A methodology for predicting motor misalignment using artificial neural networks

A methodology to predict the misalignment condition of a motor-driven dynamometer using principal component analysis and artificial neural networks is presented.

Vibration data was experimentally obtained and archived from two separate test sites, the University of Tennessee (UTK), and Oak Ridge National Laboratory (ORNL). At both test sites, the respective motor was moved into various misalignment conditions for which two minutes of vibration data was taken. Four different types of flexible couplings were used to connect the motor to the. Frequency spectra of the motor vibration signatures were reduced to their principal components via singular value decomposition and principal component analysis. The principal components extracted from the frequency spectra were then used to train a pair of artificial neural networks to predict the associated angular and offset motor misalignment conditions.

Results from applying this prediction methodology vary' widely. A modal analysis done of the UTK motor-dynamometer set-up indicates that vibration data taken at the UTK test site may not be able to allow for misalignment prediction, and further analysis of data taken from the UTK test site supports this conclusion. When the vibration data obtained at the ORNL test site is used to predict motor misalignment, results are mixed. Certain misalignment conditions from certain couplings could be predicted while others did not.

The methodology presented in this thesis can used to predict the misalignment condition of a motor driven dynamometer. However, a couple important aspects must be considered. First, the motor-system must behave in a well-characterized fashion as determined by modal analysis. Secondly, the type of flexible coupling used to connect the motor to the driven machine has a significant bearing on how well the misalignment condition of the motor can be ascertained. And lastly, advanced signal analysis routines need to be used to properly convert vibration signatures to frequency spectra.