Application Research Of Multi-view Semi-supervised Learning On Bridge Structural Health Data Classification

The condition restriction of sufficient and redundant view of Standard Co-trainingwas based on fully theoretical derivation, which made the algorithm successful; however,this condition limited the application of this algorithm. Based on the basic idea ofstandard Co-training, Co-training was improved to be practical based on theimprovement of View Segmentation method, Label method of unlabeled samples andEvaluation method of Classifier Difference, therefore, a method named Co-trainingBased on Difference was proposed, which was aimed to solve the classification problemof multi-view data more efficiently, especially that of bridge structural health data. Theimprovement of Co-training was mainly as follows:(1) According to the problem of initial dataset without natural feature segment view,The View Segmentation method was put forward, which was aiming at producing twoviews through attribute partition. This method firstly evaluated mutual informationbetween non-class attribute and class attribute through a way based on Bootstrap andhistogram, then sorted them, finally segmented all attitudes except class attribute equally.(2) Label method of unlabeled samples was introduced, which worked after the highconfidence unlabeled samples was selected. This step was added by combing agreementand disagreement to judge label veracity, hoping for selecting unlabeled samples withmore correct label.(3) Take the key role of classifier difference in co-training into condition,Evaluation method of Classifier Difference was used to determine whether stopclassification model update or not, which made fully focus on classifier difference. Andthus the classifier performance was improved definitely.In this paper, the efficiency of each improvement was validated through theoreticalanalysis and experiment. Experiment result of each improvement suggested them waspractical. Meanwhile, experiments was carried out with several datasets, theclassification performance of our algorithm was compared to other semi-supervisedlearning algorithms and supervised learning algorithms. The superior performance ofCo-training Based on Difference Algorithm was proved, and also it’s applicability to classification problem of bridge structural healthy data, which provide information ofbridge structural health for management decision.