Dynamic Load Identification at Natural Frequencies for Aircraft via Attention Based 1D-CNN

Abstract

Since the frequency response function is ill-conditioned at the natural frequency of the system, the traditional load identification method based on the system parameters is no longer applicable. Aiming at the difficulty of dynamic load identification at the natural frequency of the structural system, a dynamic load identification method at the natural frequency of the structure based on one-dimensional convolutional neural network(1D-CNN) with attention mechanism is proposed. Specifically, the high-level features in the vibration response signal are first extracted through the convolution layer. Then the weight matrix of the network is updated by backpropagation algorithm, which represents the importance of different features. The mapping relationship between response and load is established to realize the task of load identification. From the trained data, the attention module learns the contribution of features according to the different contribution of different features to load prediction. The important components in the response signal are highlighted and noise pollution is suppressed. Excitation and response signals at the natural frequency of the system were acquired using exciters and an accelerometer mounted on the GARTEUR aircraft model. Excitation and response signals at the natural frequencies of the system are obtained by an exciter and accelerometer mounted on the GARTEUR aircraft model. The responses of the model at the first three natural frequencies of 6.4Hz,35.8Hz and 48.5Hz were obtained respectively. Experimental results show that compared with the traditional TSVD load identification method, the maximum error of this method is only 3.19%. Compared with the 1D-CNN method, the proposed method has stronger robustness under 20%, 50% and 80% noise levels.