| To improve the reliability of motor-driven processes, condition monitoring of electric machines has received considerable attention from industry. For small- and medium-sized machines, the focus is on low-cost sensorless schemes that use only measured voltages and currents for fault diagnostics.; Turn faults arising from stator winding insulation deterioration account for a large percentage of motor failures. The objective of a turn-fault detection scheme is to provide a warning before the fault propagates further and results in ground current, causing irreversible damage to the magnetic material. In this work, a neural-network-based robust scheme for early detection of turn faults in induction machines is developed. The negative-sequence component of line currents is used as the fault signature, and a neural network is trained to compensate for the effects of unbalanced supply voltages and nonidealities in the machine or instrumentation. Novel training algorithms for self-commissioning and on-line training of the neural network have also been developed. Experimental results, obtained on a specially-rewound machine, are provided to demonstrate that the method is capable of early fault detection. Data memory and computational requirements are also minimal, making the scheme viable for commercial implementation. The method is also extended to turn-fault detection in open-loop inverter-fed induction machines. Data obtained from a thermally accelerated insulation failure experiment is also used to test the performance and sensitivity of the method, and to show that a turn fault can be detected before failure of insulation to ground. |
