Work

Evaluating high-speed railway intelligent construction business levels via ERT-TOPSIS and SSA-optimized BP with rigorous synthetic-data validation

Year: 2025

Type: article

Abstract: • An evaluation method combining Seq GAN and SSA-BP neural networks for intelligent construction business level evaluation. • Validation strategy employing synthetic data for training and real data for testing (TSTR). • Established mapping rules converting continuous scores to discrete grades based on fixed threshold intervals. ABSTRACT Intelligent construction has emerged as a critical pathway to enhance the quality and transformation of high-speed railway projects. However, existing evaluation approaches often focus on isolated technical aspects or static indicators, lacking a multidimensional, predictive, and interpretable model capable of comprehensively and objectively assessing intelligent construction business level. This study proposes a novel evaluation framework based on a SeqGAN-SSA-BP neural network to achieve continuous and objective assessment of intelligent construction business levels. A hierarchical indicator system is constructed across three dimensions: intelligent, construction, and other dimension. The Sparrow Search Algorithm (SSA) optimizes the weights and biases of the BP neural network, while an Extreme Random Tree (ERT)-weighted TOPSIS method is introduced to determine objective indicator weights. To mitigate data scarcity, 90 synthetic samples are generated from 10 expert evaluations using a Sequence Generative Adversarial Network (SeqGAN). The synthetic data's distributional consistency with real data is validated via t-SNE visualization and Maximum Mean Discrepancy (MMD) tests. SHapley Additive exPlanations (SHAP) analysis is employed to interpret feature importance and underlying mechanisms. The model is rigorously evaluated using a Training on Synthetic Data, Testing on Real Data (TSTR) approach with a hold-out real dataset. The results demonstrate that the proposed model significantly outperforms BP neural network, Random Forest, SVM, and XGBoost models, with MSE reduced by 2.346 × 10⁻³, 0.565 × 10⁻³, 1.535 × 10⁻³, and 2.671 × 10⁻³; RMSE reduced by 4.436 × 10⁻², 1.988 × 10⁻², 3.515 × 10⁻², and 4.759 × 10⁻²; MAE reduced by 3.788 × 10⁻², 1.734 × 10⁻², 2.962 × 10⁻², and 0.016 × 10⁻²; and R² increased by 0.312, 0.075, 0.204, and 0.356, respectively.

Source: Results in Engineering

DOI: 10.1016/j.rineng.2025.107591

Authors: Xizhen Xu, Yuming Liu, Guoliang Ou

Institutions: Beijing Jiaotong University, Shenzhen Polytechnic University

Language: English

Topic: Evaluation and Optimization Models

Subfield: Safety, Risk, Reliability and Quality

Field: Engineering

Domain: Physical Sciences

SDG: Industry, innovation and infrastructure

Open Access status: gold

Estimated APC paid: $1,940