def get_value(self, inject_at_index, ts_id):
"""
:param inject_at_index: the index to inject the anomaly at
:param ts_id: the affected time series
:return: the value to add to the current value at index 'inject_at_index'
"""
frequency = self.df.index.inferred_freq # the granularity
before_dt = next_earlier_dt(inject_at_index, frequency, 10)
after_dt = next_later_dt(inject_at_index, frequency, 10)
index_min = self.df.index.min()
index_max = self.df.index.max()
lower_boundary = max(index_min, before_dt)
upper_boundary = min(after_dt, index_max)
injection_series = self.df.loc[lower_boundary:upper_boundary, ts_id]
# remove indexes which are already an anomaly, we won't consider them for the deviation value as it would change the outcome significantly
df_anomaly_part = self.df_class.loc[(self.df_class.index.isin(injection_series.index)) & (self.df_class[ts_id] == True), ts_id]
injection_series = injection_series.drop(index=df_anomaly_part.index)
# check the remaining length
if injection_series.index.size == 0:
logging.debug(msg_injection_all_anomalies())
return 0
local_std = injection_series.std(axis=0, skipna=True, level=None, ddof=0)
return np.random.choice([-1, 1]) * self.get_factor() * local_std
def inject(self, range):
ts_id = self.get_time_series().id
inject_at_index = self.next_injection_index(range)
if inject_at_index is not None:
upper_boundary = min(next_later_dt(inject_at_index, self.df.index.inferred_freq, self.get_factor()), self.df.index.max())
trend_indexes = pd.date_range(inject_at_index, upper_boundary, freq=self.df.index.inferred_freq)
# injection
self.df_inject.loc[trend_indexes, ts_id] = 0
self.df_inject_class.loc[trend_indexes, ts_id] = 1
def inject(self, range_index):
ts_id = self.get_time_series().id
inject_at_index = self.next_injection_index(range_index)
if inject_at_index is not None:
lower_boundary = max(next_earlier_dt(inject_at_index, self.df.index.inferred_freq, 4), self.df.index.min())
upper_boundary = min(next_later_dt(inject_at_index, self.df.index.inferred_freq, 5), self.df.index.max())
level_shift_indexes = pd.date_range(lower_boundary, upper_boundary, freq=self.df.index.inferred_freq)
adjustment_value = self.get_value(inject_at_index, ts_id)
if adjustment_value != 0:
self.df_inject.loc[level_shift_indexes, ts_id] += adjustment_value
self.df_inject_class.loc[level_shift_indexes, ts_id] = 1
def inject(self, range):
ts_id = self.get_time_series().id
inject_at_index = self.next_injection_index(range)
if inject_at_index is not None:
lower_boundary_before = max(next_earlier_dt(inject_at_index, self.df.index.inferred_freq, 5), self.df.index.min())
if lower_boundary_before == self.df.index.min():
lower_boundary = next_later_dt(lower_boundary_before, self.df.index.inferred_freq, 1)
else:
lower_boundary = max(next_earlier_dt(inject_at_index, self.df.index.inferred_freq, 4), self.df.index.min())
upper_boundary = min(next_later_dt(inject_at_index, self.df.index.inferred_freq, 5), self.df.index.max())
variance_diff_indexes = pd.date_range(lower_boundary_before, upper_boundary, freq=self.df.index.inferred_freq)
variance_indexes = pd.date_range(lower_boundary, upper_boundary, freq=self.df.index.inferred_freq)
difference = np.diff(self.df.loc[variance_diff_indexes, ts_id])
# distortion
self.df_inject.loc[variance_indexes, ts_id] += (self.get_factor() - 1) * difference
self.df_inject_class.loc[variance_indexes, ts_id] = 1
def inject(self, range):
ts_id = self.get_time_series().id
inject_at_index = self.next_injection_index(range)
if inject_at_index is not None:
upper_boundary = min(
next_later_dt(inject_at_index, self.df.index.inferred_freq, 9),
self.df.index.max())
growth_indexes = pd.date_range(inject_at_index,
upper_boundary,
freq=self.df.index.inferred_freq)
slope = np.random.choice([-1, 1]) * self.get_factor() * np.arange(
len(growth_indexes))
# trend
self.df_inject.loc[growth_indexes, ts_id] += slope
self.df_inject_class.loc[growth_indexes, ts_id] = 1
# adjust remaining
self.df_inject.loc[self.df_inject.index > upper_boundary,
ts_id] += slope[-1]