Automating Ground Motion Quality Assessment Using Residual Neural Networks

The recent advancement in recording capabilities and the large number of resulting recorded motions render the conventional approach of using visual inspection unfeasible. Deep learning offers a promising solution for addressing this scalability challenge, albeit limited studies have explored ML-based approaches to automate ground motion quality assessment. Therefore, this study aims to explore residual neural networks for evaluating record quality using all data encoded in their unprocessed acceleration records. A customized residual neural network architecture was implemented, where the training workflow included data pre-processing and data augmentation using synthetic minority over-sampling techniques to address data imbalance. The results showed that the developed deep learning model can effectively classify records, achieving an accuracy, precision, recall, and F1 score of 93%, 93%, 97%, and 95% on the test set of a record database from New Zealand, respectively. Additionally, the trained model was applied to an entirely new record database, ensuring no data leakage, to test its performance for future deployment. The model achieved an accuracy of 74%, suggesting the possible generalizability of deep learning-based solutions for quality assessment applications and opportunities for further improvement by expanding the training databases.