Adversarial autoencoding Crossformer model with transfer learning for remaining service life prediction of rolling bearings

Abstract The new encoder-decoder model, namely the Crossformer, has inherent limitations such as poor noise interference resistance, weak generalization performance, and difficulty in model training. These problems result in suboptimal prediction accuracy and undesirable training cost when the Crossformer is used for predicting the remaining service life (RSL) of rolling bearings. To this end, a novel codec named adversarial autoencoding (AAE) Crossformer model with transfer learning (AAET-Crossformer Model) is proposed and applied in RSL prediction of rolling bearings for the first time. First, an AAE denoising mechanism is constructed for the Crossformer to adaptively extract the deep features reflecting the state decay process of rolling bearings, thus enhancing its nonlinear approximation ability. Moreover, this mechanism can adapt to the input form of the Crossformer and enable the Crossformer to catch the cross-dimensional and cross-temporal dependencies of multivariate time series (MTS) more effectively, thereby enhancing the MTS trend prediction accuracy. Then, a dynamic domain adaptation network is constructed in the encoding layer of the Crossformer, which can use the historical state decay features of different rolling bearings under different operating conditions to generalize the Crossformer’s new task of predicting the state decay feature trend of the current rolling bearing under current operating condition. Finally, a policy gradient-based parameter learning algorithm is developed to accelerate the training speed of AAET-Crossformer model parameters. Due to the above progressiveness of AAET-Crossformer model, the constructed AAET-Crossformer model-based RSL prediction method for rolling bearings can achieve superior RSL prediction accuracy, better generalization performance, lower training cost, and considerable computation efficiency compared to the RSL prediction method based on the Crossformer. The effectiveness and superiority of the constructed RSL prediction method are demonstrated by two experimental verification cases of RSL prediction of rolling bearings.