Author(s)

N. Charkaoui, B. Dubuisson, C. Ambroise & S. Millemann

Abstract

In recent years, the automotive industry has known great evolutions. The introduction of on-board electronic systems and the growth of electronic based functions have made the reparability very difficult for after-sales vehicle technicians. This paper introduces a decision tree classifier in order to detect and isolate after-sales failures on a vehicle. It uses binary data generated by vehicle on-board electronic control units. By selecting a one-against-all decision tree structure, we demonstrate that it is possible to classify instances which may belong simultaneously, with full certainty, to more than one class. The presence of several classes corresponds to the presence of several vehicle failures at the same time. Tree complexity is reduced and overfitting is avoided by applying Error Based Pruning method (EBP). The different tree outputs are combined in a decision system in order to classify an instance into one or several known classes or to consider the instance as belonging to an unknown class. The introduced method has been tested on a real data set. The obtained results demonstrate the good performances and the adapted structure of this classifier. Keywords: diagnosis, data mining, pattern recognition, machine learning, oneagainst- all decision tree, binary data, car failure detection and isolation. 1 Introduction One important application of data mining is to classify instances in a database. In classification, we are given a training sample. A set of classesΩ = {ω1, . . . , ωm} is defined by the training set. Each class corresponds to an operating mode of the system under consideration. The data set is composed of instances (or feature vectors) xi, i = 1. . . , n where each one is described by a set of descriptive attribute

Keywords

diagnosis, data mining, pattern recognition, machine learning, oneagainst- all decision tree, binary data, car failure detection and isolation.