Impact load reconstruction and localization based on a deep recurrent neural network

Monitoring and identification of impact loads acting on structures remains a challenging problem in engineering practice. To address this problem, a DRNN model is established in this work for simultaneous impact load reconstruction and localization. The DRNN model consists of one long-short-term memory (LSTM) layer, two bidirectional-LSTM (BLSTM) layers, and two fully connected layers to construct the complex inverse mapping between input (dynamic response) and output (excitation). A numerical test is designed to evaluate the accuracy of the DRNN in impact load identification of three different types of impact loads on a rectangular thin plate, including trapezoidal, semi-sinusoidal, and sawtooth pulse loads. The anti-noise ability of the DRNN is further examined by introducing different levels of both white noise and colored noise into the sample data. Additionally, two experimental impact tests are designed and conducted to further verify the effectiveness and applicability of the proposed DRNN model in practical applications. The DRNN is also compared against multiple benchmark deep learning models, and the influence of different structural response phases (forced vibration vs. decaying vibration) on model performance is investigated using SHapley Additive exPlanations and signal energy intensity analysis. Results indicate that the proposed DRNN model demonstrates superior identification accuracy, noise immunity, and robustness in impact load identification tasks compared to the benchmark models. The constructed DRNN model provides an efficient and robust solution for the inverse problem of impact load identification, possessing significant application potential in structural strength analysis, reliability assessment, fault diagnosis, and structural health monitoring.