Identifying poses of safe and productive masons using machine learning

Abstract This paper presents a framework to classify work poses among groups of masons during the building of a standard wall of concrete masonry units. The experience of the group composed of masonry instructors and master masons averaged five times that of the other groups, their productivity was highest, and the loads on their joints were the lowest. Thus, they were deemed experts in this paper. Inertial measurement units (IMU) and video cameras were used to collect kinematic data of the masons, from which pose clusters were identified. A Support Vector Machine (SVM) algorithm was used to classify masons' poses into expert and inexpert classes based on the relative frequency of poses in the motions used to lay each of 945 masonry units. Two classification scenarios were tested. While both scenarios achieved similar levels of accuracy, 91.23% and 92.04% respectively, the processing time for binary classification was only 13 s compared to 523 s for inter-group multiclass SVM. Like characteristic vibration frequencies in machine diagnostics and system identification, the characteristic poses identified provide insight into differing methods between expert and less experienced masons. For example, results show that experts utilize fewer and more ergonomicaly safe poses, while being more productive, which indicates lower energy expenditure (less wasted motions). The classification method and the poses identified contribute knowledge to help develop affordable mason training systems that utilize IMU and video feedback to improve health and productivity of apprentice masons.