def main(argv):
# Time the analysis:
tic = time.time()
# Get commandline arguments
try:
infile = argv[1]
granularity = int(argv[2])
training_window = int(argv[3])
forecasting_interval = int(argv[4])
except Exception:
raise RuntimeError("usage: python "
+ argv[0]
+ " <infile>"
+ " <granularity>"
+ " <training_window>"
+ " <forecasting_interval>")
print ("\nStarting analysis on %s with settings %d %d %d..."
% (infile, granularity, training_window, forecasting_interval))
# Generate outfile name based on infile
outfile = infile.rstrip('.csv') + '.results.csv'
# Separate feature timestamps, feature data, and power data
dataframe = pd.read_csv(infile)
timestamps = dataframe.ix[:, 0].values
targets = dataframe.ix[:, -1].values
features_df = dataframe.ix[:, 1:-1]
# Filter features with little or no variance
print "\nRemoving features due to low variance:"
for column in features_df.columns:
values = features_df[column].values
if (values.max() == values.min()):
features_df = features_df.drop(column, 1)
print column
#dataframe = dataframe.drop('Audio Sensor', 1)
#dataframe = dataframe.drop('Luminescence_11 (Lux)', 1)
features_df = features_df[['Audio Sensor',
'Temperature_10 (C)']]
print "\nThe following features will be used: "
for column_name in features_df.columns:
print column_name
features = features_df.values
features = scaleFeatures(features)
num_features = features.shape[1]
# Algorithm settings
algo = Algo(granularity, training_window, forecasting_interval, num_features)
# Output lists
y_time = ['Timestamp']
y_target = ['Target']
y_predict = ['Prediction']
anomalies = ['Anomaly']
print "Algorithm settings:"
# EWMA additions
# alpha is adjustable on a scale of (0, 1]
# The smaller value of alpha, the more averaging takes place
# A value of 1.0 means no averaging happens
#alpha = float(raw_input('Enter Value of alpha:'))
algo.setEMAParameter(alpha=1.0)
#algo.setEMAParameter(alpha=0.75)
#algo.setEMAParameter(alpha=0.5)
#algo.setEMAParameter(alpha=0.25)
# Recomended Severity Parameters from Paper
#algo.setSeverityParameters(w=0.53, L=3.714) # Most sensitive
#algo.setSeverityParameters(w=0.84, L=3.719) # Medium sensitive
#algo.setSeverityParameters(w=1.00, L=3.719) # Least sensitive
algo.setSeverityParameters(w=1, L=2) # Custom senstivity
# More settings
algo.setMaxPrediction(500)
#USED For F1 Calculation
detected = set()
ground_truth = set()
#==================== ANALYSIS ====================#
print "\nBeginning analysis..."
count = 0
for row in features:
# Update the user with the current timestamp
cur_time = timestamps[count]
if (count % 360) == 0:
cur_dt = dt.datetime.fromtimestamp(cur_time)
print "Trying time %s" % cur_dt.strftime(DATE_FORMAT)
row = np.append(row, targets[count])
target, prediction = algo.run(row) # Magic!
# If there is a prediction, check for anomalies
if prediction != None:
y_time.append(cur_time)
y_target.append(target)
y_predict.append(float(prediction))
if algo.checkSeverity(target, float(prediction)):
detected.add(cur_time)
anomalies.append(1)
else:
anomalies.append(0)
# NOTE: If you plan on injecting attacks, you can add the
# timestamps here to create a "ground truth" set, which can then
# be used to calculate the accuracy, precision and F1 scores.
#if cur_time >= 1338696000 and cur_time <= 1338699600:
# ground_truth.add(cur_time)
#elif cur_time >= 1338955200 and cur_time <= 1338958800:
# ground_truth.add(cur_time)
#elif cur_time >= 1339128000 and cur_time <= 1339131600:
# ground_truth.add(cur_time)
# Loop back to the beginning
count += 1
#==================== RESULTS ====================#
# Save data for later graphing
results = y_time, y_target, y_predict, anomalies
writeResults(outfile, results)
#f1_scores(detected, ground_truth) #Uncomment to show F1 scores
print_stats(y_target[1:], y_predict[1:]) #Remove header
print "Run time: ", time.time() - tic
print "Ending analysis. See %s for results." % outfile
return results
