Research On Independent-Subject Emotion Recognition Of EEG Based On Global Spatial Structure Attention

Emotional recognition of electroencephalogram(EEG)signals is the construction of automated and intelligent recognition patterns through collected EEG databases using machine learning technology.It is widely used in disease diagnosis,assisted driving,criminal investigation,and other fields.However,traditional single subject model training and testing data from the same subject have serious subject dependency issues,which limit the generalization ability and application scenarios of the model.Therefore,it is increasingly important to establish independent-subject emotion recognition models that have better generalization capabilities and make full use of the existing large-scale information in the EEG signal library.Independent-subject emotion recognition models are widely used.For example,in disease diagnosis,Independent-subject emotional recognition models can help doctors predict the emotional status and disease progression of other patients through collected patient EEG data and cases.However,there are two issues that cannot be ignored in the establishment of independent-subject emotional recognition model based on EEG:(1)EEG data errors and noise may occur due to the physiological diversity of subjects during the acquisition process,and(2)under the same environmental stimuli,psychological differences among different subjects may lead to local temporal semantic features such as premature or delayed response signals.Therefore,it is important to extract and select the most effective global temporal and spatial features from existing EEG features to train independent-subject emotion recognition models and improve the generalization ability of the models.Therefore,how to obtain unbiased and consistent EEG signal features has become a necessary and important issue in solving independent-subject recognition.In response to the above issues,this article proposes solutions,the specific content is as follows:(1)An adaptive online incremental auto-encoder correction network is proposed to achieve physiological error correction for different subjects.Firstly,the analytical expression of the output weight matrix of a single hidden layer auto-encoder network is obtained using an extreme learning machine solution process,and it is embedded as a sample in the spatial output model;Secondly,through matrix operation,the output model learns the data at the next moment and improves the training speed;Finally,an adaptive forgetting factor is introduced to the data at the previous moment to ensure that the model knowledge is updated.The network does not require the use of sample real class tags,and can self-learning and extract potentially uniform features from EEG temporal data in the embedded space,thereby eliminating noise and data errors caused by physiological differences among subjects.Independent-subject experiments were conducted in EEG emotion databases SEED,DEAP,and DREAMER to verify the effectiveness of the proposed model.The accuracy rate of full frequency classification for independent-subject emotion recognition in SEED dataset reached80.02%.This network model is of great significance for brain neuroscience research and independent-subject emotion recognition,and can provide more accurate and effective data analysis network models in the field of emotion recognition and diagnosis of neurological diseases in the future.(2)Establish a unified global spatial representation(UGSR)model for spatial attention independent-subject to avoid local temporal semantic features such as premature or delayed response signals caused by psychological differences between different subjects.Firstly,using Gram angular fields(GAF)to transform local temporal features into global continuous spatial representations,avoiding problems such as inconsistent response signal timing caused by psychological differences in the same environment.Secondly,convolutional neural networks(CNN)are established to extract local features and encode spatial locations;Finally,a global spatial attention mechanism is introduced to discover more discriminative global features under the supervision of different emotion category labels,and noise information is ignored to improve the global feature encoding ability,thereby improving image classification and performance in EEG emotion recognition tasks.The validity of the proposed model was verified in independent-subject experiments on SEED,DEAP,and DREAMER datasets.The subject-independent emotion recognition accuracy rates for full frequency features in the three datasets were 84.44%,67.83%,and 77.64%,respectively.At the same time,using ablation experiments,the necessity of the global spatial feature representation model in solving independent-subject emotion recognition problems was verified.