Electroencephalogram signal faces challenges in fatigue driving detection due to its non-stationarity and difficulties in feature extraction

and power spectral density features are difficult to be effectively utilize.To address these challenges

a deep learning-based multivariate time series classification model:multi-scale attention convolutional neural network(MACNN)is proposed.The model employs multi-scale convolution and depthwise separable convolution to enhance the modeling ability of different frequency bands and cross-channel features;integrates residual convolution and the channel attention mechanism to strengthen feature expression ability and improve training stability;and incorporates a causal convolution-supported multi-head attention mechanism

which can effectively capture temporal dependenc in electroencephalogram signal.Experimental results on the SEED-VIG dataset show that MACNN achieves an average accuracy of 95.8%

and its performance on power spectral density is superior to that of various deep learning models.Ablation experiments verify the key role of each component module in feature extraction and performance enhancement.The research demonstrates that MACNN can solve the problem of electroencephalogram signal feature extraction

and provides an effective solution for fatigue driving detection.