关键词: 不相关多源载荷识别, 频域, 多输入多输出神经网络, 模型迁移学习, 网络权值

Abstract: Aiming at the problem of low training efficiency and low accuracy load identification in frequency-domain when using a neural network model with random initial weights,a frequency-domain load identification method based on neural network and model transfer learning was proposed.The neural network model of uncorrelated multi-source load identification was trained with the historical data of a certain frequency point.The neural network parameters at this frequency were transferred to the adjacent target frequency model as the initial value of the network weight,and the neural network was fine-tuned by using the historical data of target frequency to obtain the load identification model of the target frequency.The frequency model was transferred to the next adjacent frequency model,and the process was looped until the neural network model of all frequency had been established.The experimental results on the data of cylindrical shell structure showed that the neural network model based on model transfer learning could obtain a better initial weight of the neural network model and reduce the training time effectively.Moreover,the proposed method had a higher identification accuracy than the method based on neural network method without transfer learning,the transfer function and least square generalized inverse and the multivariate linear regression.

Key words: uncorrelated multi-source load identification, frequency-domain, multiple inputs and multiple outputs neural network, model transfer learning, network weight