-
Notifications
You must be signed in to change notification settings - Fork 138
/
anomaly_detection.py
204 lines (166 loc) · 6.67 KB
/
anomaly_detection.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
import numpy as np
import time
from sklearn.metrics import f1_score, precision_score, recall_score
import bottleneck as bn
# consider delay threshold and missing segments
def get_range_proba(predict, label, delay=7):
splits = np.where(label[1:] != label[:-1])[0] + 1
is_anomaly = label[0] == 1
new_predict = np.array(predict)
pos = 0
for sp in splits:
if is_anomaly:
if 1 in predict[pos:min(pos + delay + 1, sp)]:
new_predict[pos: sp] = 1
else:
new_predict[pos: sp] = 0
is_anomaly = not is_anomaly
pos = sp
sp = len(label)
if is_anomaly: # anomaly in the end
if 1 in predict[pos: min(pos + delay + 1, sp)]:
new_predict[pos: sp] = 1
else:
new_predict[pos: sp] = 0
return new_predict
# set missing = 0
def reconstruct_label(timestamp, label):
timestamp = np.asarray(timestamp, np.int64)
index = np.argsort(timestamp)
timestamp_sorted = np.asarray(timestamp[index])
interval = np.min(np.diff(timestamp_sorted))
label = np.asarray(label, np.int64)
label = np.asarray(label[index])
idx = (timestamp_sorted - timestamp_sorted[0]) // interval
new_label = np.zeros(shape=((timestamp_sorted[-1] - timestamp_sorted[0]) // interval + 1,), dtype=np.int)
new_label[idx] = label
return new_label
def eval_ad_result(test_pred_list, test_labels_list, test_timestamps_list, delay):
labels = []
pred = []
for test_pred, test_labels, test_timestamps in zip(test_pred_list, test_labels_list, test_timestamps_list):
assert test_pred.shape == test_labels.shape == test_timestamps.shape
test_labels = reconstruct_label(test_timestamps, test_labels)
test_pred = reconstruct_label(test_timestamps, test_pred)
test_pred = get_range_proba(test_pred, test_labels, delay)
labels.append(test_labels)
pred.append(test_pred)
labels = np.concatenate(labels)
pred = np.concatenate(pred)
return {
'f1': f1_score(labels, pred),
'precision': precision_score(labels, pred),
'recall': recall_score(labels, pred)
}
def np_shift(arr, num, fill_value=np.nan):
result = np.empty_like(arr)
if num > 0:
result[:num] = fill_value
result[num:] = arr[:-num]
elif num < 0:
result[num:] = fill_value
result[:num] = arr[-num:]
else:
result[:] = arr
return result
def eval_anomaly_detection(model, all_train_data, all_train_labels, all_train_timestamps, all_test_data, all_test_labels, all_test_timestamps, delay):
t = time.time()
all_train_repr = {}
all_test_repr = {}
all_train_repr_wom = {}
all_test_repr_wom = {}
for k in all_train_data:
train_data = all_train_data[k]
test_data = all_test_data[k]
full_repr = model.encode(
np.concatenate([train_data, test_data]).reshape(1, -1, 1),
mask='mask_last',
casual=True,
sliding_length=1,
sliding_padding=200,
batch_size=256
).squeeze()
all_train_repr[k] = full_repr[:len(train_data)]
all_test_repr[k] = full_repr[len(train_data):]
full_repr_wom = model.encode(
np.concatenate([train_data, test_data]).reshape(1, -1, 1),
casual=True,
sliding_length=1,
sliding_padding=200,
batch_size=256
).squeeze()
all_train_repr_wom[k] = full_repr_wom[:len(train_data)]
all_test_repr_wom[k] = full_repr_wom[len(train_data):]
res_log = []
labels_log = []
timestamps_log = []
for k in all_train_data:
train_data = all_train_data[k]
train_labels = all_train_labels[k]
train_timestamps = all_train_timestamps[k]
test_data = all_test_data[k]
test_labels = all_test_labels[k]
test_timestamps = all_test_timestamps[k]
train_err = np.abs(all_train_repr_wom[k] - all_train_repr[k]).sum(axis=1)
test_err = np.abs(all_test_repr_wom[k] - all_test_repr[k]).sum(axis=1)
ma = np_shift(bn.move_mean(np.concatenate([train_err, test_err]), 21), 1)
train_err_adj = (train_err - ma[:len(train_err)]) / ma[:len(train_err)]
test_err_adj = (test_err - ma[len(train_err):]) / ma[len(train_err):]
train_err_adj = train_err_adj[22:]
thr = np.mean(train_err_adj) + 4 * np.std(train_err_adj)
test_res = (test_err_adj > thr) * 1
for i in range(len(test_res)):
if i >= delay and test_res[i-delay:i].sum() >= 1:
test_res[i] = 0
res_log.append(test_res)
labels_log.append(test_labels)
timestamps_log.append(test_timestamps)
t = time.time() - t
eval_res = eval_ad_result(res_log, labels_log, timestamps_log, delay)
eval_res['infer_time'] = t
return res_log, eval_res
def eval_anomaly_detection_coldstart(model, all_train_data, all_train_labels, all_train_timestamps, all_test_data, all_test_labels, all_test_timestamps, delay):
t = time.time()
all_data = {}
all_repr = {}
all_repr_wom = {}
for k in all_train_data:
all_data[k] = np.concatenate([all_train_data[k], all_test_data[k]])
all_repr[k] = model.encode(
all_data[k].reshape(1, -1, 1),
mask='mask_last',
casual=True,
sliding_length=1,
sliding_padding=200,
batch_size=256
).squeeze()
all_repr_wom[k] = model.encode(
all_data[k].reshape(1, -1, 1),
casual=True,
sliding_length=1,
sliding_padding=200,
batch_size=256
).squeeze()
res_log = []
labels_log = []
timestamps_log = []
for k in all_data:
data = all_data[k]
labels = np.concatenate([all_train_labels[k], all_test_labels[k]])
timestamps = np.concatenate([all_train_timestamps[k], all_test_timestamps[k]])
err = np.abs(all_repr_wom[k] - all_repr[k]).sum(axis=1)
ma = np_shift(bn.move_mean(err, 21), 1)
err_adj = (err - ma) / ma
MIN_WINDOW = len(data) // 10
thr = bn.move_mean(err_adj, len(err_adj), MIN_WINDOW) + 4 * bn.move_std(err_adj, len(err_adj), MIN_WINDOW)
res = (err_adj > thr) * 1
for i in range(len(res)):
if i >= delay and res[i-delay:i].sum() >= 1:
res[i] = 0
res_log.append(res[MIN_WINDOW:])
labels_log.append(labels[MIN_WINDOW:])
timestamps_log.append(timestamps[MIN_WINDOW:])
t = time.time() - t
eval_res = eval_ad_result(res_log, labels_log, timestamps_log, delay)
eval_res['infer_time'] = t
return res_log, eval_res