Abstract: The industrial field has entered the era of industrial intelligence, following the developmental stages of mechanization and informatization. The hallmark of the new industrial revolution is cyber-physical systems (CPS), which merge the digital and physical worlds. However, owing to advancements in various fields, a fully autonomous system is not achievable in the near future. In the process of deepening industrial intelligence, leveraging their respective advantages to achieve safe and efficient cooperation remains a significant issue. Many studies have shown that accidents caused by human error constitute a significant portion of safety-centric complex systems, with mental workload being the primary factor leading to such errors. As human-in-the-loop research based on transparent interfaces, such as the brain–computer interface, advances, passively and naturally integrating human cognitive models into human–machine systems has become a trend, thereby participating in the decision-making and control of the future. Mental workload is the key cognitive component related to human cognition in safety-centered complex systems research. As an important factor reflecting system performance, it is crucial to evaluate mental workload scientifically and quantitatively in future industrial human-cyber-physical systems research. Utilizing this as the starting point, this paper comprehensively reviews the development of the theoretical basis of mental workload, research progress of theoretical framework, development of mental load assessment, existing experimental paradigms, and induction methods. By visiting related factories and communicating with automation experts and job operators, the components and characteristics of industrial CPSs are analyzed, and the characteristics and core operation mode of the human-computer interaction (HCI) system of industrial CPSs are proposed. Three primary task types in the industrial context have been identified, and an experimental paradigm for mental workload research, tailored to the industrial environment, has been established. Thus far, a simulated task load paradigm has been applied in the field of industrial systems. In the modeling phase, the key electroencephalogram (EEG) sensitivity parameters for three tasks must first be extracted. Among the features that showed significant performance under different mental workload levels, the characteristics of absolute EEG power exhibit the best performance in distinguishing mental workload. Subsequently, multitype task recognition models based on EEG signals are established. Results show that in monitoring and verification, control operation, and communication scenarios, the test accuracy rates of the algorithm model were 88.14%, 94.72%, and 82.42%, respectively. Through five-fold cross-validation and mesh parameter optimization, the optimal parameters for the model are obtained. To assess the model's generalization ability, a subject-wise method is employed. Upon examination, the average recognition accuracy of the random forest model based on absolute EEG power features reached 97.85%, 96.95%, and 89.88%, and the highest recognition accuracy can reach 100%. The nonlinear entropy features in the frequency domain showed crossover, while the fusion features did not show an obvious trend beyond the absolute EEG power features. This study provides a theoretical basis for the elucidation of the relationship between operator physiology and system workload, improving the productivity of industrial CPS and operator enthusiasm and providing a new perspective for promoting the research on human-in-the-loop based on transparent interfaces, natural human-computer interactions, adaptive software, and brain–computer interfaces.