The Water Recovery System 

Context in source publication

Context 1

... transition structure on a state space, which is a cross product of two domains: (i) discrete-event and (ii) continuous-time dynamics. The interaction of discrete-event and time-based variables makes the behavior generation and analysis tasks quite challenging and computationally complex. Considerable amount of research work has been dedicated recently to the study of hybrid systems dynamics [5, 6]. The complex nature of hybrid systems limits the applica- bility of traditional optimal control techniques and supervisory control techniques that can be applied directly to hybrid systems. Several promising approaches have been proposed in the literature to deal with the complexity of hybrid systems. For example, abstraction techniques have been developed to reduce the complexity of the hybrid models while preserving features of the original model relevant to the analysis/control objectives (e.g., [7]). Supervisory control design with abstracted hybrid system models has been investigated in [8, 9]. Efficient control synthesis for reachability specifications through mode switching has been presented in [14]. Section 2 introduces the basic building blocks of our model-based FDI and fault-adaptive control schemes, and emphasizes the importance of component-based modeling and the link between fault isolation, identification, and fault-adaptation. Section 3 presents the models of two coupled components of the Water Recovery system (WRS) of the ALS that we have chosen as the test- bed for our fault-adaptive control studies. Section 4 discusses our diagnosis scheme and the hierarchical decision-theoretic control scheme to achieve optimal performance in the system given resource constraints and a set point trajectory that applies for nominal operation. Section 5 presents the results of the experiments we have conducted on the WRS, and section 6 presents the conclusions of this work. Our approach to fault-adaptive control, illustrated in Fig. 1, is centered on model-based approaches for fault detection, fault isolation and estimation, and hierarchical online supervisory control for hybrid systems. The plant is assumed to be a hybrid system [4, 5]. The heart of the Fault Adaptive Control Unit is the Hybrid Observer [11] that tracks the behavior of the plant under nominal conditions. When the Fault Detector detects a discrepancy between the measured and the expected behavior, the diagnosis units are triggered. The Hybrid Diagnosis unit combines qualitative reasoning with quantitative parameter estimation. Qualitative diagnosis is based on dynamic plant models represented as Temporal Causal Graphs [12]. The set of candidates picked by the qualitative diagnoser are passed to the Parameter Estimation Unit that reduces the candidate set to a single fault candidate by computing the degree of degradation, and re- taining that candidate that has the least prediction error. The online hybrid control approach focuses on optimal resource management and robust fault-adaptive control using a decision-theoretic control scheme. The proposed approach is designed to ensure distribution of a finite amount of resources among contending subsystems of a larger system in a way that near optimal performance may be obtained over an extended period of time. In more detail, the control algorithm is designed to achieve a set of pre-specified performance requirements for the system over finite time intervals, while simultaneously optimizing a given utility/cost function for the composite system and maintaining overall system stability. To achieve fault adaptivity, the results of fault diagnosis are used to update the system model online so that the observer may again track system behavior accurately under faulty conditions. The online supervisory controller uses the updated system model to derive a new set of performance requirements. The decision-theoretic control schemes for the individual subsystems are then applied at runtime to optimize performance in the faulty system. The ALS system is made up of multiple loosely-coupled subsystems [3], such as a Water Recovery System (WRS), an Air Revitalization System (ARS), a Biomass Production system, and a Power generation system. These subsystems comprise a number of interacting control loops, such as the fluid flow loop, the energy management loop, and the bio-regeneration and gas transfer loop. These loops also cover multiple physical domains, and operate at multiple time scales. An effective way to describe the behavior of the controlled subsystems is to model them as hybrid dynamic systems [4]. In this paper, we focus on the WRS, in particular on an experimental system that was developed and tested at the NASA Johnson Space Center (JSC) [13]. This subsystem recycles urine and wastewater into potable water. Critical requirements for such a system are that it consumes low power, minimize the use of consumable resources, and run in a fully autonomous mode for long periods of time. The WRS, as shown in Fig. 2, is com- prised of a Biological Water Processor (BWP) to remove organic compounds including ammonia, a Reverse Os- mosis (RO) System to remove particulate matter after the BWP, an Air Evaporation subsystem (AES) to purify the remaining concentrated brine that is purged from the RO system, and a post processing system (PPS) to remove the trace organic and trace inorganic compounds by ultra-violet treatment to bring the water to potable limits. The combination of the BWP and RO subsystems produce about 85% of the clean water. The remaining 15% is produced by an evaporation and condensation process in the AES from the concentrated brine that is purged to it from the RO. In this work, we focus on controllers for the RO and AES systems, and the interactions between these systems to assure desired output given limited resources, which is primarily the energy available for subsystem operation. THE RO SUBSYSTEM - This subsystem, shown in Fig. 3 is the linchpin subsystem in the WRS loop. It pulls water from the GLS (gas liquid separator) of the BWP, and delivers purified water (permeate) to the PPS and concentrated brine to the AES. The RO removes inorganic compounds and particulate matter by pushing the input water at high speed through a cylindrical membrane that acts like a molecular sieve. The clean water permeate is passed on to the PPS, and the dirty water (brine) continues to circulate in the RO loop. The RO is designed to go through six modes. The primary mode draws water into a coiled section of pipe that acts like a reservoir, while processing permeate in the outer loop. When the brine concentration increases above a preset level, the system is switched to a secondary mode, where the brine circulates faster in a smaller inner loop with the recirculation pump, therefore, it is pushed harder against the membrane. This keeps the clean water production at a reasonable rate, but the concentration of brine in the inner loop continues to increase. At some point, the concentration of brine be- comes high enough to reduce the output from the RO system significantly, so the brine is purged into the AES, a new batch of water is drawn in from the BWP, and the primary cycle starts again. Periodically, however, as particulate matter accumulates in the membrane, it needs to be cleaned by running the water backwards in the inner loop. This is known as the slough phase. The primary power consumers in the RO system are the two pumps, which circulate the water through the system. THE AES SUBSYSTEM - This subsystem contains a reservoir where the brine is collected. The brine is ab- sorbed onto a wick and evaporated using hot air blown over the wick. The evaporated water is condensed by passing it through a heat exchanger, and collected in a tank before it is sent to the PPS system. The primary power consumers in this subsystem are the blower, which moves the air through the system, and the heating unit, which heats the air to facilitate evaporation of water. MODELING THE RO & AES SUBSYSTEMS – Building models at the right level of detail is a critical first step in the success of a model-based fault-adaptive control scheme. The choice of the model representation and the level of detail included in the model determine the set of faults that are linked to model parameters, and the set of inputs that can be controlled. In the bond-graph modeling paradigm [7] that we have adopted, faults in components that are linked to parameters in the bond-graph model can be isolated, and the controlled inputs are represented as sources of flow and effort. Bond graphs (BG) define a domain-independent topo- logical modeling language that captures energy-based interactions among the different physical processes that constitute a dynamic system [14]. The vertices in the graph are components or subsystems modeled as ge- neric physical processes, such as capacities and inertias (energy storage processes), dissipators (dissipative process), transformers and gyrators (transformation between energy domains) and sources (interactions of system with environment). Component behavior can be linear or nonlinear. Additional vertices impose conservation of energy at idealized connecting points between components. Hybrid Bond graphs (HBG) are an extension of the bond graph formalism that allow some elements to have discrete states, giving the modeler the ability to create domain-independent models that can describe both continuous and discrete behaviors of a system [4]. A unique property of the HBG is the use of switching signals to turn energetic connections between HBG components on and off . Nonlinear systems are modeled by components that have time-varying parameters, i.e., their parameter values are functions of system variables. The HBG models for the RO system and the AES are shown in Figs. 5 and 6. The HBG model of the RO system was derived by decomposing the system into three principal domains of operation. Given the pump-fluid system, ...