A Novel Multiscale Feature Fusion Neural Network for Analyzing Uncertain Dynamic Characteristics of Laminated Structures

Conventional Monte Carlo simulations (MCS) face numerous challenges in addressing the unpredictable dynamic features of composite structures, such as excessive data volume, low computational efficiency, and limited applicability. By fully considering the stochasticity of geometric and material parameters, a data-driven framework was introduced to achieve highly efficient and highly accurate prediction of the key indicators of the dynamic characteristics of laminated plates employed in satellites, including the natural frequency, random response, and harmonic response. First, a multiscale feature fusion neural network (MFEFLN) was developed based on two-dimensional convolutional neural networks (2DCNNs) and gated recurrent units (GRUs). The MFEFLN system utilized convolutional operations to extract multiscale and multi-level features and employed GRU to learn sequential features. The MFEFLN was trained on a small sample set and applied to analyze uncertain dynamic characteristics. The results of MFEFLN were compared with those of MCS, back-propagation (BP) neural network, generative adversarial network (GAN), long short-term memory (LSTM), 2D CNN, and ADCNN. The results indicated that the MFEFLN system could rapidly and accuratly predict the dynamic characteristics. These results indicate that the MFEFLN system can serve as a robust and effective tool for uncertainty quantification and sensitivity analysis, while also offering an innovative methodology for the reliability assessment of satellite structures and design optimization.