Automated Frequency Domain Decomposition Method Based on Convolutional Neural Network

The Frequency Domain Decomposition (FDD) method has been extensively applied in modal analysis of structures under operational conditions. It is based on the singular value decomposition of the Power Spectral Density matrix of structural responses. The modal frequencies of the structure are picked from the peaks of the first singular value, while the mode shape vectors are estimated from the left singular vectors at the modal frequencies. However, the process of manually picking peaks from the singular value curve is cumbersome, time-consuming and the accuracy is heavily dependent on the user’s expertise. To address this problem, a fully automated peak picking method based on the convolutional neural network (CNN) is proposed in this work. Using the image of the singular value curve with two labels, i.e., well-separated peak and closely-spaced peak, as the input, a CNN model is developed to identify peaks that are generated from the structural modes. Furthermore, the CNN is followed by an auxiliary step to identify the repeated modes of the structure. A finite element model with eight degrees of freedom and a filed dataset are used to verify the proposed method. The results show that the developed CNN model can be trained only using the numerical simulation data, and the proposed method can identify structural modes with high accuracy and efficiency.