基于神经网络和模型迁移学习的不相关多源频域载荷识别  

Uncorrelated multi-source load identification in frequency domain based on neural network and model transfer learning

作　　者：陈德蕾 王成[1,2] 曾煜 李海波 赖雄鸣[3] 陈叶旺 CHEN Delei;WANG Cheng;ZENG YU;LI Haibo;LAI Xiongming;CHEN Yewang(College of Computer Science and Technology, Huaqiao University, Xiamen 361021, China;Xiamen Engineering Research Centre of Enterprise Interoperability and Business Intelligence, Xiamen 361021, China;College of Mechanical Engineering and Automation, Huaqiao University, Xiamen 361021,China)

机构地区：[1]华侨大学计算机科学与技术学院,福建厦门361021 [2]厦门市企业互操作与商务智能工程技术研究中心,福建厦门361021 [3]华侨大学机电及自动化学院,福建厦门361021

出　　处：《计算机集成制造系统》2021年第10期2961-2969,共9页Computer Integrated Manufacturing Systems

基　　金：国家自然科学基金资助项目(61502181,71571056);国家自然科学基金青年资助项目(51305142,51305143);中国博士后科学基金资助项目(2014M552429);福建省高校青年自然基金重点资助项目(JZ160409);泉州市科技计划资助项目(2018Z008,2017G019,2017G045,2018C110R,2018C114R);华侨大学研究生教育教学改革研究立项资助项目(18YJG28);华侨大学研究生科研创新能力培育计划资助项目(18013083003)。

摘　　要：针对使用模型初始权重随机设定的神经网络模型进行不相关多源频域载荷识别时训练效率低、精度低的问题,提出一种基于神经网络和模型迁移学习的不相关多源频域载荷识别方法。首先,利用某频率点的历史数据对不相关多源载荷识别的多输入多输出神经网络模型进行训练;其次,将该频率下的神经网络的模型参数迁移到相邻的目标频域的神经网络中作为网络权值的初值;再次,利用目标频率的历史数据对神经网络进行微调训练,从而得到目标频率的不相关多源频域载荷识别模型;最后,将该频率的训练好的神经网络的模型参数迁移到下一个相邻频率的模型,循环该过程直到建立所有频率点的神经网络模型。在圆柱壳声振实验数据集上的载荷识别结果表明,该方法具有较好的网络权值初值、能有效减少训练时间,比不使用迁移学习的神经网络方法、基于传递函数和最小二乘广义逆的方法、基于多元一次线性回归的方法具有更高的识别精度。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.

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 

分 类 号：O212.1[理学—概率论与数理统计]