Rotating machinery misalignment detection and prediction using motor power spectrum

This thesis explores methods to detect and predict rotating machinery misalignment using power frequency spectrum data obtained at the Oak Ridge National Laboratory Advanced Motor Testing Facility for a 3 phase 50 horsepower electric motor connected to 150 horsepower dynamometer. Three types of flexible couplings were used to create data for controlled alignment conditions observations. An elastomer type flexible coupling, a gear type flexible coupling, and a grid type flexible coupling are used to show that the alignment conditions of a motor system can be predicted from the input power frequency spectrum of an electric motor. Five types of alignment prediction methods are used and they all performed differently for the three types of couplings.

Some of the differences in the mapping procedures are due to the way that the method performs the mapping. The Multiple Linear Regression method simply fits a straight line between the inputs and outputs, so its prediction results will be highly dependent on changes in the motor system and the amount of linearity between the inputs and outputs. The Linear Correlation Coefficient (LCC) Neural Network uses an Artificial Neural Network (ANN) to train on the most highly correlated frequencies to the alignment condition to allow the non-linearities of those frequencies to be fitted. The Principal Component Analysis (PCA) method removes the collinearity from the data by mapping the data to an orthogonal space based on variability of the input data and then trains an ANN based on the important principal components (PC's). The Linear Partial Least Squares (LPLS) method maps the relationship linearly based on PC's that are determined in a supervised manner from the variability of the input and its relationship to the outputs. The Non-Linear Partial Least Squares (NLPLS) method uses an ANN to perform the mapping between the supervised PC's and the alignment condition.

The methods that performed the best for each of the couplings and the alignment conditions were as follows. The method that produced the smallest mean error for the elastomer type coupling to predict the parallel offset misalignment was the LCC Neural Network mapping with an average prediction error of 11.87 mils compared to 25.15 mils for a MLR. The best method for predicting the angular misalignment in the elastomer type coupling was the LPLS method with an average prediction error of 2.92 mils/inch compared to the MLR error of 2.99 mils/inch.

The method that produced the smallest mean error for the gear type coupling to predict the parallel offset misalignment was the PCA Neural Network mapping with an average prediction error of 2.71 mils compared to 7.34 mils for a MLR. The best method for predicting the angular misalignment in the gear type coupling was the LCC Neural Network method with an average prediction error of 0.13 mils/inch compared to the MLR error of 3.97 mils/inch.

The method that produced the smallest mean error for the grid type coupling to predict the parallel offset misalignment was the NLPLS mapping with an average prediction error of 1.59 mils compared to 3.04 mils for a MLR. The best method for predicting the angular misalignment in the grid type coupling was the LCC Neural Network method with an average prediction error of 0.31 mils/inch compared to the MLR error 0.93 mils/inch.

An online system can be implemented by placing sensors on two of the three phase voltages and currents to determine the total input power to a system. The system would then have to be trained on the operating parameters for each system that it is placed on to be able to more accurately predict the alignment conditions. A combination of these methods could be used as the predictor function, but the most robust method would use the NLPLS techniques with more independent testing and training.