As I understand you have a dataset of EEG sequences and you want to increase the number of sequences for training.
You can do so by using data augmentations to the EEG signal. The augmentations can be done in the time domain, spatial domain, or frequency domain. You may refer to the below list of selected transformations in the time, spatial, and frequency domains.
1. Instead of using non-overlapping segments for the EEG signal, you can use overlapping segments to increase the sequence count and capture temporal dependencies which benefit sequence labeling models. This is an example of time domain augmentation. You may use the "buffer" function in MATLAB to do so:
segment_length = 1024;
overlap = 512;
segments = buffer(eeg_signal, segment_length, overlap, 'nodelay');
For more information on the "buffer" function, please refer to the following MathWorks documentation link: https://www.mathworks.com/help/releases/R2024a/signal/ref/buffer.html
2. You can also add zero-mean white Gaussian noise to your EEG signal. This will help you model the measurement noise that occurs during the process of capturing the EEG signal. You can use the "randn" function in MATLAB to do so:
standard_deviation = 0.01;
noisy_signal = eeg_signal + standard_deviation * randn(size(eeg_signal));
3. EEG is a multi-channel signal; you can remove some channels to simulate situations where data is unavailable. This will make your model more robust. This is an example of a spatial domain augmentation. You may refer to the below MATLAB code snippet to do so:
channels_to_drop = [3 7];
eeg_signal(channels_to_drop, :) = 0;
4. To further augment your data in the frequency domain, you can use a bandpass filter to selectively remove a small frequency band from all channels. You may use the "bandpass" function in MATLAB to do so. You will need to set the pass-band according to your requirements.
For more information on the "bandpass" function, you may refer to the following MathWorks documentation link:
I hope this is helpful!
