A deep convolutional neural networks model for intelligent fault diagnosis of a gearbox under different operational conditions

Abstract

For intelligent fault diagnosis of a gearbox using deep convolutional neural networks (DCNNs), we performed a gearbox vibration experiment. To understand the effects of different operational conditions and a cumulative degradation of the operational process of the gearbox, we collected the vertical and horizontal vibration signals of five different degradation states under three different operational conditions (rotational speeds of 1000r/min, 1050r/min and 1090r/min). Using feature extraction methods, such as time-domain analysis, frequency-domain analysis, and wavelet packet decomposition, a feature space with 52 features was constructed. The dimensionality of the extracted features was reduced to 4 by principal component analysis (PCA), with contribution rates of 51.55%, 20.45%, 12.33%, and 5.99%, respectively. To verify the superiority of DCNNs, the performance was compared to a support vector machine (SVM) classifier, and the hyper-parameters of the SVM classifier were optimized using a grid search technique. Results show that the vertical vibration signals are correlated with the degradation of the gear, and the identification accuracy is increased by imposing a certain load. DCNNs have been shown to achieve a higher accuracy than the SVM classifier, indicating that DCNNs are a more suitable method for solving a multi-state fault identification problem. Additionally, by inputting the raw signals directly, the gearbox intelligent fault diagnosis, based on DCNNs, has also achieved a higher accuracy with a lower computational time cost.