and Mats Leksell Department of Electrical Engineering Royal Institute of Technology (KTH) Stockholm, Sweden e-mail: lilantha@ekc.kth.se Sanath Alahakoon Department of Electrical and Electronics Engineering University of Peradeniya Sri Lanka e-mail: sanath@ee.pdn.ac.lk ABSTRACT Inherent time-delays in distributed real-time control sys- tems can degrade the performance or unstable the sys- tem unless treated properly. Standard delay compensa- tion schemes usually include time stamped measurements, which require clock synchronization at different nodes. This needs periodic re-synchronization to minimize the im- pact of clock drift, consuming additional bandwidth. Fur- ther, the controller to actuator delay is unknown prior to the controller calculation, where probability distribution based complicated methods are required, to know it in advance. In this paper, a plant model based predictor is cascaded to the state observer to predict the states corresponding to the next sample. These states are then used in the controller to derive the next control signal, which is released just af- ter reading the next sample. This completely eliminates clock synchronization and complications in delay compen- sation. The proposed method is implemented in simulation in a real-time environment for open loop stable and unsta- ble systems. KEY WORDS Distributed, Real-Time, Time-Delay, Output feedback, Predictor based compensator 1 Introduction Thanks to the vast advancements in control networks, many computer controlled real-time systems are becoming dis- tributed systems offering a large reduction in wiring and hence cost and maintenance burdens. It is not uncommon for the sensor, the actuator, and the control calculations to reside in different nodes of a communication control network, as in large manufacturing systems, modern vehi- cles etc, for example. This gives rise to networked control loops. Within the individual nodes, the controllers are of- ten implemented as one or several tasks on a microproces- sor with a real-time operating system (OS). Often the mi- croprocessor also contains tasks for other functions (e.g., communication and user interfaces). The OS typically uses multiprogramming to multiplex the execution of the var- ious tasks. The CPU time and the communication band- width can hence be viewed as shared resources for which the tasks compete. Digital control theory normally assumes equidistant sampling intervals and a negligible or constant known con- trol delay from sampling to actuation. However, this can seldom be achieved in practice specifically in distributed control systems closed over communication control net- works. Within a node, tasks interfere with each other through preemption and blocking when waiting for com- mon resources. The execution times of the tasks them- selves may be data dependent or may vary due to hardware features such as caches. On the distributed level, the com- munication gives rise to delays that may or may not be de- terministic depending on the communication protocol. An- other source of temporal non-determinism is the increasing use of commercial off-the-shelf (COTS) hardware and soft- ware components in real-time control (e.g., general purpose OS such as Windows and Linux and general purpose net- work protocols such as Ethernet). These components are designed to optimize average-case rather than worst-case performance. However the motivation to use COTS over dedicated networks (ControlNet, DeviceNet, etc) and OS is justifiable in terms of compatibility, flexibility and cost issues that would run into otherwise. The temporal non-determinism can be reduced by the proper choice of implementation techniques and platforms. For example, time-driven static scheduling improves the determinism, but at the same time it reduces the flexibility and limits the possibilities for dynamic modifications. De- spite dynamic scheduling techniques, some level of tempo- ral non-determinism is unavoidable. From a control sys- tems perspective, the non-deterministic time delays add time varying phase lags giving rise to to significant perfor- mance degradations [1]. This can be in the form of higher rise times, settling times, oscillations etc. Therefore, de- lay compensations have to be implemented to avoid such degradations. Such compensations usually depend on de- lay measurements and/or delay predictions. Delay mea- surements require every node of the distributed system to be synchronized to a global clock or node clocks synchro- nized among each other. The delay predictions depend on the probability distribution usually evaluated off-line, which are not very accurate. Further it is hard to adapt any change in the network traffic conditions. Therefore a sim- pler delay compensation, which does not depend on delay measurements is presented in this paper.