A novel transfer learning approach based on deep degradation feature adaptive alignment for remaining useful life prediction with multi-condition data

Transfer learning (TL) plays an important role in the remaining useful life (RUL) prediction when the training data and testing data are collected under different operating conditions. However, the existing studies have two problems: (1) Only using the single-condition data as the source domain may encounter negative transfer, especially when the operating conditions in the training and actual usage are vastly different. (2) Traditional domain adaptation methods only reduce the discrepancy of global feature distributions of source and target, and ignore the impact of local features. To tackle these problems, this paper proposes a novel TL approach based on deep degradation feature adaptive alignment, which uses multi-condition degradation datasets as the source domains and forms multiple domain pairs with the target data. A network framework with multiple parallel sub-networks is designed to extract the degradation features of all domain pairs, and a deep degradation feature adaptive alignment mechanism is developed that can minimize marginal and conditional distribution discrepancies and adaptively adjust their calculation proportions to align the global and local features of each domain pair. In the experiment, the RUL prediction performance is verified by using the turbofan engine dataset, and its advantages are validated by comparisons with other methods.