Abstract

Crane accidents remain a critical safety issue in the construction industry, with nearly 300 fatalities reported in the United States between 2011 and 2017. These accidents commonly involve falling objects, crane collapses, and overturning, and are often caused by human error, equipment failure, and environmental conditions. They also result in significant economic losses, including equipment damage and project-related costs. Despite the availability of detailed accident reports, their analysis remains largely manual and expert-driven, making it time-consuming and potentially inconsistent. This thesis investigates two prediction problems using a nationwide dataset of 710 crane accidents collected and compiled by Jim D. Wiethorn, P.E., PhD: damage level estimation based on structured crane accident information and responsible personnel classification using narrative descriptions. Across Experiments, the work examines how heterogeneous data contributes to model performance and interpretability.
For damage prediction, a multi-stage machine learning approach is employed, incorporating SHapley Additive exPlanations (SHAP) to rank feature importance and examine the contribution of different feature types. The analysis shows that narrative information plays a critical role in prediction, with lower severity cases associated with assembly and setup-related issues, and higher severity cases linked to system-level or environmental failures.
For personnel identification, fine-tuned Bidirectional Encoder Representations from Transformers (BERT)-based models are applied to narrative accident descriptions, with data augmentation to address class imbalance. The results demonstrate that narrative data alone can effectively support the identification of responsible personnel.
This study develops a machine learning-based approach designed to serve as an initial screening tool for post-accident analysis. It enables a more cost-efficient and consistent assessment through the identification and ranking of influential crane accident factors.