Convolution Generative Adversarial Network For Bolt Anchorage System Grout Defect Detection

Anchor support has many advantages such as low cost,high efficiency,and good performance,and has been widely used in engineering.The good anchor support effect is the guarantee of engineering safety.However with anchor bolt plays a role in the project,various defects will inevitably occur.Grout is an important part of the bolt anchorage system.Once serious damage occurs,the bonding force cannot meet the requirements,it may affect the support effect and threaten the safety of the project.Therefore,it is necessary to determine whether the grout of the bolt anchorage system is defected.Most of the current defect recognition methods are based on signal processing technology,and then artificially extract features by professional analysis of the waveform,which has problems such as time-consuming,labor-intensive and difficult to fully dig features.The neural network can be applied as an intelligent classifier,and feature extraction can be performed automatically through a non-linear activation function,which provides a new research direction for defect identification of the bolt anchorage system.However,there are two problems in using the neural network to identify the defects.The first is that the defects occurred in the bolt anchorage system in actual engineering may be types that are not in the database.Therefore,they have unknown characteristics which is difficult to determine whether there are defects in the bolt anchorage system.Secondly,it is very expensive to collect the labeled data of anchor bolt.Therefore,detecting defects with fewer labeled training samples is a challenge.Aiming at the above two problems,this paper proposes a convolution generative adversarial network and uses it as a classifier for defect recognition of experimental data of bolt anchorage system which collected based on ultrasonic guided waves.The main researches are as follows:(1)Based on one good model of bolt anchorage system and three models of bolt anchorage system with grout defects,an experimental platform is set up.The ultrasonic guided wave method based on magnetostriction was used to collect data for the four types of bolt anchorage systems,which provided a data basis for the defectidentification.(2)Aiming at the problem that the shallow neural network has a poor feature extraction capability and therefore requires an additional feature extraction process,this paper proposes a defect recognition method for end-to-end based on convolutional neural network.This method has good feature learning capability,which can automatically learn and find available features from the original signal,and then identify the defect.In addition,the proposed method also draws on multi-scale strategy,which further improving the ability of feature learning.(3)Aiming at the two problems of using the neural network to identify the defects of the bolt anchorage system,firstly use the ability of the generator in the generative adversarial network to generate fake data which similar to the real data from noise to enrich the diversity of the data of the bolt anchorage system and then solve the problem of the anchoring system with unknown characteristics.Secondly,with the fake data generated by the generator,the data without the label will also have the label of real data,which can assist the labeled sample to learn the distribution of the overall data and realize semi-supervised learning.It solves the problem that it is difficult to calibrate the label of the waveform data obtained in the actual project,therefore the supervised learning method cannot fully play its role.Moreover,the data collected on the experimental platform for bolt anchorage system are used for experimental verification.The comprehensive analysis of the experimental results of different algorithms and scales proves the effectiveness of the proposed method.