2005 Annual Report Conference on Electrical Insulation and Dielectric Phenomena Detection of Insulation Failure in BLDC Motors Using Neuro-Fuzzy Systems M. A. Awadallah' and M. M. Morcos2 'Department of Electric Power and Machines, University of Zagazig, Egypt 2Department of Electrical and Computer Engineering, Kansas State University, USA Abstract: The paper presents a method for automatic detection of insulation failure in the stator winding of brushless DC (BLDC) motors using adaptive neuro- fuzzy systems. Healthy performance of the motor is obtained under balanced conditions through a discrete- time numerical model. Motor parameters are modified due to insulation failure across a number of turns on one phase of the stator winding. The electromagnetic torque is selected as the characteristic waveform to identify and locate the fault. One index extracted from the signal by the discrete Fourier transform (DFI) could perfectly characterize the number of faulty turns, while two other indices derived using the short-time Fourier transform (STFTT) could signify the faulty phase. The diagnostic process is automated through two independent adaptive neuro-fuzzy inference systems (ANFIS) trained on simulated waveforms. Testing of ANFIS shows good performance in diagnosing and locating the fault. Experimental measurements of the output torque under normal and faulty operations show acceptable matching with simulated values, which verifies the analytical results and validates the proposed methodologies. Introduction The breakdown of winding insulation is one of the most detrimental faults hitting electric machines. The fault normally starts as confined insulation degradation with partial discharge and excessive local heat. The fault is stochastic in nature and initiates due to material contamination, manufacturing deficiencies, and/or excessive thermal, mechanical, electrical, or other environmental stresses. The fault propagates randomly in either axial or helical direction; propagation in both directions is possible despite a low probability of occurrence. If not early detected, insulation failure can develop into a whole phase-to-ground or phase-to-phase short circuit. Most of the research in the area of machine fault diagnosis has been dedicated to induction motors and their drive systems. On-line estimated negative- sequence impedance of the motor helped in characterizing the insulation failure in three-phase induction motors used in some mining industry applications [1]. Root-mean-square value of the filtered summation of the instantaneous line-to-neutral voltage waveforms could identify the fault in three-phase star- connected induction motors [2]. In induction machine drives, the standard deviation of RMS values of the machine line currents was also used to diagnose stator insulation failure [3]. Elsewhere, the authors have introduced a technique based on Fourier transform to identify insulation-failure of stator windings in brushless DC motors [4]. The present paper employs the capabilities of adaptive neuro-fuzzy systems as a promising intelligent paradigm in characterizing insulation failures on the stator winding of brushless DC (BLDC) motors. The 12 V, 1000 RPM, 6 pole, BLDC motor drive considered in the paper is used in some automotive application. The three-phase machine has a permanent-magnet rotor, and is equipped with Hall sensors for rotor-position detection. The machine is fed from a 12 V supply through a filter capacitor, two-quadrant chopper circuit, and three-phase MOSFET-based inverter bridge, Fig. 1. Switching logic of the inverter devices is determined through the rotor position estimated by the Hall sensors, while the chopper circuit applies hysteresis control on the DC-link current. Such switching pattern makes the system undergo six different switching states; each has two conducting MOSFETs and two active phases. The torque constant of the motor is 0.032 N.m/A. .! _IH _|_ InveneI Figure 1: Schematic diagram of the drive system. System performance under normal operation is computed through a discrete-time numerical model [5]. The insulation failure is assumed to initiate as a point- to-point short circuit between two adjacent turns, then propagate helically to isolate a number of turns in one stator phase. The net effect of the fault is, therefore, a reduction in the effective number of turns across one stator phase. Consequently, the back EMF constant of such phase is deviated under fault, as well as self inductance of the faulty phase and mutual inductances to healthy phases. 0-7803-9257-4/05/$20.00 ©2005 IEEE Mo i *1 18 Authorized licensed use limited to: UNIVERSITY OF ALBERTA. Downloaded on December 22, 2008 at 13:42 from IEEE Xplore. Restrictions apply.