An intelligent data-driven adaptive health state assessment approach for rolling bearings under single and multiple working conditions

Abstract

The health state assessment of critical rotating components (e.g., rolling bearings) is vital to supporting lifecycle health management in complex systems. Traditional approaches that are pivotal for such assessments usually require prior knowledge and manual intervention. Additionally, the challenge amplifies when considering assessments across varying operating conditions. Nowadays, intelligent data-driven approaches have been becoming increasingly prevalent. This paper presents a novel data-driven approach to realize adaptive health assessment in complex and multiple operational conditions. The approach consists of several functions. First, automatic health state segmentation employed deep feature extraction by variational auto-encoder (VAE), feature filtering, and unsupervised clustering. Furthermore, to enable adaptive feature extraction under multiple operating conditions, an improved classification variational auto-encoder with domain confusion (CVAEDC) model was designed to adapt cross-domain feature representation. Finally, a long short-term memory neural network with reinforcement learning to optimize the network hyperparameters (RL-LSTM) was developed for health assessment under single and multiple operating conditions. The proposed approach was validated using the 2012 IEEE Challenge dataset and the XJTU-SY dataset. In the former dataset, the assessment accuracies in the given operational condition and target operating condition reached over 92.4% and 85.9%. In the latter dataset, the accuracies reached 95.8% and 87.6%. Experimental results on the two datasets demonstrated the effectiveness and robustness of the approach, notably reflected in the root mean square (RMS) curve comparisons. In summary, the approach provides an effective solution to support practical health assessments for critical rotating components in varying operational environments.