Remaining useful life prediction for bearing based on small-sample online degradation point detection and intelligent optimization transfer

Abstract

Transfer learning plays a crucial role in remaining useful life (RUL) prediction for bearings under complex operating conditions and small-sample scenarios. However, existing methods still face challenges such as unstable initial degradation point (IDP) detection, negative transfer caused by feature distribution discrepancies between the source and target domains, and difficulty in selecting appropriate hyperparameters. To address these issues, this paper proposes a bearing RUL prediction method integrating online IDP detection with intelligent optimization transfer learning. First, an adaptive continuous exceedance method is developed to identify the IDP, reducing the interference from healthy-stage data. Second, based on the identified degradation information, a transfer prediction network combining multi-layer bidirectional long short-term memory and a convolutional residual network is constructed. It captures multi-scale temporal dependencies and enhances cross-domain feature transferability. In addition, a golden sine strategy and restart mechanism are incorporated into an intelligent optimization algorithm to optimize network hyperparameters, improving convergence efficiency and prediction performance. Experiments on the IEEE PHM2012 and XJTU-SY datasets validate the effectiveness of the proposed method. Under the small-sample cross-condition transfer setting, the proposed method achieves a minimum mean absolute error of 0.028 and root mean square error of 0.060 after hyperparameter optimization, demonstrating its effectiveness and generalization capability with limited target-domain run-to-failure data.