CRNN-Based Modeling of Mental Fatigue from Multimodal Signals

Abstract:- Reliable, real-time detection of mental fatigue from EEG is critical in safety-sensitive settings. We present LogMel-CRNN, a convolutional–recurrent framework that couples a one-dimensional convolution– based Short-Time Fourier Transform (STFT) with a Mel-scale filter bank to produce log-Mel spectrograms explicitly tailored to the spectral characteristics of fatigue. The convolutional front-end enhances frequency– time resolution and noise tolerance, while the recurrent back-end captures temporal dependencies associated with the gradual onset of fatigue. We evaluate multiple architectural variants and compare against traditional machine-learning pipelines and contemporary deep baselines. Across experiments, LogMelCRNN consistently delivers superior classification performance, with notably high recall and F1 scores, demonstrating robust sensitivity to fatigue-related EEG fluctuations. Ablation analyses indicate that the logMel transformation is pivotal—aligning feature emphasis with perceptual relevance—whereas the recurrent component is essential for modeling temporal dynamics beyond frame-level cues. These findings establish LogMel-CRNN as an accurate and resilient approach to EEG-based fatigue detection, with practical potential for low-latency deployment in high-risk environments such as driver monitoring and industrial operations.

---