Research On Intelligent Monitoring Method Of Bolt Tension State Based On Convolutional Neural Network

The bolted connection has the characteristics of strong carrying capacity,easy maintenance and replacement,but the looseness of the bolted joint structure often occurs due to the detachability of the bolt.Loose bolt connection failure will not only affect the normal use of the mechanism,shorten the service life,and even cause casualties.In order to avoid catastrophic consequences,online monitoring and evaluation of bolt assembly tightness is particularly important.Convolutional neural networks have achieved good results in the intelligent diagnosis of mechanical parts such as bearings and gears.However,intelligent mechanical fault diagnosis based on convolutional neural networks has not paid sufficient attention to the convolutional neural network models of different depths on multi-sensor signals.performance.In this paper,we propose a series of algorithms based on convolutional neural networks to learn the characteristics of the excitation response signal data collected by multiple sensors,and to diagnose the tightness of the bolted structure.This paper first proposes a convolutional neural network with two layers of convolutional layers to monitor the state of the 1D original mechanical signal.The model was trained on the vibration test data of the bolt connection of the frame test bench,and the recognition rate can reach over 97%.Since the deep convolutional neural network(DCNN)is more capable of mining representative information and sensitive features from the excitation response signal than the shallow CNN without prior knowledge,this paper proposes a DCNN model.DCNN has the ability to feature extraction,feature selection,and classification,with raw data as input and identification results as output.The impact of different parameters and configurations of the network architecture on network performance is also studied.In order to train the data of different sensor points at the same time,so that the model can obtain high recognition rate in multi-point data,a residual deep convolutional neural network(RDCNN)model based on Shortcut method is proposed.The model improves the state monitoring capability of the DCNN model for multi-spot signals,and solves the degradation phenomenon of the training set accuracy and the test set accuracy when the number of model network layers is deepening.Aiming at the shortcomings of traditional feature extraction methods,which are usually sensitive to the changes of physical properties of mechanical systems,a Network in Network deep convolutional neural network(NINDCNN)model based on migration learning is proposed.The NINDCNN model has good mobility,and it automatically learns important "feature extraction filters" from mechanical data,which can achieve strong adaptability in variable points and variable tasks.In order to reduce the training difficulty of CNN,the Inception Module was introduced for CNN,which improved the accuracy and was easy to train.