Fault classification for Rolling Element Bearing in Electric Machines

Abstract

Machines never break down by chance. Their operating condition gradually deteriorates and this gradual quantitative deterioration transforms into a failure, manifesting a qualitative change. When a fault takes place, some of the machine parameters are subjected to change. The change in the machine parameter depends on the degree of faults and the interaction with other parameters. In most cases, more than one parameter is subjected to change under abnormal conditions. If these machine parameters can be measured, they give information relating to a developing fault and then maintenance can be planned for the next available shutdown. Therefore, condition monitoring requires measurement to be taken from the machine on a continuous basis and is used to indicate the current working condition of that machine. There are a number of parameters that can be considered for monitoring the machine. For induction machines, monitoring parameters are terminal voltages, line currents, stack temperature of stator, rotor speed, torque, stator frame vibration, bearing vibrations, etc. [1-4] This paper discusses the results of the laboratory investigations carried out on 10-hp cage induction motors to identify the machine health by monitoring. A data acquisition system, capable of recording vibration, line currents and line voltages, is designed and fabricated. The signals are sampled at a suitable sampling frequency using an add-on card and are saved in a computer for future use. The vibration signals are measured at a load end bearing of the test motor. The recorded vibration signals are then pre-processed both in time and frequency domain to extract the important features that contain the maximum fault information. The selected features are analysed for their class separation capability using the Fisher criterion. These features are then used as an input to the fault classifier. The fault classifier used in this study is the multi-layered perception trained with a back propagation algorithm. Two networks are trained independently, one with time domain parameters, module-TD, and the other with frequency domain parameters, module-FD. It is found that module-TD gives its best performance for incipient fault detection (100%); however, for medium degree fault, it gives an 85% efficiency. The frequency parameter-based module gives a better performance for both the cases (98 and 95% respectively). However, a combination of the two can be used to enhance the decision ability of the system.