class Press_model(object):
'''
classdocs
'''
def __init__(self,
data,
maximum,
corr_tresh=0.85,
mean_ratio=0.1,
match_rate=1.0,
filter_window=21):
'''
Create a Signature Pattern recognizer Markov based on PRESS's signature-driven method
Params
------
@param data: Training data
@type data: Array-like
@param corr_tresh: Patterns should have a Pearson-Correlation above corr_tresh
@type corr_tresh: float
@param mean_ratio:
'''
self.history = data[:]
self.data = data[:]
self.corr_thres = corr_tresh
self.mr = mean_ratio
self.window = filter_window
self.match = match_rate
self.N = data.shape[0]
self.contain_pat = False
self.warp = False
self.markov = Markov_model(data=data, maximum=maximum)
def fit(self):
'''
Fits a model: First smoothes the data using a Median filter. Then determines if a pattern exists
by calculating the dominant period and splitting the time series into pattern windows.
For each adjacent pair of the pattern widows, the Pearson-Correlation is calculated and compared
to the threshold together with the two pattern means. If the pattern exists for more than the match_ratio
a Signature pattern is calculated using the original data and pattern windows that match.
**Note: If no pattern is found, a 1st order Markov model is trained on the data instead
'''
self.markov.fit()
# Smooth data and dominant period
smooth = signal.medfilt(self.data, self.window)
Z = tsutils.findDominentSeason(self.data)
# Z=288
# Determine the number of pattern windows and split smoothed data
Q = np.int(np.ceil(1.0 * self.N / Z))
if Q > 2:
idx = range(Z, Q * Z, Z)
P = np.array(np.split(smooth, idx))
self.P = P
self.warp = P[-1].shape[0] != P[-2].shape[0]
# Check pattern exist
PC = []
MR = []
for i in range(Q - 1 - self.warp):
PC.append(stats.pearsonr(P[i], P[i + 1])[0])
MR.append(np.mean(P[i]) / np.mean(P[i + 1]))
# Test if pattern exist
if np.all(PC > self.corr_thres):
self.contain_pat = True
Pr = np.array(np.split(self.data, idx))
if self.warp:
self.sig = np.average(Pr[:-1], axis=0)
else:
self.sig = np.average(Pr, axis=0)
def update(self, newData):
'''
Updates the markov's training data and evaluates the pattern again. If a pattern does not exist
call's Markov_models update() function with the new data.
'''
self.data = np.append(self.data[len(newData):], newData)
self.contain_pat = False
self.warp = False
self.fit()
self.markov.update(newData[:])
self.history = np.append(self.history, newData)
def predict(self, fc):
'''
Uses the Signature pattern and determines the index of the last pattern window, using Dynamic Time Warping
and predicts fc samples into the future using the Signature pattern.
**Note: Will return the Markov markov predictions if no pattern was found
'''
forecasts = np.zeros((fc))
if self.contain_pat:
if self.warp:
Z = len(self.sig)
index = getWarpIndex(self.P[-1], self.sig)
if (index + fc) >= Z:
forecasts = np.array(self.sig[index:])
sub = (index + fc) - Z
forecasts = np.append(forecasts, self.sig[:sub])
else:
forecasts = self.sig[index:fc + index]
else:
forecasts = self.sig[:fc]
else:
return self.markov.predict(fc)
return forecasts
class Press_model(object):
'''
classdocs
'''
def __init__(self, data, maximum, corr_tresh=0.85, mean_ratio=0.1, match_rate=1.0, filter_window=21):
'''
Create a Signature Pattern recognizer Markov based on PRESS's signature-driven method
Params
------
@param data: Training data
@type data: Array-like
@param corr_tresh: Patterns should have a Pearson-Correlation above corr_tresh
@type corr_tresh: float
@param mean_ratio:
'''
self.history = data[:]
self.data = data[:]
self.corr_thres = corr_tresh
self.mr = mean_ratio
self.window = filter_window
self.match = match_rate
self.N = data.shape[0]
self.contain_pat = False
self.warp = False
self.markov = Markov_model(data=data, maximum=maximum)
def fit(self):
'''
Fits a model: First smoothes the data using a Median filter. Then determines if a pattern exists
by calculating the dominant period and splitting the time series into pattern windows.
For each adjacent pair of the pattern widows, the Pearson-Correlation is calculated and compared
to the threshold together with the two pattern means. If the pattern exists for more than the match_ratio
a Signature pattern is calculated using the original data and pattern windows that match.
**Note: If no pattern is found, a 1st order Markov model is trained on the data instead
'''
self.markov.fit()
# Smooth data and dominant period
smooth = signal.medfilt(self.data, self.window)
Z = tsutils.findDominentSeason(self.data)
# Z=288
# Determine the number of pattern windows and split smoothed data
Q = np.int(np.ceil(1.0*self.N/Z))
if Q > 2:
idx = range(Z,Q*Z,Z)
P = np.array(np.split(smooth, idx))
self.P = P
self.warp = P[-1].shape[0] != P[-2].shape[0]
# Check pattern exist
PC = []
MR = []
for i in range(Q-1-self.warp):
PC.append(stats.pearsonr(P[i], P[i+1])[0])
MR.append(np.mean(P[i])/np.mean(P[i+1]))
# Test if pattern exist
if np.all(PC>self.corr_thres):
self.contain_pat = True
Pr = np.array(np.split(self.data, idx))
if self.warp:
self.sig = np.average(Pr[:-1],axis=0)
else:
self.sig = np.average(Pr,axis=0)
def update(self, newData):
'''
Updates the markov's training data and evaluates the pattern again. If a pattern does not exist
call's Markov_models update() function with the new data.
'''
self.data = np.append(self.data[len(newData):], newData)
self.contain_pat = False
self.warp = False
self.fit()
self.markov.update(newData[:])
self.history = np.append(self.history, newData)
def predict(self, fc):
'''
Uses the Signature pattern and determines the index of the last pattern window, using Dynamic Time Warping
and predicts fc samples into the future using the Signature pattern.
**Note: Will return the Markov markov predictions if no pattern was found
'''
forecasts = np.zeros((fc))
if self.contain_pat:
if self.warp:
Z = len(self.sig)
index = getWarpIndex(self.P[-1], self.sig)
if (index+fc)>=Z:
forecasts = np.array(self.sig[index:])
sub = (index+fc) - Z
forecasts = np.append(forecasts, self.sig[:sub])
else:
forecasts = self.sig[index:fc+index]
else:
forecasts = self.sig[:fc]
else:
return self.markov.predict(fc)
return forecasts
