Data Mining Based Fault Isolation with FMEA Rank: A Case Study of APU Fault Identifcation Chunsheng Yang" Sylvain Letourneau" Yubin Yang 2 , and Jie Liu 3 I Information and Communication Technologies, National Research Council Canada, Ottawa, ON K IA OR6, Canada chunsheng.yang@nrc.gc.ca, sylvain.letourneau@nrc.gc.ca 2 State Key Laborator for Novel Software Technolog, Nanjing Universit, Naning, Jiangsu, 210093, China yangubin@nju.edu.cn 3 Dept. of Mechanical and Aerospace Engineering, Carleton Universit, Ottawa, ON K1S 5B6, Canada jliu@mae.carleton.ca Abstract - FMEA (Failure Mode and Effects Analysis), which was developed to enhance the reliability of complex systems, is a standard method to characterize and document product and process problems and a systematic method for fault identifcation/isolation in maintenance industry. Fault identifcation for a given failure effect or mode is a reactive process. Usually, a failure has occurred and it needs to identify which component is the root cause or to isolate the fault to a specifc contributing component. Traditional method is to conduct TSM (Trouble Shooting Manuals)-based fault isolation, which is complicated, expensive, and time-consuming. To efciently perform fault isolation, this paper proposed data mining-based framework for fault isolation by using FMEA information to rank data-driven models. In this paper, we present the proposed framework along with a case study for APU fauIt identifcation. Keywords - FMEA, data nunmg, data driven models, binar classier, fault isolation and identication, failure mode, FMEA validation. I. INTRODUCTION Failure Mode and Effects Analysis (FMEA) has been used for fault identifcation and prevention in maintenance industry as a systematic method, since it was originally developed to enhance the reliability of space program hardware [1]. FMEA provides a foundation for qualitative reliability, maintainability, safety and logistic analysis; it documents the relationships between failure cause(s) and failure effects. In particular, FMEA contains usefl inforation such as Severity Class (SC), Failure Rate (FR), and Failure Mode Probability (FMP) which indicate the probability of each failure mode and its effects on system perforance. In analyzing failure modes/effects (which is defmed as a relationship between a failure mode and its effects) for a complex system, a fault or failure is usually associated with many potential components based on FMEA documents or the Trouble Shooting Manuals (TSM). Fault identifcation for a given failure effect or mode is a reactive process. Usually, a failure has occurred and it needs to identif which component is the root cause or to isolate the fault to a specifc contributing component. Traditionally, the process is conducted based on TSM or human experiences. However, the suggested procedures fom the TSM are ofen complicated, expensive, and time-consuming. The TSM 978-1-4673-5723-4/13/$31. 00 ©2013 IEEE outlines list of possible causes, but these lists are not ranked. The work in [2] used maintenance and operational data to provide maintenance technicians with a ranked list of possible causes along with the standard TSM list. To effciently perform fault identifcation and frther enhance the troubleshooting procedures, we proposed a data mining-based famework for fault identifcation or isolation, which applies the validated FMEA to rank data-driven models and uses the operation data prior to failures as input to identif the root contributing component for a given failure mode. Starting a brief overview on FMEA validation to intoduce some background of FMEA knowledge, this paper presents the proposed data mining-based method for fault isolation. The paper also reports an application of the method through a case study: APU fault identifcation. The rest of this paper is organized as follows. Section II briefy reviews the validation of FMEA document and the updated FEMA parameters which will be used to rank data driven models; Section III presents data mining-based method for fault isolation; Section IV provides an application of APU fault identifcation; Section V discusses the limitation and provides the fture direction; the fnal section concludes the paper. II. OVERVIEW OF THE FMEA V ALTDATION Since FMEAs are produced at design time and then hardly validated afer deployment of the corresponding system, there is a risk that the information provided is incomplete or no longer accurate. The likelihood for such inaccuracies is particularly high for complex systems such as aircraf engines that operate over a long period time. In such cases, using the initial FMEA information without adequate validation could result in the introduction of irrelevant maintenance actions. To avoid such issues, the initial FMEA information needs to be validated and then updated as required. We strongly suggested that prior to use FMEA information, one must try to confrm its validity. In our previous work [3, 4], we had performed this task using real-world readily available maintenance and operational data. In particular, we had investigated validation and updating of an FMEA for an APU (Auxiliary Power Unit engine). The FMEA validation relied on an "in-house" representation of an APU FMEA prepared by domain experts