Multi-parameter data–mechanism collaborative framework for remaining useful life prediction of wet friction components based on adaptive staged Wiener process

Abstract Wet friction components are a key part of heavy-duty transmission systems, and accurate remaining useful life (RUL) prediction is essential for equipment reliability. To overcome the limitations of existing methods in adaptive stage division, multi-source feature extraction, and data–mechanism fusion, this paper proposes a convolutional neural network (CNN)–bidirectional long short-term memory (BiLSTM)–attention–Wiener RUL prediction framework. Accelerated degradation tests are conducted to obtain temperature differences, thickness, and Fe/Cu concentrations. A dynamic-threshold adaptive membership function divides the degradation stages, while CNN-BiLSTM-Attention extracts multi-source coupled degradation features. A multi-stage Wiener process is then introduced to describe the stochastic degradation evolution and prediction uncertainty. In addition, the Wiener prediction error is used as feedback, and Adam optimizes the network parameters and composite health indicator mapping, achieving collaborative updating between the data-driven and mechanism-based models. The proposed method achieves a mean absolute error (MAE) of 1.162 h, with a confidence interval of 1.076–1.248 h, demonstrating high prediction accuracy, stability, and engineering applicability.