Steady-State Visual-Evoked-Potential-Driven Quadrotor Control Using a Deep Residual CNN for Short-Time Signal...

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Classifying short-time-window steady-state visual evoked potentials (SSVEPs) is challenging due to difficulties in distinguishing frequency-domain features.
• Existing methods may struggle with the subtle differences in SSVEP signals within brief time windows.

Purpose of the Study:

• To propose a novel deep convolutional network, EEGResNet, for improved SSVEP classification.
• To leverage residual connections and time-domain feature extraction for enhanced performance.

Main Methods:

• Developed EEGResNet, a deep convolutional network incorporating residual connections.
• Employed a filter bank (FB) module with four Butterworth filters (19-50 Hz, 14-38 Hz, 9-26 Hz, 3-14 Hz) for time-domain feature extraction.
• Aggregated FB features and processed them through a six-layer convolutional neural network with residual connections.

Main Results:

• Demonstrated the effectiveness and superiority of EEGResNet through experiments on two large public datasets.
• Achieved improved classification performance for short-time-window SSVEPs.

Conclusions:

• EEGResNet offers a promising approach for accurate SSVEP classification, particularly for short time windows.
• The trained network shows potential for applications like brain-computer interface (BCI) controlled quadrotor systems, as validated by virtual simulation.