A deep-learning-based model for quality assessment of earthquake-induced ground-motion records

High-quality earthquake ground-motion records are required for various applications in engineering and seismology; however, quality assessment of ground-motion records is time-consuming if done manually and poorly handled by automation with conventional mathematical functions. Machine learning is well suited to this problem, and a supervised deep-learning-based model was developed to estimate the quality of all types of ground-motion records through training on 1096 example records from earthquakes in New Zealand, which is an active tectonic environment with crustal and subduction earthquakes. The model estimates a quality and minimum usable frequency for each record component and can handle one-, two-, or three-component records. The estimations were found to match manually labeled test data well, and the model was able to accurately replicate manual quality classifications from other published studies based on the requirements of three different engineering applications. The component-level quality and minimum usable frequency estimations provide flexibility to assess record quality based on diverse requirements and make the model useful for a range of potential applications. We apply the model to enable automated record classification for 43,398 ground motions from GeoNet as part of the development of a new curated ground-motion database for New Zealand.