Remaining useful life prediction method for rolling bearings based on hybrid dilated convolution transfer

Abstract It is difficult to effectively predict remaining useful life (RUL) due to limited training samples and lack of life labels in some operating conditions of practical engineering. When existing deep learning methods predict the RUL of equipment in such operating conditions using a model trained on other operating conditions, the poor generalization of the model caused by large distribution differences cannot be ignored. In this study, an RUL prediction method based on integrated dilated convolution transfer is proposed. This method jointly adjusts the model parameters by inverting the loss function of the RUL prediction module and the domain adaptive module, and then realizes the extraction of domain‐invariant features between different operating condition data through the feature extraction module, which provides support for transfer RUL prediction between different operating conditions. In the feature extraction module, a one‐dimensional convolution network with a large‐size kernel reduces noise in the original data, which reduces the erroneous effect of noise on the trending expression of the original data, and a hybrid dilated convolution network extracts the features of the different sensory fields of the noise‐reduced data, which increases the richness of the extracted features and thus improves the accuracy of the modeling. Next, the extracted features are fed into the RUL prediction module to predict RUL; into the classification model in the domain adaptation module to divide the source and target domains; and into the distribution difference measurement model in the domain adaptation module to identify the feature distribution differences between the source and target domains, and inversely adjust the model parameters by reducing the distribution differences. Furthermore, domain invariant characteristics of the features in different receptive fields under multiple operating conditions are obtained to enhance the model's generalization ability and achieve RUL prediction across various operating conditions. Monte Carlo (MC) dropout simulation technology is used to quantify the uncertainty of prediction results. Finally, the effectiveness and superiority of the proposed method are verified using the prognostics and health management (PHM) 2012 bearing dataset.