Early Damage Detection Based on FRF-CNN-BiLSTM Model and NOFRF

Early-stage damage has become a potential source of catastrophic structural failures, therefore, the development of effective early damage detection methods is crucial. The nonlinear output frequency response function (NOFRF) and related indexes can be considered effective detection tools. NOFRF is typically determined using an established nonlinear auto-regressive with exogenous inputs (NARX) model. However, the hyperparameters in the NARX model are difficult to determine, and the identification accuracy cannot be guaranteed. Model based on the frequency response function (FRF) can capture the linear relationship between the system's input and output channels. However, most actual systems exhibit characteristics of nonlinear behavior, leading to prediction errors in the FRF model's responses. Inspired by the powerful learning capabilities of Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory (BiLSTM) networks, this paper proposes a hybrid modeling strategy. This strategy combines the FRF model with the CNN-BiLSTM model to obtain a more accurate response of the nonlinear system. Through simulation studies using two different types of excitation signals (harmonic signals and sweep signals), NOFRF and related indexes can be accurately estimated, thereby enabling the detection of early damage. The effectiveness and generality of the proposed method have been validated.