1 Mechanical Fault Diagnosis using Wireless Sensor Networks and a Two-Stage Neural Network Classifier 1 2 P. Ballal * , A. Ramani * , M. Middleton * , C. McMurrough * , A. Athamneh ** , W. Lee ** , C. Kwan *** and F. Lewis * * Automation & Robotics Research Institute, University of Texas at Arlington, 7300 Jack Newell Blvd. S., Fort Worth, TX 76118-7115, USA 682-365-0015 pmballal@gmail.com ** Energy Systems Research Center, University of Texas at Arlington, 416 S. College St., Arlington, Texas 76019-0048 *** Signal Processing Inc., 13619 Valley Oak Circle Rockville MD 20850-3572 1 1 978-1-4244-2622-5/09/$25.00 ©2009 IEEE. 2 IEEEAC paper #1515, Version 1, Updated October 5, 2008 Abstract—This paper has three contributions. First, we develop a low-cost test-bed for simulating bearing faults in a motor. In Aerospace applications, it is important that motor fault signatures are identified before a failure occurs. It is known that 40% of mechanical failures occur due to bearing faults. Bearing faults can be identified from the motor vibration signatures. Second, we develop a wireless sensor module for collection of vibration data from the test- bed. Wireless sensors have been used because of their advantages over wired sensors in remote sensing. Finally, we use a novel two-stage neural network to classify various bearing faults. The first stage neural network estimates the principal components using the Generalized Hebbian Algorithm (GHA). Principal Component Analysis is used to reduce the dimensionality of the data and to extract the fault features. The second stage neural network uses a supervised learning vector quantization network (SLVQ) utilizing a self organizing map approach. This stage is used to classify various fault modes. Neural networks have been used because of their flexibility in terms of online adaptive reformulation. At the end, we discuss the performance of the proposed classification method. TABLE OF CONTENTS 1. INTRODUCTION.................................................................1 2. THE TEST-BED .................................................................2 3. THE TWO-STAGE NEURAL NETWORK CLASSIFIER ........4 4. IMPLEMENTATION AND PERFORMANCE ANALYSIS ........5 5. CONCLUSIONS ..................................................................7 ACKNOWLEDGEMENTS ........................................................8 REFERENCES ........................................................................8 BIOGRAPHY ..........................................................................9 1. INTRODUCTION Failure avoidance is one of the main approaches for ensuring the quality and performance of a system. There are two main types of failure avoidance in terms of maintenance, namely preventive and corrective. In preventive maintenance, all actions are taken to keep equipments in good operating condition. It should be able to indicate when a failure may occur so that actions can be taken to avoid failures. In corrective maintenance, a repair is performed after a failure has occurred. Condition-based maintenance (CBM) is an approach of preventive maintenance. The process of CBM involves monitoring the system, predicting failures and making repairs before these failures occur. A system can contain many fault modes and a decision has to be taken on the type of repair necessary for eliminating any future faults [7] [34] [43]. Monitoring of the system is done using a range of sensors which can either be wired or wireless. Wireless sensors are generally used to enable remote monitoring. Wireless Sensor Networks (WSN) provide an intelligent platform to gather and analyze data without human intervention. Typically, a sensor network consists of autonomous wireless sensing nodes that are organized to form a network. Each node is equipped with sensors, embedded processing unit, short-range radio communication module, and power supply, which is typically 9-volt battery. With recent innovations in MEMS sensor technology, WSN hold significant promise in many application domains [42]. In this paper, we concentrate on CBM for induction motors. In Aerospace applications, it is important that motor fault signatures are identified before a failure occurs. Some of the common problems in induction motors are: bearing, stator winding, and rotor bar failures. Since the bearings carry the weight of the rotor, its fault diagnosis becomes very important. A multi-layered feed forward Neural Network (NN) trained with Error Back propagation technique and an unsupervised Adaptive Resonance Theory-2 (ART2) based NN has been used for detection and diagnosis of localized faults in ball bearings in [41]. In [25] Auto-Regression Model with NN has been used for fault