Air traffic controllers communicate with flight crews by using standard language for radiotelephony communication throughout the execution of control tasks, and fatigue detection of controllers' radiotelephony communication voice can effectively assess the fatigue state of controllers and reduce the flight safety risk. Given that controller fatigue state is difficult to mark in real time and it is subjective, a controller fatigue state recognition model based on unsupervised learning was proposed. Firstly, the controller simulator speech data were used to investigate the presentation of fatigue state in speech features, and eight features with strong fatigue recognition ability were screened from 62-dimensional GeMAPS speech features by normalisation and recursive feature elimination cross validation (REFCV), and the feature values were compared between fatigue state and normal state. Secondly, the supervised classification of controller fatigue state was performed using the selected features, and the recognition accuracy reaches 82.87%, which proves the fatigue recognition ability of the selected features. Finally, the control voice data were clustered, and the cluster distribution was analysed according to the feature trend to reclassify the fatigue state, and the clustering effect was verified, obtaining a fatigue state detection accuracy of 98.78%. The research results provide a theoretical basis and experimental support for the subsequent fatigue state research using control speech data in real work scenarios.