IEEE TRANSACTIONS ON POWER ELECTRONICS, VOL. 17, NO. 6, NOVEMBER 20021089Self-Commissioning Training Algorithms for NeuralNetworks With Applications to Electric MachineFault DiagnosticsRangarajan M. Tallam, Member, IEEE, Thomas G. Habetler, Fellow, IEEE, and Ronald G. Harley, Fellow, IEEEAbstract—Themainlimitationsofneuralnetwork(NN)methodsforfaultdiagnosticsapplications aretraining dataand data memory requirements, and computational complexity.Generally, a NN is trained off-line with all the data obtained priorto commissioning, which is not possible in a practical situation. Inthis paper, three novel and self-commissioning training algorithmsare proposed for on-line training of a feed-forward NN to effec-tively address the aforesaid shortcomings. Experimental resultsare provided for an induction machine stator winding turn-faultdetection scheme, to illustrate the feasibility of the proposedon-line training algorithms for implementation in a commercialproduct.Index Terms—Algorithms, neural network, turn-fault.I. INTRODUCTIONTHE RECENT trend in electric machine condition mon-itoring is toward sensorless schemes that use only themeasured voltages and currents to extract fault information.Such methods rely on mathematical models of a machine toobtain fault signatures, which are characteristic of a particulartype of fault. Due to the effects of nonidealities in the machine,instrumentation or power supply, and modeling uncertaintiesor parameter errors, model-based methods require high faultthreshold settings and are also prone to false alarms . To de-tect a fault at an incipient stage of development, it is necessaryto compensate for nonideal effects.The negative-sequence component of line currents is themost commonly used fault signature for the detection of statorwinding turn faults. However, inherent asymmetries in themachine or instrumentation, and unbalanced supply voltagesalso affect the negative-sequence current component. Thenegative-sequence voltage component can be expressed interms of the current sequence-components as follows, whereis the slip(1)Manuscript received December 1, 2001; revised May 27, 2002. Recom-mended by Associate Editor A. M. Trzynadlowski.R. M. Tallam was with the School of Electrical and Computer Engineering,Georgia Institute of Technology, Atlanta, GA 30332-0250 USA. He is now withthe Advanced Technology Laboratories, Rockwell Automation, Milwaukee, WI53204 USA (e-mail: [email protected]).T. G. Habetler and R. G. Harley are with the School of Electrical and Com-puter Engineering, Georgia Institute of Technology, Atlanta, GA 30332-0250USA (e-mail: [email protected]; [email protected]).Digital Object Identifier 10.1109/TPEL.2002.805611The effect of asymmetries is expressed by the hybridimpedance term, which is zero for an ideal machine.