Occupational health and safety applications with machine learning

According to the latest report published by the International Labor Organization (ILO), 374 million workers are exposed to non-fatal accidents, while 2.78 million workers die annually. Most of the accidents are caused by human error, lack of experience, incomplete preparation, working while feeling uncomfortable or tired and failure to fully implement safety procedures. With the contributions of machine learning studies, which have shown very important developments in recent years, injuries and deaths can be prevented by training models that can detect and report workplace safety violations that cannot be easily detected by humans, generate alarms and take the necessary precautions. In this study, machine learning applications and types and usage areas in the field of occupational health and safety were examined, four personal protective equipment datasets with different features selected and models trained on these datasets with YOLOv5, YOLOv7 and YOLOv8 architectures and results are compared. As a result of the trainings carried out with YOLO architectures on different datasets, it was determined that YOLOv8 architecture generally gave better results than YOLOv5 and YOLOv7 architectures, while the YOLOv7 model gave better results in the dataset containing high resolution images. Hyper-parameters, which are one of the most important factors affecting the success of machine learning models, were optimized on the CHV data set in YOLOv5 and YOLOv7 architectures by using the genetic algorithm used effectively in the solution of optimization problems, for YOLOv5 person detection mAP value changed from %96.3 to %97.4, precision value changed from %92.1 to %92.6 and for the YOLOv7 the average accuracy value of mAP was changed from 93.4% to 96.3%, F1-Score value was increased from 90% to 94% and average 3.34% more successful results were obtained in all values than standard hyper-parameters.