Abstract: Excessive mental workload in complex assembly operations can reduce assembly performance and quality, and may even lead to safety accidents. Therefore, accurately measuring and evaluating the mental workload of assembly tasks to determine the level of mental workload for operators is an important basis for optimizing production systems. In order to evaluate the mental workload in complex assembly tasks, assembly task experiments based on Lego block simulation are designed. Data from 25 indicators are collected by subjective measurement, performance measurement and physiological measurement to analyze the sensitivity of each indicator to changes of mental workload. Results show that seven indicators comprised of the performance measurement indicator, subjective measurement indicator, fixation times, total fixation duration, scanning times, total scanning duration and average pupil diameter are the effective measurement ones for the mental workload in complex assembly tasks. Based on these effective measurement indicators, a synthetic evaluation model for assembly task mental workload is established by two modeling methods, namely BP neural network and Bayesian linear discrimination. Findings show that the BP neural network model, which uses factor analysis to transform the 7-D indicators into the 2-D synthetic indicators as the principal component input, and employs the normalized conjugate gradient method as the training algorithm, is the optimal synthetic evaluation model for mental workload in assembly tasks, with the accuracy of 84.80%. The proposed method for measuring and evaluating mental workload can provide reference for the evaluation and optimization of mental workload in assembly operations.