Probabilistic gear fatigue life prediction based on physics-informed transformer

Abstract

Gear fatigue testing is a complex process that is both time-consuming and labor-intensive, and often produces dispersed results. Traditional statistical analysis methods for probability-stress-life (P-S-N) and fatigue limit require large sample sizes. This study proposes a novel probabilistic gear fatigue life prediction method based on physics-informed Transformer. This method seeks to overcome the limitations of traditional statistical analysis methods and data-driven models by integrating stress-life physical knowledge with the data-driven model. Specifically, based on the obtained gear fatigue limit, this method predicts the probabilistic fatigue life at each stress level by using only a set of probability distribution characteristics of fatigue life under constant amplitude load. The proposed method was validated through multiple cases of gear bending fatigue tests with varying materials, treatments, and geometries, and prediction errors were within an error factor of 2. Compared with other traditional machine learning methods, the accuracy of life prediction was improved by 19.74% to 29.44%. Additionally, feature analysis was conducted to identify key parameters influencing the accuracy of probabilistic fatigue life prediction and provide a range of optimal parameter selection, which is expected to significantly promote the efficiency of fatigue basic data construction in the gear industry.