Study On Nondestructive Testing Method Of Bolt Anchorage Quality Based On Deep Belief Network

Rock bolt has been widely used in geotechnical,bridge,mine and other engineering fields.It is very important to test the quality of anchorage system.Since the performance and quality of anchorage system are closely related to the safety of the whole engineering system.Deep learning is a hot topic in the field of machine learning,which provides a new method for the nondestructive testing of bolt anchorage.Compared with the shallow learning algorithm,the deep structure of deep learning can learn the original signal of the rock bolt layer by layer,and extract its deep features to make the classification results of bolt defects more accurate.Based on the Deep Belief Network(DBN),the relevant data of bolt is abstracted,and the features are extracted to identify the different defects.Differential Search(DS)algorithm is applied to the DBN network to optimize the initial weights and thresholds of the network,which solves the problem of great difference between the reconstructed data and the original input caused by the improper selection of the initial parameters.The main research contents are as follows:(1)Based on an improved semi-hard soft threshold wavelet de-noising method,the noise reduction of the acceleration signal of the rock bolt collected in the experiment is carried out,which provides training samples and test samples for the recognition of the defect type of the anchorage system.(2)The network structure and training method of DBN is analyzed and the recognition model of bolt defect type based on DBN is established.The effects of hidden layer number setting,hidden layer node number setting and reverse fine-tuning on the result of defect type identification are analyzed.According to the recognition results,the optimal network structure is determined.The experimental results show that,compared with the shallow learning algorithm,the model can effectively improve the recognition rate of the defect type.(3)Since the random initialization weights and thresholds of DBN networks may lead to large reconstruction errors and fall into local optimum easily,DS algorithm is applied to optimize the initial weights and thresholds of DBN.A defect recognition model of anchorage system based on DS-DBN-SVM is established.DS-DBN model is used to extract the feature of bolt data and the extracted feature is used as the input of SVM to identify the bolt defect type.The experimental results show that,compared with the traditional DBN network model,the model improves the accuracy of bolt defect type identification.(4)Aiming at the problem that the classification accuracy is limited by insufficient data of rock bolt,a model based on EDBN-SVM(Ensemble DBN-SVM)is established to identify the defect type of anchorage system.In this model,DBN-SVM with good classification performance is combined with ensemble learning,and based on different feature sets,the individual classifiers with differences are trained.The experimental results show that,compared with the single classifier model,the model can obtain a higher recognition rate.