Differential evolution optimization for resilient stacked sparse autoencoder and its applications on bearing fault diagnosis

Abstract

The rolling element bearing is an important component in most rotating machinery. The unexpected failure of a bearing may cause the whole mechanism to break down. Hence, research has focused on developing effective intelligent fault diagnosis to generate more accurate and robust diagnostic results. Bearing fault diagnosis based on stacked sparse autoencoder (SSAE) architecture is proposed in this study. SSAE is capable of providing a featureless methodology for bearing fault diagnosis. However, the architecture of SSAE is greatly influenced by its hyperparameter settings and there is no standard method of determining the optimal hyperparameter values. In addition, the standard learning algorithm used in SSAE architecture is time-intensive. In this paper, a method that combines differential evolution and a resilient back-propagation approach is proposed to improve the performance of SSAE networks in bearing fault classification. The differential evolution approach optimised SSAEs hyperparameters such as the hidden nodes number, weight decay parameter, sparsity parameter, and weight of the sparsity penalty term, that are associated with each hidden layer of SSAE networks. An increase in the hidden layers of SSAE will further complicate the hyperparameter selection process. The resilient back-propagation training algorithm is used to train the SSAE network due to its low computation cost. Results from analysis of three databases demonstrate that the proposed model achieved 99% performance accuracy in bearing fault diagnosis. The proposed model is found to be more user-friendly and effective in handling multi-condition of bearing faults compared to the original autoencoder.

A Novel Method for Early Gear Pitting Fault Diagnosis Using Stacked SAE and GBRBM

“…Sparse autoencoder (SAE) [ 33 , 34 ] is an unsupervised learning network mainly used for data dimensionality reduction and feature extraction. The SAE includes three layers: input layer ( n + 1 neuron), hidden layer ( m + 1 neuron, m < n ), and output layer ( n neurons).…”

A Novel Order Analysis and Stacked Sparse Auto-Encoder Feature Learning Method for Milling Tool Wear Condition Monitoring

“…As the difference between ρ andρ increases, the value of the penalty factor will rise sharply [30]. Furthermore, to avoid overfitting, a weight decay term J Weight is applied:…”

Low-Speed Bearing Fault Diagnosis Based on ArSSAE Model Using Acoustic Emission and Vibration Signals

“…Based on studies reviewed by Ab Wahab et al, DE is among the best optimization methods [43]. The details of the DE and resilient algorithm implementation in the SSAE network can be referred to the following works [28]. To solve the hidden layer problem discussed in the preceding section, we developed a stacking layer of SAEs depending on the feature size as an initial reference configuration.…”

“…In addition, there is no standard procedure to determine the number of hidden nodes and layers in an SSAE network. At present, the SSAE's hyperparameter and hidden node number has been successfully optimized using metaheuristic algorithm as mentioned in the following research [28]. Wang et al mentioned the effects of autoencoder hidden layer numbers on the model performance [29].…”