Stress Identification Of Steel Components Based On Ultrasonic Technique And Deep Learning

The stress of steel components is a key parameter in the design,construction and service of steel structures.Traditional stress testing methods have drawbacks of high cost,inconvenient operation,and low accuracy,which limit their application in stress detection.A new method based on deep learning and ultrasonic technique is proposed to accurately and efficiently obtain the stress of steel components.A one-dimensional convolutional neural network(1-D CNN)model is used to establish the corresponding relationship between the ultrasonic signals of steel components and its stress value.The one-dimensional CNN takes the raw ultrasonic timedomain signal of the steel components as input,and outputs the stress value of the steel components.The flow of this method is as follows.Firstly,ultrasonic signals of steel components with single thickness and steel components with different thicknesses under different stress levels were collected by uniaxial compression tests and used to build the training,validation and test datasets.The stress levels in the test dataset are different from those in the training and validation datasets.The number of ultrasonic samples for the training,validation and test datasets are 3072,768 and 288 for steel components with single thickness,and 2400,600 and 270 for steel components with different thicknesses,respectively.Secondly,three one-dimensional convolutional neural networks with different number of convolutional layers(4,6,and 8 layers)were trained,validated,and tested with the two prepared datasets to determine the optimal 1-D CNN architecture for structural stress identification.The optimal architecture of 1-D CNNs for stress identification of steel components was determined.The results showed that the 1-D CNN model with six convolutional layers achieves a good balance between stress identification accuracy and computational efficiency,and is therefore the optimal architecture for stress identification of steel components with single thickness and steel components with different thicknesses.Finally,the parameters of the 1-D CNN with the optimal architecture were optimized and then used to identify the stress of the test dataset.The results showed that the optimal hyperparameter combination of the 1-D CNN model is MSE for the loss function,Adam for the training optimizer,and 64 for the batch size for steel components with single thickness,and the average stress identification error is 3.50%;for steel components with different thicknesses,the optimal hyperparameter combination of the 1-D CNN model is MSE for the loss function,Adam for the training optimizer,and 64 for the batch size,and the average stress identification error is 3.83%,which indicates that the stress identification accuracy of the proposed method basically meets the actual engineering requirements.The stress detection method proposed in this thesis has the advantages of low cost,high accuracy,easy operation and minimum environmental affect,making it a viable reference for stress detection of steel components in practical engineering.