Sensor-based activity recognition for construction activities on site using stacking ensemble method

Automatically recognizing workers' activities can improve productivity, safety, and management in construction. This study aimed to collect data using sensors from workers in a real construction site environment to compare various machine learning models for recognizing workers' activities. Additionally, it sought to develop a novel meta-ensemble machine-learning model to enhance prediction accuracy. For this purpose, formwork, rebar, concrete, walling, roughcast, gypsum, painting, and tiling activities were analyzed. The XGB model had the highest prediction accuracy at 96.14%. This study created a stacking meta-ensemble model to increase these prediction accuracy rates further. As a result of trying various variations, the Stacking model, which was formed by combining the SVM, RF, GBM, and XGB models and choosing the XGB model as the meta-learner, reached 98.53% prediction success. The results demonstrate that ensemble machine-learning models surpass basic machine-learning models' ability to predict outcomes accurately. By using this proposed stacking ensemble model, it is expected to create an automatic system that will calculate the productivity, risks, and fatigue of workers in future studies.