{"title":"A confident learning-based support vector machine for robust ground classification in noisy label environments","authors":"Xin-Yue Zhang , Xiao-Ping Zhang , Hong-Gan Yu , Quan-Sheng Liu","doi":"10.1016/j.tust.2024.106128","DOIUrl":null,"url":null,"abstract":"<div><div>Geological labels obtained from field exploration have potential errors due to technique limitations and subjective interference, leading to noisy labels when developing ground-machine interaction models for TBM tunneling. The present study proposes a novel confident learning-based support vector machine (CL-SVM) to eliminate label noise, thereby improving the accuracy and credibility of ground classification. The proposed model optimizes confidence values for each label and recognizes those with low confidence values as potential noise. Its effectiveness and superiority are confirmed through a noise test. The results indicate that the maximum acceptable noise ratio of the CL-SVM is 35%, while that of the conventional SVM is only 10%. In addition, the CL-SVM consistently emerges as a superior performer compared to the SVM in noisy label environments. The CL-SVM is further verified through its application on a class-imbalanced dataset collected from a metro tunnel project in Wuhan, China. Here, the accuracy metric <em>F1-score</em> for the most noise-interfered class is significantly improved from 0.7273 to 0.88. To enhance the model’s practical value, a confidence criterion is established to evaluate the credibility of individual predictions, which requires reliable predictions to have higher confidence values than specified thresholds. Without prior knowledge of true sample labels, this criterion distinguishes mispredictions from correct predictions with a remarkable precision of 99.08%. In summary, the proposed CL-SVM exhibits significantly better robustness to noisy labels than conventional models, demonstrating great potential for ground perception in tunnel projects.</div></div>","PeriodicalId":49414,"journal":{"name":"Tunnelling and Underground Space Technology","volume":"155 ","pages":"Article 106128"},"PeriodicalIF":6.7000,"publicationDate":"2024-10-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Tunnelling and Underground Space Technology","FirstCategoryId":"5","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S0886779824005467","RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"CONSTRUCTION & BUILDING TECHNOLOGY","Score":null,"Total":0}
引用次数: 0
Abstract
Geological labels obtained from field exploration have potential errors due to technique limitations and subjective interference, leading to noisy labels when developing ground-machine interaction models for TBM tunneling. The present study proposes a novel confident learning-based support vector machine (CL-SVM) to eliminate label noise, thereby improving the accuracy and credibility of ground classification. The proposed model optimizes confidence values for each label and recognizes those with low confidence values as potential noise. Its effectiveness and superiority are confirmed through a noise test. The results indicate that the maximum acceptable noise ratio of the CL-SVM is 35%, while that of the conventional SVM is only 10%. In addition, the CL-SVM consistently emerges as a superior performer compared to the SVM in noisy label environments. The CL-SVM is further verified through its application on a class-imbalanced dataset collected from a metro tunnel project in Wuhan, China. Here, the accuracy metric F1-score for the most noise-interfered class is significantly improved from 0.7273 to 0.88. To enhance the model’s practical value, a confidence criterion is established to evaluate the credibility of individual predictions, which requires reliable predictions to have higher confidence values than specified thresholds. Without prior knowledge of true sample labels, this criterion distinguishes mispredictions from correct predictions with a remarkable precision of 99.08%. In summary, the proposed CL-SVM exhibits significantly better robustness to noisy labels than conventional models, demonstrating great potential for ground perception in tunnel projects.
期刊介绍:
Tunnelling and Underground Space Technology is an international journal which publishes authoritative articles encompassing the development of innovative uses of underground space and the results of high quality research into improved, more cost-effective techniques for the planning, geo-investigation, design, construction, operation and maintenance of underground and earth-sheltered structures. The journal provides an effective vehicle for the improved worldwide exchange of information on developments in underground technology - and the experience gained from its use - and is strongly committed to publishing papers on the interdisciplinary aspects of creating, planning, and regulating underground space.