(251f) Adaptive Hierarchical Monitoring of Distributed Processes with Multiple Modes of Operation

Multivariate statistical monitoring techniques have been used for monitoring large scale distributed chemical processes with many variables. These techniques rely on the availability of normal operation data that is used in the statistical model building to detect the deviations from normal operation. However, there are some drawbacks related to using such modeling techniques for monitoring complex systems with multiple operating modes and slowly drifting dynamics.

The parameters in the statistical models such as the number of principal components, the system order used and the scaling technique are dependent on the data used. In addition, normal operation may be subject to change and data used in model building should be updated to reflect this change. These are some of the problems that need to be considered for automated monitoring approaches. The automated system should tailor the parameters of the monitoring methods according to the data, and should update the model to increase the effectiveness of fault detection.

Effective process monitoring and reliable and timely fault detection of spatially distributed nonlinear processes with multiple operating points or with slowly drifting dynamics is difficult. Most approaches proposed for monitoring processes with multiple operating modes involve building separate statistical models for each operating mode and switching between these models when the operating mode switches. This approach enhances the effectiveness of the monitoring technique; however, it requires the operating modes to be known a priori and can be cumbersome with many modes of operation. For slowly drifting processes, static statistical models are rendered ineffective for monitoring and fault detection, since static models cannot adapt to the drift.

The aim of this study is to provide an adaptive hierarchical monitoring framework with agent- based systems for large, spatially distributed complex processes. Various monitoring agents construct statistical models using different techniques, adjust the models parameters automatically and collaborate with the control agents to update the statistical models in order to accommodate changes in the operating modes.

The hierarchical agent framework is made of two interconnected layers. The lower level of the framework consists of multiple monitoring agents that provide local and global monitoring for distributed systems. The higher level, called agent management layer, contains agents that monitor the performances of the monitoring agents, and evaluate their performances for different fault scenarios, such as sensor faults or process faults with various magnitudes. They automatically improve the effectiveness of monitoring through adaptation by changing the model parameters or updating the model data.

In this paper, an agent-based adaptive monitoring approach based on cooperation and communication between monitoring and control agents responsible for each process unit is proposed and the effectiveness of the combined agent architecture is demonstrated on a CSTR network exposed to several fault scenarios. The cooperation between control and monitoring manager agents initiate the adaptation since new statistical models might be required to monitor the new in-control process. This framework is implemented by using the Monitoring, Analysis, Diagnosis and Control with Agent-Based Systems (MADCABS) suite developed at IIT. MADCABS contains agents that implement various process monitoring and control techniques. It also contains agents that use the results of these techniques with built-in heuristic rules to provide local intelligence to complete their assignments and reach their objectives.

The monitoring agents in MADCABS use statistical techniques from literature such as principal components analysis (PCA), dynamic PCA (DPCA) and multi-block PCA (MBPCA). For spatially distributed systems separate PCA and DPCA agents build models in the beginning of the analysis for each operating unit in the processes. A single multi-block model is also built for the whole processes, where the blocks consist of data coming from individual operating units. This distributed monitoring structure makes up the lower level of the proposed hierarchical architecture. In earlier studies, the cooperation and consensus decision making among different monitoring agents has been demonstrated on a CSTR network. Using multiple agents for monitoring, and their context dependent reliability weighing over each other provided effective monitoring and fault detection in terms of detection times and false and missed alarm rates. In this study, the earlier framework is further enhanced via the addition of the higher level monitoring manager agents that automate adaptation.

The CSTR network used in this study is a complex distributed process with multiple steady state operating regimes. Three product species that are using the same resource coexist in the CSTR network in a way such that only one of the species is dominant in each reactor and the others are found in trace amounts. The overall network concentrations are set on demand and are required to be kept constant despite faults entering the system. The aim is to satisfy the global objective which is to keep the overall network exit concentration at the desired level. Because of the nonlinear dynamics, the faults may irreversibly disturb the process, and the network concentration deviates from the set point. Control agents cooperate with the monitoring agents in MADCABS to find the control strategy that will confine the fault giving minimal disturbance to the rest of the network, and in the mean time satisfy the global network concentration despite the fault. However, depending on the severity of the fault introduced to the process, it is possible to satisfy the global objective but it may not be possible to restore the original concentrations in each reactor, since the operation point might have changed in some of the reactors. In other words, one reactor might have moved to a new steady state where a new species is dominant. This change is compensated by another reactor, which will produce more or less of another species to satisfy the global objective.

A case study is performed in the CSTR network 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 between control and monitoring agents and adaptation of the agents is demonstrated.