(740d) A Fault Tolerant Control Framework in Reactor Networks with Agent-Based Systems

Control of distributed and networked chemical plants is challenging due to highly interacting process units, size of controlled system, great difficulty in obtaining an accurate system model using first principles, and nonlinearities. There are two major ways of implementing plantwide control: centralized control approach and distributed control approach. The centralized control framework uses a centralized controller that gathers all the information, makes the decision with the consideration of the overall condition, and sends the control commands (setpoints) to all the low level controllers; while in distributed control, every controller generates control actions according to their operating states and the local or regional information provided by the other controllers in the interacting units.

The centralized control could achieve a best performance, and react to the fault in a most efficient way, since it has the comprehensive understanding of the whole system. However, the centralized approach needs a large amount of information transmission, much more computational power, and complex action logic because it has to deal with all possible controller state transitions. Although distributed control often performs sub-optimally, its distributed nature enables it to react faster, and has more flexibility. The local controllers generate local control actions to handle disturbances without disturbing the entire system. In case of a fault, a distributed approach is more robust than a centralized approach since the normal operating subsystems would still be in control. It is easy to add or remove a controller to or from a networked controller system, when it is under a distributed control framework.

An accurate enough model of the process that captures the relations between system inputs and outputs is necessary to enable accurate estimation and develop the model-based controllers. For modeling complex systems, we use data-based system identification tools rather than first principle models that necessitate significantly more resources. In our research, subspace system identification method is used for developing the models for the distributed model that is required for distributed control more conveniently. Furthermore, in order to maintain control with minimum disturbance to the entire network under fault scenarios, the networked controllers also need to be informed of the possible abnormalities in the system. Thus, a combined distributed system identification, fault detection, diagnosis and control methodology is needed.

MADCABS (Monitoring, Analysis, Diagnosis, and Control with Agent-Based Systems) is a software platform developed at Illinois Institute of Technology to provide real-time process supervision and control environment for distributed and networked processes. It incorporates a hierarchical multi-agent system philosophy to make an adaptive, decentralized, hierarchical supervision system with distributed artificial intelligence. Currently we have all the main agents that are needed for a comprehensive distributed, fault-tolerant and adaptive control. to the presentation will focus on how the agents generate system models, execute process monitoring and fault diagnosis, communicate with each other, and how all this information is synthesized in distributed controllers to solve the overall control problem. The distributed control strategy has a hierarchical architecture: supervisor controller agents, regional controller agents and local controller agents. The system has the flexibility to add or remove layers as needed.

An autocatalytic CSTR network is studied for illustrating the performance of the distributed control network. Several CSTRs are put in a rectangular grid network. Each reactor has its own feed flow and exit flow, and there are interaction flows between the neighboring reactors. Three different products are produced in the network using the same feed resource. The product ratio at the output from the network needs to be maintained at a desired set-point. The reactors in the network can experience process abnormalities that may affect the desired product ratio or the product grade may be changed (set-point change to new product ratio).

The aim of the combined fault detection and diagnosis (FDD) and distributed control framework is to detect and identify the potential abnormality in the system and effectively reject the disturbances or provide fault tolerant control to maintain the production at the desired level. MADCABS FDD agents use statistical process monitoring methodologies, such as PCA, dynamic PCA and multiblock PCA, and effectively detect the abnormalities in the system using a performance-based consensus mechanism. Linear discriminant analysis methods and contribution plots are used by the diagnosis agents for fault classification and identification. The object-oriented environment of MADCABS provides the flexibility of monitoring and control of each reactor in the network. Once a potential fault is identified, the control agents generate the corrective actions (using the agents own decision privileges and through negotiations with other control agents) and provide fault-tolerant control to maintain the desired network product ratio.

In the control hierarchical architecture, the local controller agents in the network have some autonomy to dynamically switch between available manipulated variables according to the fault information provided by the diagnosis agents, and retune controller parameters according to controller performance evaluation results. However, since the local controller agents just have limited authority of only perceiving information in their neighborhood, conflicts may happen when different local controllers want to take over the same resource, or when the local decision is not complying with the higher level controller, thus such conflicts should be resolved by either certain coordination mechanism, or by following some supervisory rules.

In our presentation, the combined fault detection, diagnosis and control agent structure will be discussed. Identification of the control strategies that can be used by different control agents, evaluation of agents performances using different strategies, selection of the best strategy for the given fault condition and the effectiveness of the combined framework will be illustrated with various cases in the CSTR network. The performance of distributed control based on multi-agent philosophy will be assessed.