Study On Identification Method Of Bolt Anchorage Defect Type Based On Convolutional Neural Network

Bolt has been widely used in engineering fields such as highways,tunnels,and buildings.The quality of the bolt and its supporting performance are closely related to the safety and reliability of the whole project.Therefore,it is very important to detect the quality of bolt anchorage system.It is also an important task to identify the defect types of bolt anchorage system.The traditional recognition method is based on signal processing technology,which requires manual selection of features.It is time-consuming,laborious,and difficult to excavate new features.As a new method of machine learning,deep learning provides a new research direction for the type identification of bolt anchorage system.Compared with the traditional recognition method,deep learning can learn the original signal of the bolt layer by layer,and obtain more abstract high-level features,so as to achieve more accurate identification of bolt anchorage defect type.The convolutional neural network(CNN)is used to identify the bolt anchorage system with different defect types.In order to improve the speed and accuracy of effective feature extraction in the CNN model,firefly algorithm(FA)and sparse auto-encoder(SAE)network are combined to pre-train CNN's initial convolution kernel.The main research contents are as follows:(1)The experimental platform for signal acquisition of bolt is set up,and the acceleration signals acquisition of four different types of bolt anchorage systems is carried out by using stress wave reflection method,which provides a data basis for the subsequent identification of bolt type.(2)The related theories and training methods of CNN network are elaborated,and the model of defect type recognition for bolt based on one-dimensional convolutional neural network(1D CNN)is established.The optimal network structure is determined by analyzing the influence of the number of convolutional layers,the size and the number of convolution kernel on the recognition results.Experimental results show that this model has better recognition effect than the shallow learning method.(3)Since randomly initializing the convolution kernel of 1D CNN will cause problems such as slow speed and low accuracy of network effective feature extraction.An identification model based on FA-SAE-1D CNN for bolt defect type recognition is established.The initial convolution kernel of 1D CNN is pre-trained with SAE network optimized by FA.In order to further improve the accuracy of feature extraction of the model,dynamic sampling is adopted in the sampling layer of the model.The results show that compared with 1D CNN,the improved model achieves higher recognition accuracy.(4)In view of the problem that the 1D CNN model only learned the time-domain characteristics of the bolt signal but ignored the frequency-domain characteristics,a model for identifying bolt defect types based on short-time Fourier transform(STFT)is established.The time-frequency spectrum obtained by the STFT of the original acceleration signal of the bolt is taken as the input of the model,and the defect type of the bolt anchorage system is finally recognized by convolution and sampling operations.The results show that the STFT-CNN model has a higher recognition rate than the CNN model with time domain or frequency domain data as input.