Wang Z K,Tan H M,Gao Z B. 2024. Classification evaluation of construction sites with thick overburden based on machine learning. Acta Seismologica Sinica,46(3):1−13. doi: 10.11939/jass.20220176
Citation: Wang Z K,Tan H M,Gao Z B. 2024. Classification evaluation of construction sites with thick overburden based on machine learning. Acta Seismologica Sinica,46(3):1−13. doi: 10.11939/jass.20220176

Classification evaluation of construction sites with thick overburden based on machine learning

  • In response to the problem that the category of the site is easily changed due to slight changes in a single factor caused by measurement and other errors in the calculation of equivalent shear wave velocity, a large amount of relevant field test data such as standard penetration value, depth, and shear wave velocity were collected under thick covering soil layers of Yancheng area of Jiangsu Province. Machine learning methods were used for training and modeling, and the ability of multi eigenvalue models to solve site classification problems under thick covering soil layers was studied. The results showed that through feasibility analysis, the accuracy of the logistic regression model, the support vector machine model and the random forest model was 0.809, 0.939, 0.951, respectively. Considering the accuracy gap between the logistic regression model and the support vector machine model and the random forest model, the support vector machine algorithm and the random forest algorithm were selected as the optimal algorithms for building the model. In order to consider the integrity of the entire borehole as much as possible, this paper proposes a parameter called as “equivalent coefficient of variation”, which effectively improves the accuracy of the model. Subsequently, when establishing the support vector machine model, the classification performance of using linear, polynomial, and Gaussian kernels was compared horizontally, and the Gaussian kernel function was ultimately selected for model building. The accuracy of the obtained support vector machine model was 0.951. When establishing a random forest model, the classification performance of the model was tested by setting different numbers of decision trees. Finally, 150 decision trees were selected to build the model, and the accuracy of the obtained random forest model was 0.977. From the results, the accuracy of the support vector machine model and the random forest model are 95.1% and 97.7%, respectively, with recall rates of 98.2% and 97.3%. The AUC values of both models are 0.98. Therefore, while the classification performance of the random forest model is not inferior to that of the support vector machine model, it has a higher adaptability to the sample data, and the recall and accuracy of the random forest model are similar, that is, the model’s judgment of the sample population is more balanced. In summary, the above random forest model is optimal to solve the problem studied in this paper, and can provide reliable basis for determining the category of sites with thick covering soil layers. Therefore the random forest model was used to determine the site category of 75 sets of data in the critical sample of this study. The results showed that 61 sets were consistent with the judgment results of the exploration report, while 14 sets were different from the judgment results of the exploration report. Moreover, the model’s judgment of class Ⅲ sites was completely consistent with the exploration report. The judgment of the random forest model for class Ⅲ sites is completely consistent with the exploration report, and the judgment results are mostly consistent with the exploration report results, proving that the model not only has excellent judgment ability in non-critical situations, but also maintains good judgment ability when used to solve problems in critical situations. Therefore, this model can make secondary judgments for similar engineering problems and provide effective reference basis. Based on the random forest model, the judgment results are output in sequence, and are organized and verified according to the original drilling information. It is found that in the judgment of the critical sample, the model correctly judged eight drilling holes, and only two drilling holes had different site classification judgments from the exploration report, all of which were classified as class Ⅳ drilling sites in the report and were classified as class Ⅲ drilling sites in the model. In practical engineering, judgments made on site for safety reasons are often conservative, and such judgments are magnified into two different site classification results near the boundary. This can explain the significant divergence between the model and exploration report’s judgments of class Ⅳ sites.
  • loading

Catalog

    Turn off MathJax
    Article Contents

    /

    DownLoad:  Full-Size Img  PowerPoint
    Return
    Return