XIX IMEKO World Congress Fundamental and Applied Metrology September 6−11, 2009, Lisbon, Portugal ANT-BASED SEARCH STRATEGY FOR INDUSTRIAL MULTIPLE-FAULT DIAGNOSTICS Pasquale Arpaia 1,2 , Carlo Manna 3 , Giuseppe Montenero 2 1 Dipartimento di Ingegneria, Università del Sannio, Benevento, Italy, arpaia@unisannio.it 2 CERN, European Organization for Nuclear Research, Geneva, Switzerland, giuseppe.montenero@cern.ch 3 Dipartimento di Ingegneria Elettrica, Università degli Studi di Napoli Federico II, Napoli, Italy, massimo.manna@isib.cnr.it Abstract − A swarm-intelligence solution to industrial problems of automatic multiple-faults diagnostics is proposed. In particular, drawbacks of swarm-based algorithms in heuristic search strategy related to the mutual dependence of solutions are overcome by a likelihood-based trail intensity modification of ant-colony optimization. Experimental results of comparison tests with an evolutionary state-of-the-art solution of a case study on an advanced industrial system for remote monitoring, diagnostics, and maintenance are reported. Keywords: Artificial Intelligence, Industrial Diagnostics, Multiple Faults Diagnosis. 1. INTRODUCTION Swarm intelligence is a biologically-inspired soft computing technique based upon the study of collective behaviour in decentralized and self-organized systems [1]. These systems typically consist of numerous autonomous simple agents whose movements through a continuous space are governed by various local forces exerted by other nearby agents or by the environment. Although normally a centralized control structure dictating how individual agents should behave is missing, local interactions among such agents often lead to the emergence of an interesting global behavior. The vast majority of applications of swarm intelligence have involved modelling movements in a 2D or 3D physical space, especially in the simulation of biological populations, computer graphics, and robotic control [1]-[4]. Inspired by successes in these applications, recently several efforts have been carried out to generalize swarm methods to higher dimensional abstract spaces. In particular, particle swarms [5] and bacteria-inspired chemotaxis algorithms [6] have shown their applicability as general-purpose numerical optimization methods. Among high-dimensional numerical problem, Multiple Faults Diagnosis (MFD) problem [7] is a formidable task. However, MFD features cannot be easily faced by a classical swarm intelligence approach [3]. In particular, MFD defines a solution composed by a set of possible faults, where the occurrence of each fault is not conditioned to other faults. Conversely, most promising swarm intelligence methods, such as classical Ant Colony Optimization (ACO) algorithms [3], resolve numerical optimization problems (e.g. the traveling salesman problem), with optimal solutions composed by a set of mutually-conditioned choices influencing the final solution. In this paper, a novel swarm intelligence method, inspired to a likelihood-based trail intensity modification of the classical algorithm ACO aimed at overcoming the above drawback, is proposed for MFD problems (and similar) of industrial diagnostics. In particular, after stating analytically the MFD problem in Section2, the state of the art of the proposed solutions is analyzed in Section 3. In Sections 4 and 5, the and the implementation of the proposed method are outlined, respectively. Finally, in Section 6, the preliminary experimental results of the proposed approach application to a case study of industrial diagnostics, with performance comparison to a classical evolutionary solution, are reported. 2. MULTIPLE-FAULTS DIAGNOSIS PROBLEM Multiple Fault Diagnosis problems are characterized as the 4-tuple [7]: <D, M, C, M + > where: • D is a finite nonempty set of faults, • M is a finite set of symptoms, • C is a relation, which is a subset of DxM, pairing faults with associated symptoms such that (d,m) ∈ C means that the fault d may cause the symptom m, • M + is a subset of M identifying the observed manifestations. A diagnosis DI (i.e. a subset of D) identifies the disorders eventually responsible for the symptoms in M + . A prior probability p j is associated to each fault d j in D. Values are assumed to exist and faults in D are assumed to be independents each other. Associated with each “causal association” in the matrix C is a causal strength c ij representing how frequently a fault d j causes the symptom m i . The causal strength represents the conditional probability P(d j causes m i |d j ). An example of prior probability and C matrix 3x3 one-half dense is reported in Tab. 1. 852 ISBN 978-963-88410-0-1 © 2009 IMEKO