Experimental prediction model for full life friction performance of wet clutch via attention-based LSTM network

Effective performance prediction under different working conditions is crucial for the reliability evaluation and health management of wet multi-disc clutches throughout the service life. In light of the difficulty in data acquisition and dynamic assessment of degradation status within the clutch life cycle, a machine learning-based service period performance prediction method that does not rely on offline data is proposed. Firstly, the clutch lifecycle experiments are designed and conducted under different working conditions to verify the phased decline in friction performance, where changes in friction surface roughness lead to the continuous deterioration of tribological behavior. Then, based on the extracted average coefficient of friction (COF), engagement time, peak torque attenuation coefficient (K _torq), and chaotic feature parameters including correlation dimension (D₂) and standard deviation of distance matrix (SD-DM), the attention-based deep learning model (Att-LSTM) is established to predict the performance degradation. Finally, the predictive performance of the proposed model is validated by using different historical data volumes. The results show that the proposed Att-LSTM model has an average MAPE of 7.70% and 6.54% on a limited historical dataset of 40% and 60%, respectively, demonstrating superior accuracy. This work is conducive to promoting the health management and conditional maintenance for clutch system during the on-duty phase.