关键词: CFRP薄壁C柱, 轴压响应, LSTM自编码器, 鲸鱼优化算法, BP神经网络

Abstract: Aiming to predict the force-displacement responses of carbon fiber reinforced plastic (CFRP) thin-walled C-columns in the aircraft sub-cargo area under quasi-static axial crushing, an intelligent prediction model (WOA-BP-LSTM autoencoder model) integrating the Whale Optimization Algorithm (WOA), Back Propagation (BP) neural network, and Long Short-Term Memory (LSTM) autoencoder was proposed. The reliability of the finite element model of CFRP thin-walled C-columns was validated through quasi-static axial crushing tests, with axial crushing response evaluation indicators showing errors within 10%. A dataset comprising 700 force-displacement response samples with variable cross-sectional geometric parameters was constructed based on the model. The LSTM autoencoder was employed for dimensionality reduction and reconstruction of the force-displacement responses. Subsequently, the BP neural network was used for force-displacement responses prediction, with WOA optimizing the neural network parameters. The results showed that the LSTM autoencoder achieved high-precision reconstruction of force-displacement responses, where the errors for initial peak crushing force and energy absorption in the test set were both less than 3%, and 80% of the samples had errors within 1%. The optimized prediction model significantly improved prediction accuracy, reducing the test set’s mean absolute error (MAE) by 17.55%, mean squared error (MSE) by 31.77%, and root mean squared error (RMSE) by 17.47%. Prediction errors for the initial peak crushing force and energy absorption were both less than 8%, with 80% of samples showing errors within 5%. This model enables rapid and accurate prediction of axial crushing responses for variable cross-section CFRP thin-walled C-columns while reducing computational costs, providing an efficient parameter-performance mapping tool for the study of its axial crushing response.

Key words: CFRP thin-walled C-columns, axial crushing response, LSTM autoencoder, whale optimization algorithm, BP neural network