Predicting Building Damage with AI Models

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Jul 29 2024

In a paper published in the journal Scientific Reports, researchers investigated a model for predicting damage intensity in buildings on mining terrains using a database of 185 masonry residential buildings. They applied convolutional neural network (CNN) and support vector machine (SVM) methodologies, comparing accuracy, precision, recall, and F1 score.

Both models achieved high prediction accuracy despite minor damage intensities. The CNN model outperformed in F1 score and precision, justifying its effectiveness for this prediction task.

Related Work

Past work on damage to residential buildings in Poland has primarily used statistical methods and traditional machine learning (ML) classifiers like SVMs. Recent advances in deep learning (DL), particularly CNNs, show promise for high accuracy in classification tasks. Challenges include accurately predicting damage due to multifactorial influences from construction defects and environmental impacts and effectively preprocessing numerical data into quasi-images for CNN analysis.

Methodology Overview

Solving problems with deep NN (DNN) requires selecting the appropriate type and architecture. The architecture should be optimal to handle input data effectively. CNN networks typically start with image data fed into the hidden layers, which include convolutional, pooling, and fully connected layers. These layers learn optimal filters through backpropagation and gradient descent, ultimately making predictions at the output layer.

The convolutional layer, the core of CNNs, consists of filters (kernels) that extract features by sliding over the input volume, creating activation maps. Users specify hyperparameters such as padding, strides, and filter sizes, which are optimized during training. The pooling layer, used to reduce noise by down-sampling feature maps, also requires user-defined hyperparameters. Common pooling types include maximum, minimum, and average pooling.

The fully connected layer connects each neuron from the previous layer, assigning weights to determine class probabilities during learning. This layer has as many output nodes as there are classes. Non-linear activation functions, crucial for forming decision boundaries, enhance the network's predictive performance. The learning algorithm significantly influences CNN's overall effectiveness.

The study also used SVM, a reliable classification and regression algorithm, to benchmark CNN results. SVM classifies data into two or more classes by finding a decision boundary (hyperplane) that maximizes the separation margin between classes. During training, the SVM algorithm identifies this optimal hyperplane for better generalization and classification of test data.

Building an SVM model involves determining parameters such as σ (regularization constant) and γ (curvature of the kernel function). These parameters guide the learning process, with σ influencing the loss function and γ defining the kernel function's curvature. This study used Bayesian optimization to ascertain the optimal values for these parameters.

Model Performance Comparison

Data were split into training (75%) and test (25%) sets, with performance evaluated using accuracy, precision, recall, and F1 score. Classification accuracy and qualitative criteria were used to categorize results, and generalization ability was assessed by comparing accuracy differences between the sets.

The CNN model, trained using the stochastic gradient descent with momentum (SGDM) algorithm, achieved nearly 80% accuracy after 15 iterations, demonstrating high generalization ability and varied precision and recall across different damage categories. The SVM model was built using radial basis functions (RBF) as the kernel function, optimizing parameters σ and γ through Bayesian optimization. The model achieved an accuracy of approximately 80% for both training and test sets, indicating high generalization ability.

The SVM classifier reduced the number of support vectors by 35%, showing good generalization properties. Precision and recall were high for the first damage category, good for the second, and acceptable for the third. Differences in classification accuracy between the training and test sets were minimal, with some exceptions in precision and recall for the second damage category.

Comparing CNN and SVM models demonstrated high generalization abilities and similar performance levels. SVM achieved slightly higher training set accuracy, while CNN showed higher average precision and F1 scores. Despite CNN's additional time cost for converting numerical data into quasi-images, it provided opportunities for introducing more information into the analysis. SVM had lower hardware requirements for numerical data analysis but could face challenges with image data. Both methods were effective, with CNN's data preparation offering potential advantages in this context.

Conclusion

To sum up, the research aimed to create a model for predicting damage to masonry residential buildings in mining areas, comparing CNN and SVM methods. The CNN model, trained with the SGDM algorithm, achieved 78.26% accuracy for the test set, showing high generalization ability. Optimized using Bayesian methods and RBF kernels, the SVM model also demonstrated high generalization with similar accuracy. Both models had minor differences in performance, with CNN requiring more initial data preparation but offering the potential for automating damage prediction processes.