This academic paper addresses the challenges related to the limited accuracy and practicality of laboratory simulation data in driving tests. The main goal is to investigate driving behavior under real-world vehicle conditions. To achieve this, the paper proposes a dual time window approach that effectively captures time series lengths. Moreover, it extracts the intersection of operational behaviors and employs the verification recursive feature elimination method to identify the optimal fatigue feature indices. These indices are then combined with whale optimization and gate-cycle control neural network techniques to detect fatigued driving more effectively. To enhance recognition accuracy, an attention mechanism is introduced. Experimental results using vehicle data demonstrate that the proposed method achieves an 89.84% accuracy rate in detecting three-level fatigued driving, with a false negative rate of 10.99%.