Active Learning Supported Metadata Extraction from NVH Simulations

Simulations can only be searched, reused and leveraged as trainings data for machine learning methods if they possess suitable metadata. Manually obtaining this metadata is time-consuming and requires expert knowledge. Consequently, there is often a lack of metadata, and this prohibits the reutilisation of simulation data. Therefore, automated frame works for metadata extraction are essential to quickly, effortlessly and cost-efficiently obtain this information. At present, there are no existing toolboxes for Finite-Element-Simulation data and current Large Language Models are not usable. Nevertheless, machine learning methods are a promising solution for this task. However, classical supervised machine learning methods cannot be used due to the high effort for labelling. Therefore, classical rule-based extraction algorithms are an alternative for fundamental metadata extraction. For more enhanced tasks, where more context is necessary, they are not usable. Active Learning is the optimal technique to overcome this contradiction. Here only as much data points as necessary are labelled and furthermore only data points with a high gain of knowledge are requested for labelling in an iterative loop. Therefore, it is possible to use machine learning methods for the metadata extraction. In this work, the specific advantages and challenges of metadata and their extraction are shown for NVH-Simulations. Rule-based extractors are presented for fundamental metadata extraction and as a benchmark. The workflow for active learning is depicted and applied on NVH models. An effort analysis and comparison of the different extractors complete the study. Based on the results, recommendations for employing these techniques are provided.