A Model for Assessing Lethal Resistance Levels of Various Buildings Based on Improved Genetic Algorithm + BP Neural Network Optimization

Zhang Jie ¹ Bin Tan ^1* Xia Chaoxu ² Wenbin Yan ³ Yuan Tao ¹ Ben Ma ¹

• ¹ Anhui Earthquake Administration, Hefei, China
• ² Institute of Geology, China Earthquake Administration, Beijing, China
• ³ Jinan Big Data Research Association, jinan, China

Assessing the lethal resistance levels of buildings during earthquakes is crucial for reducing disaster losses and human casualties. This study proposes a novel model that integrates an improved genetic algorithm (IGA) with an optimized backpropagation neural network (OBPNN) to address data imbalance in classifying building types for lethal resistance levels assessment. The Synthetic Minority Class Oversampling Technique was applied to balance class distributions in the training set by oversampling minority classes. To address overfitting, L2 regularization was combined with a genetic algorithm to optimize the backpropagation neural network (BPNN)'s weights and biases, enhancing global search capability and classification accuracy. Momentum parameters and the Adam optimizer were incorporated to smooth gradient updates, prevent oscillations during training, and accelerate convergence. Additionally, domain adaptation techniques were employed to improve test set performance through feature adaptation, enhancing the model's robustness under varying data distributions and its generalization ability. The experimental results show that the proposed improved model achieves excellent performance in classifying the level of lethal resistance levels of buildings, with an accuracy of 97% and an AUC value of 1, which indicates that the model's generalization and discriminative abilities are more excellent.