关键词: 多点频域振动响应预测, 多源未知载荷, 多输入多输出神经网络, 模型迁移学习, 网络权值

Abstract: To address the problems of long training time and low prediction accuracy due to the need to build a neural network model for each frequency point independently and to select the initial value of the neural network model independently and randomly for multipoint frequency-domain vibration response prediction by a Multiple Input Multiple Output Artificial Neural Network (MIMO-ANN),a multi-point frequency-domain vibration response prediction method based on MIMO-ANN and model transfer learning was proposed.The data-driven multipoint frequency-domain vibration response prediction problem under unknown uncorrelated multi-source loads was described formally,and its differences was compared with the data-driven multi-point frequency-domain vibration response prediction problem under known uncorrelated multi-source loads.The self-power spectra known vibration response at a certain frequency point were constructed as the input and the self-power spectra of with unknown vibration response as the output,and the unknown point vibration response prediction model at that frequency was built using MIMO-ANN;based on the continuity of the transfer function in the frequency domain,the weights of the trained MIMO-ANN at this frequency were used as the initial values of the MIMO-ANN at neighboring frequency;the historical data at this neighboring frequency were used for training to obtain the prediction model at this frequency;the process was repeated until all the models at all frequencies were trained.This method solved the matrix pathological inverse problem,obtained better initial values of the neural network model,was not easy to fall into local optimum and speeded up the convergence of the neural network.The results of multi-point response prediction on the experimental data set of cylindrical shell acoustic vibration showed that the proposed method had higher prediction accuracy and better training efficiency than the methods based on no transfer learning ANN,multiple linear regression and transfer function under the condition of unknown load from multiple sources.

Key words: multi-point frequency domain response prediction, unknown multi-source load, multiple inputs and multiple outputs neural network, model transfer learning, network weights