DEEP LEARNING OPTIMIZATION TO CLASSIFY THE USE OF PERSONAL PROTECTIVE EQUIPMENT AT BUILDING CONSTRUCTION AREA

• Main Author: Zakaria, Muhammad
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/51366
• Institution: Institut Teknologi Bandung

In this day, artificial intelligence technology is growing up rapidly that we can utilize this technology in some area such as to supervise workers with personal protective equipment (PPE) at building construction field. The purpose of using PPE is to minimize an accident probability that can lose a life. At 2018, work accident number reach 170.000 in Indonesia and this number keep increasing for every year. This number occurred due to the lack of awareness of construction workers and the ineffectiveness of their supervisory duties. There have been some efforts to decrease this number by doing induction about safety awareness and manual monitoring task by safety health environment staff. Meanwhile a few of solutions have been introduced by researcher such as PPE detection system and high risk pose worker. This research purposes an artificial intelligence system that can detect and classify worker’s PPE usage status at building construction area. This classification will be divided into two classes, namely safe and not safe where safe category captures the worker that using full PPE meanwhile for not safe category captures the worker that not using full PPE. For implementation, surveillance camera will be installed in the work areas to be monitored, captured image from surveillance camera will be processed to detect worker’s position and will be an input to Convolutional Neural Network (CNN) model. In order to improve performance in general, activation function layer at CNN model will be optimized resulting in better classification error from CNN model. The use of the CNN model generally aims to classify or recognize patterns from an image. In this study, the CNN model is used to process the image of the worker so that the image of the worker can be classified based on the use of their PPE. Meanwhile, the activation function layer of the CNN model will use the non-zero slope activation function where the slope of the activation function will be optimized to increase the performance of the model. 1062 images were taken with each class having 531 images. The dataset was taken with several variations in the position of the worker, the color of the worker PPE, the work area and the camera's viewing point. The system that has been created successfully detects and classifies workers with an accuracy value of 91.4% and a loss of 23.2%. These results were obtained by using the Exponential Linear Unit (ELU) activation function which has been optimized using the Particle Swarm Optimization (PSO) method. This result is much better than the model using the Relu activation function, where the model with the Relu activation function achieves an accuracy value of 87.9% and a loss of 32.3%.