Supervision and control of complex chemical processes with agent-based systems.

摘要

It is highly desirable to have a combined process supervision and control framework that automates the simultaneous operation of fault detection, diagnosis and control without operator intervention, provides a flexible environment for adaptive operation, and easily scales to large, spatially distributed processes.;An adaptive agent-based supervision and control framework is developed in this thesis and implemented as part of a multi-agent hierarchical, autonomous, distributed decision making system for monitoring, analysis, diagnosis, and control with agent-based systems (MADCABS).;Multiple alternative methodologies such as principal component analysis (PCA), dynamic PCA (DPCA), and multi-block PCA (MBPCA) are implemented for use by a group of intelligent monitoring and fault detection agents simultaneously. Diagnosis agents use contribution plots, partial least squares (PLS), and Fisher's discriminant analysis (FDA) techniques. Different consensus mechanisms are utilized for fault detection and diagnosis. The decisions of agents are summarized in a consensus using different criteria from voting-based criteria to more complex context-dependent performance-based criteria.;Fast and reliable fault detection and accurate fault diagnosis are required for effective fault-tolerant control which utilizes information about the current process condition, the type of the fault in effect, and the availability of process sensors and actuators, to dynamically switch to an alternative control strategy that is applicable to the current process state.;The novel findings of this agent-based supervision and control framework are the dynamic evaluation of the performances of agents under changing operating conditions, learning and adaptation on the basis of experience, and the improvement of the overall performance of the combined framework for fault detection, diagnosis and control with performance-based consensus building and multi-level adaptation.;The effectiveness of using different consensus criteria for fault detection and diagnosis is demonstrated with case studies on a distributed continuous stirred tank reactor (CSTR) network using multiple system disturbances of various magnitudes. Adaptive performance-based consensus-building yields fewer missed alarms in fault detection and fewer misclassifications in fault diagnosis over time than a voting-based criterion.