The monitoring of vibrations of large rotating machinery can give an indication of the machine's condition and aid in preventive maintenance. To simulate this in a more controllable environment, vibrations were measured on a small Bently-Nevada machine consisting of a frame, an electromotor with adjustable speed, a shaft and a flywheel. Two faults were introduced: shaft rub was applied using a vertical screw and shaft whirl was introduced by the placement of small weights on the flywheel. Horizontal(x) and vertical(y) vibrations were measured using two accelerometers at a variety different motor speeds from 70 Hz to 100 Hz. These vibration time series indicate the displacement of the shaft centre and can be used to plot its orbit. The objective of this work was to train an artificial neural network to classify the machine condition from the time series data. This represents an alternative approach to what was presented in COMADEM-94.

2 EXTRACTION OF FEATURES

Selection of appropriate features is essential if a neural network is to be of any use. The orbit plots of the four different conditions have distinctive characteristics:

• (NN) No faults applied: the shaft displacement is small.

• (NR) Only rub is applied: the shaft orbit follows a highly irregular path.

• (WN) Only a Weight is added: the shaft orbit is an approximate circle.

• (WR) Both rub and the weight applied: the shaft follows an irregular path, but the average displacement from the centre is larger than without the weight.

The horizontal and vertical shaft positions can be combined to give z(t) = x(t) + jy(t) as the complex time series describing the motion of the shaft centre. The mean value of |z(t)| gives an indication the weight being present. The variance of |z(t)| gives some indication of the presence of rub. Higher order central moments also contain information and the usefulness of this was investigated.

The orbit followed by the WN condition is approximately circular, therefore dz(t)/dt, will also be app roximately circular. When rub is present, the shaft centre follows a far more complicated path, the speed of this motion varying widely. Consequently, the probability density function of |dz(t)/dt| for a rub condition is flat and well spread out whereas the p. d. f. of |dz(t)/dt| with no rub has a much sharper peak By integrating z(t), a smoothed orbit is obtained. This attenuates the higher frequency components of z(t) which caused some NR and WR orbit plots to be similar in appearance.

3 Summary of Results.

The success rates for classifying the conditions using the different time series for all the motor speeds are shown in the table below.

4 Conclusions

The moments of the magnitude of the complex time series of machine vibrations provide useful features which characterise the orbit of the shaft centre and which are sufficiently distinct that an artificial neural network can classify them with reasonable success. The use of the derivative and integral of the time series enhances high and how frequency components in the time series respectively, aiding the classification process to produce a very reliable system. Details of the structure of the ANN and training features will be presented in the full paper.

Times Series Used	Conditions Classified	Success Rate
|z(t )|	All	80%
|dz(t)/dt|	All	70%
|z(t) |	Weight-No Weight	96%
|dz(t)/dt|	Rub-No Rub	92%
|z(t) | and |dz(t)/dt|	All	91%
|z(t)dt|	Weight-No Weight	100%
|z(t) | and |dz(t)/dt |and |z(t)dt|	All	99%