English：

To solve the problem of quality control in the mixed construction system, a closed⁃loop researchmethod of “ feature screening, modeling, interpretability, and scoring ” is constructed based onengrineering practice. By expanding the collected construction logs, quality inspection reports, and field⁃measured data to form 1 000 groups of valid samples, nine categories of core features, such as aluminumformwork support, composite plate hoisting, and concrete performance, etc. , were extracted. Theprediction performance of three machine learning prediction algorithms was compared around the threekey quality indicators of aluminum formwork verticality deviation, composite plate elevation deviation,and joint leakage risk. The results indicate that the random forest model has the optimal predictionaccuracy. The root mean square error (RMSE) and R2of verticality deviation prediction are 2. 1mm and0. 962, respectively. The RMSE and R2of composite plate elevation deviation prediction are 1. 8mm and0. 958, respectively. The risk prediction F1value of the joint leakage risk is 0. 91. After eliminating theredundant features, the three core influencing parameters of aluminum mold back corrugated spacing,composite plate support spacing, and concrete pouring strength were further screened out. Based on thismodel, a comprehensive scoring system for construction quality is constructed to realize quality gradingcontrol, which can provide technical support for early warning and parameter optimization of constructionquality of similar projects.