Study And Application Of Bridge Structural Health Classification Model Based On Semi-supervised Learning

Bridge structural health monitoring with multi-parameter data is a necessary meansto safeguard the normal operation of the bridge. Static structural parameter ischaracteristic of irregular waving value and unbalanced distribution and rich of thebridge structural information. A bridge structural health classification model built bymachine learning methods could obtain static and dynamic force behavior which reflectsthe actual structure status of the bridge. In recent years, with a growing size and highupdating speed of static data, traditional supervised classification model requires a largenumber of labeled samples. However, it will cost lot of manpower and material resourcesto label all the complex bridge structural monitoring data. Semi-supervised learningcombines the few labeled samples with a large number of the unlabeled for a patternclassification. This method absorbed the advantages of supervised learning classificationperformance and reduced the sample label cost as well. Therefore a classification modelbased on semi-supervised learning can be the effective solution to this problem.The paper is under the background of bridge structural monitoring and collaborativetraining in semi-supervised learning. The research work is consists of the followingaspects:1. Summarized the research in bridge structural health monitoring field, discussedthe data analysis methods of bridge structural health monitoring according to feature ofthe bridge structure parameters.2. A deeply study on the basis theory of semi-supervised learning especially theco-training algorithm.3. Proposed a semi-supervised classification algorithm based ondifference-increasing strategy. The experimental results on UCI standard data sets showthat compared with the original co-training algorithm, the algorithm has a much higheraccuracy and stronger generalization ability.4. The bridge structural attribute samples were constructed through the real staticdata and a collaborative classification model was established for bridge structural healthmonitoring. Experimental results show that semi-supervised collaborative classificationmodel is suitable and practical for the bridge structure data analysis. The error rate of theproposed algorithm was much lower compared to supervised algorithms and othersemi-supervised learning algorithm, which is a further evidence for its excellent classification performance.5. Proposed the principle in the selection of sample size, unlabeled sample ratio andclassifier while establishing a bridge classification model, which provides with theinformation for the relevant departments in evaluation of the bridge overall structurehealth status.6. Developed the software of semi-supervised learning module in WEKA which wasintegrated of the semi-supervised collaborative classification model.