Is mental fatigue a nonlinear accumulation process? Evidence from field EEG-based brain network analysis

Mental fatigue is a difficult to detect risk state in high-risk industries and is closely associated with sustained cognitive load and prolonged work duration. To overcome the limitations of post-incident management, research has introduced wearable EEG technologies, promoting a shift in safety management from passive response to feedforward warning. However, most existing studies are based on laboratory experiments, treating cognitive load and work duration as independent variables and inducing fatigue over short durations, which limits the characterization of fatigue under extended work periods. To address these issues, this study conducted a field experiment at a drilling site, combining wearable EEG functional connectivity and brain network analyses with a PCA Transformer framework for automated fatigue recognition. The results show that as task duration increased, functional connectivity analysis indicated that mental fatigue generally increased and then decreased; meanwhile, brain network analysis revealed a transition from local strengthening to global integration, accompanied by reduced network stability. The Transformer model demonstrated excellent classification performance, achieving recognition accuracies of 90.7% and 91.0% under low and high load conditions, respectively. Overall, these findings indicate that shift scheduling should not rely solely on cumulative work duration but should instead pay closer attention to changes in risk across different work periods. Based on this, safety management practices may consider reducing the concentration of high cognitive load tasks and critical decision making activities during later stages of work, while machine learning based fatigue state recognition offers feasible technical support for feedforward safety management.