Development of a hybrid intelligent system for on-line real-time monitoring of nuclear power plant operations

Abstract

A nuclear power plant (NPP) has an intricate operational domain involving systems, structures and components (SSCs) that vary in scale and complexity. Many of the large scale SSCs contribute to the lost availability in the operation of the NPPs, when their malfunctions cannot be detected in a timely manner. The lost availability can result in millions of dollars economic loss. Currently, one of the main reasons of the incapacity for avoiding such critical system failures is the lack of an appropriate health monitoring system (HMS). A comprehensive HMS can help prevent the system failures, by analyzing the large amount of information for determining the performance status of the SSCs and for providing decision support in the NPP operations. The immediate goal of this work is to design the methodology for the cognition system of an automated multi-faceted HMS to be implemented at NPPs. The tasks of this system are providing efficient and reliable fault diagnosis, failure prediction, and decision support in NPP operations. The ultimate goal of this work is to enhance the NPP operations by increased availability, and consequently, further improved reliability and safety. This work presents the design of the cognition system of the HMS as a unique hybrid intelligent system. In this hybrid structure, we use the Bayesian network (BN) and neural network (NN) techniques in conjunction, for the first time, in order to provide complementary probabilistic performance status estimates and fault diagnosis concerning the monitored SSCs. This strategy makes the real-time implementation of this diagnostic model feasible in large scale, complex problem domains.

(cont.) It permits an important step in the applied artificial intelligence field. We use the influence diagram (ID) technique for extending the Bayesian network structure to provide decision support to the plant personnel. The advisory model using the ID technique derives recommendations concerning the optimal alternative actions to be executed during incidents. Finally, we adopt an analytical model to estimate the mean time to occurrence of the initiating faults and failures of the monitored systems. This model considers the multi-state dynamic behavior of the SSCs. The ultimate configuration of these models for various tasks of the HMS is named the HIS-N, hybrid intelligent system for NPP operations. The HIS-N has been developed, tested and verified on a test bed that mimics the bearing system of a horizontal charging pump. The results from the implementation of the HIS-N on the target system show consistent behavior in fault diagnosis, decision support and failure prediction tasks. The features of our HMS methodology gives us new capabilities to enhance the NPP operations, in order to reach its ultimate goal. The new monitoring method with the HIS-N developed in this work is not restricted to NPP operational applications. We recommend that an identical approach be adopted for similar purposes also in other large scale, complex process domains.