Deep regression adaptation networks with model-based transfer learning for dynamic load identification in the frequency domain

摘要

Frequency-domain dynamic load identification methods based on neural network (NN) models construct models independently at each frequency, but are inaccurate and inefficient to train. To address these problems, a deep regression adaptation network (DRAN) with model-transfer learning is proposed for identifying dynamic loads in the frequency domain. The aim is to take advantage of the similarity of uncorrelated multi-source dynamic loads and multi-vibration response at adjacent frequencies. First, a DRAN model for load identification is established using the historical data for a specific frequency. Second, the trained DRAN parameters are transferred to the DRAN for the target frequency as the initial parameter values. Next, the transferred DRAN is fine-tuned with the historical data of the target frequency to obtain the load identification model of the target frequency. Finally, the trained DRAN parameters of the current target frequency are transferred to the next target frequency. This process is iterated until a DRAN model for all frequencies is established. Because a frequency response function is a continuous function varying with frequency, the relationships between the dynamic loads and response at adjacent frequencies are similar. DRAN can adapt the historical data of different frequencies to one neural network for training, and then extract the common feature information of different frequencies to improve the accuracy of the model. Moreover, instead of setting the initial weights randomly and training them independently for each DRAN model, model-transfer learning is used to obtain better initial weights from the trained weights of DRAN models of adjacent frequencies. The proposed method was evaluated on the experimental data of a cylindrical shell structure under acoustic vibration joint excitation. The results show that the proposed method can obtain better initial weights, higher accuracy, better noise robustness, and shorter training time than a neural network.