Novel ANOVA-Statistic-Reduced Deep Fully Connected Neural Network for the Damage Grade Prediction of Post-Earthquake...

Area of Science:

• Structural Engineering
• Artificial Intelligence
• Data Science

Background:

• Seismic damage assessment in buildings is complex due to environmental uncertainties.
• Accurate and timely damage grading is vital for effective repair and accident prevention.
• Current methods for damage categorization are time-consuming and labor-intensive.

Purpose of the Study:

• To develop an accurate model for seismic damage grade categorization of concrete buildings.
• To enhance damage prediction accuracy through effective feature selection.
• To introduce a novel ANOVA-Statistic-Reduced Deep Fully Connected Neural Network (ASR-DFCNN) model.

Main Methods:

• Utilized a dataset of 762,106 damaged buildings with 26 attributes.
• Applied Analysis of Variance (ANOVA) for feature selection, followed by Principal Component Analysis (PCA) to reduce dimensionality.
• Developed and implemented the ASR-DFCNN model using a sequential Keras architecture with ReLU and tanh activation functions, NADAM optimizer, and L2 regularization.

Main Results:

• Initial ML classifiers without feature selection showed lower performance.
• The Bagging classifier achieved 83% accuracy on the reduced dataset.
• The proposed ASR-DFCNN model demonstrated superior performance, reaching 98% accuracy in damage grade categorization.

Conclusions:

• Feature selection significantly enhances the accuracy of seismic damage grade categorization.
• The ASR-DFCNN model provides a highly accurate and efficient approach for assessing earthquake-induced building damage.
• This AI-driven methodology offers a promising solution for improving post-earthquake response and safety.