Data Analysis On Spatial-Temporal EEG Via Multi-Scale Cascaded Conv-GRU

Electroencephalography is a kind of physiological data collected by braincomputer interface devices that can directly reflect the internal cortical activity.Due to its diversity,efficiency and easy training characteristics,researchers can use EEG for muscle-based control and induce electronic devices.These designs have attracted extensive attention in the fields of artificial intelligence,equipment manufacturing,etc.Brain science experimental paradigms such as motor imagery and steady-state visual evoking are commonly used in EEG signal recognition tasks.The steady-state visual evoking experimental paradigm requires external stimulation to induce the cerebral cortex for producing associated brain wave signals.However,the motor imagery experimental paradigm does not require stimulation to generate event desynchronization or event synchronization signals in the cortical area.Therefore,the motor imagery experimental paradigm is more easily nested on the brain-computer interface measurements and control equipments.Moreover,with the advancement of signal acquisition technology and synchronization recognition technology,a wide range of engineering applications will be developed.Briefly,there are the following problems in the research of brain wave analysis and recognition: Firstly,the collected brain waves are limited by the existing collection technology with electrical noise and power frequency interference factors.The blinking and muscle artifact in the electroencephalography,interfered by fatigue,attention,environmental artifacts,etc.Besides,brain wave signals have non-stationary electrophysiological properties,and signals collected by brain-computer interfaces have characteristics of the low signal-to-noise ratio.Secondly,feature engineering relies on specific domain knowledge of human expertise,resulting in low efficiency of information analysis.Thirdly,the classification models of deep neural networks are mostly focused on non-real-time visualization of static data,and need to be strengthened in medical interpretability.This work interests in the data sets which are related to motor imaging experimental paradigm and builds a framework of spatio-temporal analysis and recognition :The original brain wave data in electrophysiological is decomposed by the IIR filter.Then,we use a dynamic window and energy image mapping to enhance feature representation.Inspired by montage theory: when different pictures are stitched together,they often produce special meanings that a single picture does not have.Therefore,the energy image mapping layer can map the data of point-in-time,and combine the images of the same task along the time axis.Finally,the cascaded convGRU neural network classifies the action intentions of the motor imagery task.The results show that the proposed framework in this paper has low dependence on the scale of the training data,and it is superior to the benchmark methods in related research in multi-intention recognition.