Addressing environmental challenges in blasting: A dual-method data generation framework for peak particle velocity prediction

The monitoring of peak particle velocity (PPV) during tunneling addresses the pressing environmental challenge of blast-induced ground vibrations, a significant concern for project managers and stakeholders as excessive PPV can jeopardize structural stability. Accurate PPV prediction is therefore central to sustainable engineering practices and responsible resource management. Soft computing methodologies, particularly machine learning, are now regarded as an effective method to predict and control PPV. However, these models are data-hungry and perform poorly when data availability is limited. This study tackles this challenge by proposing a dual-method data generation framework to augment the original PPV data. The first strategy is empirically generated data, which leverages the United States Bureau of Mines (USBM) formula to produce synthetic PPV values based on sampled distributions of key blasting parameters. The second, a core soft computing strategy, uses a deep learning-based model, the Tabular Variational Autoencoder (TVAE), to learn complex patterns and relationships within the original dataset to generate realistic synthetic PPV data. By integrating these two strategies, robust and diverse augmented datasets are acquired to develop and enhance soft computing PPV prediction models. The findings show that the augmented datasets significantly improve the predictive performance of various machine learning models compared to models trained solely on raw data. Applying augmented datasets not only combines the physical insights of empirical models but also the flexibility of deep learning techniques. This study highlights the potential of combining empirical and generative techniques to address data scarcity challenges in blasting vibration prediction and provides a robust framework that enhances multi-criteria decision-making for engineers. • A dual-method data augmentation framework for PPV prediction. • Enhanced machine learning model performance through data augmentation. • Practical applicability of augmented PPV data in engineering scenarios.