Random load identification of cylindrical shell structure based on multi-layer neural network and support vector regression

Abstract

A new-type identification method in the frequency domain by combining a multi-layer neural network and support vector regression is proposed to identify random load of a complex cylindrical shell structure. The kernel function of support vector regression has a great influence on the prediction accuracy of machine learning model, and it is effective to employ the linear function. As the penalty factor is large, the identification accuracy of the Gaussian kernel function is close to the linear kernel function. In the process of random load identification, the prediction accuracy of the neural network using the L-BFGS method is higher than the traditional Adam method. The number of hidden layers of the neural network has little effect on the L-BFGS algorithm, but a great effect on the Adam method. Different levels of noise are introduced to verify the robustness of the machine learning model. Both the support vector regression with linear kernel function and neural network model based on the L-BFGS method have strong robustness. For the noise percentage of 1%, the support vector regression has better prediction accuracy than the neural network, yet the case is contrary for the noise percentage greater than 5%. The verification shows that the neural network model trained by simulation data has better identification accuracy for real load at some frequencies. The load identification method is proposed based on the frequency points which may establish the machine learning model. The mean absolute percentage error shows that the method based on a multi-layer neural network and support vector regression has high identification accuracy.