Associative memories, stochastic activity networks and their application to sensor validation systems of nuclear power plants

Abstract

In this paper, the problem of designing an advanced sensor validation system (SVS) which is robust and fault-tolerant under faulty conditions is considered. Associative memories, which provide robust pattern recognition are investigated as an information processing technology that can be applied to sensor validation. Studies of Binary Associative Memories (BAM) and Continuous Associative Memories (CAM) yield many results including (1) the stability condition of exemplars and spurious memories in BAMs, (2) the formula of choosing diagonal weights and bias that eliminates spurious memories most effectively in BAMs, (3) the convergence theory of CAMs that have asymmetric weight matrix with non-zero diagonal elements and non-monotonically increasing activation functions, (4) the energy function that explores the convergence behavior of CAMs, and (5) the hybrid learning algorithm that reduces spurious memories effectively in CAMs. The concept of performability is introduced to the evaluation of SVS. A set of important performability variables is introduced. Stochastic Activity Networks are used as a modeling tool to evaluate the performability of SVS. An illustration example, the evaluation of the pressurizer SVS of a PWR, is provided.