Condition diagnosis of multiple bearings using adaptive operator probabilities in genetic algorithms and back propagation neural networks

Condition diagnosis of bearings is one of the most common plant maintenance activities in manufacturing industries. It is essential to detect bearing faults early to avoid unexpected breakdown of plant due to undetected faulty bearings. Many meta-heuristics techniques for condition diagnosis of single bearing systems have been developed. The techniques, however, are not effectively applicable for multiple bearing systems. In this paper, a new hybrid technique of genetic algorithms (GAs) with adaptive operator probabilities (AGAs) and back propagation neural networks (BPNNs), called AGAs-BPNNs, is proposed specifically for condition diagnosis of multiple bearing systems. In this technique, AGAs are integrated with BPNNs to attain better initial weights for the BPNNs and hence reduce their learning time. We tested the proposed technique on a two bearing systems, and used ten extracted features from the system's vibration signals data as input and sixteen bearing condition classes as target output. The experimental results show that the AGAs-BPNNs technique obtains much higher classification accuracy in shorter CPU time and number of iterations compared with the standard BPNNs, and the hybrid of standard GAs and BPNNs.