Wear indicator construction for rolling bearings based on an enhanced and unsupervised stacked auto-encoder

• Published in: Soft computing (Berlin, Germany) Vol. 28; no. 15-16; pp. 8835 - 8848
• Main Authors: Zeng, Wenhui, Yu, Lisha, Xu, Fan, Huang, Zhelin, Zhou, Shengwen, Guo, Shunsheng, Du, Baigang
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.08.2024; Springer Nature B.V
• Subjects: Application of Soft Computing; Artificial Intelligence; Artificial neural networks; Bearings; Coders; Computational Intelligence; Construction; Control; Deep learning; Depth indicators; Engineering; Engineers; Fault diagnosis; Frequency domain analysis; Indicators; Kurtosis; Labeling; Labels; Life prediction; Machine learning; Mathematical Logic and Foundations; Mechanical systems; Mechatronics; Neural networks; Physical work; Robotics; Roller bearings; Smoothness; Time-frequency analysis; Unsupervised learning; Useful life; Deep learning; Stacked auto-encoder; Wear indicator; Sine function; Roller bearings
• DOI: 10.1007/s00500-023-09068-x
• ISSN: 1432-7643

The degradation state of a bearing can be monitored effectively by using a wear indicator (WI). A WI curve having smoothness and monotonicity can lay a good foundation for predicting the remaining useful life of the bearing. Most traditional models for bearing WI construction, such as time–frequency indicators and signal decomposition, are complicated; for example, some WI construction models need several models to select from, and their fusion depends on the manual experience of engineers. For example, single and mixed traditional time–frequency indicators, such as the root mean square (RMS), Kurtosis, multiple time–frequency domain fusion. However, the mentioned-above time–frequency domain indicators are difficult to adaptively reflect the operating status of the equipment when the operating conditions of the mechanical system change. Some signal decomposition models are combined with other models and rely on manual experience to extract WI, such as the selection of effective intrinsic mode function components function, parameter setting of empirical mode decomposition and ensemble empirical mode decomposition model, etc. Deep learning models, such as stacked auto encoder (SAE) and convolution neural network, have been widely used in bearing health monitoring and WI construction, because its powerful learning and feature extraction capabilities of multiple hidden layer structures. But these deep learning models are designed to use output labels. Particularly when the data volume is large, it requires manpower, material resources, and experienced engineers to label the data, or it is difficult to label and distinguish the categories of data samples. Therefore, to solve these problems and eliminate the need for manual labor, such as labeling data and selecting models for fusion, we propose using SAE without an output label layer to extract WI from original signals directly. However, an extracted WI curve without good monotonicity (Mon) will result in a poor remaining useful life prediction accuracy. To improve the monotonicity of the extracted WI and reduce the complexity of the WI construction model, we propose an unsupervised enhanced SAE without an output layer, named SINSAE, by adding a sine function of an average value which is calculated form start time to current at each hidden layer to eliminate concussion. Moreover, to demonstrate that our proposed model is better than other models, such as the RMS, Kurtosis, multiple time–frequency domain fusion, SAE, SAE without an output layer, and signal decomposition models, the Mon indicator in this study is used to compare the monotonicity of the extracted WI. Lastly, the results of our experiments using different bearing datasets and various working conditions show that the smoothness and monotonicity of the WI curve extracted by the SINSAE is better than that of other models. Moreover, compared to the traditional commonly used single and multiple time–frequency domain indicators, supervised deep learning and basic unsupervised deep models, the unsupervised SINSAE model can increase the Mon indicators from [0.1, 0.8], [0.02, 0.1], [0.1, 0.8] to above 0.9, respectively.