Acoustical modeling of construction jobsites with multiple operational machines for activity recognition and productivity analysis

Construction job sites are dynamic environments with various resources operating simultaneously. For most construction projects, a significant expense is the budget allocated to acquiring and renting heavy equipment. Carefully analyzing heavy equipment productivity rates and monitoring productive times are significant factors in the success of construction projects. Traditional methods for construction equipment performance monitoring are through direct observations and surveys. These methods are labor-intensive and prone to error, making them impractical for larger job. As a result, there is an increasing demand for efficient and systematic solutions for productivity analysis of heavy equipment.Construction equipment productivity rates are directly associated with activities the machine performs during routine operations. Recognizing these activities is the first step toward analyzing efficiency rates. Recent technological advancements motivated researchers to develop automated techniques for automated equipment activity detection in construction job sites.This dissertation aims to use an audio-based method to develop an acoustical model of construction job sites with multiple machines for activity recognition and productivity analysis. In the first phase, the author proposes a method by investigating the feasibility of integrating two major sources of data, kinematic and acoustic, to address the distinct weaknesses of these existing methods. In the second phase, the author focuses on hardware-based methods by investigating several beamformers to separate equipment sounds for the multiple-machine scenario using microphone arrays. In the first step of the third phase, the author utilizes software-based methods and binary Time-Frequency Masking (TFM) to separate equipment sound using single-channel microphones. In the second step of the second phase, the author improves the framework to generalize it for more than two machines by proposing a data augmentation method and Convolutional Neural Network (CNN) for multiple-equipment activity recognition. Finally, the author proposes a method to calculate the productivity rates and cycle times using the recognized activities. This study has been tested on various case studies and the results show that the final activity recognition method recognizes multiple-equipment activities with accuracies up to 98.1% and 88.1% for synthetic and real-world mixed sound data, respectively, demonstrating the capability of this method for progress monitoring of construction equipment.