A Perceptron Neural Network for Asymmetric Comparison-Based System-Level Fault Diagnosis

Abstract

The system-level fault diagnosis problem aims at answering the very simple question "Who's faulty and who's fault-free?", in systems known to be diagnosable. In this paper, we answer such a question using neural networks. Our objective is to identify faulty nodes based on an input syndrome that has been generated using the asymmetric comparison model. In such a model, the system, which is composed of interconnected independent heterogeneous nodes, is modeled using an undirected comparison graph. Tasks are assigned to pairs of nodes and the results of executing these tasks are compared. Based on the agreements and disagreements among the nodes' outputs, the diagnosis algorithm must identify faulty nodes. In general, it is assumed that faults are permanent, and that at most t nodes can fail simultaneously. The new solution we introduce in this paper uses a perceptron neural network to solve the fault identification problem. The neural network is first trained using various input syndromes with known fault sets. Extensive simulations have been conducted next using randomly generated diagnosable systems. Surprisingly, the neural network was able to identify all the millions of faulty situations we have tested, including those that are unlikely to occur. Simulations results indicate that the perceptron-based diagnosis algorithm is a viable addition to present diagnosis problems.