Complex system anomaly detection based on discrete event sequences
View Patent ↗A method detects anomalies in a system having sensors for collecting multivariate sensor data including discrete event sequences. The method determines, using a NMT model, pairwise relationships among the sensors based on the data. The method forms sequences of characters into sentences on a per sensor basis, by treating each discrete variable in the sequences as a character in natural language. The method translates, using the NMT, the sentences of source sensors to sentences of target sensors to obtain a translation score that quantifies a pairwise relationship strength therebetween. The method aggregates the pairwise relationships into a multivariate relationship graph having nodes representing sensors and edges denoted by the translation score for a sensor pair connected thereto to represent the pairwise relationship strength therebetween. The method performs a corrective action to correct an anomaly responsive to a detection of the anomaly relating to the sensor pair.
1. A computer-implemented method for automatic anomaly detection in a physical hardware system having a plurality of sensors, comprising:
determining, by a hardware processor using a Neural Machine Translation (NMT) model, pairwise relationships among the plurality of sensors based on the multivariate sensor data including discrete event sequences obtained from the plurality of sensors;
forming sequences of characters into sentences on a per sensor basis, by treating each discrete variable in the discrete event sequences as a character in natural language;
translating, by using the NMT, the sentences of source sensors, from among the plurality of sensors, to sentences of target sensors, from among the plurality of sensors, to obtain a translation score that quantifies a pairwise relationship strength therebetween;
aggregating the pairwise relationships into a multivariate relationship graph having nodes representing respective ones of the plurality of sensors and edges denoted by the translation score for a sensor pair connected thereto to represent the pairwise relationship strength therebetween; and
performing a corrective action to correct an anomaly responsive to a detection of the anomaly relating to the sensor pair based on a detected violation of the pairwise relationship strength therebetween determined using the multivariate relationship graph relative to input testing multivariate sensor data.
2. The computer-implemented method of claim 1 , wherein said translating step uses same parameter setting to train the NMT model among each sensor pair formed from the plurality of sensors.
3. The computer-implemented method of claim 1 , wherein a higher translation score implies a stronger relationship between two sensors in the sensor pair from among the plurality of sensors, while a lower score implies a weaker relationship.
4. The computer-implemented method of claim 1 , further comprising tracing the detected anomaly through the multivariate relationship graph to determine a root cause of the detected anomaly.
5. The computer-implemented method of claim 1 , wherein the pairwise relationships are determined during normal system operation of the physical hardware system.
6. The computer-implemented method of claim 1 , wherein an edge weight is denoted by the translation score.
7. The computer-implemented method of claim 1 , wherein the pairwise relationship strength denotes a pairwise logical relationship.
8. The computer-implemented method of claim 1 , wherein the pairwise strength denotes a pairwise physical relationship.
9. The computer-implemented method of claim 1 , wherein the multivariate relationship graph is generated during an offline training phase, and is subsequently used for said performing step in an online testing phase.
10. The computer-implemented method of claim 1 , wherein said forming step comprises mapping event records in the discrete event sequences into an alphabet that forms an artificial sensor language.
11. The computer-implemented method of claim 1 , further comprising generating the sentences from groups of m words using a sliding window of n words, wherein m and n are user selectable integers, and wherein n>m.
12. The computer-implemented method of claim 1 , wherein for two sensor pairs that are similar based on similarity criteria, a single representative sensor pair from among the two sensor pairs is selected with a remaining sensor pair from among the two sensor pairs being ignored.
13. The computer-implemented method of claim 1 , wherein the corrective action comprises automatically disabling the sensor pair having the anomaly relating thereto and enabling a backup sensor pair in place thereof.
14. A computer program product for automatic anomaly detection in a physical hardware system having a plurality of sensors, the computer program product comprising a non-transitory computer readable storage medium having program instructions embodied therewith, the program instructions executable by a computer to cause the computer to perform a method comprising:
determining, by a hardware processor using a Neural Machine Translation (NMT) model, pairwise relationships among the plurality of sensors based on the multivariate sensor data including discrete event sequences obtained from the plurality of sensors;
forming sequences of characters into sentences on a per sensor basis, by treating each discrete variable in the discrete event sequences as a character in natural language;
translating, using the NMT, the sentences of source sensors, from among the plurality of sensors, to sentences of target sensors, from among the plurality of sensors, to obtain a translation score that quantifies a pairwise relationship strength therebetween;
aggregating the pairwise relationships into a multivariate relationship graph having nodes representing respective ones of the plurality of sensors and edges denoted by the translation score for a sensor pair connected thereto to represent the pairwise relationship strength therebetween; and
performing a corrective action to correct an anomaly responsive to a detection of the anomaly relating to the sensor pair based on a detected violation of the pairwise relationship strength therebetween determined using the multivariate relationship graph relative to input testing multivariate sensor data.
15. The computer program product of claim 14 , wherein said translating step uses same parameter setting to train the NMT model among each sensor pair formed from the plurality of sensors.
16. The computer program product of claim 14 , wherein a higher translation score implies a stronger relationship between two sensors in the sensor pair from among the plurality of sensors, while a lower score implies a weaker relationship.
17. The computer program product of claim 14 , further comprising tracing the detected anomaly through the multivariate relationship graph to determine a root cause of the detected anomaly.
18. The computer program product of claim 14 , wherein the pairwise relationships are determined during normal system operation of the physical hardware system.
19. The computer program product of claim 14 , wherein an edge weight is denoted by the translation score.
20. A computer processing system for automatic anomaly detection in a physical hardware system having a plurality of sensors, the computer processing system comprising:
a memory device including program code stored thereon;
a hardware processor, operatively coupled to the memory device, and configured to run the program code stored on the memory device to
determine, using a Neural Machine Translation (NMT) model, pairwise relationships among the plurality of sensors based on the multivariate sensor data including discrete event sequences obtained from the plurality of sensors;
form sequences of characters into sentences on a per sensor basis, by treating each discrete variable in the discrete event sequences as a character in natural language;
translate, using the NMT, the sentences of source sensors, from among the plurality of sensors, to sentences of target sensors, from among the plurality of sensors, to obtain a translation score that quantifies a pairwise relationship strength therebetween;
aggregate the pairwise relationships into a multivariate relationship graph having nodes representing respective ones of the plurality of sensors and edges denoted by the translation score for a sensor pair connected thereto to represent the pairwise relationship strength therebetween; and
perform a corrective action to correct an anomaly responsive to a detection of the anomaly relating to the sensor pair based on a detected violation of the pairwise relationship strength therebetween determined using the multivariate relationship graph relative to input testing multivariate sensor data.