1993 Volume 29 Issue 1 Pages 94-101
Bearings are the components of rotary machinery, which are easiest to fall into faults. This paper describes how to identify standard numbers of the ball bearings and diagnose their failure modes from scream when they are rotating.
We apply techniques of the neural networks to the identification and the diagnosis. Here we present an architecture of the hierarchically structured multi-layer neural networks. The tasks of identification of the number of bearing and diagnosis of the failure mode are allotted to each hierarchical stages, by which the efficiency of learning and recognition by the network can be improved.
The standard number of a bearing is identified by directly feeding the line power spectrum under the frequency normalized by the revolution frequency to the first stage of the network. The revolution frequency is estimated from the sound of scream. Failure is diagnosed by the networks in the second stage by the information of the standard number of ball bearing, the repetitive frequency of the spectrum and the stationary or nonstationary of the scream.
The ability of associative memory of the neural network can reduce the effect of noise and error in the signal.
The identification and diagnosis of bearings were carried out both by simulation and experiment. We found the hierarchical neural network proposed could effectively identify the standard number and diagnose the faults.