A dual-regressor adversarial framework with staged-wise training strategy for cross-domain remaining useful life prediction

Abstract Traditional cross-domain remaining useful life (RUL) prediction methods mitigate distribution discrepancies and label scarcity in machinery data through global domain alignment, which leads to fine-grained distribution discrepancies being insufficiently captured. To this end, a novel method based on a dual-regressor architecture is proposed for cross-domain RUL prediction, which incorporates a stage-wise training strategy to enhance adaptation performance. The dual-regressor architecture is designed to measure the prediction discrepancies and identify samples with significant domain shifts in the target domain. Meanwhile, cosine similarity is integrated to quantify the prediction discrepancies between regressors for enhancing the training stability. During stage-wise training, an adversarial mechanism between the regressors and feature generator is formed to force the generator to learn degradation-sensitive and domain-invariant features, thereby achieving fine-grained feature alignment. To further alleviate the distribution shift between the source and target domains, the maximum mean discrepancy with multiple kernels is incorporated into the model to work in conjunction with the adversarial mechanism. Finally, experimental results on the CMAPSS dataset demonstrate that the proposed method outperforms the second-best method source-free domain adaption by 12.07% and 43.31% on the evaluation metrics, highlighting its superior cross-domain prediction capability.