def create_all_psd():
f = pylab.linspace(0, 1, 4096)
pylab.figure(figsize=(12,8))
# MA model
p = spectrum.pma(xx, 64,128); p(); p.plot()
"""
#ARMA 15 order
a, b, rho = spectrum.arma_estimate(data, 15,15, 30)
psd = spectrum.arma2psd(A=a,B=b, rho=rho)
newpsd = tools.cshift(psd, len(psd)//2) # switch positive and negative freq
pylab.plot(f, 10 * pylab.log10(newpsd/max(newpsd)), label='ARMA 15,15')
"""
# YULE WALKER
p = spectrum.pyule(xx, 7 , NFFT=4096, scale_by_freq=False); p.plot()
# equivalent to
# plot([x for x in p.frequencies()] , 10*log10(p.psd)); grid(True)
#burg method
p = spectrum.pburg(xx, 7, scale_by_freq=False); p.plot()
#pcovar
p = spectrum.pcovar(xx, 7, scale_by_freq=False); p.plot()
#pmodcovar
p = spectrum.pmodcovar(xx, 7, scale_by_freq=False); p.plot()
# correlogram
p = spectrum.pcorrelogram(xx, lag=60, NFFT=512, scale_by_freq=False); p.plot()
# minvar
p = spectrum.pminvar(xx, 7, NFFT=256, scale_by_freq=False); p.plot()
# pmusic
p = spectrum.pmusic(xx, 10,4, scale_by_freq=False); p.plot()
# pmusic
p = spectrum.pev(xx, 10, 4, scale_by_freq=False); p.plot()
# periodogram
p = spectrum.Periodogram(xx, scale_by_freq=False); p.plot()
#
legend( ["MA 32", "pyule 7", "pburg 7", "pcovar", "pmodcovar", "correlogram",
"minvar", "pmusic", "pev", "periodgram"])
pylab.ylim([-80,80])
def create_all_psd():
f = pylab.linspace(0, 1, 4096)
pylab.figure(figsize=(12, 8))
# MA model
p = spectrum.pma(xx, 64, 128)
p()
p.plot()
"""
#ARMA 15 order
a, b, rho = spectrum.arma_estimate(data, 15,15, 30)
psd = spectrum.arma2psd(A=a,B=b, rho=rho)
newpsd = tools.cshift(psd, len(psd)//2) # switch positive and negative freq
pylab.plot(f, 10 * pylab.log10(newpsd/max(newpsd)), label='ARMA 15,15')
"""
# YULE WALKER
p = spectrum.pyule(xx, 7, NFFT=4096, scale_by_freq=False)
p.plot()
# equivalent to
# plot([x for x in p.frequencies()] , 10*log10(p.psd)); grid(True)
#burg method
p = spectrum.pburg(xx, 7, scale_by_freq=False)
p.plot()
#pcovar
p = spectrum.pcovar(xx, 7, scale_by_freq=False)
p.plot()
#pmodcovar
p = spectrum.pmodcovar(xx, 7, scale_by_freq=False)
p.plot()
# correlogram
p = spectrum.pcorrelogram(xx, lag=60, NFFT=512, scale_by_freq=False)
p.plot()
# minvar
p = spectrum.pminvar(xx, 7, NFFT=256, scale_by_freq=False)
p.plot()
# pmusic
p = spectrum.pmusic(xx, 10, 4, scale_by_freq=False)
p.plot()
# pmusic
p = spectrum.pev(xx, 10, 4, scale_by_freq=False)
p.plot()
# periodogram
p = spectrum.Periodogram(xx, scale_by_freq=False)
p.plot()
#
legend([
"MA 32", "pyule 7", "pburg 7", "pcovar", "pmodcovar", "correlogram",
"minvar", "pmusic", "pev", "periodgram"
])
pylab.ylim([-80, 80])
from scipy.io import loadmat
import spectrum
from matplotlib import pylab as pl
import numpy as np
import sys
if __name__ == "__main__":
d = loadmat(sys.argv[1])
o1 = signal.detrend(d['raw'][0])
o2 = signal.detrend(d['raw'][1])
diff = o1 - o2
avg = (o1 + o2) / 2.0
data_labels = ["O1", "O2", "O1-O2", "(O1+O2)/2"]
datas = [o1, o2, diff, avg]
fig, axs = pl.subplots(nrows=1, ncols=4, sharey=True)
time = d['data']['time'][0][0][0][0]
for i in range(4):
psd = spectrum.pburg(datas[i], order=8, NFFT=256, sampling=128.0)
psd.run()
p = 10 * np.log10(psd.get_converted_psd('onesided'))
axs[i].set_title(data_labels[i])
axs[i].plot(psd.frequencies(), p)
pl.show()
def _calc_pburg_psd(rri, fs, order=16, nfft=None):
burg = pburg(data=rri, order=order, NFFT=nfft, sampling=fs)
burg.scale_by_freq = False
burg()
return np.array(burg.frequencies()), burg.psd
def feature_band_selection(data,
labels,
sfreq,
step=1,
band_range=(0, 0),
band_size=(0, 0, 0),
channel=('EEG:C3', 'EEG:Cz', 'EEG:C4'),
features_type=0):
# AR model parameters
ar_order = 12
nfft = 1000
subject_optimal_frequency_bands = dict()
# mu band selection
for subject in data:
if features_type == 1: # compute AR Model PSD based on burg algorithm
ar_psd = dict()
freq_flag = True
for channel_name in channel:
ar_psd[channel_name] = list()
for idx in range(len(labels[subject])):
for channel_name in channel:
x = data[subject][channel_name][idx]
p = pburg(x,
order=ar_order,
NFFT=nfft,
sampling=sfreq,
scale_by_freq=True)
if freq_flag:
ar_psd['frequency'] = np.array(p.frequencies())
freq_flag = False
ar_psd[channel_name].append(p.psd)
f_score = list()
optimal_band = list()
for band in band_size:
for num_windows in range(
int((band_range[1] - band_range[0] - band) / step)):
lowcut = band_range[0] + num_windows * step
highcut = lowcut + band
optimal_band.append((lowcut, highcut))
if features_type == 1:
ar_freq = ar_psd['frequency']
psd_idx_start = np.where(ar_freq >= lowcut)[0][0]
psd_idx_end = np.where(ar_freq >= highcut)[0][0]
left_features = list()
right_features = list()
for idx in range(len(labels[subject])):
features = list()
for channel_name in channel:
# BP features
if features_type == 0: # 5th butterworth filter
filtered_data = butter_bandpass_filter(
data[subject][channel_name][idx],
lowcut,
highcut,
sfreq,
order=5)
elif features_type == 1: # AR model PSD
filtered_data = ar_psd[channel_name][idx][
psd_idx_start:psd_idx_end]
else:
raise Exception(
"feature type wrong!\n band pass features: features_type=0\n "
"AR PSD features: features_type=1")
features.append(math.log10(np.var(filtered_data)))
if labels[subject][idx] == 1:
left_features.append(features)
elif labels[subject][idx] == 2:
right_features.append(features)
left_mean_val = np.mean(left_features, axis=0)
right_mean_val = np.mean(right_features, axis=0)
left_var = np.var(left_features, axis=0)
right_var = np.var(right_features, axis=0)
f_score.append(
sum(np.square(left_mean_val - right_mean_val)) /
sum(left_var + right_var))
# get optimal frequency corresponding to max F-score
subject_optimal_frequency_bands[subject] = optimal_band[f_score.index(
max(f_score))]
# pause = input("pause")
return subject_optimal_frequency_bands