Abstract: To address the issues of variable compaction time and single evaluation index based on dry density assessment of compaction quality, a new method of vibratory compaction control for high-speed railway graded aggregate (HRGA) based on coarse particles embedding point is proposed. Firstly, the vibration compaction evaluation system is improved by combining the mechanical indexes of dynamic stiffness Krb and modified foundation coefficient K20. The index of compaction control “embedded locking point” Tlp is then proposed, and the mechanical properties and applicability of graded aggregates before and after Tlp are investigated through indoor tests. Secondly, the relationship between Tlp and various performance indexes of HRGA is established through vibratory compaction test, and the master control features of Tlp are analyzed using grey relation analysis (GRA) algorithm. Finally, the Tlp prediction model is proposed based on the machine learning (ML) method, the best Tlp prediction model is selected using the three-level preference system, and the best ML model is interpreted using SHapley Additive exPlanations(SHAP) interpretable method. The results show that the optimal vibration time can be determined based on Tlp, thereby controlling the compaction quality. The master control features of the Tlp are maximum particle size of filler dmax, grading parameter b, grading parameter m, flat elongated particles Qe and Los Angeles abrasion LAA based on the GRA algorithm. The comprehensive evaluation index (CEI) of each Tlp prediction model is calculated as follows: artificial neural networks for improved particle swarm optimization (IPSO-ANN) model > support vector regression for improved particle swarm optimization (IPSO-SVR) model > random forests for improved particle swarm optimization (IPSO-RF) model, with the IPSO-ANN model being optimal. The overall importance values  based on SHAP method are ranked as follows: dmax(17.31) > b(13.93) > m(6.59) > Qe(2.17) > LAA(1.54), which corroborates with the results obtained from the GRA algorithm, indicating that the SHAP method can improve the comprehensibility of the ML model. The research results can provide new ideas for quality assessment of vibratory compaction, and also provide strong theoretical support for intelligent control of vibratory compaction.

Key words: vibratory compaction, high-speed railway graded aggregate, optimal vibration time, master control features, machine learning, interpretable