| Emotion is an important and complex physiological and psychological state in human life,which has a huge impact on thinking,communication,judgment and other behaviors.Electroencephalogram(EEG)is an electrophysiological signal acquired from the cerebral cortex,which contains rich information that can be used for the detection and recognition of human physiological and psychological states.Compared with other input signals such as images and text,EEG signals are not easy to disguise.Therefore,affective computing based on EEG signals can better reflect the real emotional state of the human body.Due to differences in physiological and psychological conditions,acquisition equipment.and experimental settings,there are large differences in the multi-source EEG signals from different individuals and different experimental scenarios.This is one of the main obstacles to real-time emotional brain-computer interfaces.With the rapid development of neuroscience,artificial intelligence and other fields,affective computing based on multi-source EEG signals has broad and important application prospects.Emotional brain-computer interface systems usually include a series of complete processes from EEG signal acquisition,signal analysis to final application.This study starts with the design and implementation of a highprecision EEG acquisition equipment,and conducts transfer learning research based on raw multi-source EEG signals,which can complete the entire process of signal acquisition and analysis tasks.Specifically,the main contributions of this paper are as follows.1.This study designed and developed a high-precision multi-channel EEG acquisition device,which can acquire 4 kHz EEG signals under the 32-channel setting.In the underlying structure of the device,the program execution framework is designed by using techniques such as direct memory access,first-in-first-out,and interrupt to ensure the stability and precision.A long period of continuous data acquisition evaluation was performed with an average maximum delay of 0.7 s/h and without data loss.Moreover,the device is open source,and designed as an easy-to-build structure,so that it can be quickly rebuilt and deployed.In addition,it is also compatible with other electrophysiological signal acquisitions such as Electrocardiogram,Electrooculogram,and Electromyogram.2.This study proposed a frequency spatial information extraction network based on raw EEG signals.A frequency-dependent convolution structure is designed to extract low-frequency and high-frequency related feature maps in the raw EEG signal,and the feature maps of different frequencies are further weighted and aggregated through the attention structure.Compared with traditional artificial features,this method does not require prior knowledge,but can automatically extract and aggregate frequency and spatial features of EEG signals.This method has been verified on several publicly available common EEG affective computing datasets.Compared with the latest methods,it has achieved better results.Among them,the 2.68%accuracy and 3.31%MF1 were improved on the SEED-V dataset for five categories of emotion recognition.3.This study proposed a multi-dimensional transfer learning framework for cross-subject and cross-dataset scenarios.In the data source dimension,the framework performs data alignment in Euclidean space on the raw EEG signals to reduce the variance.In the network training dimension,the framework uses domain classifiers to jointly optimize network parameters to blur the differences in high-order features from different data sources.Cross-subject transfer experiments have been carried out under multiple datasets,and good results have been achieved.Among them,at least 3.85%accuracy and 6.84%MF1 were improved under the SEED-Ⅴ dataset.In multiple cross-dataset transfer experimental scenarios,good results have been achieved.Among them,in the transfer scenario of the SEED-Ⅳ dataset and the SEED dataset,at least 4.46%accuracy and 5.32%MF1 were improved. |
