基于机器学习的路基压实质量评价研究

Research on Subgrade Compaction Quality Evaluation Based on Machine Learning

摘要: 针对现有的智能压实技术应用于粗粒土填料时的适配性不足，以及在多变施工场景下压实度预测准确性不足的问题方面，本文成功开发了一种依据压实度的增量原则进行优化的人工神经网络模型。研究结果显示，通过将粗粒土压实质量的关键评价指标逐一加入模型，显著提升了模型的预测准确度，预测误差随着输入参数数量的增加而显著减少，并最终达到稳定状态。包含四个参数(2次谐波分量、CMV、加速度上峰值和THD)的综合评价模型，在粗粒土路基压实质量评估中表现最佳，其平均绝对误差(MAE)和均方误差(MSE)分别降至0.63和0.79，最大偏差控制在1.82%以内。这些改进确保了模型输出的压实度与现场实测值高度吻合，实现了对路基压实质量的高效、实时智能评估。

Abstract: In response to the insufficiency of adaptability of existing intelligent compaction technologies when applied to coarse-grained soil fillers, and the lack of accuracy in compaction degree prediction under variable construction scenarios, this paper successfully developed an artificial neural network model optimized based on the incremental principle of compaction degree. The research findings indicate that by incrementally incorporating key evaluation indicators of coarse-grained soil compaction quality into the model, the predictive accuracy of the model was significantly enhanced, with the prediction error markedly decreasing as the number of input parameters increased, eventually reaching a stable state. The comprehensive evaluation model, which includes four parameters (A_2Ω, A_u, CMV, THD), demonstrated the best performance in the assessment of compaction quality for coarse-grained soil subgrades, with the Mean Absolute Error (MAE) and Mean Squared Error (MSE) reduced to 0.63 and 0.79, respectively, and the maximum deviation controlled within 1.82%. These improvements ensure that the model's output of compaction degree closely matches the field-measured values, achieving efficient and real-time intelligent assessment of subgrade compaction quality.