Time series retrieval for analyzing and correcting system status
View Patent ↗Methods and systems for detecting and correcting anomalous behavior include generating a joint binary embedding of each of a set of historical time series sequences. A joint binary embedding of a recent time series sequence is generated. A ranked list of the plurality of historical time series sequences is generated according to respective similarities of each historical time series sequence to the recent time series sequence based on the respective joint binary embeddings of each. Anomalous behavior of a system associated with the recent time series sequence is determined according to a label of a top-ranked historical time series sequence in the ranked list. A corrective action is performed to correct the anomalous behavior.
1. A method for detecting and correcting anomalous behavior, comprising:
generating a joint binary embedding of each of a plurality of historical time series sequences;
generating a joint binary embedding of a recent time series sequence;
generating a ranked list of the plurality of historical time series sequences according to respective similarities of each historical time series sequence to the recent time series sequence based on the respective joint binary embeddings of each;
determining anomalous behavior of a system associated with the recent time series sequence according to a label of a top-ranked historical time series sequence in the ranked list; and
performing a corrective action to correct the anomalous behavior.
2. The method of claim 1 , wherein generating a joint binary embedding of the plurality of historical time series sequences and of the recent time series sequence comprises:
encoding a time series sequence with a long short term memory unit to generate a first representation of temporal dynamics; and
encoding pairs of time series sequences with convolutional neural networks to generate second representations for correlations between different pairs of time series within a same time segment.
3. The method of claim 2 , wherein generating a joint binary embedding of the plurality of historical time series sequences and of the recent time series sequence further comprises:
concatenating the representations; and
mapping the concatenation to a value between 1 and −1.
4. The method of claim 1 , wherein generating the ranked list comprises determining a similarity of each of the plurality of historical time series sequences to the recent time series sequence according to a Hamming distance between the respective joint binary embeddings.
5. The method of claim 1 , wherein generating the ranked list comprises optimizing an objective function that includes a ranking loss term and a regularization term.
6. The method of claim 5 , wherein optimizing the objective function comprises optimizing an approximated objective function:
(
𝒟
triplet
,
W
)
_
=
(
⌊
M
s
⌋
·
∑
j
=
1
s
σ
(
V
qi
-
V
qj
)
)
1
r
+
λ
2
W
F
2
where D triplet is a triplet of segments, W is a parameter matrix, M is a total of choices of triplets, s is a number of different segments in the triplets, Σ s j=1 σ(V qi −V qj ) is a soft-approximated rank of historical time series yi with respect to query y q , r is a penalization parameter, and λ is a parameter controlling a tradeoff between ranking loss and regularization.
7. The method of claim 5 , wherein the objective function is an approximate objective function base on a stochastic gradient descent.
8. The method of claim 1 , wherein each time series sequence represents a set of sensor values ordered by time of recording.
9. The method of claim 1 , wherein the corrective action is selected from the group consisting of changing a security setting for an application or hardware component of a monitored system, changing an operational parameter of an application or hardware component of the monitored system, halting or restarting an application of the monitored system, halting or rebooting a hardware component of the monitored system, changing an environmental condition of the monitored system, and changing status of a network interface of the monitored system.
10. A system for detecting and correcting anomalous behavior, comprising:
a joint embedding module comprising a processor configured to generate a joint binary embedding of each of a plurality of historical time series sequences and to generate a joint binary embedding of a recent time series sequence;
a ranking module configured to generate a ranked list of the plurality of historical time series sequences according to respective similarities of each historical time series sequence to the recent time series sequence based on the respective joint binary embeddings of each; and
a correction module configured to determine anomalous behavior of a system associated with the recent time series sequence according to a label of a top-ranked historical time series sequence in the ranked list and to perform a corrective action to correct the anomalous behavior.
11. The system of claim 10 , wherein the joint embedding module is further configured to encode a time series sequence with a long short term memory unit, to generate a first representation of temporal dynamics, and to encode pairs of time series sequences with convolutional neural networks, to generate second representations for correlations between different pairs of time series within a same time segment.
12. The system of claim 11 , wherein the joint embedding module is further configured to concatenate the representations, and to map the concatenation to a value between 1 and −1.
13. The system of claim 10 , wherein the ranking module is further configured to determine a similarity of each of the plurality of historical time series sequences to the recent time series sequence according to a Hamming distance between the respective joint binary embeddings.
14. The system of claim 10 , wherein the ranking module is further configured to optimize an objective function that includes a ranking loss term and a regularization term.
15. The system of claim 14 , wherein the ranking module is further configured to optimize the objective function as an approximated objective function:
(
𝒟
triplet
,
W
)
_
=
(
⌊
M
s
⌋
·
∑
j
=
1
s
σ
(
V
qi
-
V
qj
)
)
1
r
+
λ
2
W
F
2
where D triplet is a triplet of segments, W is a parameter matrix, M is a total of choices of triplets, s is a number of different segments in the triplets, Σ s j=1 σ(V qi −V qj ) is a soft-approximated rank of historical time series yi with respect to query y q , r is a penalization parameter, and λ is a parameter controlling a tradeoff between ranking loss and regularization.
16. The system of claim 14 , wherein the objective function is an approximate objective function base on a stochastic gradient descent.
17. The system of claim 10 , wherein each time series sequence represents a set of sensor values ordered by time of recording.
18. The system of claim 10 , wherein the corrective action is selected from the group consisting of changing a security setting for an application or hardware component of a monitored system, changing an operational parameter of an application or hardware component of the monitored system, halting or restarting an application of the monitored system, halting or rebooting a hardware component of the monitored system, changing an environmental condition of the monitored system, and changing status of a network interface of the monitored system.