Faults diagnoses of rotating machines by using neural nets: GRNN and BPN

Abstract

Rotating machines such as compressors, fans, and motors are the most important objects in plant maintenance. Like the finger print or the voice print of a human, each abnormal vibration has its own characteristic feature in its power spectrum. We make feature vectors from the power spectra of vibration signals, and applied them as inputs to the neural nets. The general regression neural network (GRNN) has several advantages over the backpropagation network (BPN) such as very short training time (one-pass learning) and guaranteed performance even with sparse data. Further one can easily modify or upgrade GRNN according to the specific needs of the machine conditions or environments. We compared the performances of GRNN versus BPN using the same feature vectors made from a vibration test bench. The experimental results show us that GRNN outperforms BPN.