Study on Human-Machine Collaborative Systems using Agent-Based Simulation Incorporating Human Performance and Reliability Analysis

Abstract

The systems in our lives are rapidly changing due to Industry 4.0 technologies such as artificial intelligence, robotics, and digital twins, many of which are becoming automated. However, human characteristics and creativity cannot be replaced, and not all parts of these systems work automatically. Therefore, most systems can be described as human-machine collaborative systems, where humans and machines work together to perform tasks. However, the inherent uncertainty of human behavior significantly impacts the performance and reliability of these systems. For this reason, analyzing a system requires analyzing the humans within it. This study introduces human performance modeling and human reliability analysis for system analysis. Additionally, agent-based modeling and simulation were used to reflect individual human characteristics and the complex interactions within the working environment. Through modeling and simulation, researchers can assess the performance and reliability of new systems or various scenarios and gain insights into the system by analyzing simulation data using artificial intelligence. In this study, agent-based modeling and simulation were used for system modeling and simulation due to their many advantages. To model human behavior, an affordance-based finite state automata model was used, and for human analysis, human performance modeling and human reliability analysis were conducted. To effectively analyze the synthetic data generated by the simulation, several machine learning models, including linear regression, random forest, and support vector regression, were used to identify the importance of each variable and determine which elements of the system affect the outcome. For models with unknown variable importance, Shapley additive explanations, a method of explainable artificial intelligence, were used to extract information about feature importance. To validate this framework, data on workload and time from virtual reality-based human-in-the-loop experiments were referenced. Based on them, equations for average workload, workload weight, and working time weight were assumed. This framework was then applied to the control room of a thermal power plant and an additive manufacturing factory. To understand the relationship between input and output variables, sensitivity analyses were conducted by training machine learning models with simulation data. The analyses confirmed that workload weight is strongly positively correlated with workload, and working time weight is strongly positively correlated with task completion time. It was also found that while human-related performance shaping factors are important for the mistakes made by operators, environment-related performance shaping factors are more significant. Scenario analyses were then conducted to see how the system changes as the work environment changes. By running many simulations after altering system components such as the presence of a supervisor, the level of automation, and the number of workers or machines, the simulation data were statistically analyzed to observe the system’s behavior. The results showed that a supervisor delayed task completion time to reduce the mistakes made by operators, that the workload of operators decreased as the level of automation increased, and that the system’s performance improved as the number of operators and machines increased, but the system’s reliability decreased. This study makes three key contributions. First, it integrates human performance modeling and human reliability analysis with agent-based modeling and simulation to analyze humans in the system from multiple perspectives, reflect individual worker characteristics, and implement complex interactions in the work environment. Second, modeling and simulation can generate synthetic data for many systems, including new systems that do not yet exist or various scenarios that are difficult to replicate in reality, providing valuable knowledge and insights into the system. Third, it determines which factors are important to the system through sensitivity analysis using artificial intelligence, which can be used for decision support when designing new systems or improving existing ones.