Deep Transfer Learning Based Uncorrelated Multi-source Load Identification In Frequency-domain

Frequency domain load identification technology based on data-driven has a wide range of applications in modern engineering design,reliability test,vibration control and so on.However,in the process of load identification,there are often ill posed problems,which lead to the decline of identification accuracy.Neural network can well alleviate the ill posed problems.However,the load identification method based on neural network has the problems of long training time,low efficiency and low accuracy.According to the continuity of transfer function in frequency domain,transfer learning is proposed to improve the training efficiency and identification accuracy of neural network load identification model in target domain.The main research contents include:(1)The load identification method based on neural network randomly selects the initial value of model parameters and trains the model independently at each frequency,which leads to the problem of long training time and low accuracy,taking advantage of the similarity of transfer function between most adjacent frequencies and the same network structure and objective function of neural network models with different frequencies,a frequency-domain load identification method based on neural network and model transfer learning is proposed.Firstly,the neural network model of load identification is trained by using the historical data of a frequency point;secondly,the model parameters of the neural network at this frequency are migrated to the neural network in the adjacent target frequency domain as the initial value of the network weight;thirdly,the neural network is fine-tuned by using the historical data of the target frequency,thus the neural network can be obtained.Uncorrelated multi-source frequency domain load identification model of target frequency;finally,the model parameters of the trained neural network with this frequency are transferred to the next adjacent frequency model and the process is cycled until the neural network models of all frequency points are established.The load identification results on experimental data set of cylindrical shell show that the method has better initial value of network weight,can effectively reduce training time,and has higher recognition accuracy than the neural network method without transfer learning,the method based on transfer function and least square generalized inverse,and the method based on multiple primary linear regression.(2)To solve the problem of low similarity between adjacent frequency data and how to select the source domain when using model transfer learning to identify the load in frequency domain,a method of selecting the source domain by using MMD distance and solving the corresponding Traveling Salesman Problem model(MMD-TSP)is proposed.Firstly,a neural network load identification model is established based on the relationship between load and response.Then,MMD distance is used to measure the difference of sample data under different frequencies,and the MMD distance matrix between different frequencies is obtained.Then,genetic algorithm is used to solve the TSP problem of the sum of MMD distances needed to access all frequencies.Finally,genetic algorithm is used to obtain the minimum cost of MMD distance matrix.The experimental results on experimental data set of cylindrical shell show that compared with other sequential transfer learning neural networks,the mmd-tsp transfer order can get a better neural network model at high amplitude frequency and effectively improve the accuracy of the model.(3)To further utilize the information of historical data of different frequencies and extract the useful information of data between different frequencies,this chapter proposes the method of combining Deep Regression Adaptation Networks(DRAN)with model migration for load identification based on the principles of deep adaptation networks and multi task learning.DRAN can adapt the historical data of different frequencies to a network for training,and then extract the information of common features in different frequencies,so as to further improve the generalization ability and accuracy of the model.There are two data inputs in the DRAN,which are the auxiliary frequency data in the source domain and the target frequency data in the target domain;there are several adaptation layers in the DRAN to calculate the adaptation error of the network,and then the total error training network is added,in which the output of the adaptation layer is the error between the predicted load and the actual load of the auxiliary frequency.The experimental results on experimental data set of cylindrical shell show that the combination of DRAN and model transfer method can further improve the generalization ability and accuracy of the model.