The experimental study of the expert system for diagnosing unbalances by ANN and acoustic signals

Abstract

The expert system based on the backward propagation neural network (BPN) has been developed and tested for diagnosing mass unbalance of rotational machines. The system adopts the acoustic signals as input features. In order to minimize the distance and background noise effects, the so-called d-normalization was introduced. The d-normalization is similar to the loudness in speech synthesis. By utilizing the normalized power spectra together with the rectified statistic moments of higher order of the acoustic signals, the system is found to be very successful. However, it was found that the system still could not discriminate those faults near the natural frequencies. The main reason may stem from the system non-linearities even though they are small.