def approximate_entropy(self, x, m=None, r=None):
"""
As in tsfresh \
`approximate_entropy <https://github.com/blue-yonder/tsfresh/blob/master/tsfresh/feature_extraction/\
feature_calculators.py#L1601>`_
Implements a `vectorized approximate entropy algorithm <https://en.wikipedia.org/wiki/Approximate_entropy>`_
For short time-series this method is highly dependent on the parameters,
but should be stable for N > 2000, see :cite:`Yentes2013`. Other shortcomings and alternatives discussed in \
:cite:`Richman2000`
:param x: the time series to calculate the feature of
:type x: pandas.Series
:param m: Length of compared run of data
:type m: int
:param r: Filtering level, must be positive
:type r: float
:return: Approximate entropy
:rtype: float
"""
if m is None or r is None:
m = 2
r = 0.3
entropy = feature_calculators.approximate_entropy(x, m, r)
logging.debug("approximate entropy by tsfresh calculated")
return entropy
def get_global_feature(self):
"""
获取时域全局特征，包含最大值、标准差、平均值
:param hadcropped:
:return:
"""
square_data, square_energy, square_azrate = self.pre_process(method='hanning', ifcrop=True)
func = lambda x: [
# feature_calc.autocorrelation(norm(x), 5),
np.std(x),
feature_calc.approximate_entropy(norm(x), 5, 1),
feature_calc.cid_ce(x, normalize=True),
feature_calc.count_above_mean(x),
feature_calc.first_location_of_minimum(x),
feature_calc.first_location_of_maximum(x),
feature_calc.last_location_of_maximum(x),
feature_calc.last_location_of_minimum(x),
feature_calc.longest_strike_above_mean(x),
feature_calc.number_crossing_m(x, 0.8*np.max(x)),
feature_calc.skewness(x),
feature_calc.time_reversal_asymmetry_statistic(x, 5)
]
# global features I want to get
upper_rate = self.get_upper_rate(square_energy)
feature = np.hstack([
[np.mean(norm(square_energy))],
[upper_rate],
func(square_azrate),
func(square_energy)
])
return feature
def TS_features5(signal):
# ts features with
mts = 2
rts = 6
entropy = ts.approximate_entropy(signal, mts, rts)
max_langevin = ts.max_langevin_fixed_point(signal, mts, rts)
return entropy, max_langevin
def get_mfcc_feature(self, hadcropped=False):
'''
calculate Mel-frequency cepstral coefficients in frequency domain and extract features from MFCC
:return: numpy array
'''
assert self.frame_per_second not in [32, 64, 128, 256], \
Exception("Cannot operate butterfly computation ,"
"frame per second should in [32, 64, 128, 256]")
hanning_kernel = self.get_window(method='hanning')
windowed = self._add_window(hanning_kernel, self.meta_audio_data) # [num_frame, kernel_size]
hanning_energy = self.get_energy(self.meta_audio_data, hanning_kernel)
if not hadcropped:
boundary = self.get_boundary(hanning_energy)
cropped = windowed[boundary[0]: boundary[1] + 1, :]
frequency = np.vstack([fft.fft(frame.squeeze()) for frame in np.vsplit(cropped, len(cropped))])
else:
frequency = np.vstack([fft.fft(windowed)])
frequency = np.abs(frequency)
frequency_energy = frequency ** 2
low_freq = self.sr / self.num_per_frame
high_freq = self.sr
H = self._mfcc_filter(self.mfcc_cof, low_freq, high_freq)
S = np.dot(frequency_energy, H.transpose()) # (F, M)
cos_ary = self._discrete_cosine_transform()
mfcc_raw_features = np.sqrt(2 / self.mfcc_cof) * np.dot(S, cos_ary) # （F，N)
upper = [self.get_upper_rate(fea) for fea in mfcc_raw_features.transpose()]
assert len(upper) == mfcc_raw_features.shape[1]
func = lambda x: [
# feature_calc.autocorrelation(norm(x), 5),
np.std(x),
feature_calc.approximate_entropy(norm(x), 5, 1),
feature_calc.cid_ce(x, normalize=True),
feature_calc.count_above_mean(x),
feature_calc.first_location_of_minimum(x),
feature_calc.first_location_of_maximum(x),
feature_calc.last_location_of_maximum(x),
feature_calc.last_location_of_minimum(x),
feature_calc.longest_strike_above_mean(x),
feature_calc.number_crossing_m(x, 0.8*np.max(x)),
feature_calc.skewness(x),
feature_calc.time_reversal_asymmetry_statistic(x, 5)
]
mfcc_features = np.hstack(
[func(col) for col in mfcc_raw_features.transpose()]
)
return mfcc_features
def feature_vector_fun(data, isFun=False, test=False):
trimmed_data = trim_or_pad_data(data, TRIM_DATA_SIZE_FUN)
rX = trimmed_data['rightWrist_x']
normRawColumn = universal_normalization(rX, trimmed_data, x_norm=True)
normRawColumn = general_normalization(normRawColumn)
# Area under curve
auc = np.array([])
auc = np.append(auc, abs(integrate.simps(normRawColumn, dx=5)))
# Absolute Sum of Consecutive Differences
scd = fc.absolute_sum_of_changes(normRawColumn)
# Entropy
entropy = fc.approximate_entropy(normRawColumn, 2, 3)
# AutoCorrelation
ac = fc.autocorrelation(normRawColumn, lag=5)
# Count Above Mean
cam = fc.count_above_mean(normRawColumn)
# Count Below Mean
cbm = fc.count_below_mean(normRawColumn)
featureVector = np.array([])
featureVector = np.append(featureVector, auc)
featureVector = np.append(featureVector, scd)
featureVector = np.append(featureVector, entropy)
featureVector = np.append(featureVector, ac)
featureVector = np.append(featureVector, cam)
featureVector = np.append(featureVector, cbm)
if TRIM_DATA_SIZE_FUN - 1 > featureVector.shape[0]:
featureVector = np.pad(
featureVector,
(0, TRIM_DATA_SIZE_FUN - featureVector.shape[0] - 1), 'constant')
featureVector = featureVector[:TRIM_DATA_SIZE_FUN - 1]
if not test:
if isFun:
featureVector = np.append(featureVector, 1)
else:
featureVector = np.append(featureVector, 0)
return featureVector
def function(x):
return approximate_entropy(x, m=self.m, r=self.r)
def approximate_entropy(current_observation: pd.DataFrame, raw_key: str):
# m Length of compared run of data
# r filtering level
return tsf.approximate_entropy(
current_observation[raw_key], 3,
peak_to_peak_value(current_observation, raw_key))
