Designing an Associative Memory via Optimal Training for Fault Diagnosis

In this paper, the authors discuss a new synthesis approach to train associative memories, based on recurrent neural networks. They propose to determine the weight vector as the optimal solution of a linear combination of support patterns. The proposed training algorithm maximizes the margin between the training patterns and the decision boundary. The design problem considers: (1) obtaining of weights via an optimal hyperplane algorithm for SVMs and (2) obtaining conditions to reduce the total number of spurious memories. This algorithm is applied to the synthesis of an associative memory for fault diagnosis in fossil electric power plants.