摘要
Abstract
The real-time and accurate assessment of railway subgrade compaction quality is critical for ensuring long-term track service performance.Traditional intelligent compaction prediction methods often rely on simplified empirical indices,while the pure data-driven models(such as CNN and LSTM)suffer from limitations of poor gen-eralization,lack of interpretability,and dependence on large datasets.In response,this paper proposes a dynamic inversion framework for compaction parameters based on physics-informed neural network(PINN)and vibration time-domain signals.Based on the roller's vibration acceleration time-domain signals collected by sensors,a dual-net-work structure consisting of a solution network and a parameter network is constructed.As strong physical con-straints,the dynamic differential equations of the roller-soil system are embedded into the loss function to achieve a dual constraint of data-driven fitting and physical laws.The results demonstrate that the proposed model can accu-rately invert the equivalent stiffness values and fluctuation trends of the test section.Small-sample sensitivity analysis and 5-fold cross-validation further validate that the model maintains stable prediction accuracy and good general-ization capability even in the case of scarce training data.Therefore,the proposed PINN dual-network framework provides a new paradigm for intelligent compaction with high precision,strong interpretability,and low data depen-dency.关键词
智能压实/物理信息神经网络/铁路路基/刚度预测/动态反演Key words
intelligent compaction/physical-informed neural network/railway subgrade/stiffness prediction/dy-namic inversion分类
交通工程
