Enhanced genetic algorithm-based back propagation neural network to diagnose conditions of multiple-bearing system

Abstract

Condition diagnosis of critical system such as multiple-bearing system is one of the most important maintenance activities in industry because it is essential that faults are detected early before the performance of the whole system is affected. Currently, the most significant issues in condition diagnosis are how to improve accuracy and stability of accuracy, as well as lessen the complexity of the diagnosis which would reduce processing time. Researchers have developed diagnosis techniques based on metaheuristic, specifically, Back Propagation Neural Network (BPNN) for single bearing system and small numbers of condition classes. However, they are not directly applicable or effective for multiple-bearing system because the diagnosis accuracy achieved is unsatisfactory. Therefore, this research proposed hybrid techniques to improve the performance of BPNN in terms of accuracy and stability of accuracy by using Adaptive Genetic Algorithm and Back Propagation Neural Network (AGA-BPNN), and multiple BPNN with AGA-BPNN (mBPNNAGA- BPNN). These techniques are tested and validated on vibration signal data of multiple-bearing system. Experimental results showed the proposed techniques outperformed the BPPN in condition diagnosis. However, the large number of features from multiple-bearing system has affected the complexity of AGA-BPNN and mBPNN-AGA-BPNN, and significantly increased the amount of required processing time. Thus to investigate further, whether the number of features required can be reduced without compromising the diagnosis accuracy and stability, Grey Relational Analysis (GRA) was applied to determine the most dominant features in reducing the complexity of the diagnosis techniques. The experimental results showed that the hybrid of GRA and mBPNN-AGA-BPNN achieved accuracies of 99% for training, 100% for validation and 100% for testing. Besides that, the performance of the proposed hybrid accuracy increased by 11.9%, 13.5% and 11.9% in training, validation and testing respectively when compared to the standard BPNN. This hybrid has lessened the complexity which reduced nearly 55.96% of processing time. Furthermore, the hybrid has improved the stability of the accuracy whereby the differences in accuracy between the maximum and minimum values were 0.2%, 0% and 0% for training, validation and testing respectively. Hence, it can be concluded that the proposed diagnosis techniques have improved the accuracy and stability of accuracy within the minimum complexity and significantly reduced processing time.