Aiming at the problem of insufficient feature extraction and low efficiency of dynamic time series modeling caused by the nonstationarity of driver electroencephalogram (EEG) signals in driving scenarios, this paper constructed an innovative framework that integrates wavelet packet decomposition (WPD) and bidirectional long short‐term memory (BiLSTM) to solve the problem. First, the original EEG signal frequency band is subdivided into 32 sub‐bands through a 5‐layer Daubechies 4 (db4) complete decomposition tree, and the optimal sub‐band time‐frequency energy entropy features of different frequency bands are extracted based on the Shannon entropy criterion to suppress high‐frequency noise and enhance frequency band specificity. In particular, the focus is on the γ frequency band (30–45 Hz), which is closely related to high‐order cognition and attention regulation, and its energy entropy features can effectively characterize abnormal neural activity in the state of distraction. Second, a dual 128‐unit BiLSTM layer is constructed, and the forward and backward dependencies of the feature sequence are analyzed through the bidirectional gate control mechanism bidirectional gate control mechanism (BiGCM) to achieve dynamic modeling of attention drift and continuous focus state. Then, a cross‐modal stacking strategy (features from different modalities are concatenated along the feature dimension to form a joint feature representation, thereby fusing frequency‐domain and time‐domain information) is used to splice the 256‐dimensional BiLSTM state and the 2‐dimensional WPD energy entropy feature into a joint feature matrix, which is input into the support vector machine (SVM) classifier for attention state discrimination. Finally, experiments based on the SEED–VIG dataset show that the Fisher discriminant ratio of the method used in this paper in the γ band reaches 0.87, which is 24.3% higher than the traditional power spectrum entropy. The median delay of attention state transition detection is 0.78 s ( Q 1 = 0.65 s, Q 3 = 0.91 s), which is 48.7% lower than the unidirectional long short‐term memory (LSTM) and 64.2% lower than the convolutional neural network (CNN). The classification accuracy is 93.7%, and its monitoring performance accuracy is significantly superior to baseline models such as unidirectional LSTM (89.1% ± 0.9%), CNN (85.5% ± 1.2%), and SVM (82.3% ± 1.5%), with specificity and sensitivity reaching 92.1% ± 0.7% and 94.3% ± 0.8%, respectively. The framework uses a graphics processing unit (GPU) to accelerate the efficient extraction of WPD features in 5.2 ms, and the 5‐s sequence BiLSTM inference takes 42.7 ms, which meets the needs of real‐time vehicle monitoring and provides a highly robust solution to reduce the risk of traffic accidents.
Wang et al. (2026) studied this question.