Condition Diagnosis of Bearing System Using Multiple Classifiers of ANNs and Adaptive Probabilities in Genetic Algorithms

Condition diagnosis in bearing systems needs an effective and precise method to avoid unacceptable consequences from total system failure. Artificial Neural Networks (ANNs) are one of the most popular methods for classification in condition diagnosis of bearing systems. Regarding to ANNs performance, ANNs parameters have important role especially connectivity weights. In several running of learning processes with the same structure of ANNs, we can obtain different accuracy significantly since initial weights are selected randomly. Therefore, finding the best weights in learning process is an important task for obtaining good performance of ANNs. Previous researchers have proposed some methods to get the best weights such as simple average and majority voting. However, these methods have some limitations in providing the best weights especially in condition diagnosis of bearing systems. In this paper, we propose a hybrid technique of multiple classifier-ANNs (mANNs) and adaptive probabilities in genetic algorithms (APGAs) to obtain the best weights of ANNs in order to increase the classification performance of ANNs in condition diagnosis of bearing systems. The mANNs are used to provide several best initial weights which are optimized by APGAs. The set optimized weights from APGAs, afterward, are used as the best weights for condition diagnosis. Our experiment shows mANNs-APGAs give better results than of the simple average and majority voting in condition diagnosis of bearing systems. This experiment also shows the distinction of maximum and minimum accuracy in mANNs-APGAs is lower than the two existing methods.