A Novel Generalized-Comparison-Based Self-Diagnosis Algorithm for Multiprocessor and Multicomputer Systems Using a Multilayered Neural Network.
ABSTRACT We consider the system-level self-diagnosis of multiprocessor and multicomputer systems under the generalized comparison model (GCM). In this diagnosis model, a set of tasks is assigned to pairs of nodes and their outcomes are compared by neighboring nodes. The collections of all comparison outcomes, agreements and disagreements among the nodes, are used to identify the set of faulty nodes. We consider only permanent faults in t-diagnosable systems that guarantee that each node can be correctly identified as fault-free or faulty based on a valid collection of comparison results (the syndrome) and assuming that the number of faulty nodes does not exceed a given bound t. Given that comparisons are performed by the nodes themselves, faulty nodes can incorrectly claim that fault-free nodes are faulty or that faulty nodes are fault-free. In this paper, we introduce a novel neural networks-based diagnosis approach to solve this fault identification problem. The new diagnosis approach exploits the off-line learning phase of neural networks to speed up the diagnosis algorithm. We have implemented and evaluated the new diagnosis approach using randomly generated diagnosable systems. The new neural-network-based self-diagnosis approach correctly identified most of the faulty situations forming hence a viable addition or alternative to solve the GCM-based fault identification problem.
[Show abstract] [Hide abstract]
ABSTRACT: A modified Hop field neural network is introduced to solve the comparison-based system-level fault diagnosis problem when only partial syndromes are available. We use the generalized comparison model, where a set of tasks is assigned to pairs of nodes and their outcomes are compared by neighboring nodes. To identify the set of permanently faulty nodes, the collections of all agreements and disagreements, i.e., the comparison outcomes, are used. First, we show that the new diagnosis approach works correctly when t-diagnosable systems are considered. Then, we show the main contribution of this new diagnosis approach which is its capability of correctly identifying the set of faulty nodes when not all the comparison outcomes are available to the diagnosis algorithm at the beginning of the diagnosis phase, i.e., partial syndromes. The simulation results indicate that the modified Hop field neural network-based fault identification algorithm provides an effective solution to the system-level fault diagnosis problem even when partial syndromes are available.01/2011; DOI:10.1109/ICPADS.2011.8