Attention-Based Generative Adversarial Network Model for RC Column Backbone Curve Identification

Backbone curves can reflect the force and deformation characteristics of structures under seismic action and provide a basis for designing structures ready for construction and evaluating the residual seismic capacity of existing structures. To achieve fast, accurate, and cost-effective identification of the backbone curves of reinforced concrete (RC) columns, the main factors affecting the backbone curves of RC columns are determined by discretization of the continuous problem combined with the maximal information coefficient method, and based on the 253 RC columns pseudo-static experimental data, taking into account the hidden correlation between the points of the backbone curve and between different backbone curves, an RC column backbone curve identification model is proposed based on an improved generative adversarial network and taking into account the attention mechanism (BC-ARGAN). Compared to backbone curve identification models based on other machine learning methods, such as the multilayer perceptron, long short-term memory, bidirectional long short-term memory, attention mechanism combined with BiLSTM, and existing identification methods, the BC-ARGAN model has higher accuracy and stability, with MAE, MAPE, RMSE, and $R^{2}$ on the testing set of 12.473 kN, 8.438%, 20.400 kN, and 0.9926, respectively, which can better serve engineers in evaluating the seismic performance of structures.