(140b) Supervision of Distributed Processes with Agent-Based Systems: Agent Management and Adaptation
Multiagent systems provide a powerful framework for developing real-time process supervision and control systems for distributed and networked processes by automating adaptation and situation-dependent rearrangement of confidence to specific monitoring and diagnosis techniques. An agent-based framework for monitoring, analysis, diagnosis, and control with agent-based systems (MADCABS) is developed and tested by using detailed models of chemical reactor networks. MADCABS is composed of three main hierarchical layers, the physical communication layer, the supervision layer and the agent management layer. The supervision layer consists of agents and methods for data preprocessing, process monitoring, fault diagnosis, and control. The agent management layer conducts the assessment of agent performances to assign the priorities for selecting the most useful methods of process supervision for specific types of situations.
There are strong reasons for distributing the activities and intelligence in software for supervision of distributed process operations:
? The complex layout of a manufacturing process yields a problem that is physically distributed,
? The supervision problem is distributed and heterogeneous in functional terms,
? The complexity of the supervision problem dictates a local point of view that contributes to the development of system-wide decisions that may force reexamination of local decisions,
? The supervision system must be able to adapt to changes in the structure or environment of the supervised process or network.
Agents are capable of acting, communicating with other agents, perceiving their environment, and determining behavior to satisfy their objectives. They are endowed with autonomy and they possess resources. However, the MAS framework offers challenges as well: Agents have only partial information about their environments, they may act ?selfishly' or initiate actions that may conflict with actions of other agents. This may lead to undesirable or harmful behavior in MADCABS or the supervised process, compromising its profitability and safety.
The nature of the supervision problem dictates the use of multiple layers of agents where lower-level agents perform local well-defined tasks such as information validation from sensors and higher-level agents perform more global tasks over wider regions of the supervised system. Several agents can be used to perform a specific task, each using different methods to enable not only decision by consensus-building but also to reduce the influence of the weaker methods over time. Intelligence and adaptation is provided both at agent and at system level.
MADCABS is composed of three main hierarchical layers, the physical communication layer, the supervision layer and the agent management layer. The physical layer is where two-way information communication between the process and MADCABS takes place. The supervision layer consists of agents and methods for data preprocessing, process monitoring, fault diagnosis, and control. Data preprocessing agents filter the process data, check for outliers and missing data, and provide estimates for them. Monitoring agents detect deviations from normal operation and trigger the fault detection and diagnosis (FDD) agents. When abnormal process operation is validated, FDD is carried out using contribution plots, statistical methods, and process knowledge. Control agents range from simple local PI controllers to decentralized plant-wide grade transition agents. Some agents collaborate to help each other, while others work in a competing manner to satisfy a global objective. The performances of different agents and methods are evaluated in the topmost agent management layer. The assessment of agent performances guide the priorities assigned to select the most useful methods of process supervision for specific types of situations. In this paper, MADCABS modules for monitoring and fault detection, and the information flow among them is summarized and the elements and operation of the agent management layer is discussed. The interaction between the supervision and agent management layers is illustrated with case studies.
Practice has shown that some monitoring and diagnosis methods perform better for specific situations and worse for others and that it is difficult to have a single method that would give the best results under any circumstances. Therefore, in MADCABS there are multiple methods and different agents that use these methods. The performances of different agents and methods are stored as history along with values of state metrics and process status. The performance history is then used as a reference in estimating a method's performance when a similar situation arises. When two methods are available to perform the same or similar tasks, they are compared based on their historical performances and one method is chosen or its decision is given more weight or priority than the competing method. When a method is not selectrd for a long time because of poor performance, it is motivated to update or adapt to become better. Based on the results of the performance evaluation, the agents update their built-in knowledge and/or the methods they are using.
The current state of the system is compared with the recorded states and the agent or the method that has been the best performing agent or method for similar states in the history is selected. A distance measure is used in the estimations of performances for each method for the current state. The agent or method that gives the highest performance estimation is selected or given a higher priority or reliability and is used accordingly with other less reliable methods, and the combined performance is observed in the latter case. The state metrics for each evaluation are chosen such that they are differentiating and are relevant for the comparison. The state metrics define the situation when the performances were measured. Possible state metrics define if the system was at steady state or at a transient state or if there was a fault in the process when the performance was observed or they can be the estimated fault severity or degree of catastrophe or for networked systems state metrics can be the average neighborhood size and similarity. The performance measurement, on the other hand, can be a single criterion or a composite criterion. The performances of agents and methods are measured at the end of their performance episodes, when they complete their task or when they use up the time allocated for their performance. The common performance criteria are the proximity to the local or the global objective; the cost such as aggressive input manipulation, task completion time or for networked system the disturbance to the network; sum of squared error or in monitoring the number of false and missed alarms. Before the performance analysis, the agents and the methods that need to be evaluated, the state metrics that define the system and the performance criteria should be determined. After the performance evaluation, each agent or method updates its performance values, reliabilities or priorities. And a new performance episode begins, after which the historical performance space is updated with the current performances. This cycle repeats itself for each performance episode.
The case study that illustrates the operation of the management layer will involve grade transition in a reactor network with multiple products with and without simultaneous process disturbances.