"""
The monitoring module contains the PerformanceMonitoring class used to run
quality control tests and store results. The module also contains individual
functions that can be used to run quality control tests.
"""
import pandas as pd
import numpy as np
import logging
none_list = ['','none','None','NONE', None, [], {}]
NoneType = type(None)
logger = logging.getLogger(__name__)
def _documented_by(original):
def wrapper(target):
docstring = original.__doc__
old = """
Parameters
----------
"""
new = """
Parameters
----------
data : pandas DataFrame
Data used in the quality control test, indexed by datetime
"""
new_docstring = docstring.replace(old, new) + \
"""
Returns
----------
dictionary
Results include cleaned data, mask, and test results summary
"""
target.__doc__ = new_docstring
return target
return wrapper
### Object-oriented approach
class PerformanceMonitoring(object):
def __init__(self):
"""
PerformanceMonitoring class
"""
self.df = pd.DataFrame()
self.trans = {}
self.tfilter = pd.Series()
self.test_results = pd.DataFrame(columns=['Variable Name',
'Start Time', 'End Time',
'Timesteps', 'Error Flag'])
@property
def mask(self):
"""
Boolean mask indicating data that failed a quality control test
Returns
--------
pandas DataFrame
Boolean values for each data point,
True = data point pass all tests,
False = data point did not pass at least one test (or data is NaN).
"""
if self.df.empty:
logger.info("Empty database")
return
mask = ~pd.isnull(self.df) # False if NaN
for i in self.test_results.index:
variable = self.test_results.loc[i, 'Variable Name']
start_date = self.test_results.loc[i, 'Start Time']
end_date = self.test_results.loc[i, 'End Time']
if variable in mask.columns:
try:
mask.loc[start_date:end_date,variable] = False
except:
pass
return mask
@property
def cleaned_data(self):
"""
Cleaned data set
Returns
--------
pandas DataFrame
Cleaned data set, data that failed a quality control test are
replaced by NaN
"""
return self.df[self.mask]
def _setup_data(self, key):
"""
Setup data to use in the quality control test
"""
if self.df.empty:
logger.info("Empty database")
return
# Isolate subset if key is not None
if key is not None:
try:
df = self.df[self.trans[key]]
except:
logger.warning("Undefined key: " + key)
return
else:
df = self.df
return df
def _generate_test_results(self, df, bound, min_failures, error_prefix):
"""
Compare DataFrame to bounds to generate a True/False mask where
True = passed, False = failed. Append results to test_results.
"""
# Lower Bound
if bound[0] not in none_list:
mask = (df < bound[0])
error_msg = error_prefix+' < lower bound, '+str(bound[0])
self._append_test_results(mask, error_msg, min_failures)
# Upper Bound
if bound[1] not in none_list:
mask = (df > bound[1])
error_msg = error_prefix+' > upper bound, '+str(bound[1])
self._append_test_results(mask, error_msg, min_failures)
def _append_test_results(self, mask, error_msg, min_failures=1, use_mask_only=False):
"""
Append QC results to the PerformanceMonitoring object.
Parameters
----------
mask : pandas DataFrame
Result from quality control test, boolean values
error_msg : string
Error message to store with the QC results
min_failures : int (optional)
Minimum number of consecutive failures required for reporting,
default = 1
use_mask_only : boolean (optional)
When True, the mask is used directly to determine test
results and the variable name is not included in the
test_results. When False, the mask is used in combination with
pm.df to extract test results. Default = False
"""
if not self.tfilter.empty:
mask[~self.tfilter] = False
if mask.sum(axis=1).sum(axis=0) == 0:
return
if use_mask_only:
sub_df = mask
else:
sub_df = self.df[mask.columns]
# Find blocks
order = 'col'
if order == 'col':
mask = mask.T
np_mask = mask.values
start_nans_mask = np.hstack(
(np.resize(np_mask[:,0],(mask.shape[0],1)),
np.logical_and(np.logical_not(np_mask[:,:-1]), np_mask[:,1:])))
stop_nans_mask = np.hstack(
(np.logical_and(np_mask[:,:-1], np.logical_not(np_mask[:,1:])),
np.resize(np_mask[:,-1], (mask.shape[0],1))))
start_row_idx,start_col_idx = np.where(start_nans_mask)
stop_row_idx,stop_col_idx = np.where(stop_nans_mask)
if order == 'col':
temp = start_row_idx; start_row_idx = start_col_idx; start_col_idx = temp
temp = stop_row_idx; stop_row_idx = stop_col_idx; stop_col_idx = temp
#mask = mask.T
block = {'Start Row': list(start_row_idx),
'Start Col': list(start_col_idx),
'Stop Row': list(stop_row_idx),
'Stop Col': list(stop_col_idx)}
#if sub_df is None:
# sub_df = self.df
for i in range(len(block['Start Col'])):
length = block['Stop Row'][i] - block['Start Row'][i] + 1
if length >= min_failures:
if use_mask_only:
var_name = ''
else:
var_name = sub_df.iloc[:,block['Start Col'][i]].name #sub_df.icol(block['Start Col'][i]).name
frame = pd.DataFrame([var_name,
sub_df.index[block['Start Row'][i]],
sub_df.index[block['Stop Row'][i]],
length, error_msg],
index=['Variable Name', 'Start Time',
'End Time', 'Timesteps', 'Error Flag'])
self.test_results = self.test_results.append(frame.T, ignore_index=True)
def add_dataframe(self, data):
"""
Add data to the PerformanceMonitoring object
Parameters
-----------
data : pandas DataFrame
Data to add to the PerformanceMonitoring object, indexed by datetime
"""
assert isinstance(data, pd.DataFrame), 'data must be of type pd.DataFrame'
assert isinstance(data.index, pd.core.indexes.datetimes.DatetimeIndex), 'data.index must be a DatetimeIndex'
temp = data.copy()
if self.df is not None:
self.df = temp.combine_first(self.df)
else:
self.df = temp
# Add identity 1:1 translation dictionary
trans = {}
for col in temp.columns:
trans[col] = [col]
self.add_translation_dictionary(trans)
def add_translation_dictionary(self, trans):
"""
Add translation dictionary to the PerformanceMonitoring object
Parameters
-----------
trans : dictionary
Translation dictionary
"""
assert isinstance(trans, dict), 'trans must be of type dictionary'
for key, values in trans.items():
self.trans[key] = []
for value in values:
self.trans[key].append(value)
def add_time_filter(self, time_filter):
"""
Add a time filter to the PerformanceMonitoring object
Parameters
----------
time_filter : pandas DataFrame with a single column or pandas Series
Time filter containing boolean values for each time index
"""
assert isinstance(time_filter, (pd.Series, pd.DataFrame)), 'time_filter must be of type pd.Series or pd.DataFrame'
if isinstance(time_filter, pd.DataFrame):
self.tfilter = pd.Series(data = time_filter.values[:,0], index = self.df.index)
else:
self.tfilter = time_filter
def check_timestamp(self, frequency, expected_start_time=None,
expected_end_time=None, min_failures=1,
exact_times=True):
"""
Check time series for missing, non-monotonic and duplicate
timestamps
Parameters
----------
frequency : int or float
Expected time series frequency, in seconds
expected_start_time : Timestamp (optional)
Expected start time. If not specified, the minimum timestamp
is used
expected_end_time : Timestamp (optional)
Expected end time. If not specified, the maximum timestamp
is used
min_failures : int (optional)
Minimum number of consecutive failures required for
reporting, default = 1
exact_times : bool (optional)
Controls how missing times are checked.
If True, times are expected to occur at regular intervals
(specified in frequency) and the DataFrame is reindexed to match
the expected frequency.
If False, times only need to occur once or more within each
interval (specified in frequency) and the DataFrame is not
reindexed.
"""
assert isinstance(frequency, (int, float)), 'frequency must be of type int or float'
assert isinstance(expected_start_time, (NoneType, pd.Timestamp)), 'expected_start_time must be None or of type pd.Timestamp'
assert isinstance(expected_end_time, (NoneType, pd.Timestamp)), 'expected_end_time must be None or of type pd.Timestamp'
assert isinstance(min_failures, int), 'min_failures must be of type int'
assert isinstance(exact_times, bool), 'exact_times must be of type bool'
logger.info("Check timestamp")
if self.df.empty:
logger.info("Empty database")
return
if expected_start_time is None:
expected_start_time = min(self.df.index)
if expected_end_time is None:
expected_end_time = max(self.df.index)
rng = pd.date_range(start=expected_start_time, end=expected_end_time,
freq=str(int(frequency*1e3)) + 'ms') # milliseconds
# Check to see if timestamp is monotonic
# mask = pd.TimeSeries(self.df.index).diff() < 0
mask = pd.Series(self.df.index).diff() < pd.Timedelta('0 days 00:00:00')
mask.index = self.df.index
mask[mask.index[0]] = False
mask = pd.DataFrame(mask)
mask.columns = [0]
self._append_test_results(mask, 'Nonmonotonic timestamp',
use_mask_only=True,
min_failures=min_failures)
# If not monotonic, sort df by timestamp
if not self.df.index.is_monotonic:
self.df = self.df.sort_index()
# Check for duplicate timestamps
# mask = pd.TimeSeries(self.df.index).diff() == 0
mask = pd.Series(self.df.index).diff() == pd.Timedelta('0 days 00:00:00')
mask.index = self.df.index
mask[mask.index[0]] = False
mask = pd.DataFrame(mask)
mask.columns = [0]
mask['TEMP'] = mask.index # remove duplicates in the mask
mask.drop_duplicates(subset='TEMP', keep='last', inplace=True)
del mask['TEMP']
# Drop duplicate timestamps (this has to be done before the
# results are appended)
self.df['TEMP'] = self.df.index
#self.df.drop_duplicates(subset='TEMP', take_last=False, inplace=True)
self.df.drop_duplicates(subset='TEMP', keep='first', inplace=True)
self._append_test_results(mask, 'Duplicate timestamp',
use_mask_only=True,
min_failures=min_failures)
del self.df['TEMP']
if exact_times:
temp = pd.Index(rng)
missing = temp.difference(self.df.index).tolist()
# reindex DataFrame
self.df = self.df.reindex(index=rng)
mask = pd.DataFrame(data=self.df.shape[0]*[False],
index=self.df.index)
mask.loc[missing] = True
self._append_test_results(mask, 'Missing timestamp',
use_mask_only=True,
min_failures=min_failures)
else:
# uses pandas >= 0.18 resample syntax
df_index = pd.DataFrame(index=self.df.index)
df_index[0]=1 # populate with placeholder values
mask = df_index.resample(str(int(frequency*1e3))+'ms').count() == 0 # milliseconds
self._append_test_results(mask, 'Missing timestamp',
use_mask_only=True,
min_failures=min_failures)
def check_range(self, bound, key=None, min_failures=1):
"""
Check for data that is outside expected range
Parameters
----------
bound : list of floats
[lower bound, upper bound], None can be used in place of a lower
or upper bound
key : string (optional)
Data column name or translation dictionary key. If not specified,
all columns are used in the test.
min_failures : int (optional)
Minimum number of consecutive failures required for reporting,
default = 1
"""
assert isinstance(bound, list), 'bound must be of type list'
assert isinstance(key, (NoneType, str)), 'key must be None or of type string'
assert isinstance(min_failures, int), 'min_failures must be of type int'
logger.info("Check for data outside expected range")
df = self._setup_data(key)
if df is None:
return
error_prefix = 'Data'
self._generate_test_results(df, bound, min_failures, error_prefix)
def check_increment(self, bound, key=None, increment=1, absolute_value=True,
min_failures=1):
"""
Check data increments using the difference between values
Parameters
----------
bound : list of floats
[lower bound, upper bound], None can be used in place of a lower
or upper bound
key : string (optional)
Data column name or translation dictionary key. If not specified,
all columns are used in the test.
increment : int (optional)
Time step shift used to compute difference, default = 1
absolute_value : boolean (optional)
Use the absolute value of the increment data, default = True
min_failures : int (optional)
Minimum number of consecutive failures required for reporting,
default = 1
"""
assert isinstance(bound, list), 'bound must be of type list'
assert isinstance(key, (NoneType, str)), 'key must be None or of type string'
assert isinstance(increment, int), 'increment must be of type int'
assert isinstance(absolute_value, bool), 'absolute_value must be of type bool'
assert isinstance(min_failures, int), 'min_failures must be of type int'
logger.info("Check for data increment outside expected range")
df = self._setup_data(key)
if df is None:
return
if df.isnull().all().all():
logger.warning("Check increment range failed (all data is Null): " + key)
return
# Compute interval
if absolute_value:
df = np.abs(df.diff(periods=increment))
else:
df = df.diff(periods=increment)
if absolute_value:
error_prefix = '|Increment|'
else:
error_prefix = 'Increment'
self._generate_test_results(df, bound, min_failures, error_prefix)
def check_delta(self, bound, key=None, window=3600, direction=None,
min_failures=1):
"""
Check for stagnant data and/or abrupt changes in the data using the
difference between max and min values (delta) within a rolling window
Parameters
----------
bound : list of floats
[lower bound, upper bound], None can be used in place of a lower
or upper bound
key : string (optional)
Data column name or translation dictionary key. If not specified,
all columns are used in the test.
window : int or float (optional)
Size of the rolling window (in seconds) used to compute delta,
default = 3600
direction : str (optional)
Options = 'positive', 'negative', or None
* If direction is positive, then only identify positive deltas
(the min occurs before the max)
* If direction is negative, then only identify negative deltas
(the max occurs before the min)
* If direction is None, then identify both positive and negative
deltas
min_failures : int (optional)
Minimum number of consecutive failures required for reporting,
default = 1
"""
assert isinstance(bound, list), 'bound must be of type list'
assert isinstance(key, (NoneType, str)), 'key must be None or of type string'
assert isinstance(window, (int, float)), 'window must be of type int or float'
assert direction in [None, 'positive', 'negative'], "direction must None or the string 'positive' or 'negative'"
assert isinstance(min_failures, int), 'min_failures must be of type int'
assert self.df.index.is_monotonic, 'index must be monotonic'
logger.info("Check for stagant data and/or abrupt changes using delta (max-min) within a rolling window")
df = self._setup_data(key)
if df is None:
return
window_str = str(int(window*1e3)) + 'ms' # milliseconds
min_df = df.rolling(window_str, min_periods=2, closed='both').min()
max_df = df.rolling(window_str, min_periods=2, closed='both').max()
diff_df = max_df - min_df
diff_df.loc[diff_df.index[0]:diff_df.index[0]+pd.Timedelta(window_str),:] = None
def update_mask(mask1, df, window_str, bound, direction):
# While the mask flags data at the time at which the failure occurs,
# the actual timespan betwen the min and max should be flagged so that
# the final results include actual data points that caused the failure.
# This function uses numpy arrays to improve performance and returns
# a mask DataFrame.
mask2 = np.zeros((len(mask1.index), len(mask1.columns)), dtype=bool)
index = mask1.index
# Loop over t, col in mask1 where condition is True
for t,col in list(mask1[mask1 > 0].stack().index):
icol = mask1.columns.get_loc(col)
it = mask1.index.get_loc(t)
t1 = t-pd.Timedelta(window_str)
if (bound == 'lower') and (direction is None):
# set the entire time interval to True
mask2[(index >= t1) & (index <= t),icol] = True
else:
# extract the min and max time
min_time = df.loc[t1:t,col].idxmin()
max_time = df.loc[t1:t,col].idxmax()
if bound == 'lower': # bound = upper, direction = positive or negative
# set the entire time interval to True
if (direction == 'positive') and (min_time <= max_time):
mask2[(index >= t1) & (index <= t),icol] = True
elif (direction == 'negative') and (min_time >= max_time):
mask2[(index >= t1) & (index <= t),icol] = True
elif bound == 'upper': # bound = upper, direction = None, positive or negative
# set the initially flaged location to False
mask2[it,icol] = False
# set the time between max/min or min/max to true
if min_time < max_time and (direction is None or direction == 'positive'):
mask2[(index >= min_time) & (index <= max_time),icol] = True
elif min_time > max_time and (direction is None or direction == 'negative'):
mask2[(index >= max_time) & (index <= min_time),icol] = True
elif min_time == max_time:
mask2[it,icol] = True
mask2 = pd.DataFrame(mask2, columns=mask1.columns, index=mask1.index)
return mask2
if direction == 'positive':
error_prefix = 'Delta (+)'
elif direction == 'negative':
error_prefix = 'Delta (-)'
else:
error_prefix = 'Delta'
# Lower Bound
if bound[0] not in none_list:
mask = (diff_df < bound[0])
error_msg = error_prefix+' < lower bound, '+str(bound[0])
if not self.tfilter.empty:
mask[~self.tfilter] = False
mask = update_mask(mask, df, window_str, 'lower', direction)
self._append_test_results(mask, error_msg, min_failures)
# Upper Bound
if bound[1] not in none_list:
mask = (diff_df > bound[1])
error_msg = error_prefix+' > upper bound, '+str(bound[1])
if not self.tfilter.empty:
mask[~self.tfilter] = False
mask = update_mask(mask, df, window_str, 'upper', direction)
self._append_test_results(mask, error_msg, min_failures)
def check_outlier(self, bound, key=None, window=3600, absolute_value=True,
min_failures=1):
"""
Check for outliers using normalized data within a rolling window
The upper and lower bounds are specified in standard deviations.
Data normalized using (data-mean)/std.
Parameters
----------
bound : list of floats
[lower bound, upper bound], None can be used in place of a lower
or upper bound
key : string (optional)
Data column name or translation dictionary key. If not specified,
all columns are used in the test.
window : int or float (optional)
Size of the rolling window (in seconds) used to normalize data,
default = 3600. If window is set to None, data is normalized using
the entire data sets mean and standard deviation (column by column).
absolute_value : boolean (optional)
Use the absolute value the normalized data, default = True
min_failures : int (optional)
Minimum number of consecutive failures required for reporting,
default = 1
"""
assert isinstance(bound, list), 'bound must be of type list'
assert isinstance(key, (NoneType, str)), 'key must be None or of type string'
assert isinstance(window, (NoneType, int, float)), 'window must be None or of type int or float'
assert isinstance(absolute_value, bool), 'absolute_value must be of type bool'
assert isinstance(min_failures, int), 'min_failures must be type int'
assert self.df.index.is_monotonic, 'index must be monotonic'
logger.info("Check for outliers")
df = self._setup_data(key)
if df is None:
return
# Compute normalized data
if window is not None:
window_str = str(int(window*1e3)) + 'ms' # milliseconds
df_mean = df.rolling(window_str, min_periods=2, closed='both').mean()
df_std = df.rolling(window_str, min_periods=2, closed='both').std()
df = (df - df_mean)/df_std
else:
df = (df - df.mean())/df.std()
if absolute_value:
df = np.abs(df)
df.replace([np.inf, -np.inf], np.nan, inplace=True)
if absolute_value:
error_prefix = '|Outlier|'
else:
error_prefix = 'Outlier'
#df[df.index[0]:df.index[0]+datetime.timedelta(seconds=window)] = np.nan
self._generate_test_results(df, bound, min_failures, error_prefix)
def check_missing(self, key=None, min_failures=1):
"""
Check for missing data
Parameters
----------
key : string (optional)
Data column name or translation dictionary key. If not specified,
all columns are used in the test.
min_failures : int (optional)
Minimum number of consecutive failures required for reporting,
default = 1
"""
assert isinstance(key, (NoneType, str)), 'key must be None or of type string'
assert isinstance(min_failures, int), 'min_failures must be type int'
logger.info("Check for missing data")
df = self._setup_data(key)
if df is None:
return
# Extract missing data
mask = pd.isnull(df) # checks for np.nan, np.inf
missing_timestamps = self.test_results[
self.test_results['Error Flag'] == 'Missing timestamp']
for index, row in missing_timestamps.iterrows():
mask.loc[row['Start Time']:row['End Time']] = False
self._append_test_results(mask, 'Missing data', min_failures=min_failures)
def check_corrupt(self, corrupt_values, key=None, min_failures=1):
"""
Check for corrupt data
Parameters
----------
corrupt_values : list of int or floats
List of corrupt data values
key : string (optional)
Data column name or translation dictionary key. If not specified,
all columns are used in the test.
min_failures : int (optional)
Minimum number of consecutive failures required for reporting,
default = 1
"""
assert isinstance(corrupt_values, list), 'corrupt_values must be of type list'
assert isinstance(key, (NoneType, str)), 'key must be None or of type string'
assert isinstance(min_failures, int), 'min_failures must be type int'
logger.info("Check for corrupt data")
df = self._setup_data(key)
if df is None:
return
# Extract corrupt data
mask = pd.DataFrame(data = np.zeros(df.shape), index = df.index, columns = df.columns, dtype = bool) # all False
for i in corrupt_values:
mask = mask | (df == i)
self.df[mask] = np.nan
self._append_test_results(mask, 'Corrupt data', min_failures=min_failures)
### Functional approach
[docs]@_documented_by(PerformanceMonitoring.check_timestamp)
def check_timestamp(data, frequency, expected_start_time=None,
expected_end_time=None, min_failures=1, exact_times=True):
pm = PerformanceMonitoring()
pm.add_dataframe(data)
pm.check_timestamp(frequency, expected_start_time, expected_end_time,
min_failures, exact_times)
mask = pm.mask
return {'cleaned_data': pm.df, 'mask': mask, 'test_results': pm.test_results}
[docs]@_documented_by(PerformanceMonitoring.check_range)
def check_range(data, bound, key=None, min_failures=1):
pm = PerformanceMonitoring()
pm.add_dataframe(data)
pm.check_range(bound, key, min_failures)
mask = pm.mask
return {'cleaned_data': data[mask], 'mask': mask, 'test_results': pm.test_results}
@_documented_by(PerformanceMonitoring.check_increment)
def check_increment(data, bound, key=None, increment=1, absolute_value=True,
min_failures=1):
pm = PerformanceMonitoring()
pm.add_dataframe(data)
pm.check_increment(bound, key, increment, absolute_value, min_failures)
mask = pm.mask
return {'cleaned_data': data[mask], 'mask': mask, 'test_results': pm.test_results}
[docs]@_documented_by(PerformanceMonitoring.check_delta)
def check_delta(data, bound, key=None, window=3600, direction=None, min_failures=1):
pm = PerformanceMonitoring()
pm.add_dataframe(data)
pm.check_delta(bound, key, window, direction, min_failures)
mask = pm.mask
return {'cleaned_data': data[mask], 'mask': mask, 'test_results': pm.test_results}
[docs]@_documented_by(PerformanceMonitoring.check_outlier)
def check_outlier(data, bound, key=None, window=3600, absolute_value=True,
min_failures=1):
pm = PerformanceMonitoring()
pm.add_dataframe(data)
pm.check_outlier(bound, key, window, absolute_value, min_failures)
mask = pm.mask
return {'cleaned_data': data[mask], 'mask': mask, 'test_results': pm.test_results}
[docs]@_documented_by(PerformanceMonitoring.check_missing)
def check_missing(data, key=None, min_failures=1):
pm = PerformanceMonitoring()
pm.add_dataframe(data)
pm.check_missing(key, min_failures)
mask = pm.mask
return {'cleaned_data': data[mask], 'mask': mask, 'test_results': pm.test_results}
[docs]@_documented_by(PerformanceMonitoring.check_corrupt)
def check_corrupt(data, corrupt_values, key=None, min_failures=1):
pm = PerformanceMonitoring()
pm.add_dataframe(data)
pm.check_corrupt(corrupt_values, key, min_failures)
mask = pm.mask
return {'cleaned_data': data[mask], 'mask': mask, 'test_results': pm.test_results}