Using support vector machines to solve the comparison-based system-level fault diagnosis problem. Issue 2 (3rd March 2016)

• Record Type: Journal Article
• Authors: Elhadef, Mourad; Nayak, Amiya
• Is Part Of: International journal of parallel, emergent and distributed systems. Volume 31:Issue 2(2016)
• Page Start: 176
• Page End: 200
• Publication Date: 2016-03-03
• Subjects: fault tolerance -- system-level self-diagnosis -- comparison models -- partial syndromes -- support vector machines; Parallel computers -- Periodicals; Electronic data processing -- Distributed processing -- Periodicals; Computer algorithms -- Periodicals; 004.35
• DOI: 10.1080/17445760.2014.958824
• Languages: English
• ISSNs: 1744-5760

Abstract

Abstract : This paper introduces a new system-level fault diagnosis approach using support vector machines (SVMs). The objective of the fault diagnosis problem is to identify the set of permanent faulty nodes when at most nodes can fail simultaneously. We consider both the asymmetric and symmetric comparison models which assume that nodes are assigned a set of tasks and their outcomes are compared. Based on the comparison outcomes, the diagnosis algorithm must identify all faulty nodes. In both models, it is assumed that two fault-free nodes give matching results, while a faulty and a fault-free node give mismatching outcomes. The two models differ in the assumption on comparisons involving pairs of faulty nodes. In the asymmetric model, two faulty nodes always give mismatching outputs, while in the symmetric model, both comparison outcomes are possible in this case. For the asymmetric comparison model, we show that a linear SVM is sufficient to solve the diagnosis problem, while for the symmetric model a nonlinear SVM is developed. The new SVM-based diagnosis is first trained using various input syndromes with known fault sets. Then, it is extensively tested using randomly generated diagnosable systems of different sizes and under various fault scenarios. Results from the thorough simulation study demonstrate the effectiveness of the SVM-based fault diagnosis approach, in terms of diagnosis correctness, latency and scalability. We also conducted extensive simulations using partial syndromes, i.e. when not all the comparison outcomes were available before initiating the diagnosis phase. Simulations showed that the SVM-based diagnosis performed efficiently, i.e. diagnosis correctness was around 99% even when at most half of the comparison outcomes were missing, making it a viable addition or alternative to existing diagnosis algorithms.