Chronicle Discovery for Diagnosis from Raw Data: A Clustering Approach. Issue 24 (2018)

• Record Type: Journal Article
• Authors: Sahuguède, Alexandre; Corronc, Euriell Le; Lann, Marie-Véronique Le
• Is Part Of: IFAC-PapersOnLine. Volume 51:Issue 24(2018)
• Page Start: 1
• Page End: 8
• Publication Date: 2018
• Subjects: Fault Diagnosis -- Machine Learning -- Clustering -- Temporal Pattern Mining -- Chronicle Discovery; Automatic control -- Periodicals; 629.805
• DOI: 10.1016/j.ifacol.2018.09.521
• Languages: English
• ISSNs: 2405-8963

Abstract

Abstract: Chronicles are temporal patterns well suited for an abstract representation of the behavior of dynamic systems. For fault diagnosis, chronicles describe the nominal and faulty behaviors of the process. Powerful algorithms allow the recognition of chronicles in the flow of observations of the system and appropriate actions can be taken when a faulty situation is recognized. However, designing chronicles is not a trivial thing to do. The increasing complexity and capacity of data generation of highly-advanced processes cause the acquisition of a complete model difficult. This paper focuses on the problem of discovering chronicles that are representative of a system behavior from direct observations. A clustering approach to this problem is considered. The chronicle discovery algorithm proposed here designs chronicles with minimal knowledge of the system to diagnose. Furthermore, unprocessed data obtained directly from the system can be used in this clustering algorithm. Finally, the chronicle discovery algorithm proposed in this paper is illustrated on a sport performance monitoring device for a diagnosis of movement deviations in the temporal domain, in the event domain, or both, considered as faults for the athlete.