Abstract: To address the core bottleneck that the prediction accuracy in gear transmission fatigue life and reliability assessment is constrained by small-sample data, this study overcomes the limitations of conventional approaches by developing a deep neural network model that integrates physical mechanisms with experimental data, aiming to achieve high-confidence probabilistic life prediction. This model embeds stress-life (S-N) physical constraints into the network architecture and establishes a joint optimization objective integrating data-driven approaches and physical laws, thereby achieving accurate prediction of gear fatigue life under multiple stress levels under small-sample conditions. This method effectively overcomes the limitations of traditional P-S-N curves, which rely on extensive tests, involve high costs, and suffer from insufficient extrapolation accuracy. It provides a probabilistic life prediction framework with both physical interpretability and engineering practicality for the reliability assessment of gear transmission systems, significantly improving the accuracy and efficiency of system reliability assessment in small-sample scenarios.

Key words: reliability assessment, life prediction, physics driven, Gear Test, Machine Learning