| Driver fatigue is currently recognized as one of the primary causes of traffic accidents,and using the electroencephalogram(EEG)signals to detect driver fatigue is a viable method.However,existing methods for cross-subject classification of EEG signals often require a substantial amount of target domain data for calibration in order to achieve satisfactory classification performance.In this paper,a novel fatigue driving recognition model is proposed,which is based on improved Euclidean space alignment and multi-feature fusion transfer learning.The primary objective of this model is to ensure robust cross-subject classification performance for fatigue driving,even when limited calibration data is available.Notably,this model takes into consideration the individual differences and low signal-to-noise ratio characteristics of EEG signals,resulting in enhanced cross-subject classification performance even under limited calibration data conditions.To mitigate the individual differences between subjects and enhance the accuracy of cross-subject classification in transfer learning,we employ the Euclidean-space alignment method for data alignment.To overcome the limitations of traditional Euclidean-space alignment methods with limited calibration data,we introduce two improved methods: Average Euclidean-space Alignment(AEA)and Weighted Average Euclidean-space Alignment(WAEA),for zero calibration and limited calibration data scenarios,respectively.AEA estimates the reference matrix of the target domain by averaging the reference matrices of the source domain data,while WAEA calculates the weighted average of the source domain reference matrices using cosine similarity between matrices and combines the source and target domain reference matrices with a weighting parameter to achieve better alignment results with limited calibration data,thereby improving the cross-subject classification performance.Test results demonstrate that under a calibration ratio of5%,the two alignment methods achieve 5% and 10% higher cross-subject classification accuracy than commonly used machine learning and deep learning methods without alignment,and the accuracy of model transfer and fine-tuning of all convolutional layers is improved by 20% with WAEA,highlighting the necessity of WAEA in transfer learning.To further enhance the classification accuracy of the model for low signal-to-noise ratio signals,we incorporate a multi-scale feature fusion method to enhance the model’s feature extraction capability,in combination with WAEA and model transfer learning to improve cross-subject fatigue driving recognition performance.Firstly,we conduct in-depth research on the pre-training and fine-tuning strategies of the model based on WAEA and select three commonly used convolutional neural networks in the field of EEG.We use different pre-training and fine-tuning strategies to obtain a common feature extraction network and freeze its network parameters as the pre-trained model.Then,we connect multiple feature extraction networks with different depths after the common feature extraction network,concatenate the common features with the characteristic features,and utilize a multi-layer perceptron for classification,and fine-tune the model with calibration data.Finally,we compare the proposed method with single-scale feature classification algorithms and other transfer learning algorithms.The results demonstrate that the proposed method achieves the best performance at different calibration ratios,with an accuracy of 95.08%,an f1-score of 95.05%,a precision of96.11%,and a recall of 94.34% at a calibration ratio of 5%,outperforming the compared methods.This indicates that the proposed method can achieve superior classification performance with limited calibration data. |
