Multisource dual unsupervised domain adaptation method for remaining life prediction under multiple operating conditions

Aero engines, with their complex structures and intricate gas paths, are a crucial component of the aviation system. Traditional machine learning-based methods for predicting the remaining useful life (RUL) of rolling bearings often struggle to provide adequate feature representation and adaptive extraction due to complex operating conditions and challenges in obtaining degradation data. Given that traditional machine learning methods require substantial amounts of labeled data, they frequently fail to yield satisfactory RUL prediction results. To overcome these limitations, domain adaptation methods have been employed, facilitating the transfer of knowledge from a well-labeled source domain to an unlabeled target domain. Additionally, to augment the available data, a multisource domain transfer approach is proposed, which transfers data from multiple source domains to the target domain intended for prediction. To enhance the effectiveness of this multisource domain transfer method, a dual alignment strategy is introduced for aligning features at both the feature and semantic levels, thereby improving predictive performance. The model incorporates a convolutional neural network (CNN) and long short-term memory (LSTM) as its backbone, leveraging CNN's robust feature extraction capabilities combined with LSTM’s proficiency in processing time-dependent data. Experiments conducted on the CMAPSS dataset, along with comparisons and ablation studies against current popular multisource domain adaptation methods, demonstrate the superiority of the proposed approach.