Learning Methods for Distributed Diagnosis: a Failure Classification Methodology in Discrete Event Systems

Abstract

In the dependability context, the diagnosis is fundamental for identifying and classifying failures to reduce system malfunctions and accidents that can cause serious damage to property and people. Currently, systems are generally complex and the distribution approach of the diagnostic function makes it possible to better managing the systems to be diagnosed. This article proposes a methodology for the distributed diagnosis of Discrete Event Systems, which are dynamic systems widely spread. The proposed approach, based on a generic model, is formed of two sequential steps. The first one consists in the system modeling giving a mean to understand the real behavior of the studied system, through a distributed recovery of signals. The second step is to standardize and process the obtained data for the classification of failures. To achieve this challenge, two learning classification methods are adopted: the Learning Algorithm for Multivariate Data Analysis (LAMDA) coupled with the K-Nearest Neighbor algorithm (K-NN). To validate the approach, the generic model is applied on a particular discrete event system, a railway transport network following a real case of railway line and the obtained results showed the improvement of the classification by recognition. The proposed framework of distributed diagnosis, based on the two learning methods, seems an efficacy and robust methodology for classify any failures of systems.