(439b) Agent-Based Distributed Monitoring and Fault Detection

Authors: 
Perk, S. - Presenter, Illinois Institute of Technology
Cinar, A. - Presenter, Illinois institute of technology


In spatially distributed, large chemical processes, holistic monitoring and fault detection using the highly correlated, nonlinear process data is difficult and may lead to too many false and missed alarms that would render the monitoring process as unreliable. The performance of the monitoring method is also dependent on the process operating region, the fit of the statistical model to the process data and on the performances of the monitoring statistics for that method. Ideally, a monitoring method should explain the process data very well and its monitoring statistics should neither be too strict nor loose.

The aim of this study is to provide an agent-based adaptive distributed monitoring and fault detection framework, which calculates the performance of each agent under different process conditions, keeps this performance history for future reference, and enables the betterment of each agent through adaptation and coordination. In this scheme, using object oriented programming, monitoring and fault detection are embedded as two interdependent layers, where multiple local monitoring and fault detection agents are located. In this paper, an agent-based distributed monitoring and fault detection scheme is proposed for a spatially distributed reactor network.

To increase the performance of the monitoring method multiblock principal component analysis (PCA) has been proposed for large processes, where meaningful division of the process into separate monitoring blocks is possible. Since, multiblock PCA provides statistical process information both at the local and holistic level, fast fault detection and fault localization is provided. Building multiple principal component analysis (PCA) and dynamic PCA in a decentralized manner for each of the above mentioned monitoring blocks provides extra flexibility for each model, such as different number of principal components can be employed if necessary.

Here, the monitoring layer consists of a series of agents that simultaneously perform decentralized PCA, dynamic PCA for each reactor in the network and an agent that perform multiblock PCA for the whole network, and monitoring organizers that track and store the performances of each of the agents for different parts of the network, and for different operation states. The fault detection layer consists of a series of fault detection agents, namely statistical prediction error (SPE) and T2 agents, associated with each monitoring method for each reactor, and a series of fault detection organizers, that keep track of the performances of each of the local fault detection agents and provide a consensus fault flag for the parts of the network they are located. The fault detection agents' performances are updated online as a weighted moving average in time, where the instantaneous performances are based on the payoffs specified for their actions. A missed alarm is considered as the worst performance, and is followed by false alarm and correct detection, latter being the best performance. The overall performance is also dependent on the performances of other coordinating fault detection methods working on the same part of the network. If a monitoring method's fault detection agents are doing well, this implies that the method itself is also performing well. If the monitoring methods performance is degrading in time because its fault detection methods are performing badly, this requires either an update of the monitoring method or update of the fault detection statistic's limits. This approach helps identify if the fault detection agents are irresponsive or highly aggressive or if the monitoring method is not sufficient for the data provided.

The reactor network model for multiple interconnected CSTRs, each having feed and exit streams, as well as multiple connections to their neighbors, is written in C and linked with Repast Simphony via the Java Native Interface (JNI). The framework is built in Java using Repast Simphony, which is an open source agent modeling and simulation environment, and using COLT distribution. Repast Simphony is a java-based framework for agent simulation and provides features such as an event scheduler and 2D, 3D visualization tools. The agents created with Repast Simphony interact with virtual representations of the physical reactor network. The virtual network objects map the states of the physical system to objects that can be manipulated by the agents in the framework. COLT distribution is used to implement the statistical model building and statistical testing methods and is connected to the agent simulation for online monitoring using Repast Simphony event scheduler. The COLT distribution consists of several free Java libraries, for user convenience bundled up under a single name and provides an infrastructure for scalable scientific and technical computing in Java. It contains, among others, efficient and usable data structures and algorithms for off-line and on-line data analysis, linear algebra, multi-dimensional arrays, statistics, histogram computation and Monte Carlo simulation and parallel and concurrent programming.

A case study is performed in the network of reactors by introducing a disturbance to the network. The effectiveness of the methods to detect the fault, the performance evaluation of the methods, and the cooperation and adaptation of the agents is demonstrated.