High-speed railway dispatcher plays an important role in ensuring railway traffic safety. Taking the working state of dispatchers in different emotions as the research content, an experimental mode is designed to collect the physiological data of high-speed railway dispatchers performing dispatching tasks under different emotional states. The emotion-induction experiment induces the schedulers to enter four working states respectively. Then eye movement data are collected and the K-nearest neighbor classification model and the classification and regression tree decision tree classification model are established for state identification. The identification accuracy of the former is higher, reaching more than 90%. The CNN-GRU fusion model is built to improve the prediction effect of fatigue degree.