Scaffolding Acceleration-based Deep Learning Modeling for Fall-related Behavior Monitoring

Abstract

Falls are the leading cause (e.g., 30–35%) of work-related fatalities during construction. However, conventional sensing approaches to recognizing workers’ fall-related movements may have the following limitations: (1) wearable sensors may cause physical interference and privacy issues, and (2) the commonly used algorithms may be unsuitable for classifying undefined classes. Thus, a convolutional neural network (CNN) is proposed to learn unsafe action patterns based on a scaffold’s accelerations due to workers’ movements, along with four modeling strategies to recognize predefined precursor signals while rejecting undefined classes. These models achieved detection and classification F1 scores of 72–78% and 93–97%, respectively, implying that a scaffold’s accelerations could include sufficient information on workers' movements to recognize their actions, and the modeling strategies could seamlessly classify predefined precursors and differentiate between predefined and unseen classes. This approach could reduce fall-related incidents by constantly monitoring workers at elevated locations and providing proactive feedback.