Deyla Viola Natalia Soegiono 國立臺灣科技大學營建工程系(指導教授 :Min-Yuan Cheng)
Fall accidents significantly contribute to the fatality rate in construction. Accurately predicting risk level classifications allows for proactive mitigation of these risks, thereby reducing the number of accidents. However, due to the complexity and vast amount of data, conventional methods may be insufficient. Therefore, this study combines the proficiency of NNs with high-dimensional data and the sequential data processing capabilities of GNNs to enhance predictive models for construction falls. Specifically, for this research's objectives, NNs process time-independent variables, while GNNs handle time-dependent variables. To further improve performance, the Optical Microscope Algorithm (OMA) is employed as an optimization process. OMA aims to optimize the hybrid NN-GNN’s architecture parameters and output weights. The proposed model classifies fall risk levels into low risk, medium risk, and high risk on construction sites based on hazardous areas and metabolic heat load severity. The factors contributing to metabolic heat load severity include the worker’s physique, heart rate, and Wet Bulb Globe Temperature (WBGT). The classification performance evaluation metrics—accuracy, precision, recall, and F1 score—obtained by the hybrid NN-GNN achieved higher values than its base models, which are the original NN and original GNN, at 91.65%, 91.99%, 91.65%, and 91.61%, respectively. However, employing OMA to optimize the hybrid NN-GNN model improves the performance to a higher level, with an accuracy of 92.98%, precision of 93.17%, recall of 92.98%, and F1 score of 92.96%. These results demonstrate that OMA-NN-GNN outperforms its base models and is reliable for fall risk level detection, thereby reducing fall-related accidents on construction sites.
Keywords: Fall Accidents; Classification; Deep Learning; Graph Neural Network; Optimization