【信号识别-脑电分类】基于EASRC和ELM算法实现音频脑电图 (EEG) 数据分类附matlab代码.zip

# Classification-of-Pitch-Imagery-from-EEG This project is based on classification of audio electroencephalogram (EEG) data for 3 pitches: C4, D4, E4. Classification was done using Extreme Learning Machine Model in Matlab. The data available was preprocessed according to the following description: 1. Original EEG raw data, 32 channels, .edf format, 5 subjects, 3 pitches. 2. Pre-processed data ran on EEGLab, extension from Matlab on the following process, saved in .set & .txt formats: - Retain of channels for auditory imagery: F7, F8, AF3, AF4 - Retain of channels for auditory perception: T7, T8, TP7, TP8 - Passband filter to remove noise: 0.1-100Hz - Notch Filter: 60Hz - Epoch Extraction at time -0.25 to 1.5 seconds - Baseline Removal [-252 0] - Artifact Removal using ICA - Alpha band IIR filter (8 - 15Hz) - Automatic Epoch Rejection of Voltage over 100uV ## Approach for Data preprocessing and extraction: .edf files for each pitch, c4,d4,e4 was respectively put into seperate folders and all converted to .mat files using the code file attached ‘readedf.m’. Usage: ``` [header, data] = readedf('filename.edf'); ``` These mat files were then loaded into matlab and data for ⅘ subjects were concatenated and labelled for training dataset. Labels are such as C4 => 0, D4 => 1, E4 =>2. The data for 8/32 channels was then extracted for channels of auditory imagery: F7, F8, AF3, AF4 and Auditory perception: T7, T8, TP7, TP8. ``` d = filepath; ff = what(d); con=[]; for j =1:length(ff.mat) fname = strcat(d, ff.mat{j}); M=load(fname); temp=M.data; %for concatenating individual files con=[con, temp]; %for concatenating sets %con=vertcat(con, temp); end con = con'; %comment it out for concatenating sets ``` These files were all combined together to make the final training dataset attached as ‘Training_dataset.mat’. Rest of the ⅕ subject’s data was concatenated, labelled and extracted on 8 channels successively and put into ‘Test_dataset.mat’. ``` %labelling data = load(filepath); label = 2; data = (data.data)'; [r,c]= size(data); z = label.* ones(r,1); data=horzcat(z,data); ``` Each dataset has the structure of labels in column 1 and data points of 8 channels in next 8 columns. Data of each subject is concatenated vertically. ## Training and testing of model: ELM implementation of Mathworks was utilized and modified for our problem. The code files are available in the EA_SRC.zip along with the datasets. The code file ‘Demo_Classification.m’ takes our training and test datasets to perform classification on the ELM and Regularized ELM. ``` load('Training_dataset.mat'); %loading Training_dataset con = con'; %transpose of dataset into column vectors trainlabel=con(1,:); %extract labels from first column traindata=con(2:9,:); %extract data values from rest ``` ``` load('Test_dataset.mat'); %same as above steps for test dataset con = con'; testlabel=con(1,:); testdata=con(2:9,:); ``` **ELM and RELM was performed using an Neural network trained on our labels.** ``` method = {'ELM','RELM'}; nn.hiddensize = 200; %number of hidden neurons nn.activefunction = 's'; %sigmoid nn.inputsize = size(traindata,1); nn.method = method{1}; %pick ELM or RELM nn.type = 'classification'; ``` ## Improvements: The results can be improved by: - Increasing the number of hidden layers on the neural networks. Currently it uses 200 layers. - Using a larger training dataset. This dataset was of only 4 subjects (approx. #50000) reduced to 8 channels and 5th subject was used as test data. The results can be much better with a larger dataset. - Using a deep convolutional network for multi label classification.